<%BANNER%>

Algorithms for Solving Multi-Level Optimization Problems with Discrete Variables at Multiple Levels

University of Florida Institutional Repository
Permanent Link: http://ufdc.ufl.edu/UFE0024784/00001

Material Information

Title: Algorithms for Solving Multi-Level Optimization Problems with Discrete Variables at Multiple Levels
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Taskin, Zeki
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: constraint, decomposition, edge, graph, imrt, integer, matrix, partition, programming, search
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: In this dissertation, we investigate a class of multi-level optimization problems, in which discrete variables are present at multiple stages. Such problems arise in many practical settings, and they are notoriously difficult to optimize. Benders decomposition, which is a well-known decomposition method for solving large-scale mixed-integer programming problems, cannot be utilized for the class of problems that we consider due to the existence of discrete variables at lower levels. Cutting plane algorithms such as those proposed by Laporte and Louveaux have been designed for use in bi-level integer programming problems with binary variables at both levels. However, these are based on generic cuts, which do not utilize any problem specific structures, and hence often result in weak convergence. Our goal in this dissertation is to propose novel formulation and solution strategies for several multi-level optimization problems to solve these problems to optimality within practical computational limits. We first consider an edge-partition problem. The motivation for this study is provided by a Synchronous Optical Network (SONET) design application. In the SONET context, each edge represents a demand pair between two client nodes, and the weight of each edge represents the number of communication channels needed between the client nodes. We consider a stochastic version of the problem, in which the edge weights are not deterministic, but their underlying probability distribution is known. The problem is to design a set of SONET ``rings'' at minimum cost, while ensuring that the resulting network can handle the random demand with a prespecified level of service. We first model the problem as a large-scale integer program, and attempt to solve it via a bi-level decomposition approach, in which both levels contain binary variables. We then propose a three-level solution approach for the problem, which is based on a hybrid integer programming/constraint programming decomposition algorithm. We show computationally that the hybrid algorithm significantly outperforms the other approaches. Next, we focus on a matrix decomposition problem, which arises in Intensity Modulated Radiation Therapy (IMRT) treatment planning. The problem input is a matrix of intensity values that needs to be delivered to a patient, which must be decomposed into a collection of apertures and corresponding intensities. In a feasible decomposition the sum of binary shape matrices multiplied by corresponding intensity values is equal to the original intensity matrix. We consider two variants of the problem: (i) a variant in which the shape matrices used in the decomposition have to satisfy the ``consecutive-ones'' property, and (ii) a variant in which the shape matrices have to be rectangular. For the first variant, we start by investigating an integer programming model proposed in the literature, and show how the formulation can be strengthened. We then formulate the problem as a bi-level optimization problem that has discrete variables at both stages, and suggest a hybrid integer programming/constraint programming decomposition algorithm similar to our algorithm for the edge-partition problem. Our tests on data obtained from patients show that our algorithm is capable of solving problem instances of clinically relevant dimensions within practical computational limits. We then turn our attention to the second variant of the matrix decomposition problem. We formulate the problem as a mixed-integer program, and investigate a decomposition method for solving it. Unlike the first variant, the second-level problem turns out to be a linear programming problem, and therefore we are able to derive a Benders decomposition algorithm for solving this variant. Finally, we investigate a class of graph search problems. In this class of problems, an intruder is located at an unknown node on the input graph, and a group of searchers needs to be coordinated to detect the intruder within a limited amount of time. This problem arises in settings such as search-and-rescue and search-and-capture operations, and patrol route design. We investigate three variants of the problem: (i) a hide-and-seek version, in which a stationary intruder is hiding at an unknown node, (ii) a pursuit evasion version, in which the intruder can move across the edges of the graph to avoid being detected, and (iii) a patrol problem, in which the searchers are assigned to recurring patrol routes to protect the graph from intrusion. We model each problem as a large-scale integer program, and propose a branch-cut-price algorithm to find the minimum number of searchers needed, and a route for each searcher. In our formulation, both the master problem and the subproblems corresponding to the searchers and the intruder contain binary variables. We model the master problem as a set covering problem and propose a solution approach that is based on dynamic column and row generation.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Zeki Taskin.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Smith, Jonathan.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024784:00001

Permanent Link: http://ufdc.ufl.edu/UFE0024784/00001

Material Information

Title: Algorithms for Solving Multi-Level Optimization Problems with Discrete Variables at Multiple Levels
Physical Description: 1 online resource (158 p.)
Language: english
Creator: Taskin, Zeki
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: constraint, decomposition, edge, graph, imrt, integer, matrix, partition, programming, search
Industrial and Systems Engineering -- Dissertations, Academic -- UF
Genre: Industrial and Systems Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: In this dissertation, we investigate a class of multi-level optimization problems, in which discrete variables are present at multiple stages. Such problems arise in many practical settings, and they are notoriously difficult to optimize. Benders decomposition, which is a well-known decomposition method for solving large-scale mixed-integer programming problems, cannot be utilized for the class of problems that we consider due to the existence of discrete variables at lower levels. Cutting plane algorithms such as those proposed by Laporte and Louveaux have been designed for use in bi-level integer programming problems with binary variables at both levels. However, these are based on generic cuts, which do not utilize any problem specific structures, and hence often result in weak convergence. Our goal in this dissertation is to propose novel formulation and solution strategies for several multi-level optimization problems to solve these problems to optimality within practical computational limits. We first consider an edge-partition problem. The motivation for this study is provided by a Synchronous Optical Network (SONET) design application. In the SONET context, each edge represents a demand pair between two client nodes, and the weight of each edge represents the number of communication channels needed between the client nodes. We consider a stochastic version of the problem, in which the edge weights are not deterministic, but their underlying probability distribution is known. The problem is to design a set of SONET ``rings'' at minimum cost, while ensuring that the resulting network can handle the random demand with a prespecified level of service. We first model the problem as a large-scale integer program, and attempt to solve it via a bi-level decomposition approach, in which both levels contain binary variables. We then propose a three-level solution approach for the problem, which is based on a hybrid integer programming/constraint programming decomposition algorithm. We show computationally that the hybrid algorithm significantly outperforms the other approaches. Next, we focus on a matrix decomposition problem, which arises in Intensity Modulated Radiation Therapy (IMRT) treatment planning. The problem input is a matrix of intensity values that needs to be delivered to a patient, which must be decomposed into a collection of apertures and corresponding intensities. In a feasible decomposition the sum of binary shape matrices multiplied by corresponding intensity values is equal to the original intensity matrix. We consider two variants of the problem: (i) a variant in which the shape matrices used in the decomposition have to satisfy the ``consecutive-ones'' property, and (ii) a variant in which the shape matrices have to be rectangular. For the first variant, we start by investigating an integer programming model proposed in the literature, and show how the formulation can be strengthened. We then formulate the problem as a bi-level optimization problem that has discrete variables at both stages, and suggest a hybrid integer programming/constraint programming decomposition algorithm similar to our algorithm for the edge-partition problem. Our tests on data obtained from patients show that our algorithm is capable of solving problem instances of clinically relevant dimensions within practical computational limits. We then turn our attention to the second variant of the matrix decomposition problem. We formulate the problem as a mixed-integer program, and investigate a decomposition method for solving it. Unlike the first variant, the second-level problem turns out to be a linear programming problem, and therefore we are able to derive a Benders decomposition algorithm for solving this variant. Finally, we investigate a class of graph search problems. In this class of problems, an intruder is located at an unknown node on the input graph, and a group of searchers needs to be coordinated to detect the intruder within a limited amount of time. This problem arises in settings such as search-and-rescue and search-and-capture operations, and patrol route design. We investigate three variants of the problem: (i) a hide-and-seek version, in which a stationary intruder is hiding at an unknown node, (ii) a pursuit evasion version, in which the intruder can move across the edges of the graph to avoid being detected, and (iii) a patrol problem, in which the searchers are assigned to recurring patrol routes to protect the graph from intrusion. We model each problem as a large-scale integer program, and propose a branch-cut-price algorithm to find the minimum number of searchers needed, and a route for each searcher. In our formulation, both the master problem and the subproblems corresponding to the searchers and the intruder contain binary variables. We model the master problem as a set covering problem and propose a solution approach that is based on dynamic column and row generation.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Zeki Taskin.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Smith, Jonathan.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024784:00001


This item has the following downloads:


Full Text
xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20110216_AAAAII INGEST_TIME 2011-02-17T00:58:33Z PACKAGE UFE0024784_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 29740 DFID F20110216_AABVSA ORIGIN DEPOSITOR PATH taskin_z_Page_120.QC.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
2abe1891774d0a1726e01938ffb549e9
SHA-1
304d80fccf044bc7b8073c3ab2d3f4095045200c
30686 F20110216_AABVRL taskin_z_Page_043.QC.jpg
d5c2168aebcd5451f1c53841299ffe7b
492603759ef7c493c1029644db80da2e7eddcee3
33493 F20110216_AABVQX taskin_z_Page_056.QC.jpg
061e73cd22e60467f8a9e913d93348d1
510bead98f564248805d2453545f0588b717e488
112265 F20110216_AABUOJ taskin_z_Page_028.jpg
9f1fcc223583bff3d9a51aba1650c653
e3f71f35fb56a22f9413b7ef342a673928358d45
27759 F20110216_AABUNV taskin_z_Page_095.QC.jpg
1c61ea05217c3f045058da432de5468f
c156cc0e7d99a3fd7700656b50fbbd2678f3418b
35760 F20110216_AABVSB taskin_z_Page_126.QC.jpg
b361497661d40b720fde89e1743b3359
f3a2db560d5da61eb56791681c6e8a1e6ca84f14
34910 F20110216_AABVRM taskin_z_Page_045.QC.jpg
9bcd7b5dac81287835bbff3ccb35e535
bb01d4568ec19284dfb56edbcfb1510108592e2a
26390 F20110216_AABVQY taskin_z_Page_035.QC.jpg
7e9f83b13ca75f5179306fb98cdc5e58
f563970d3563c40363c68e50cae95795d474b972
1051945 F20110216_AABUOK taskin_z_Page_122.jp2
74935a9c523876a9eccc70d909ba1b54
ea037281de47330cddb0cae4532d91f2ec859081
6908 F20110216_AABUNW taskin_z_Page_005thm.jpg
4b03e121b1597ab50f41821abd002e0b
6da1b41fd97fd68c4c505fecec9993dfea41a49b
30322 F20110216_AABVSC taskin_z_Page_129.QC.jpg
f45da6f0561140793ad3dd2b0ae4cac4
b064f5efd350b9e42902b9af7a4d31a2328e6544
34595 F20110216_AABVRN taskin_z_Page_048.QC.jpg
bad935747fd0127ae8c89e5dd53d03f1
267fca1d5a26dc2175b440de038f75575204f55c
6482 F20110216_AABVQZ taskin_z_Page_101thm.jpg
8b73090581d63127b53c6ccae62ef484
c15b21dd9b159d2bd437b8a1220d3da8fa95ca1d
1053954 F20110216_AABUOL taskin_z_Page_002.tif
2ecebf60af8da2afb5e2f6bd48483893
29c593ea3ea2cb6563ba6582d032591c353d9c15
7395 F20110216_AABUNX taskin_z_Page_129thm.jpg
d81694dcaa60c9f948c907e39c0bb1de
2bb1054bcfa9a3c8bfeec51d2daaf0ec53b4e6cc
120062 F20110216_AABUPA taskin_z_Page_016.jpg
f1dbd1030cf1044ff423dcecc311cf8e
74a119c598f1c76ce470c3ec22d7f7b77de137d7
28140 F20110216_AABVSD taskin_z_Page_131.QC.jpg
f9067057dc8e354c6a7f8c2f33444f36
c6d03ff96950badf3b8296966adfe715e1489efe
32914 F20110216_AABVRO taskin_z_Page_053.QC.jpg
87e04e67e58d96298fcd63cfd0114058
ee0274178cf753f64cc67a98796228f63dd4fde9
1051955 F20110216_AABUOM taskin_z_Page_140.jp2
7c506dff1720003e18227e6ef029aa78
36c315d4c369ccde44851029871abb0bdf363a08
1698 F20110216_AABUNY taskin_z_Page_003.QC.jpg
6d03c50347c83422971ffeecc053eb68
2852cfae881d4a569402a499e8685edfbfe91005
29950 F20110216_AABUPB taskin_z_Page_044.QC.jpg
ba06c0ff3b29e169da7acfa101a5fc28
61f50b94cdb31a4099b188d812c296cdb20e309e
26559 F20110216_AABVSE taskin_z_Page_145.QC.jpg
16e13be1ae12fa82c34b9623d287873a
fa7725e7a063d3f5eb6645bae00433c73e9a9079
37350 F20110216_AABVRP taskin_z_Page_071.QC.jpg
0caec56d118a3887570d9a657f49790c
b6188ebe6b880906ec5d047ca36890ac06c30914
8463 F20110216_AABUON taskin_z_Page_016thm.jpg
95fee03a30e83c61a03123d0e54d29f9
201dff13a0f4959ca0f969ee5811fa8482c970fe
120623 F20110216_AABUNZ taskin_z_Page_140.jpg
0628212106619a86998f84e90c304357
7858f2e56b149f0aeef45f64e891a656bc4840a4
73143 F20110216_AABUPC taskin_z_Page_047.pro
864e2c862dbb4e99c67b2a7ee1f29542
856a8d138ae75aeda72a63e9dcdadcc649324187
35259 F20110216_AABVSF taskin_z_Page_151.QC.jpg
ec9e030d10f02164e230e0d1afda02c6
04fe23b3133fae5431453e225bceffbc391c25c9
33654 F20110216_AABVRQ taskin_z_Page_073.QC.jpg
95403738303663a9f4c8793f11dfa8d1
ae970336873ee33c6cd3fc88f9890f70b09b1547
7727 F20110216_AABUOO taskin_z_Page_040thm.jpg
9b84834a8ac4fbb82449f0c40092ee09
3f3ccad7496e60e2aa0d411df08fa274a24c8099
1980 F20110216_AABUPD taskin_z_Page_157thm.jpg
4427d5098e9f029a3b8300527ff91be1
5203af5eb6ff5f06ff8a6c1c2355643cbef51e29
13229 F20110216_AABVSG taskin_z_Page_158.QC.jpg
a9cc5290f6f622bbf2cfb118c4b4523a
85085c0ea7f81ef0b30f6f2006160dbdfc3f2b08
29519 F20110216_AABVRR taskin_z_Page_091.QC.jpg
485685b2641726b6e3176ac50482e272
55d75e224340918552c53d99577256c9554b0a39
2079 F20110216_AABUOP taskin_z_Page_058.txt
3a6dfdd11beb07203c70666f40850d77
a74389ebea6d5059ab456cdcbec57606e3101f2c
25271604 F20110216_AABUPE taskin_z_Page_093.tif
60230827d3508ef2f274907312bcd6dc
8e3c908d82729dd1a40ab133bfccc5051f01f855
3672 F20110216_AABVSH taskin_z_Page_007thm.jpg
5a52cc6cda1c7c90c078cbcdc77914d6
7f802d13ab2a442649f551eb41defa9a151a562f
31115 F20110216_AABVRS taskin_z_Page_092.QC.jpg
1dfc17e4e53ff94d9cf7783cc7a7edbc
144368f28221110786f79532ece252a9ae23d23d
29481 F20110216_AABUOQ taskin_z_Page_132.QC.jpg
14ac4102cfcbdd976001603eacb94e77
dbaa59fb805e72b6e699d29ddc4364cd6cee8e73
77332 F20110216_AABUPF taskin_z_Page_148.jp2
31b34c971c1367f72894fe5559a92547
07654bb300afdcc0f82fc98d766be2a217080a0e
7134 F20110216_AABVSI taskin_z_Page_010thm.jpg
3958a28300ea764859ba46a79062411b
a8a508ab9074e975070486236b792804700583fc
30845 F20110216_AABVRT taskin_z_Page_094.QC.jpg
f6a4b66bf41bfc29471bf905b1eca4c0
73256dcdcdeaa1017a71bc8202fa1be6ee446326
6311 F20110216_AABUOR taskin_z_Page_077thm.jpg
601f73cab1adba19e1426f3322a607fe
f7bf201ea1cd7522b7022eaf9a3b138068bb9e45
1051973 F20110216_AABUPG taskin_z_Page_107.jp2
0003c14214d53a8200b3fe9edffc9ec3
392dad74e0e50d72037bffe43e429fce6ecf6fcb
3881 F20110216_AABVSJ taskin_z_Page_012thm.jpg
93599126c674dc0925a3d2e12bbf7872
1c2f5d592dc490f5b7bf6ce4b9dbccd55a678a3a
26867 F20110216_AABVRU taskin_z_Page_098.QC.jpg
4fbb609904b969072fab654cef02ad49
4fc6fc97a8eca75481ad717a3ad8adf17f21b56c
F20110216_AABUOS taskin_z_Page_135.tif
99007e527874429166f2a86c6673631a
1637d0d52f265d717a0e77e44f8d04862342471c
1014395 F20110216_AABUPH taskin_z_Page_132.jp2
4d6b0390a99fd970d106962eb63a68a1
503bc4af3165e1f414da755fc359cc233d378597
35842 F20110216_AABVRV taskin_z_Page_102.QC.jpg
ecbdd80e6db7ce6ea55131e2b217a4a5
6ccbbe19998d2047fb7098c61d0041b90aebfd6a
33134 F20110216_AABUOT taskin_z_Page_146.pro
c0a422b508ba48b15200609f3c2c277b
4650de85e0bd8670be3eac35dd6ec9370429d8ea
8328 F20110216_AABVSK taskin_z_Page_015thm.jpg
904915a8dab94821042dfe00b166bf2c
82c04776326bad1e2905f8fc7f305807e6f16b69
24192 F20110216_AABVRW taskin_z_Page_110.QC.jpg
e137bd75aa873ebff588113816b1d6fa
0daf9348423f1ade665f5b7f03646929f47d538c
6585 F20110216_AABUOU taskin_z_Page_057thm.jpg
f47da5d9958ff51c8104f53704441b7a
d10141525235255ebd328c4c39c920dae8395f1f
7435 F20110216_AABUPI taskin_z_Page_060thm.jpg
26a05918c1745f4d0ec0ed59fac93e72
32352502b662823bd95fb578bd583e4d8151c96f
7314 F20110216_AABVTA taskin_z_Page_132thm.jpg
1193e5abb118dbd1a0dfed3af23952f2
5d4d527dd2d85abff38a44b41caa06a6c5b959ea
7588 F20110216_AABVSL taskin_z_Page_025thm.jpg
600d33d9fbaf8bb62537ff6c4d024f47
f33a11d945f292fc5e85002f109a0b0541a0c4e0
37740 F20110216_AABVRX taskin_z_Page_111.QC.jpg
a99aff9292a7459465b8456bb851564d
1db3fb215e27bf214f6af48a7dd5dfa863a07906
30394 F20110216_AABUOV taskin_z_Page_147.QC.jpg
ddefb92cd70f21763fa04326ff904d0a
1b22f44f86fde0cb5e1b986d8310a23b5dd8d96b
752590 F20110216_AABUPJ taskin_z_Page_146.jp2
9534dce76e7544e1dedddc8091d1368d
ca9016df12fab7544c13192293e22d252e6e2c00
4527 F20110216_AABVTB taskin_z_Page_133thm.jpg
c539600685b3df9ae2f4568cac5a6191
321c496e14d7f33d7961640217e9b809ef44accc
6550 F20110216_AABVSM taskin_z_Page_026thm.jpg
a35b37effe3486d69ed88a6d1a2bcbb5
35319d6e7859631b06d25450be65f7fe1afa2cff
34584 F20110216_AABVRY taskin_z_Page_115.QC.jpg
d6b14c3cc8fd963e1eabe35947058f9b
bc11d1629056642bafa88c390dbf53237365bf79
897927 F20110216_AABUOW taskin_z_Page_026.jp2
5088ecd634ab273ba95f528410a85eeb
f4a3db38eb762a94c890fc66903d1b5ba6e79981
120523 F20110216_AABUPK taskin_z_Page_011.jpg
0a71943e98f87c4896ec9822dbabf0c9
f737c187908b1a4b74f2a2cbeee3ed5ff94bbcae
8442 F20110216_AABVTC taskin_z_Page_139thm.jpg
b289c648abb05aa2a719f33168429f79
5723bc1f2dd63fd41d5dfea326b5a297c951eaf7
6036 F20110216_AABVSN taskin_z_Page_027thm.jpg
07dac8f7b676f4490ad76af33833a4ed
e2e53bf241d6d624b34d49fd5216ef668f05b2bd
30260 F20110216_AABVRZ taskin_z_Page_119.QC.jpg
b1b75feeae3ad73b99669630b21cca35
057dde64c2288add956bebe3953b6e8bb5565473
2345 F20110216_AABUOX taskin_z_Page_017.txt
b555e64300908db8ee49d9f362b2f175
e8bd92fe234196baaafccdf29dd019be2faeb7e4
1051964 F20110216_AABUQA taskin_z_Page_016.jp2
c00641125f934b400df3b8cec1fdafbf
63800fe2bd6185a1dad1c9aca027236a9c0746cc
76454 F20110216_AABUPL taskin_z_Page_006.pro
47e07d98477963d550d09eaddae0c43a
b7140e11731a0f462e8994fe1bedfa635d2da5c7
7064 F20110216_AABVTD taskin_z_Page_142thm.jpg
53c6543b6846afbe11b875ee9e3b6fa4
768c8966e5bc3c463d7284301748c8f732910fb8
8388 F20110216_AABVSO taskin_z_Page_028thm.jpg
33fc7a6a52f3d8ea37a7018b5617886d
00e7c0a246e99dddb49b50b566e7a78b6f68f497
F20110216_AABUOY taskin_z_Page_101.tif
120977a6614690947bbd5aacc55eb745
e330ae123077ac7e3a71df197e6f6bac3fd4ea12
6626 F20110216_AABUQB taskin_z_Page_087thm.jpg
c89383fa05c9ecbba09458c4baf4cd43
4e6313c5c591ae360e8564220447cc9e60aceaec
41210 F20110216_AABUPM taskin_z_Page_101.pro
795b5b1bd2a249663529282e42c11d4f
c274b9d5e39e17e8514e081298d8e2cc9f3c97cb
1082 F20110216_AABVTE taskin_z_Page_143thm.jpg
daf34f6d7c5e4ff53751e53bdd416bcd
63cfddea099a4f711a37bf4d526ef0a9fda17012
7880 F20110216_AABVSP taskin_z_Page_038thm.jpg
3bb2785a0d8ac8594c8a670b98693161
0590b7c626a54e2f14a9b779b8fc2d2239c52cdb
1075 F20110216_AABUOZ taskin_z_Page_012.txt
d189f7fb782da426be5dc86c1a6a0e94
e296e71297baba44150535dd84bf44f05f92c839
8596 F20110216_AABUQC taskin_z_Page_140thm.jpg
84f9ba903ef06b0ea89afc757ae751db
9b6901ba5c506958912992b71f06ae322a1bf147
F20110216_AABUPN taskin_z_Page_073.tif
a1d0690e3592a608d6b7427c339e7394
8c2f5fc830ddbc7b01cdcfa06541f6a0699b4130
6762 F20110216_AABVTF taskin_z_Page_145thm.jpg
7c966561035d0720edc9de8ad4427fa1
0ea80eb52e116d66a98c3aa2acf01b1ea8ff2886
7962 F20110216_AABVSQ taskin_z_Page_047thm.jpg
6b8a8aa2c64678cc3b1d8474ee48e989
8891565c08112b1004d8f487482cb38c94cf6160
27456 F20110216_AABUQD taskin_z_Page_089.QC.jpg
255114b5055302828ab3070eb4dd1bb8
14673399b872aa71d6156514e903fdad9f46994d
F20110216_AABUPO taskin_z_Page_156.tif
63ebc19e2bb3c443b2fb9f97bdf75898
89a97c567017563af9f6610e4b84413e7d4b1c5a
7522 F20110216_AABVTG taskin_z_Page_147thm.jpg
b938c82339536b2b0640237ff0ecf0bd
0104abcb42de18795ee15629ccf4aef2d288fcde
7340 F20110216_AABVSR taskin_z_Page_050thm.jpg
4fc4dd16451a5abe500f94bbdf72f526
097039a1acb7ae0be0517194106ef401a6ca80cb
1051971 F20110216_AABUQE taskin_z_Page_086.jp2
9e549efbed7a050a517d6fd7b41a9fc1
8d2bfc78e307f48ec94d9d57af9bf595a0351f8f
8173 F20110216_AABUPP taskin_z_Page_075thm.jpg
567508362f3889d27b51998681ae508c
4c4dce3175a51295c3b90e16d947ea4bab29f4c2
8531 F20110216_AABVTH taskin_z_Page_153thm.jpg
c9a1d5a0a4a1f5588b2f4dc9cc4503bf
2582bf9247888d67c49757d008e90744e138f035
7609 F20110216_AABVSS taskin_z_Page_053thm.jpg
a1fed8a550978ea0afff190640e6bfd8
be0f17372a28daee5abca73d049bd906f51204ea
3247 F20110216_AABUQF taskin_z_Page_050.txt
e2c36f2de2d9edd0edd963542f83312a
075e5985128a81e51dfeeab4580549814d736a88
1003 F20110216_AABUPQ taskin_z_Page_007.txt
1b15c6ac5c3cd7ba05fa7ab6281d2b45
cc1452bef0bc2b9c2244248a8dbe7eb815fd912a
6295 F20110216_AABVST taskin_z_Page_058thm.jpg
75187b04345989e1ed2c208ff5fb3c8c
8b732d701888ccb339cf2ac4a3d4216d64ff5b1b
92095 F20110216_AABUQG taskin_z_Page_032.jpg
2b5e454e22da1c89d0fd1446c5e0a20d
68758ff7ae1ec3c729cb46f84f24aaf2c0bc55a5
7861 F20110216_AABUPR taskin_z_Page_048thm.jpg
e6e2df21e1b7b1e74e182a02ca0a8225
6aa6e1248208f5b00b0bb23ce955dfcba461aa02
7806 F20110216_AABVSU taskin_z_Page_074thm.jpg
f1b7b7fcf95ba3501eb564dbc9a29d63
bb1543c27f042553358c68600988cd4f1d3173ed
50048 F20110216_AABUQH taskin_z_Page_092.pro
7bd08e1f5a0b3a0989e8ffb094cb7570
5a193392e9618ff7b87d89b7d639b7021c8628f3
2280 F20110216_AABUPS taskin_z_Page_042.txt
e6fcd0dce21c642e99cf92d0249d492c
9dd691e1e66179cdb329cd65b4ff62f49b74a3df
7682 F20110216_AABVSV taskin_z_Page_090thm.jpg
40b3ac55d571ad6ef2c8138957e41593
70a36341eb4f306a6e139d248b90d5fb19baa9e3
58197 F20110216_AABUQI taskin_z_Page_013.pro
cab033a620b24a57c30faa6a67122da8
5cc802518d60d25fc72704cab1d1b15459ed73db
68059 F20110216_AABUPT taskin_z_Page_156.pro
7837f38cffc11e1e46a42790ba200fb4
d2b36dc69081e1aa6ab323d0e7286e904989d66e
7333 F20110216_AABVSW taskin_z_Page_091thm.jpg
c38978aae0682339ce6faea7842d7d41
20c30bf92b576ab4385742619da3a8486d0f6450
72799 F20110216_AABUPU taskin_z_Page_148.jpg
d3e50bd864d04cc084b42e7a17d2de96
749638bd1f75d3a769031e5e7ab1edad793b16f4
7079 F20110216_AABVSX taskin_z_Page_098thm.jpg
df134dc2f1396f5f944a3d981ad8fe2b
3cd00f00b829f9e5796e80688024bbababa3cdce
4885 F20110216_AABUQJ taskin_z_Page_081.txt
287cf15a422ffa2c2281306e3c2faac4
7f3904ca602b4959165937d7052c2d7321dc840d
2299 F20110216_AABUPV taskin_z_Page_132.txt
1925e45fa0972cbebfcf943e606eed99
915057b71a8dc9785a751c588a9e7cd08309ad54
8053 F20110216_AABVSY taskin_z_Page_109thm.jpg
7b8e7636c1f455d3ebc8c71bb42a7e9f
cb7a4e6926cec5c169647cef6cdb7270bb82ce59
36843 F20110216_AABUQK taskin_z_Page_054.QC.jpg
464a2e69d8b0579a09295b49b8876674
6146ae10cc25379c7a9c746bd960423337fcc660
56791 F20110216_AABUPW taskin_z_Page_040.pro
aaa054f18cd0a09c773a4783c4d876ab
9c77dea92004562313bc5b5583c99673dadfec5f
7274 F20110216_AABVSZ taskin_z_Page_119thm.jpg
e05d9c74f702bc78d47861c795f6c18b
d8498226e601e4a89bd5169a4122fad2981e1b20
69479 F20110216_AABURA taskin_z_Page_083.jpg
59ade60cce792cfb3cb2b12d0aabc77b
b7b9f3b5ffac5e318874a48cf38177754fa8baf4
834 F20110216_AABUQL taskin_z_Page_002.pro
f38db01bd7b400af9bf72c05123d1815
18f43bce15486785a22580c2e933000a6f5383d7
126131 F20110216_AABUPX taskin_z_Page_153.jpg
63c9a3925d3eb973f30a07b6f79e1dda
134ab62d329d29ccc93453421c96e646ae87327f
F20110216_AABURB taskin_z_Page_141.tif
961dfb6b4de3830445ee6acb25c06c43
6352719f6c91d42b50ce3a01a3cc99406d044f8b
1051962 F20110216_AABUQM taskin_z_Page_045.jp2
e297849aad62946953dbb790d6313e66
451cbc53d0da772b9bf9951a9847227c38b027e5
23113 F20110216_AABUPY taskin_z_Page_076.QC.jpg
dc7ebc3b7637b446c4ee10ca73847594
0d3c65275acc66ca4d4722f094bae825d9fd8575
28763 F20110216_AABURC taskin_z_Page_032.QC.jpg
08e4227a1ba94d6f1315ed9930a2215d
93458dade5b3d26a3b67f5dd67b87db8bd2fa86a
7734 F20110216_AABUQN taskin_z_Page_073thm.jpg
ee72886971de36ba19c055333a52dee8
0bd407cfeb85d05e34b11fde220b5a387c164b41
F20110216_AABUPZ taskin_z_Page_157.tif
833da12e6f78dbd28b8ddbadd0fdfa54
b271e65d49f0ad4fd8593a27a43632ecc6ad695c
108677 F20110216_AABURD taskin_z_Page_080.jpg
1dafb17ad4595fb915165e158bca6f65
78dd6f1eec7109ca430c74c8759e99116ea53686
44783 F20110216_AABUQO taskin_z_Page_025.pro
44478a98ab65717b42bfcd54bfe491ee
c19cfaea35ae3e0a9c98090921cb5a912a30af22
2400 F20110216_AABURE taskin_z_Page_151.txt
0cef54e3bb37bd4dde6761c5f08b987b
95dbc9f6bbae56ca4c500743fcce28db5518c531
31761 F20110216_AABUQP taskin_z_Page_149.QC.jpg
1cfed0a5b281bc90d931af51c2eee610
76a1081483e7bfe232db0c5ef8f6792b3c5a2589
780633 F20110216_AABURF taskin_z_Page_110.jp2
f0171c3e0b5823a70a81bef1b5d82521
ee3553194e52ff6d7007d52e37d1b8898f1b6e06
34029 F20110216_AABUQQ taskin_z_Page_064.QC.jpg
f808fb592d209aa26e577ba4b6bd4f7e
8556fcbeeb58a5739be5a99fdf6f1803ea5548b5
F20110216_AABURG taskin_z_Page_152.tif
aca05441d33793f6dfa0ce6792a03d4f
8e6818aaf818142669062d2ab4cfca169cc4c750
31471 F20110216_AABUQR taskin_z_Page_010.QC.jpg
36f5842efb6075b9785f0cf880f338fd
e729a2210e99d79e3d1805422858d2d463228413
35897 F20110216_AABURH taskin_z_Page_082.QC.jpg
001bb9a93210628e20f29a7a6beded0e
baa281f1af1fee48c8a6cbdff99ba56fcd2d0d23
F20110216_AABUQS taskin_z_Page_006.tif
cf15cec7e04ac2400ba67ae3e6724fda
74d33736c7747b109a686fe764a6306b7dd3df71
7275 F20110216_AABURI taskin_z_Page_120thm.jpg
2b62a1d77b81a9570409c2c699d246e9
4f593d748db3eb8e34fb4d1f5865c08c76d1fe24
94084 F20110216_AABUQT taskin_z_Page_120.jpg
b711f40d55386b07051983092cecc576
310f127e7768060d528485138d32b88daa598a84
2338 F20110216_AABURJ taskin_z_Page_054.txt
3ad4fcf28a4d5b1f9e07b875d8ba22e1
3aa149acd4aae0a81bc13c5fa7a863955a21e477
F20110216_AABUQU taskin_z_Page_024.tif
9061b6873840e12275955a04e99ff6d1
d86b4ee958f10000cdf1f4538baa8346f5927f0a
2229 F20110216_AABUQV taskin_z_Page_068.txt
bf3ef1fafc199758fb8cb03b9543694e
13b323c2507d1725d190299084d41531f0b77121
97492 F20110216_AABURK taskin_z_Page_121.jpg
47b2b57f6423625cf1ddaed37890e38a
720a385612eed416581451b5dead1f94b54e0b91
122521 F20110216_AABUQW taskin_z_Page_109.jpg
5a552d4817c8905e1a9f07bf5224f7d1
6399a5eded14a619f16688043e3a2647cc45165c
1051925 F20110216_AABURL taskin_z_Page_069.jp2
1a8c60d0fceadb2bd3f07b676692e366
61de53bc3271f7246391152d1c2546324864ea0f
29424 F20110216_AABUQX taskin_z_Page_090.QC.jpg
05b147eb03a0f4ba54a412d7e264e5e9
73299e388bd7f75180341fc0afacd9dadd813cef
25347 F20110216_AABUSA taskin_z_Page_087.QC.jpg
fcc7df9c4db10b5b97c82c9d2f33f712
e014723edcd9a5515029867324c81096d57f8b9e
30214 F20110216_AABURM taskin_z_Page_121.QC.jpg
d662f46ba26defc1cb64dfd1122d6f0a
d6cc2c900473963bdeb316b96e9fd3696d6f44da
46668 F20110216_AABUQY taskin_z_Page_120.pro
137d6923dd6934267e349258e9bf0af8
fcfa1f103798e7875a6f3fb12f680c6959964436
34990 F20110216_AABUSB taskin_z_Page_006.QC.jpg
bb4ba65db3fa60f1515473c35ee41e54
818ca29b65e53ab107fb71d78df32c8a2e045843
F20110216_AABURN taskin_z_Page_146.tif
ade3944353e1116b52e90d9193e0aded
a492e35dea487f08db8d4df099dfbcd846cb340f
33660 F20110216_AABUQZ taskin_z_Page_124.QC.jpg
ebfd725d49fda0817ab410ed496bcac0
eff78ceaaa27aab099dfe084ce3a08468c62fc8b
F20110216_AABUSC taskin_z_Page_020.tif
53d5c8bf8680aa6fbb1274192699b2df
5ce702b2ae8308e4457f3afe0762d3ccef677818
F20110216_AABURO taskin_z_Page_019.tif
663f2f3cc05cecbf25a4009c31ae667a
92d5b4732828d441cc4ddf9dc4ca368512dde57f
1051927 F20110216_AABUSD taskin_z_Page_097.jp2
6508f7eca40d5b36c2baec65447c5005
52f8ebfed255064a2c4d9a2e61f6837a2f793294
5098 F20110216_AABURP taskin_z_Page_146thm.jpg
085388bda527c7bfcca8a180aa2b6afa
a07687e3e59f9911b8e36a52e6dca073e1b0dc84
8399 F20110216_AABUSE taskin_z_Page_054thm.jpg
a2afac82c7028e51fba5907542aa428c
fb818da727931cb068f9029baa5ae6273ea497b7
120074 F20110216_AABURQ taskin_z_Page_055.jpg
f561465155abe9fdbe21514dc8f3d548
e6185492f2f6224b7b81e3ccd5b3b72d770870db
770 F20110216_AABUSF taskin_z_Page_112thm.jpg
6602d4519b8b1102bdbb48991ab56617
d6246ec2440611e9dfff747bce51a198de59ebc6
8504 F20110216_AABUSG taskin_z_Page_107thm.jpg
065f8a867f48e2ca38e99748dd0b48e3
11731f4ebc119b180af68cf67932f99889ed1714
2275 F20110216_AABURR taskin_z_Page_102.txt
c8bab6922aad23b14a1400c10a1a8bce
82e044949cbf52bebc953180bdadafb2478c02b9
25870 F20110216_AABUSH taskin_z_Page_058.QC.jpg
9620d84253393fb2400b2d7fb5310766
6e8e23a3d20c0e508e3c420139f94b46b34facec
49100 F20110216_AABURS taskin_z_Page_145.pro
e87b2f24566252c9ef83c8775c2fd9fc
b6a863f647c62403f3f19e212613b489b7615ee6
92945 F20110216_AABUSI taskin_z_Page_145.jpg
f16a5712a127b7d43f66312bb74259e6
7bee813c046ac259b377960349e1a54e9ff106e7
6764 F20110216_AABURT taskin_z_Page_079thm.jpg
f9db40afe4c37860f3a833c750d5b6f1
81f155cec4d2e61f523d1187284476feaccde2aa
2336 F20110216_AABUSJ taskin_z_Page_084.txt
b7087c82971ffab3c0de7151e17cf0d4
2d9de938ad6de207ae7fe7ab4d9a4a9b7b2e7ac6
7397 F20110216_AABURU taskin_z_Page_033thm.jpg
1783793ecbfcd3c59fdbbc3fc5ca1db3
9ef7200e48dd626ad1ee4f2f462d0a540049ceed
104075 F20110216_AABUSK taskin_z_Page_100.jpg
969ae9973a56e998bf7bdb1f9cb1e9d6
a8e091155d251a9103d31139b79ee78551e016f1
F20110216_AABURV taskin_z_Page_091.txt
c68b69f7f9917d4e85064c6b865365db
f2b7f41c0282c1db288956c15165a4a8db128dd3
56388 F20110216_AABURW taskin_z_Page_086.pro
d0c1ad0c77ce3b8a41c47fbd0c7a8b7f
ad79b36ae842f8f4d0b567281e036ac78477aba2
F20110216_AABUTA taskin_z_Page_140.tif
89e04e3587645407d5c980bdaa4fcccd
204e2d5b1edab932fc888081f6da14c9149793f0
1051961 F20110216_AABUSL taskin_z_Page_032.jp2
8c89d4d9c27a88898b94408c056e4005
578333a2120fc2ee0cdc26146cbc44db193b13ac
F20110216_AABURX taskin_z_Page_017.tif
1ec0ca25ca0c48e9dc512a53d23f70bd
218181abf7815e3f9973c36e31aba1a9097d47ca
F20110216_AABUTB taskin_z_Page_142.tif
ecab58ed350f11a899457988b72c4a15
f25f384b266b9d1879f8196faeb6e1fc55dd8add
2384 F20110216_AABUSM taskin_z_Page_150.txt
a96c9cec501a4cfe80b45ff31c901f29
257eecb7e7a39f75a124a2d5cf461a6f5339d48c
32688 F20110216_AABURY taskin_z_Page_117.QC.jpg
3702d1270cdd7b2536d29f329216fa12
b15522bd4a496505055fc6d683ac13b4c5b6d69a
57840 F20110216_AABUTC taskin_z_Page_136.pro
d483d889ef4ae1899f063ab80a2dbe3f
edb5e1ab2d60efb6de301794e6bbe1f97a3ba3d9
18415 F20110216_AABUSN taskin_z_Page_133.QC.jpg
71406bd5487d04d578f9102c613c4ec3
1f0a8dd36c338a856ceb6b420d1c63a3df73184d
34793 F20110216_AABURZ taskin_z_Page_040.QC.jpg
a5da2365584d1aa5e5b2c59eaadb8df3
fabb3e9748825bfcdb19de24409fb84cef002e1c
85044 F20110216_AABUTD taskin_z_Page_095.jpg
67d90a89b940a54a937679d3b0b457e7
70f0c600dc4b983b43698411bdf05cfd7964c07a
110873 F20110216_AABUSO taskin_z_Page_040.jpg
82f576f35593d7bd4c4e9069f2eb592e
4e9159e972333442456cfed2e3f6818855ab516a
46524 F20110216_AABUTE taskin_z_Page_093.pro
9816ca57987c98928ca25113d674f0e5
73436a0ddaee212c7779bbed73b0e1ed3edf74fc
117201 F20110216_AABUSP taskin_z_Page_046.jpg
e9b79d50902b13e847888df45e0cc790
96158339fa29b5b2e770afc6070888585887bc4c
32328 F20110216_AABUTF taskin_z_Page_034.QC.jpg
91936c0ac6e956fc7d14b15966c31d96
2b0a9c2c38fc5a5a3aa5b6bf036ebf34c417986c
56474 F20110216_AABUSQ taskin_z_Page_144.pro
bf89219fe4091be69fc394106b5e1e9b
abcd9cb37c495c33b7229c47f365c293095426df
8383 F20110216_AABUTG taskin_z_Page_155thm.jpg
ebfe803b5e4912d9951779d02ff4cbde
d0a2b3cddedac9e55f10a16dc96095d730bad995
2309 F20110216_AABUSR taskin_z_Page_130.txt
397880bb302b46eb800222f8f1398ff1
5ce7a6e4cb52bf13269714a162fb909f1ee750a4
F20110216_AABUTH taskin_z_Page_051.tif
70f5f7351ec008c4bba7f41290e164f3
79f2ea014703f3030318ead8a6b0142ac5fdd1cf
58157 F20110216_AABUSS taskin_z_Page_063.pro
659d316eb5105985ef37718721e322fc
684f22579f3549d41607c46e54ef7b278bb090ed
69477 F20110216_AABUTI taskin_z_Page_027.jpg
e3bb53f1a91abfee1a81598ba8b0c26f
679f7f53519b615b30576d462f31772fee6a1c1b
F20110216_AABUST taskin_z_Page_132.tif
b67f4b22cba947c5dab9eee8110d52fb
7772344856f0836e62c78fab5c74dde30221f2d0
1051981 F20110216_AABUTJ taskin_z_Page_123.jp2
0e7c9a41fc929bc0f706f916ae2f0a49
8d551d86c27549d7e2faa859bb410e18e040d092
2093 F20110216_AABUSU taskin_z_Page_099.txt
c9d73867b6e8b2cedb0957d8b55b8cdd
41b7b178d2f7d839dff37f926128f6a2504d3402
7661 F20110216_AABUTK taskin_z_Page_003.jp2
07b29ad36eedaa6c9943e0031bc00207
4164241a7989a0c57fbb25adf75c4ca7362d6102
1051972 F20110216_AABUSV taskin_z_Page_063.jp2
8c30c8d08ab5b635aeef1591dc420866
f6a6c3d37ba5bd1feac339d6417f28baa770006e
98223 F20110216_AABUTL taskin_z_Page_092.jpg
c36627972fb16e9739f4966348503735
682cbf9315b7c268912e982637e578e69f64f974
2153 F20110216_AABUSW taskin_z_Page_089.txt
88cf2e8e65ef17819914f3007c5f7c7a
eeed8bb2b1281999e4a6ced655dac8940d88004e
8144 F20110216_AABUSX taskin_z_Page_022thm.jpg
69a6c4f701859f915f5d43eea9ee9e76
15c2c986f4f90e40b7c9958fa201305f83e5c344
1051954 F20110216_AABUUA taskin_z_Page_085.jp2
1fd6abacf41c4cedc3c9c34dd2588b59
c6d4699793dd6100254e3b4307c12db2a18689cd
25189 F20110216_AABUTM taskin_z_Page_105.QC.jpg
fc3b98ef55d338825def89cbe783a666
61a3c33e7fc094cb6d370b6bce8a55db16c532d8
41662 F20110216_AABUSY taskin_z_Page_125.pro
912b586cc7b59ce4ee74d0d427f3ba2d
259107121eb996b2ac9eb27dc700d15cd076eea9
8269 F20110216_AABUUB taskin_z_Page_031thm.jpg
6e8bd5839edbf088c70e32a84adbcc4f
1de469915b8577bd69277391ce8c65c74006bd42
110120 F20110216_AABUTN taskin_z_Page_010.jp2
f72528d4ac459d3d349a313f2365456c
f172a5fbe62255361f0074e835f7a8e029db6af4
1051978 F20110216_AABUSZ taskin_z_Page_130.jp2
5d04281b5e8b53eb6b28ea17274535a3
08c294751c487628081ee8e16f832fa5e7ef1852
2462 F20110216_AABUUC taskin_z_Page_011.txt
f0aab7d739242615f6c4475db9ab4068
b0828fee2095a631001c300190b442113fb5f12d
50381 F20110216_AABUTO taskin_z_Page_019.pro
412e885d042181ec1272bb948ce473fe
4797e17656f0c02f6950a195034b8bbb6fbb4b18
6575 F20110216_AABUUD taskin_z_Page_125thm.jpg
9d3237e1ae3a5fc0382330bc7b45c8bf
5ef91d7b1fb3a9b5d4679b54fe89a47b23b1085f
2420 F20110216_AABUTP taskin_z_Page_127.txt
ff860e7d284c96a843ecb014dfd90c2c
e958fa411217d99f0d6216221679382baef41650
F20110216_AABUUE taskin_z_Page_099.tif
18a02f1e84757b68a8152684d27645f0
7ba5156e3c5b11622eed3b43a51d729d45d1b10f
2355 F20110216_AABUTQ taskin_z_Page_013.txt
4916ba0f62690201e626291361a71ece
519baabe11508d6ec76b81b5550ef4566077506e
2357 F20110216_AABUUF taskin_z_Page_021.txt
a4a70d9effdfa78e766e1656ba88422c
8826075fd45b4d9d90417417b178b94af3107ce0
1051947 F20110216_AABUTR taskin_z_Page_119.jp2
1ec40050b6310fbc130e2b00cd7785b1
c6ef2c6834e5002e6e8eabbed9343ce45c519442
F20110216_AABVAA taskin_z_Page_112.tif
5dc628d3b73f9adf4689d63fa7d60765
5fad59974b1ab8f3cf2f9ce0f37b59ab0ef85993
F20110216_AABUUG taskin_z_Page_022.tif
00d4771e8870534fd41f20e7301971c8
947883be1a0f8b94ecaf913e5e28da5b5fb9e7ca
51016 F20110216_AABUTS taskin_z_Page_033.pro
77d79c42e4a2fde8316e6a1bda0808c2
20a55c3901d21b0bc1cd07d1031fd3918f14b70a
F20110216_AABVAB taskin_z_Page_113.tif
1d72a0c09fc6ad1a2493ac65a91b1866
00892e33fa1c2a428b33658d36488649d9e44943
88072 F20110216_AABUUH taskin_z_Page_004.jp2
43299968997b05f43572cd3be58e6319
6bc6d622176298f8df28f7267c6b6f4be3d46cb6
F20110216_AABUTT taskin_z_Page_054.tif
0ad38d02167c0518c8376571ff680077
27c2c3763aa0163aa47479563f3e17389877b8c3
F20110216_AABVAC taskin_z_Page_116.tif
8d28f452588f833529a8b049b0a22281
b6d3043c4ecd3766a505f8936e39b2711c97270b
8449 F20110216_AABUUI taskin_z_Page_052thm.jpg
9b6633517a7e2b62ae7bd06f3f68759d
3fabbce007d98cb112a2004dfce47af540cc4613
108977 F20110216_AABUTU taskin_z_Page_047.jpg
d81ea0f327dcb521ebf7e575d74808c2
35efdc6e1eeb9d0bccd51d50a934cb23dc4dcd3c
F20110216_AABVAD taskin_z_Page_120.tif
6bd45644ca39063d5378c8d3df02b96f
14ff3ceb6bd3d990d3cd52132bb5886bb758873a
2176 F20110216_AABUUJ taskin_z_Page_059.txt
d27d57b745b44df381d7564f4b1f7977
c902032ff8682d49761620eda7a53c294a5ba423
954542 F20110216_AABUTV taskin_z_Page_142.jp2
fea56556f3ba53298068a5870bacd7f6
e941b008e96b3026b6e3b9f20572321450b64255
F20110216_AABVAE taskin_z_Page_121.tif
c247ee8d28289106c96176adf72a23af
9f00dd93a59deefe2f08bc3d5bce2b7cf03cc58b
F20110216_AABUUK taskin_z_Page_158.tif
147a97f9446bee702beefc7b9c1f5c69
3f748eb94131a2efc0fbb8fa71288f73d9caba3b
1051974 F20110216_AABUTW taskin_z_Page_017.jp2
425187a093e59e02e7de0a3709a9c18d
f1ed4784b56f4285c72059dc9a4c498e7b2a2d95
F20110216_AABVAF taskin_z_Page_122.tif
924e29e65591f8a6b16bbbb263c56e07
9d67a817f7f5bf164503f16b251ee59c7dee3746
1093 F20110216_AABUUL taskin_z_Page_118.txt
301564d09aabc320918f28c48050a1b9
31402af2532af739b27c8ebc26912f3201ae6f44
84308 F20110216_AABUTX taskin_z_Page_057.jpg
e9288aa296bfe4016307d98c1cee858e
2f3a07c3b6301e7c5cd6a008cc8b83037d0344a8
F20110216_AABVAG taskin_z_Page_123.tif
696a44832784246bb7558b6a3d2ed13e
8a059e49046ae38757e71f5155d3afaf565ccab6
78128 F20110216_AABUVA taskin_z_Page_058.jpg
c7265c662ab3c48b063f02ad298206ab
e77e8d2edd6562cb73bcd06e8b1f2ee182aa5faf
8207 F20110216_AABUUM taskin_z_Page_138thm.jpg
13c72144097653e23dc616bd6f12c699
87e7efc99146e30735571743e1d5ff44f0654211
43017 F20110216_AABUTY taskin_z_Page_030.pro
abb460c2fa3b515e6911e4cccf6c189a
aa832c69ef515898df9b30f39f45877da7105ebe
F20110216_AABVAH taskin_z_Page_124.tif
6675bda096e583974d00043dac690f7e
a19bb5fbdb67efc2ea1e3387e4810b7d845cc5e2
F20110216_AABUVB taskin_z_Page_027.tif
955078d6fc8c2a225a4946d42a522776
a37e184dd71d8e2e8653893e7dd0c2e39cbfa168
108144 F20110216_AABUTZ taskin_z_Page_056.jpg
886e4bcea19bad3f283a5a7d1b8583be
0d2987d203b00c6e420ee7942b9095e9f6514394
F20110216_AABVAI taskin_z_Page_126.tif
4ebd7a299a7224024b693c03ef9f6f9c
8854c2976157137aa51445ecfe18ff5d82d23c91
24832 F20110216_AABUVC taskin_z_Page_078.QC.jpg
691fcd9c5acf013cc2474086d1656ca0
c52146c1f80878078d1f5b2b464cc3107b792377
2056 F20110216_AABUUN taskin_z_Page_100.txt
4d0a11a3bf64303f2738dc6801931027
c3630d8e787ddd2080677ddb2a6419c90a9da96b
F20110216_AABVAJ taskin_z_Page_131.tif
6df6228213c3d2b86be0b5e4e916eeed
f4f9446fd24588f0213c71b225b834364e2e1470
119439 F20110216_AABUVD taskin_z_Page_005.jpg
562fd2ed6de57e1dea466e3d20e9183d
ea882eb8f246962b86fdefc487c560671602c407
1051980 F20110216_AABUUO taskin_z_Page_141.jp2
2a300c074b380c8a67e72c5aee702de9
0552831f0f6d7a741e756377377be58f1e7ab32a
F20110216_AABVAK taskin_z_Page_134.tif
4644d2ace94f00d871c8a8015b3aa715
f0039e204823a9692df27fb0d7ef25ddf3ece38b
1051889 F20110216_AABUVE taskin_z_Page_072.jp2
6e00ed35a3f9ef4a2ce767679d39de2e
6481fa7fa2fa2fd39fef5fedb4d0c23eda52ad3a
F20110216_AABUUP taskin_z_Page_015.jp2
f9e201a4ea4d8d9c04cbe1e2eb4b659e
ccac81dca6bc1fa63b28e8167bb1982b26f1bebc
F20110216_AABVAL taskin_z_Page_136.tif
d7a6853cfb3b1c9fe488ea09f0df2dd1
3c24c803d35f08f27e9a32129f68d1d1c36e46cf
101236 F20110216_AABUVF taskin_z_Page_123.jpg
e06a29b7bcab94cf141ddd8831c9dcb7
a015e8c5ada659d7119d05099ad4757aeae0f692
1051968 F20110216_AABUUQ taskin_z_Page_139.jp2
34e812d2ce631375b1f276f7b0887a98
26ac821687f86d7b0e0919cf068b12b82129f4b0
F20110216_AABVAM taskin_z_Page_138.tif
5e487a6f9c9d22a468621c22c7b1b3db
5ab9532f3be893896fe05a84f41547996ba7ecb3
2279 F20110216_AABUVG taskin_z_Page_055.txt
b42d69ded7bd8b007a99214edf3598f4
1e50eede84ffe2df6fa83737e1d26d6d5cecf0b4
F20110216_AABUUR taskin_z_Page_069.tif
468f58d4fb69acde9f29fd8b8cdfbe1c
03e05c12f240d2282c4f33654cc282210f60b778
1487 F20110216_AABVBA taskin_z_Page_009.txt
8ea45cddf5d60eb6ae315456facdaa5d
82184ba25f03e6ac162151a099b18a609295450c
F20110216_AABVAN taskin_z_Page_143.tif
29396a409811a715b9185fe76afd8f32
73ee93f01e798244d3e4ea9bc6b96ba915c11ede
38920 F20110216_AABUVH taskin_z_Page_105.pro
70b93d50032858a4926cf990e91eef66
26ee82b84b0ed9a39cf0e5fe551436fc393aa6cc
43312 F20110216_AABUUS taskin_z_Page_026.pro
2d2505d1b95bea6388dc1b7ef32f49d8
ac93968e40144ee8b975ed8e91a6b995af7df828
2417 F20110216_AABVBB taskin_z_Page_014.txt
0097661fb239bf627cab1cdbdc0f27a3
2cccb2e45950ecbf7f1c16deb114edb385e2f205
F20110216_AABVAO taskin_z_Page_144.tif
64ce9cd16e4c6781749b4f63c04768b0
7ca13878b0a698ddcee66d59bca0941ed29806d8
1958 F20110216_AABUVI taskin_z_Page_095.txt
e3db3c69c833142c80fe2bf8418b69d1
04fcbca61013b835bca840b1f6abfc65650449e3
F20110216_AABUUT taskin_z_Page_145.txt
3cea0962559359431e8e8cd5d5f9fe1b
d11c01c269414e3b91a0c522eb96b616d4bd4fb6
2322 F20110216_AABVBC taskin_z_Page_015.txt
4f36bfaee6263e73d9fead6c3dd53410
fde73dee381715c5bcf3b0b252e140cacdfd03b7
F20110216_AABVAP taskin_z_Page_145.tif
14bb1d5de50a4ff3e1dc0337ca4f16a1
489ffd4d9e5af2b7eab4ae8c1c0dd694c5d540a6
30918 F20110216_AABUVJ taskin_z_Page_076.pro
78e258fe07071b790176c1cffbd93841
529b83b535f51876fde8d238fc378a819269974f
1029677 F20110216_AABUUU taskin_z_Page_147.jp2
3aced3c0548efc8de2c0c30df11c10bb
a88c30c28d4c7f96301b6bfdee726da36286391d
2414 F20110216_AABVBD taskin_z_Page_016.txt
3836a8a90b6cbdc84e6e78c81499fca8
2bbe37e5e911e0d96f7c0d2c50ad339c775a4c0d
F20110216_AABVAQ taskin_z_Page_147.tif
ad8954d31bc11a3c2feb802485cd3ddd
94dc98d2ccb88c6a67117dfb2d6186ac5514dcb3
101873 F20110216_AABUVK taskin_z_Page_062.jpg
820a567661927e91ba1b4cebb792485e
a3c855c6864b745419603a46069583aca9b0f867
F20110216_AABUUV taskin_z_Page_029.tif
30a0b3c9c3ced0f4b82cf4acdc2952d4
808d5d65237fc4fad0fe15a29c1877b70eca69af
2274 F20110216_AABVBE taskin_z_Page_019.txt
5bfb27a29f6ce4b00731eb42d5727a71
e23b58781f4022078222cdf6744b63edd0a255f5
F20110216_AABVAR taskin_z_Page_148.tif
14cd66979285d751442e047f22b6f339
79c3a02955127a502a20720c85609be4a1ffc33e
29454 F20110216_AABUVL taskin_z_Page_037.QC.jpg
337efbb23c0359d0fb07f8bcf2ae4d54
8a13fcc1797f40595bf35f4bce1f349c0b68b071
37070 F20110216_AABUUW taskin_z_Page_108.QC.jpg
7e7fe3c95a67d54bc121657d70f2c7e4
a98d5e3de0332e4cb52a65d304838fc989443c03
2276 F20110216_AABVBF taskin_z_Page_020.txt
c17a29ac82e6765dd03dce4f960a02d1
b8522ec26825c6d6a9d17c1b36b6917957b6e3ec
7965 F20110216_AABUWA taskin_z_Page_013thm.jpg
d3511508e30e00a22156ba6874c9601b
9dd9926bcee58882a5cf7414af63c14d2f5895fb
F20110216_AABVAS taskin_z_Page_149.tif
247d9904bbf26bf8a2145714141964da
2678faa52e56d2a752192fa88bf27887cc6026d6
32623 F20110216_AABUVM taskin_z_Page_062.QC.jpg
62dfed124622869739c0685511e43bf9
109fa877b482cac602580e7bee6ea3885ba77784
110058 F20110216_AABUUX taskin_z_Page_041.jpg
dc23eb89df7a87a9dc8643c1eff73f43
6bf4600745014bde959621c27ee7775446f366ad
1858 F20110216_AABVBG taskin_z_Page_025.txt
e3873091767643acb2aa29a236d8d51f
c6caef76e070cc423530845afe5c42eccdf2e7dd
7730 F20110216_AABUWB taskin_z_Page_049thm.jpg
7473754d293a5c144f545781e3c1bc1b
9b8fe259e81872e542002d846b27317e90c6c357
F20110216_AABVAT taskin_z_Page_151.tif
61aa4a3b8d5becf33ae07e463e3da533
90a053be7ad43a21ae7e0987b0752f44cd36d5d1
1982 F20110216_AABUVN taskin_z_Page_001thm.jpg
abcc51d73ea5815d05ea90e22ecae92f
a2471c09df2224302f6fa1c34a0c77b5d8dc01bc
61250 F20110216_AABUUY taskin_z_Page_071.pro
ab31c2ac7eb1ce953f49b854139048e0
8f30dbccc1d06fa5afaece73e941b857f94bc5cc
2302 F20110216_AABVBH taskin_z_Page_028.txt
ab28c4a8981857515fbd9bf37f8cb178
08a3ba771fc5a21818b12603e4abc17c67b8744d
7433 F20110216_AABUWC taskin_z_Page_032thm.jpg
404dd6e325163b8b75341fb274d583f9
c1e9f555d66dfbed9bb724b3b4c621f50e38a852
F20110216_AABVAU taskin_z_Page_155.tif
f77e2d416e4a9ddec4f1a8380391442a
ab646950863465400ca5417acdf8bbc3f6db082a
2412 F20110216_AABUUZ taskin_z_Page_107.txt
4f3ed78963afcad284e535a8e83b3a4d
98ba3ae93c6a45344b7e67c74c8f2d5f9803a968
2110 F20110216_AABVBI taskin_z_Page_030.txt
3040d2174d3656b093a0e97ef793069e
9786f9c3b11e50ccce16e7d11529ebd6835b4faf
7923 F20110216_AABUWD taskin_z_Page_041thm.jpg
3f788a58e7900a3d8779b0274417a7de
cd0e3494436428ed0354b02b486bd33ba3188edc
493 F20110216_AABVAV taskin_z_Page_001.txt
0ba8dd14d24f857b1777a07f911c7e21
9ac50e7297439bdcc80162d8e43e728f3eb5e5f6
F20110216_AABUVO taskin_z_Page_092.tif
11acab33802e19892c79c7a3d407005b
bd42562922e868a01e9eb6a46ac4a8a4b6761584
2190 F20110216_AABVBJ taskin_z_Page_033.txt
d9fcf7eb8e88304babac8b3d9b1157b4
42251020971ace44ee74c7a22de9ab4425a8b322
122881 F20110216_AABUWE taskin_z_Page_107.jpg
997793bdf36a295d694c3cab2b19fcd7
ef80b795f4fae758ffa631d9207e11ad90398fb7
134 F20110216_AABVAW taskin_z_Page_003.txt
900cace0139c97265eee25af9bb8dab6
925fd77b3ae9e9acf609e5044bdf38695df5abc3
7458 F20110216_AABUVP taskin_z_Page_144thm.jpg
0d2b2d8d62094a531127b78dae74cd7b
5e6d4153674830bf62900c2105ebe2b21ddab119
2233 F20110216_AABVBK taskin_z_Page_035.txt
5104db2420a6cac9d21aab849f02ddc3
bc7caf6ff45108a9d74401b7bbab3b4388ea1472
741769 F20110216_AABUWF taskin_z_Page_077.jp2
86bf19a2f3d39a93712992ee302ac154
e65c9ef1cfb0a2fa41b4b6643f9f53aa69f03596
3041 F20110216_AABVAX taskin_z_Page_005.txt
a4a852c5d3e848c6a69ef90d795b9a18
b6cd2cf078dba145434571b2ab5448898e155a71
F20110216_AABUVQ taskin_z_Page_048.tif
8f36f4fa784a5641584c69d0bc246e81
393ea75563f3e99b5b03004de6ea282c90891bc8
1900 F20110216_AABVBL taskin_z_Page_037.txt
a278ce8511b4001ec9380d525a092e2e
b4ab4172b88faa6b3aa3f42af2681f53f10665f0
101897 F20110216_AABUWG taskin_z_Page_097.jpg
104aba9c5bee04d1c4412c1d380c5429
51136f5b1469ef6bdf4d2a9a63b041be73e89613
7884 F20110216_AABUVR taskin_z_Page_124thm.jpg
4684a98f5a15e943b9d365a0d81e20b1
33b634f69050edf93c9a6f42fda916ffe33082aa
3097 F20110216_AABVCA taskin_z_Page_075.txt
1e7b48a3b478d688555a7005e2b29256
175bf7834de92bdaf82992bb128041ebc771c92c
2049 F20110216_AABVBM taskin_z_Page_043.txt
09efb53fd691d6192111f60e4a631275
56d5007b4c791ee75537caa04554ed9166165648
874826 F20110216_AABUWH taskin_z_Page_101.jp2
b2fac873c356350d0d4660a0564ab747
661fb0ff5e54a7ae8fae058537c0d902053b1909
3450 F20110216_AABVAY taskin_z_Page_006.txt
61a6da45ce9e03acc544821c6431b790
0b28f16df75f7b0121acae03ccf03a196a67e67e
269 F20110216_AABUVS taskin_z_Page_018.txt
ad9042b126ae562c14466c031b591ba3
9b647fb124fee41f3aacf3de2787e8154e74e06c
1705 F20110216_AABVCB taskin_z_Page_077.txt
b71ff0e9ddca5795a71d39e4f3f7b44c
45d9871ce2fdf79930c3620ec2047da63b41791e
2062 F20110216_AABVBN taskin_z_Page_044.txt
85286295552f0f2dba9cabeb67f9cddc
69c2ed115ddc5914a8dfe26fc51b057152d83e9c
1051986 F20110216_AABUWI taskin_z_Page_106.jp2
a80df7183412d4d7f4af5009b869fdae
8d3af178db9bb60b860e278ba16b861a6e03c791
2416 F20110216_AABVAZ taskin_z_Page_008.txt
52d0973b3b8f5a96df7642a650813131
d93facb771508f7fff8f6f211781fae35736243b
56146 F20110216_AABUVT taskin_z_Page_028.pro
b7e41f39d2d8bef5ed17ab722c731510
0e876a71ad75f44a771bbbdd5f734e47345c3b81
1738 F20110216_AABVCC taskin_z_Page_079.txt
edb8072ce1efca3dd0a1fb7221cb0681
632a68e8dd4c8fad5ebfac9b1fb8e136230a1965
2511 F20110216_AABVBO taskin_z_Page_046.txt
45aab1366f05468086e5e0a3ee656f7d
64b070e1826b6aa1211f0737d0edb12d33fa1a0e
2326 F20110216_AABUWJ taskin_z_Page_122.txt
3042a4ab1d33df767d7a806df57d5bab
dbdf85ffe4f5a1fcd03d8e7a02289c2917b59423
124046 F20110216_AABUVU taskin_z_Page_155.jpg
94a22189487487d71c04c4913f77042d
d21d8306c40c124a2d3e4001b764d860117e9c52
2171 F20110216_AABVCD taskin_z_Page_080.txt
2918c6368c0e6763c32db75107f95a56
a6a6efcb2db1e3e45ced2c427c8da4f235671e4d
3169 F20110216_AABVBP taskin_z_Page_047.txt
b8e4ece47df20acad5321d6620b78fc8
d393816cd683f83edf11625a205be749fd595020
66847 F20110216_AABUWK taskin_z_Page_005.pro
3ae7d1653c42517b66b2549f1403ffc6
c477fe0f60ad622dea9edf64767c8ca6af822ca5
2268 F20110216_AABUVV taskin_z_Page_069.txt
daa23ce842a5d2282a3c2be4edc94391
ac11f13c8314c04d9c393d130b02a0c29cbcf5ce
2290 F20110216_AABVCE taskin_z_Page_082.txt
6fc90b3319da72690fb2109515269a0f
64a837d79ebbdeea489782fdaa24d433b4a301ef
3254 F20110216_AABVBQ taskin_z_Page_048.txt
900130ebcdf7049d99b24e903b875672
84ab54625c507c7a5273f5179beb127863dcc6d9
7118 F20110216_AABUWL taskin_z_Page_068thm.jpg
c8e6f4f5e64eaa86b21304e458d4b7a4
27615058b03816f512182998c9721e1352ba2bf5
2228 F20110216_AABUVW taskin_z_Page_067.txt
a8b14c2ef4684a14bb2bf1d56c1c3e80
72068267afb64b54a5b66d02cacb299a6ca1c609
2872 F20110216_AABVCF taskin_z_Page_083.txt
85c27313d3245295a56a56c3dc29318c
e4464a318a1b6ca6574906076983399d42475bd7
3929 F20110216_AABVBR taskin_z_Page_049.txt
53e01df0125ddb1f367b1bd73942de00
2c69d0cfd9ede7d046672011891cad275b5d9393
6712 F20110216_AABUWM taskin_z_Page_018.pro
622369d9e1b8bd4e5a3e6f8c93754267
a108c4b8cfa85c4c0bbaf317b26120b22faea2cb
143343 F20110216_AABUVX taskin_z_Page_156.jp2
463d2728bdf4cd7cdc932792c6cbf05e
4bd032a60566519fd75c870e59b54e1bab686c1c
2395 F20110216_AABVCG taskin_z_Page_085.txt
5cea69c9096404d9def4f6bcafa6b8af
dbea092671936750cf1f83ff8a3772cc4586167e
35088 F20110216_AABUXA taskin_z_Page_155.QC.jpg
751470f598d02886f3419da7cae53876
bee7b3fedebc11370d200c49b33df2cbc4ee6ea8
2175 F20110216_AABVBS taskin_z_Page_056.txt
3eff02836e29d95fa61ed311e1f56193
e5eb28e73c11d8adff51cca396f2680b8552c04b
8261 F20110216_AABUWN taskin_z_Page_062thm.jpg
d13af78dab8858b613282baa5709c0dd
50f2f7c267a9bb94a38a74b95434c080944f881c
F20110216_AABUVY taskin_z_Page_032.tif
328c361f06b55c627f0afda985e223c8
e45e89be5e888a2f28ecc147fad097832ad4d2cd
2243 F20110216_AABVCH taskin_z_Page_086.txt
c6624c51ca3bb8671c9b4f9dc3d5e603
fb536699eca50deffbc9852235987246d745f5f6
8186 F20110216_AABUXB taskin_z_Page_130thm.jpg
b1b7c29186b4a567e728b3aa03b36109
525ed1da1850f0a1039d1afeaa76e5b619a1d951
2052 F20110216_AABVBT taskin_z_Page_057.txt
7aa4897f4cfce2d43bcb59feb7185692
d8a5f81bfe381c251613da1e6c0799685a14d71f
60788 F20110216_AABUWO taskin_z_Page_083.pro
1ff0382d77e12595c732c24ae31cadd3
50dcbbf2a029994bd5a9a6a19ab57861c2f387e3
1051965 F20110216_AABUVZ taskin_z_Page_040.jp2
1070b6fa1b0fe82317dbb26404fa4a85
72fd0c430be02223fa9b953c879421b5b16218db
2038 F20110216_AABVCI taskin_z_Page_087.txt
d22204503b67878074f1df9d80607a3a
f747d0c1338165947e5aa108b78a4405f4d0ca57
116961 F20110216_AABUXC taskin_z_Page_017.jpg
a44481d50d58ed63dedc1c248d759fc2
dd0607ee9786a34a590de71248b8e7a21e9418bf
1992 F20110216_AABVBU taskin_z_Page_061.txt
7ba0ea4fbfe51513c81e76cb074b2ddb
6599900114b9e06e2d8593a6878f2a40df0018a1
1949 F20110216_AABVCJ taskin_z_Page_101.txt
edb0c031df1d75df058ae62b53397653
05f889815b2c3c5cbfe49191d9856f9912952c2a
44622 F20110216_AABUXD taskin_z_Page_095.pro
27dc54075dce95e500f425c73b0c667b
292e2dfecd2cf5538f8e316ce1b933c54e85a8cc
2348 F20110216_AABVBV taskin_z_Page_063.txt
90435a3c8a090f30860fecaf8004d42a
bd3166e217ed244f27df917d8239328ceab28fc5
1920 F20110216_AABVCK taskin_z_Page_103.txt
7307071ab93e11ad9278ef48a63d421b
d869bf92987c77647e27f06ae89db10e32658f9e
35804 F20110216_AABUXE taskin_z_Page_069.QC.jpg
31482b2161a3cdea0e8ac1a9aac145e6
c65096e4f34f54f9639ed6f9d033f36daefa0338
1584 F20110216_AABVBW taskin_z_Page_065.txt
44d1ffb15a14311c056cd9b1e87635cd
fb7bdf1b65eab8966f07036d47da3cec3fd47054
808815 F20110216_AABUWP taskin_z_Page_087.jp2
a1990093276b70c5476b402d04ab90ca
4633a3e8c5b7d14c36309ea38d70a9789ff61d05
2004 F20110216_AABVCL taskin_z_Page_104.txt
748f30ddbb17f48deacfacc0015c2782
b1e031edeff68a05969e48b62460f92b17de598e
55155 F20110216_AABUXF taskin_z_Page_056.pro
9bcfa8bdbd8cfeef9d380de98b2a0d4e
d4a94a8f80b93cf9b570a87000e623fcf914ddaf
2567 F20110216_AABVBX taskin_z_Page_066.txt
23b2e7b173012c1f5bf8a51e6f2b5ab5
513d86e5f455dcf2a979c52e414c2f9166ab285d
F20110216_AABUWQ taskin_z_Page_107.tif
d0dabf9a29d674cbd932795f0b4b3218
961baf01bc03de0e61475ca7380beea951280b66
F20110216_AABVDA taskin_z_Page_136.txt
31f9a893b1412474e8635aabb95565f0
795366539fdea6edec299530fc8512907247a9ae
1976 F20110216_AABVCM taskin_z_Page_105.txt
7696043c59aabde6fb26bb5ce8da3569
c1fb33b32d6ec1b60893aacdb140afbb0a2b68b9
8056 F20110216_AABUXG taskin_z_Page_108thm.jpg
f1b3743024e8b32c0bb1113ab7d81ad6
4afba871f7fea8a54608b566e36c0b5f6d1f3988
2399 F20110216_AABVBY taskin_z_Page_071.txt
7e131d87b7019f9b1a04e6dcc84b611c
7589a012e66e7de9ce3bb9ed372028668249fad3
26015 F20110216_AABUWR taskin_z_Page_026.QC.jpg
7d6d7e2e3cc2e0c06523c2ef306351c9
a4df32558e73730059dbebaeed9b7161b415752a
2386 F20110216_AABVDB taskin_z_Page_137.txt
37de3e6ca527ed5ddf2b2cf13845e2f6
beba9516676065e2b27647ec2dad382bea0bf3ca
2173 F20110216_AABVCN taskin_z_Page_106.txt
f921dc85b5740a1dd16abfab76e70a23
f6e238f6c6bf5da26efd6d72c93e226e33f42361
2065 F20110216_AABUXH taskin_z_Page_032.txt
39f24aa1d80e2fe266f9b3f3610972ea
3e5c7190a71e4cb11a25df2e2fe52d02c57a4589
107304 F20110216_AABUWS taskin_z_Page_124.jpg
b1de73c6baa3c862364911a1ad66fd2c
450136e0f4746f153956b153c463537fd1d2862e
2308 F20110216_AABVDC taskin_z_Page_139.txt
8aa525ad2001b32532f834538484a4b0
abfee61196d1e2f523eb61be082e4d3a4bf95a19
3740 F20110216_AABVCO taskin_z_Page_108.txt
e9e56869ba555dae7c0e31284f587508
6c1c6d16725d905785536f974c3f5939bb4939fe
F20110216_AABUXI taskin_z_Page_071.jp2
00c8275052c2cb773886a13a312ea33b
1ca03d581131b6e760a7ab8bb142dcdeb4f9bd2e
2371 F20110216_AABVBZ taskin_z_Page_072.txt
a4488bc8989cbc5588d743bc4d41ae7e
dcc3341512761e41cc0cb309273d7b52117c17f9
36407 F20110216_AABUWT taskin_z_Page_020.QC.jpg
92e3af8eb27c044586fe6625caae51dd
1ab398ba9d867bec70bafc7056f9e8b15eff2ec1
2435 F20110216_AABVDD taskin_z_Page_140.txt
971d90e89e4767522225906ecc1668d6
3a3fe86e51977eac8597f485ca6adc25502ec38e
3686 F20110216_AABVCP taskin_z_Page_109.txt
51261599c88aac3c11751e63887f5d37
5f472ed12a22d24d78c7b38ae8ffd393816aa835
102105 F20110216_AABUXJ taskin_z_Page_051.jpg
5cf3f76854a595b352b36d9407225366
8ca809ed2f1fd6ebb012766246d3ca7a58e19d29
1051952 F20110216_AABUWU taskin_z_Page_041.jp2
bf8942416f463a3f78901402deed8ded
5dc324f080dbfd40410c15ca935ce916395451bd
1917 F20110216_AABVDE taskin_z_Page_142.txt
bf0dd95ba8478a54eba4ad7469c8426c
0a91e480445f56edc52236171cbe3fc5e18dbbe1
1664 F20110216_AABVCQ taskin_z_Page_110.txt
14054a8ea040b19281e775455637caac
1038fe1b3f425669049c8899c27115716bda9cf6
2240 F20110216_AABUXK taskin_z_Page_040.txt
c14fb715068affc343759e52b5a9b713
99f1a2b7be42ff297b1c0a0ec967c1d1c79d0369
61860 F20110216_AABUWV taskin_z_Page_152.pro
fac4478da6f840563fe05c0bf56ba4e2
7d81547e60bbbc602c519978d81e8ce770c74269
1889 F20110216_AABVDF taskin_z_Page_147.txt
6fcaf833a0f205b6fef592e63ee279b1
4fe14ba5bbe88c95d43c1aaf352947daa4fd472f
3954 F20110216_AABVCR taskin_z_Page_111.txt
2cb55b989d2050238c469495ab8fead3
6a1e055905fcfa18f6ba67ca52d0051d15483cc4
56808 F20110216_AABUXL taskin_z_Page_038.pro
b8f9d8f3356974d0b51f27201e97f1b0
3f62dc41f72269f5e003360e1982b1510f5a4bbb
79484 F20110216_AABUWW taskin_z_Page_079.jpg
a223781155d6f1413b91c318904619a6
fcbb7979b1258e223e0de32bdd7a0cda4236769b
1552 F20110216_AABVDG taskin_z_Page_148.txt
e6d25303fd5143f1509463dc142fdaef
e0c611ff8f438e8d5589b099c09fe882a0a732f8
2301 F20110216_AABVCS taskin_z_Page_113.txt
6e1f929ed3d87701bc8ed11b9a6bcaf1
aa12112e55568f3bd169a54aa2b92ebc01e34ca1
34916 F20110216_AABUXM taskin_z_Page_070.QC.jpg
8e4ce378044dbc2c6ea49c47f4f010aa
79685b08323a123f0c7191be7f9d5ae762c4bd03
2006 F20110216_AABUWX taskin_z_Page_097.txt
a90dc4dd2e695ea565272a3c440496d7
3b7b934cacb121d434f804beaea2155448d6d009
2235 F20110216_AABVDH taskin_z_Page_149.txt
7437735e48cafa6bcb1acd25398b1a94
6ed3466130098ab29227e7a68369f703ce9f5358
F20110216_AABUYB taskin_z_Page_003.tif
1687e08a927ba580f1997c7b105b3ce4
5bd3a322df75ffbbe62c5d3376c63c18a6411985
1963 F20110216_AABVCT taskin_z_Page_119.txt
807ac9d08eccb09b5e7e66f49664106e
4171c2d0acee68d07f1f1ad31f13c9fe7bf6ded1
8086 F20110216_AABUXN taskin_z_Page_122thm.jpg
625e3024decf59bfd510d6390ed48b65
0e469de33f4136af5ddea84183a51a92e9dda0cf
1027879 F20110216_AABUWY taskin_z_Page_044.jp2
93c7aa2df80a03052216916e51b4c1f4
2670a7064d1a50d784a6e97bb427e91049e5fae3
2496 F20110216_AABVDI taskin_z_Page_154.txt
fdd99cb0261a636600c0f0bc647e2694
cdbcb574c6dbf5e927d1ead347d282ec1b524c5e
F20110216_AABUYC taskin_z_Page_008.tif
aff64398b47eb95eb7ba935fc11a7c24
4d2ea23bc57d9d2727b9ff12578c156126aed3da
2043 F20110216_AABVCU taskin_z_Page_120.txt
98a8eed4272c721d2b71a2c06b600687
eb0020f2fbd455c06e1286b8727f1bb9b0a10568
34973 F20110216_AABUXO taskin_z_Page_042.QC.jpg
ed55d158492bcd93fd5a2a5317b28598
47e403f4dc9fecd9d8af61028afad17e4c5a8ad8
F20110216_AABUWZ taskin_z_Page_119.tif
0e20273526b0fa5e182015fd7db50305
0ae7d4c32ea25759f1711d62d9f484071114a485
475 F20110216_AABVDJ taskin_z_Page_157.txt
d8715695d1a84af5d0c29c44609084ac
2989d05ed416e0b79bc8d6a88acf01d102cb5c21
F20110216_AABUYD taskin_z_Page_009.tif
7fc01ed0a9d0fe435008415b9c22b03f
0e86ddf87c9d8096fe91a237d788d722464b3e6a
2378 F20110216_AABVCV taskin_z_Page_121.txt
cf08d5ceb4c07e317b3dc4947ffb53e8
4d59563e82398409c1279440addfd2a1c3f7b7bf
36616 F20110216_AABUXP taskin_z_Page_008.QC.jpg
3e118be20029831d1b91672aad8b5fe7
ebe3cb339b2021f1f3b529024967e3517b4f85fa
794 F20110216_AABVDK taskin_z_Page_158.txt
9e65a4ecf768975bff7b9fba5c41d865
a2de4c94751743c91ce98c0816763b194aff7219
F20110216_AABUYE taskin_z_Page_011.tif
5000c8944b5869723dd09e43544a99ae
663dec51c5583489852945079c7038cb9adafc5c
2293 F20110216_AABVCW taskin_z_Page_126.txt
8cba09e308430b7822fda2988867a509
74e57333e1b67822df2bed85a1f8026c71a9b12e
8762 F20110216_AABVDL taskin_z_Page_001.pro
87e33a1846627c2f6a510bb6aa22bb36
ec2e88a5459855e58b29d3bf5ed3a39d6f2ca08f
F20110216_AABUYF taskin_z_Page_012.tif
47180811ebfa76f40994b6c0a3b8667b
34c4f476a148bba15a2a8fdfc0e88fce24f40eaa
2197 F20110216_AABVCX taskin_z_Page_128.txt
1bc9e5601ec12da61f7df67b49002ab7
abd79eaf3b07a1f940f090a6987e6e04079d6c28
2184 F20110216_AABUXQ taskin_z_Page_123.txt
c1f040fd2998972f6cbf8b4f1109a239
c7752b0a747c49177d327517f5c01d346ff408b8
F20110216_AABUYG taskin_z_Page_014.tif
92fa1e89006a4370110f420d5b88663d
c275e55a05f76368f73dff78f2ca7f5d239e6274
2054 F20110216_AABVCY taskin_z_Page_129.txt
35d51adf6891084dc4d3cd98cfd893d8
e50402e48e0735232ec1994caf9c257eeda65a92
1051984 F20110216_AABUXR taskin_z_Page_113.jp2
6bae27d2f814b72e9f10fd450f64638e
a02b7d077c7e9755b5083ca7d35a448b6263d6c3
47453 F20110216_AABVEA taskin_z_Page_037.pro
ec2752413bf1e490fddec5679f9f3a9b
08aaad127c20da810b489b9c1a788e350b4af936
41124 F20110216_AABVDM taskin_z_Page_004.pro
6f6a8bd481823a1ee8b354fecb003332
7c740c415b3815791a857dcb6b2aa1116bfb1c40
F20110216_AABUYH taskin_z_Page_018.tif
abe967f3b2b5feb42cb6e0f9c470c1f2
ee012d0c98f8735891d1fd0343b26753e65d7e24
1388 F20110216_AABVCZ taskin_z_Page_133.txt
dee1fed5a3a9a2560931dbe947287efa
bf0f11fd545ed8b6f36f25be385d66257c359624
1051963 F20110216_AABUXS taskin_z_Page_033.jp2
d2266828865f201418175d35ef8762dd
5ddb68c7c53cfc41b038f534838479b94dc5e0b0
50171 F20110216_AABVEB taskin_z_Page_039.pro
d3589618878f38fdda5ffbacc332066a
ca011237253eb42ea7806d99770f3d1f3349d281
60268 F20110216_AABVDN taskin_z_Page_008.pro
decef70b56c3b1dc9695083dd6fd6b27
a8d0cf3608e80396bf3158395ceedbf0e5c30559
F20110216_AABUYI taskin_z_Page_021.tif
675537e16cfaf1e1cdef6ebf7d650002
e7fa733736ca07d90aafc5bb864827ebe04f54d1
F20110216_AABUXT taskin_z_Page_095.tif
36028fb653ce3d07e542dacabd9f3ef0
77b133e2c2c6394f20f8c75290671b69c8d2281a
48779 F20110216_AABVEC taskin_z_Page_043.pro
a3f442b16600437174b79de9a4ca6893
dd7ca30398e9c80f3598c188fb2a0118d31eb94c
35907 F20110216_AABVDO taskin_z_Page_009.pro
85b0d35c4eaba795902b73f7e18d9352
7bc75c7e3c6fe550ae3efc58951edb32a1c50882
F20110216_AABUYJ taskin_z_Page_025.tif
d97ad157bc0d85963c3d189edb210c81
e0cf0243cb86ba34495af918f8b14439dd3382b5
29517 F20110216_AABUXU taskin_z_Page_088.QC.jpg
0aac904e319d7eae97b061400535b89e
9b5f476926a32aaad98073ab85b74d146dcae28c
48469 F20110216_AABVED taskin_z_Page_044.pro
3d2d13c221286a93ac12c7e4d34b2f60
c898f79446fbf7c149b0364fd7a379194df59f51
27102 F20110216_AABVDP taskin_z_Page_012.pro
5b67f658e7cee1ba070b5247606217f2
23b463348a8b0edeca17f52e31c52b6d68ff7632
F20110216_AABUYK taskin_z_Page_026.tif
fe84b9944f407b5ee53c6b640aae40e0
3304bdd866024e29d4770638596dd5e07e82969b
45790 F20110216_AABUXV taskin_z_Page_147.pro
eb4060d3ba3d1d93bf7c8204655cc063
7ab11fd46a6a0b7b4f597fada2e8968314c388db
57059 F20110216_AABVEE taskin_z_Page_045.pro
0a115e47f4fe2ea0f57f204353b63fba
20d3572e9fb5b3ef1a7b6f89e70d91942f5fe631
61785 F20110216_AABVDQ taskin_z_Page_014.pro
489f49df4654d8c2b10bf4517c1921de
e014847daed5488cb2ce8c57e5ca1afdcf726414
F20110216_AABUYL taskin_z_Page_028.tif
d0f58214a4faca66df030b67bdd8634e
fed248aada1737d5fbbef75684e7208f8cc4c82d
7170 F20110216_AABUXW taskin_z_Page_089thm.jpg
aa72048c27e2e21052baed259f790c9f
55800857883a12d072d050063a632c47787bf694
75149 F20110216_AABVEF taskin_z_Page_048.pro
a0cda8fe74ba0d4640bb0da16365abc2
e2308ba2d42e90ebbd4f18fef153b3ccf88bbd39
59240 F20110216_AABVDR taskin_z_Page_015.pro
3777b0d3efeeda9efbf74ad48d565d8b
10e622d96f01a81ef75f0e3e6cd328b89bb757dc
F20110216_AABUYM taskin_z_Page_031.tif
472cf230f748eb978b9e3360f8c5cd12
4ddbdeca252d63e451e6d2f6dbfc05785e8404df
902743 F20110216_AABUXX taskin_z_Page_091.jp2
5732948a3af04e29513fc4aca6dfbca8
21233adfd76f623439de363aeb046f57763b001a
86238 F20110216_AABVEG taskin_z_Page_049.pro
55485c12362421e59006d6caf57bc904
9d99a9ed7f2140a88dccecbd7720a6ee315a1812
F20110216_AABUZA taskin_z_Page_061.tif
aa9769c69758f0d18fd46bc4f99855cc
9b8073b32d2f211be3eeb05d1a5aa9ec98400919
61740 F20110216_AABVDS taskin_z_Page_016.pro
64b101cab6c38b6335dc9675fc29a1e1
3be920a50757e22ef2fac3805e9e41860d3734d4
F20110216_AABUYN taskin_z_Page_033.tif
c81e26a19f31e9728c66fbc6385d0064
ab52bd0b73322904680b7d8a461da577aca12e35
186033 F20110216_AABUXY UFE0024784_00001.mets FULL
c81f3de0ed84b1f22f582f9930158111
85770d0e13ea1691fe4f69f89a9d3956fb0c6c58
44790 F20110216_AABVEH taskin_z_Page_051.pro
93091ef7c0ea780b8dd8c61881858d60
8c09c2e910eafd596637acb267227eb7403e5431
F20110216_AABUZB taskin_z_Page_063.tif
17d417d69a3825d4695d9f5513500944
72a076a15558dafcfebefbbe53e3e167b9d72e6b
58046 F20110216_AABVDT taskin_z_Page_020.pro
4ed893aa9dba1d2b1656da9f54bd634b
364fc47ffd2716cf7d2e7cd957be9da008771769
F20110216_AABUYO taskin_z_Page_034.tif
89692139e239e5fd4fb837e73b552900
b7171781eae569d77248e2dc196e5390d45b629f
57809 F20110216_AABVEI taskin_z_Page_052.pro
b6726bcb339c54d63953b3c0d88c609b
629250e9a992c2517e24fc496ca07a5144c9b7fe
F20110216_AABUZC taskin_z_Page_065.tif
adda2ab68c803c91ceab40a5d8dadbf1
2b7f14ce564da59023619d3f780753331c02355f
57708 F20110216_AABVDU taskin_z_Page_022.pro
8361d8a053b44e78f91cae7a3a4c1afa
cda830d276990d42c15e5fc2403247382b12378a
F20110216_AABUYP taskin_z_Page_035.tif
4c5d3337665a1bfa9d6fdaf5fce90045
dd0c27d0b34c69a344a0d03231ad70599a4fdb80
57282 F20110216_AABVEJ taskin_z_Page_055.pro
1241487534015203bae668126864bf2c
c8666783ccec997c29b9ec7d36bb6d5fb421ecfd
F20110216_AABUZD taskin_z_Page_066.tif
6daaea9c0e0332a4fde7057cc7dcd8f1
65140d3d0be73e23d4ad64c556200ce9bbba803a
42917 F20110216_AABVDV taskin_z_Page_024.pro
c37ad938173b4b58b34411613aea473e
da9c153d51d9389ce856a1c269a80aac252c21dc
F20110216_AABUYQ taskin_z_Page_038.tif
e1e1413771f51cf0f96d8f082bd0d893
f7573088fb9850fe077162f58e54541a123a23ff
45211 F20110216_AABVEK taskin_z_Page_057.pro
e25d35e1c6b5e2ee22f804dec0ff527d
7741970d36ca5b6c797a4b4f8c4635cdb85fdbff
F20110216_AABUZE taskin_z_Page_068.tif
9871b6a7995c48c68f547f04d5595bc1
db2718b582914079f18205f86977bba3fd52cdb3
55089 F20110216_AABVDW taskin_z_Page_031.pro
567f0c0b798a1962cac13ab5c232dce4
676f9ace2041ec05ce38aeaeea52e7b022aba449
42444 F20110216_AABVEL taskin_z_Page_058.pro
c46757f20cfc9aa928b93da0817304df
edaa214cc4c7b53798f0225a101d12cc0df1b86e
F20110216_AABUZF taskin_z_Page_071.tif
2d627bce54e5f20dd227b8e5b5f1e1a4
48d4bea2c1e3ccc65f0fa388f8a8ebedb4b40620
46382 F20110216_AABVDX taskin_z_Page_032.pro
6ace947d45823d9b01c48ac767f07a23
08b5396bcf5c936aa14a5bc4afc99157f08dea60
F20110216_AABUYR taskin_z_Page_039.tif
0aa95197854d27765dc01c078ad3c084
acd06bc10738dc9595ad1d9adb8033bb78bd01d8
92959 F20110216_AABVFA taskin_z_Page_081.pro
105a44ad1232a4cbe694b9991f00e0f8
93d54e5e5335a9ba6fa60a6204cc7fd683bee1ab
43663 F20110216_AABVEM taskin_z_Page_059.pro
5719c7b0e8d3e8bafbeab16a8920d491
cd55ba2db4879776dd0ecdaaafc7c9e40ecf3911
F20110216_AABUZG taskin_z_Page_072.tif
28501c8a655c55b74802f47d366445e1
e525de6e8475dfcfe20ed2106a0acbf54c7a9676
56545 F20110216_AABVDY taskin_z_Page_034.pro
03020e389607948cb17c62b0e81579f9
2e7d86536befd8833b0436c8d32028c371dcfbeb
F20110216_AABUYS taskin_z_Page_040.tif
54041aaad8ceea954668b4e86ebfb3a9
cfc8a8300c7097b868cde211b8b400e343d78a7e
58155 F20110216_AABVFB taskin_z_Page_082.pro
f03632b1529cb357f0cabaa06d715f98
01d2105713abfbcf301fe416ed657907e00f7cfd
49913 F20110216_AABVEN taskin_z_Page_060.pro
6c28119bc666ef7a5aa7e29b2e50b0b5
d033b0be53a7671a7637a95bc423a45257ea753a
F20110216_AABUZH taskin_z_Page_075.tif
b37fb090ca4d4547d685f45ca713ba91
49469efc8c09907a74fb25ff71f4457c64117cc2
44809 F20110216_AABVDZ taskin_z_Page_035.pro
ca5ba9325e6d38b6dbe924f0b5b5a1cf
5a2d61d9603f57c4587f2031196277e16182b636
F20110216_AABUYT taskin_z_Page_041.tif
c539caff9b8276a75a27a2d378b5562a
01fc80a52ed6ef2dc98e706de8a8c00933ae8c26
57715 F20110216_AABVFC taskin_z_Page_084.pro
1e8d2bd5de7eb3c7b4ac86670662dc59
1df45516386c32168bea0bb904c0c67a59c7214f
49749 F20110216_AABVEO taskin_z_Page_061.pro
aa8abce6ecaf8c0224a81a458719d581
703b0657c85addc9fb8b8d7c9555af7544893c14
F20110216_AABUZI taskin_z_Page_076.tif
b2f747852ed221d05d82ae89867ce385
9db3907f734c09934b6ea6ccfd582b59cbdd3183
F20110216_AABUYU taskin_z_Page_047.tif
028bcb34297ba5abb141f8e3084ce232
2dd54751cf8cf5f7155acc6a93390f10fbad8293
61258 F20110216_AABVFD taskin_z_Page_085.pro
90d5d03fe308d7de5e2f9474324deef8
68b9d74da72d983b960b64b6860cbf87b1998d5a
52837 F20110216_AABVEP taskin_z_Page_062.pro
941db142d18aec8e566ede2d6d246c4c
159553f113d98a62470968d3d08911cb7b6c2135
F20110216_AABUZJ taskin_z_Page_077.tif
1e50f8c69c54f39790c0c40b57fde3e1
81db53759db78fbe288ba41dea01dc19a9ab7c35
F20110216_AABUYV taskin_z_Page_049.tif
3ab2ab911554480538e7564fc9420ea9
e33d7fbfca5129d8fcf48d5de9319ee0060ca2b0
40095 F20110216_AABVFE taskin_z_Page_087.pro
b15987c00201512a8561a9319425c6da
3020d233be89963fb76fb0a4ffdca1d97a2b52c2
57667 F20110216_AABVEQ taskin_z_Page_064.pro
0a677666c317318846dc93534b48bb82
97b2f0b9120333b4c00f389f30761e7dc0f1ad67
F20110216_AABUZK taskin_z_Page_079.tif
c74c2e2ab0f8df51745669e121182e95
af3962a669ef9bc667539ec157ab4bba1c6a2da1
F20110216_AABUYW taskin_z_Page_050.tif
fa31b0409cae76a8ce17cc56b69dfb80
19e4f1cb3cf450ddbf8c1c03fa29eefc605633ef
48877 F20110216_AABVFF taskin_z_Page_088.pro
fbadd6f93c037597e07680bdff253513
11155340f717d1b07116b7b3ff32d49a94d04572
63062 F20110216_AABVER taskin_z_Page_066.pro
11315673c52be3649f47f053969a98c3
d127d1d03b84ecd5088277d6b0321a54717aff78
F20110216_AABUZL taskin_z_Page_081.tif
cb3b5407c573d09f9fbeac0f71dd20de
ad531ab78952bebbf00f59ecdd9d140264f887f9
F20110216_AABUYX taskin_z_Page_053.tif
31f4a6f089f20dce6be1531ba8e68961
467282d87a5af9b2a28bbdb230f0fe2f09a8167a
46415 F20110216_AABVFG taskin_z_Page_091.pro
a4d770f99a4019433122aa7df1986a7d
83b74357335c4110a6d0faeddc330e9141d9e7b7
44011 F20110216_AABVES taskin_z_Page_067.pro
1a185ade3f7d461eb2ab848049cfb88a
d0066cb485ce01fb40e1f840aab8e4b422a68b28
F20110216_AABUZM taskin_z_Page_083.tif
d90a8e86ac6ee7b5aea1b7167c364399
fb6ae71a2bbea7eac5a14e98f24856720927ecbc
F20110216_AABUYY taskin_z_Page_059.tif
3870e549e160ecae7153384d79c6ba4d
d0ae6535a47c9f244e5aad20d2f4f07dab57d5bd
48168 F20110216_AABVFH taskin_z_Page_094.pro
ca5d583506d04829e98ca504eb2aba8d
cbab423ba674f926e2fdcdc8c10ad7551ad2d15d
42198 F20110216_AABVET taskin_z_Page_068.pro
0f1c7c24df1d8481b602a70a861b1642
1ccd6025227f5ef242bc88a1d1b038700381a516
F20110216_AABUZN taskin_z_Page_085.tif
8ce18b89b6c96d21ebb9de2bfdaa044e
df32c0d611f9215616b3113d235799bdc6227cd3
F20110216_AABUYZ taskin_z_Page_060.tif
a52ee56fa49de06e22e933c8e12c46a5
8207c6ca49197c73a0ecc3f389548035c447af18
41744 F20110216_AABVFI taskin_z_Page_096.pro
c119ad5b4daa7abaf0ed30d05217ff54
4eebd0ca782ba0d72b6e7d51788712885f8d3bef
55489 F20110216_AABVEU taskin_z_Page_070.pro
212fe0890fee8efee59e5cd292028a36
0d9b40ca8da27ae0c43d78cda4c0ed886ddf675e
F20110216_AABUZO taskin_z_Page_087.tif
1d4011805d57b645520afa3f1fedca0f
b65464d0405a7a737938e87feb062338ca19a44e
50477 F20110216_AABVFJ taskin_z_Page_097.pro
98409ee2359e5416149796ff43a60da1
a614670b0aa9d534fe2080653d0f5c3b5daaa13c
60493 F20110216_AABVEV taskin_z_Page_072.pro
b7110d2cd9924405b7f05d32e1382e6b
7cc41fc328e435cb6f983154d6421bd7b95a2764
F20110216_AABUZP taskin_z_Page_088.tif
a779d0100f9c74cbdf7a3c1aebbf2ba6
4bf2beed347842bfec0615b987991f911aa51480
52802 F20110216_AABVFK taskin_z_Page_099.pro
ed3ee6e85d1b8eba001b268168ce5602
6af232e91761096ee93a51fd2cbbe2c608385a4a
75296 F20110216_AABVEW taskin_z_Page_075.pro
e10cb6f1ff1ba7f91cebcbda9823555f
ef6d5b8f2dc7a965ebc2e43a762c1e3893cfd8fb
F20110216_AABUZQ taskin_z_Page_089.tif
ac68b1cd70e4a63e54304bfb179d30a0
5503c39c2aadd540bac25455ca64b78013d9dd8a
51990 F20110216_AABVFL taskin_z_Page_100.pro
b466630bf51450fc8bdfb9fc221bfb93
122c7ad0cb6b8d3a04d33d2f4bb9a16230806fc5
31906 F20110216_AABVEX taskin_z_Page_077.pro
fc8e5adde4792c1c56e7a6dc804ec878
1aac078b0ead2745c10eee9335d70ca5f8592d02
F20110216_AABUZR taskin_z_Page_090.tif
fd62b56e08738fb95acc84c9866b9c2f
45d5bec489031c47a5a58980a4e7a0da1e4f76a4
43314 F20110216_AABVFM taskin_z_Page_104.pro
b8b4cb0113c9b1d69bb72f9157b6817a
d27ce3736dfa158a79198ed7a9c9247f00974cbe
33431 F20110216_AABVEY taskin_z_Page_078.pro
2ac0f638362cd012b327064606874519
a627e56a4eea1bf19f612fc5328a69935a8273f4
50852 F20110216_AABVGA taskin_z_Page_131.pro
d01d07a961fd3d28abd72a8c47ef777d
d01c60fd31143c5fc5cea496e3178a5546e573d2
52798 F20110216_AABVFN taskin_z_Page_106.pro
e22393bf266a7b82016052557bf15270
33f9b38e585d042e0953b2a1d4a6909dc950e02a
33443 F20110216_AABVEZ taskin_z_Page_079.pro
6596ee04dc20723556ea0e9a78abc523
f59bb64cb3574a835027ffa1f7958250c237688f
F20110216_AABUZS taskin_z_Page_094.tif
03dbcf92fe806ec93b7871b802c5f8b2
cf085ff8c6014907490f63ad84686d03ba2d36ce
53752 F20110216_AABVGB taskin_z_Page_132.pro
4a0349e58df74646dc1a80080aaa67c7
0bd02f1cbacd1886757d1d8e0ba52e5b2b7bdebc
61573 F20110216_AABVFO taskin_z_Page_107.pro
cc810221299760a9927da55ebe53ea68
d466380d1337df8a037ea7c7ac24ec1ee42fd340
F20110216_AABUZT taskin_z_Page_096.tif
19c66ff0a8f4e6dc5dfce329151a73a3
57aaaffa5ea4f991bbf56b86fee1cc3f7ff92823
35965 F20110216_AABUDA taskin_z_Page_156.QC.jpg
8307b8dee415105ce750939d94764ae6
530c0edf03fa20eb2e1089419d2c3ba615252a16
32430 F20110216_AABVGC taskin_z_Page_133.pro
1d9024859877b0a7a04c034dc24dd8a5
bc823d42805bd3886804054c3c8ceb395773d0c6
88395 F20110216_AABVFP taskin_z_Page_108.pro
aa6e3a862a68db775fb375d96d5028a9
b9b5ffd35e9554ea5ee2304f0c1bf09d221155ed
F20110216_AABUZU taskin_z_Page_102.tif
7435cf2cf0235e76922695a35d925eb1
ad2dfc287904c3dae6a3a08a6a89d826c1ccf31c
23472 F20110216_AABUDB taskin_z_Page_077.QC.jpg
78058651a0abe60f2b5b83a9e327be95
a14f35aba2b2d6ddcd4d59f250c1ec27b0b78ea9
54798 F20110216_AABVGD taskin_z_Page_134.pro
a931726b36cc185c0b7bbdb8516094b2
9aa23c1f238056b65f5c68819f5bfa2eaf6e3dc3
36014 F20110216_AABVFQ taskin_z_Page_110.pro
4d907ba15109b940b61ee8b04684ae05
efe106551ef7b60bf24678c1699b0c08890184a6
F20110216_AABUZV taskin_z_Page_103.tif
e9eea38efa094636a5864c18c4122962
485602c7fbcafe41bfaa190dc7d74f1c4438a56d
2344 F20110216_AABUDC taskin_z_Page_064.txt
135c2052d7030dd512a973a1d62adf20
946e5233a7d1c2535b20d9ea84aff54429c95a50
61028 F20110216_AABVGE taskin_z_Page_137.pro
4a3866efc817d57f034526164a43ebc4
04760898540feac1ade3031989586de464c3ca09
57587 F20110216_AABUCP taskin_z_Page_054.pro
020bdbf5e16b151ec1001c724fefae8f
6e078efaa79672362f2de4a125a992d4c2d2ad5f
93169 F20110216_AABVFR taskin_z_Page_111.pro
f8ae12f511fd49cd75c937ce7d0fe857
40d4cb01468cb880aac750c9ccf1378810d713b7
F20110216_AABUZW taskin_z_Page_105.tif
539723962009319dcaec57fb3767a0b9
58185e4508fc4a3ba69844189983041f871cc91e
34063 F20110216_AABUDD taskin_z_Page_047.QC.jpg
8495ba181f98812f2855be6003c83492
73d9a598cbe001508c98158788e6355b332634a4
62286 F20110216_AABVGF taskin_z_Page_140.pro
a4256ec5cb1ca2e1040448d35c5188de
9b64a7614d74b62f49e08185912d6161a29019a4
90227 F20110216_AABUCQ taskin_z_Page_131.jpg
29f0213ce37586e436a588503293dd2d
e7f92e4160d16d0439a9c0f430872bc1061ff1ed
52465 F20110216_AABVFS taskin_z_Page_117.pro
872a833d08fd24e5231eb6e1c2bee704
be0c90ae084f1c125f519dab5e1d6e1409a2dd9a
F20110216_AABUZX taskin_z_Page_106.tif
b82f7f6f30ba40d2244c8ca59015a2a2
8b27e046a5184465b8b311e91fb17d43621567e1
22450 F20110216_AABUDE taskin_z_Page_027.QC.jpg
575b9c4791e8f023bac6845e785034d7
9a2bb8dd35b1f4ad496ff85300d7a42ebc63313c
4669 F20110216_AABVGG taskin_z_Page_143.pro
6f2e68c1f4d02ea3170bd16dd97e4dc3
1b45e0f8601d7f3f31ec0f9c298f272c00f6c4d8
F20110216_AABUCR taskin_z_Page_005.tif
1446ed0c5b3e44e983316cae263a7429
41929aa82b5708311d9ac994354c9b1d27c6ecd5
24522 F20110216_AABVFT taskin_z_Page_118.pro
574e5847d602ef553ebf6127318188e5
bfb946ec7f9e06ce1b7384e037554f20121a3ce4
F20110216_AABUZY taskin_z_Page_108.tif
e0887f64fbe305a192f35755c2112ae2
f8e149834953220f015e7a7eb13e276c0a3ea445
F20110216_AABUDF taskin_z_Page_056.tif
f28827df1ca32cfd235052333d813468
f7903b0570d527722bb275f07c1226121913194a
38560 F20110216_AABVGH taskin_z_Page_148.pro
8febbf8e4e3de02338669aff2019e2f8
e67103a938cb88ff75e0a53b3c5492aeca70902b
30554 F20110216_AABUCS taskin_z_Page_025.QC.jpg
50e1253b501b0608a495566e74cdc989
07a306aa3ec1cc1e34ca644fa3888ac47d03ca38
48265 F20110216_AABVFU taskin_z_Page_119.pro
4a3584eb11645c54df0333dc6cc1d415
26344613375e5299633a484c0749159d6d1c9fa9
F20110216_AABUZZ taskin_z_Page_111.tif
bc71796b031272aad114873c7f6b861b
675ac4c2b96d83bde304a574f27ffa0ae4cf6d75
F20110216_AABUDG taskin_z_Page_044.tif
edd0c2502d65c2ef9b9bcc6abe168b0b
c187744031be6c6f7c5f39c757addefa16cf1b09
56639 F20110216_AABVGI taskin_z_Page_149.pro
f2a430dd9258017924bea0a48e4523e6
029f1556191094b3292e78b4928c9885c42cc693
790871 F20110216_AABUCT taskin_z_Page_065.jp2
dda1621c6274f0e514e188851c8233a3
cfb53340d8e69c1eae0d26e0f626093a458be11f
58632 F20110216_AABVFV taskin_z_Page_122.pro
482018f3f75727b8bc7450e9d46a0243
55ca34f6fb9b9cb95013eb2d7234163e395fdfdb
26815 F20110216_AABUDH taskin_z_Page_068.QC.jpg
ac5bd845dce002cd90f91381cebf9e04
63ed11dd70e258f124aa40f6555c940bebd6aa6e
60902 F20110216_AABVGJ taskin_z_Page_151.pro
0d613f87ff8528c0fa672db5a24c948e
7bc44be36756cfe43a7b5651ddf64c4c5bc6e425
8309 F20110216_AABUCU taskin_z_Page_020thm.jpg
f267ecc87ce1062a00d04ff2a1b6d71a
083e6fe647082dc44ec318531218944e8034fcc8
58260 F20110216_AABVFW taskin_z_Page_126.pro
311bc97984574c4a90af8ba7b234a27f
473fad9add50683c30fceb03aecfe1997de27fa5
132408 F20110216_AABUDI taskin_z_Page_152.jp2
6426b50f0083479aeaa42c00228e3acc
7fe074b7cd649cdefde0fd6adc455ce8b3cb7305
63419 F20110216_AABVGK taskin_z_Page_154.pro
8834698f624c3121f8714d77565022c5
91f0352c6055e5203259d3ad5763f1500d13d52f
57908 F20110216_AABUCV taskin_z_Page_138.pro
77628667f53ee568bbbddc244a13b24d
15ac674a76fcec60e6956fecd2c52e7e41e7a71b
59773 F20110216_AABVFX taskin_z_Page_127.pro
e2b266366952032351816ed7ba6eea1f
16778ecef2a8a55061dce7c89cf4f745cd57590e
114231 F20110216_AABUDJ taskin_z_Page_126.jpg
77efec3fe4ecf9bf1064e07f808390bf
987853f687cdc928097decce82a669472af077c7
11880 F20110216_AABVGL taskin_z_Page_157.pro
70a8a2ae058da7d9c807aa5fa17c349d
ed2a0f206ca16a858e2f1b3f676e79c08d088840
86719 F20110216_AABVHA taskin_z_Page_024.jpg
925161513f41896350e34071af63c6d5
b5ac658341609c60092cb1b3041b2226955474db
25311 F20110216_AABUCW taskin_z_Page_079.QC.jpg
9e6c0b970f5be14553f2007688dcacb2
a794fc861807e28c21b64d3eca908939d3736e23
49027 F20110216_AABVFY taskin_z_Page_129.pro
2515fdd263eada3dd93a20444359d27f
463ce756218bb626d99978511b0127494410d050
8122 F20110216_AABUDK taskin_z_Page_113thm.jpg
2410ac062f6064a6d9318e40db876ab1
a8a06ec4a6ba77096ad4684dd39ac9fb28c5d2a0
19064 F20110216_AABVGM taskin_z_Page_158.pro
f6e1d53f28ac8ffe3f03c81df36120da
b70877572927b113889327fea5a0dc98fa15dd1d
93044 F20110216_AABVHB taskin_z_Page_025.jpg
a30588cacdb4ea7468d9551eb52313ec
24c6499c507f12078b7041309051e8448c83cee3
2337 F20110216_AABUCX taskin_z_Page_070.txt
62d8a582a0ddd78e15c8fa507e7b7a5b
7f35a7e0f5a9cdecb85091ac9c1e1c12ac0d1eb4
58880 F20110216_AABVFZ taskin_z_Page_130.pro
34309859a7d98efffcc55528564aaad4
bc2e8223d3fe1ad006ad54abf7f35f3ba7fda343
63802 F20110216_AABUDL taskin_z_Page_155.pro
53dc5408ba96730e93fc497f23056f5f
5e11cdaa48336398190440d16a8544d96bc2daa7
28443 F20110216_AABVGN taskin_z_Page_001.jpg
38bad5291801c5388717b029b4a5c7f4
a82dc6e1d6b1c33f708dfb4dcbce2d10de9e4258
80881 F20110216_AABVHC taskin_z_Page_026.jpg
e5680c785028a918c51fc406c4fff4a8
5399e2be55bb7c9970e8c4b1142c43ed428de8ba
2440 F20110216_AABUCY taskin_z_Page_152.txt
b92ebf040c88f7b619c226949be2c113
93ed39e49e4fbc915259110c2630ea9fb0d97674
107669 F20110216_AABUEA taskin_z_Page_099.jp2
1c04acbeeecdfb59c85885dc0ae008ae
34c2669f38d689acd4ad32f3211b7f3b95cc2a1e
7934 F20110216_AABUDM taskin_z_Page_106thm.jpg
46ed0cccd9d43d5c01ca02dc21667fc9
5fcd930f8773ad6c0ff593ef17e57e893582c42b
6607 F20110216_AABVGO taskin_z_Page_003.jpg
3d5137442e4bc46a49efecf05d8e6afa
49056592871f49f37348457e5203bdcec600643e
84490 F20110216_AABVHD taskin_z_Page_030.jpg
46c67c713f03c7a12fc78b8d594028c5
d82c82665e9d8df6d92a837285164d3f77ea8580
110379 F20110216_AABUCZ taskin_z_Page_134.jpg
1db1d3f2736c3de640b48d2837199494
c1ad0662be6d724076ebf45e402d58bb8f35044c
F20110216_AABUEB taskin_z_Page_004.tif
b7185c7406c0df8fcffaa7ad9d5860c1
06e18e674aa86a1876cd2dc539033470f16aad3e
F20110216_AABUDN taskin_z_Page_127.tif
db21a684b1fd2183247f95327f143a6a
38f5dd6dcc0682ec2b505c1c0204437e9e33c8fd
82501 F20110216_AABVGP taskin_z_Page_004.jpg
c27470263222d0f6fb30fb2349906e80
bdec0dc70ceb56ab4ec0e2b3bc8e02881ed5d4ce
108244 F20110216_AABVHE taskin_z_Page_034.jpg
7764171ca30a62e4a05fa06cbfe697e8
e7b038c4161a7870b17ba859262eaf2702d78514
6699 F20110216_AABUEC taskin_z_Page_078thm.jpg
536bad1184864f3479290d75bcfcb67b
a4cace316622482bcb67cbbde0b4724ad5a92b18
7979 F20110216_AABUDO taskin_z_Page_042thm.jpg
88f88be650b1506461d3792e67ba7ebb
b9ae6e04ce9456ead0c67dd21ea0e2dac2e39563
129696 F20110216_AABVGQ taskin_z_Page_006.jpg
425438bad4d0b59c686c31aeeb18d188
9b44e48a3a912434772e92a456bd759d9fcd5d74
83568 F20110216_AABVHF taskin_z_Page_035.jpg
70a48d7568c397a5f1ce5000411d7d25
6b4d1ed7638c6a09d152a3b24d45e946aed727f5
1051985 F20110216_AABUED taskin_z_Page_009.jp2
be2c293b3ca7b81133bda72279c7d737
a63087c5616c61e7637d2d074851f69b483be79a
F20110216_AABUDP taskin_z_Page_064.tif
30b5b9668176bd8b8c6a7cfdd3ad160e
9eca4085d36c566c18a0a38f4ca155c32ec46cb1
55179 F20110216_AABVGR taskin_z_Page_007.jpg
de037ecc631c0e56dc38f3a69b306937
22fe85a9aa6597711b8af04cfc0477fa28739a99
93558 F20110216_AABVHG taskin_z_Page_037.jpg
73174d208a92eb0fae3ca17ea019de57
9c8e18071534c50297d4e665bb4df9bb0834ae32
27979 F20110216_AABUEE taskin_z_Page_024.QC.jpg
5e2cfcb113b9682ccec81cc7ee95dcf7
9198fe594e6fd97bd5ba174f5af628ece0cb42b6
2122 F20110216_AABUDQ taskin_z_Page_024.txt
21da8f32e2a0da254954c4f7bb253551
9201c316fccd6ef1b865e80d4471993cc545c585
71708 F20110216_AABVGS taskin_z_Page_009.jpg
d984697224cc935180b633528753306f
93ff307ae1b76c9270dd3f55d88411bba309deca
98388 F20110216_AABVHH taskin_z_Page_039.jpg
18668dffbe828011e378bfc19334d01c
653d227a21b32b6c6e786b0b5ffe4f8a79ffab06
F20110216_AABUEF taskin_z_Page_074.jp2
1b9afffd322884f985b5aa4324b9d473
7dfb9746504cd5168794161669e7d35ed1d6f774
2487 F20110216_AABUDR taskin_z_Page_144.txt
92e88fd9edfd09f09e9c7e16a3c2c766
1f7bd5acccb3a2b0ee47b61cf205d35e41842049
102512 F20110216_AABVGT taskin_z_Page_010.jpg
a750326a34b252dd8154091b6dc2a480
ca46422f9b2a1e759fe58fe5951d98b91155f4bd
109785 F20110216_AABVHI taskin_z_Page_042.jpg
54edadb2003f20b11d5e658907cd0cf8
eb49712551aed1aee9cf0876db0efa6ee40b87c7
2294 F20110216_AABUEG taskin_z_Page_060.txt
0cd37070fb99038abfbf2b7be8360861
a5388e68fb3fb7fc559abb666b2d8be7add0525e
F20110216_AABUDS taskin_z_Page_125.txt
5b27d49476d2bb1966b75d7724708a23
69ffa1344200adfcb2eb0742b44ba140315b2535
121095 F20110216_AABVGU taskin_z_Page_014.jpg
030e258b14247f40bfcb94b88e32dfb3
06dad83debee60972a3e32c405faf0f0a612a7ae
95675 F20110216_AABVHJ taskin_z_Page_043.jpg
e6a0b44d86239535f0c8d368ee7c8f30
87e0b29bc1e1fabdc2112adcca40d8819727cdae
14170 F20110216_AABVGV taskin_z_Page_018.jpg
24e35f68de73458dd158255c6839eca9
8a7a4c4f2897647f59208dbfd630d06ab0aea6e4
F20110216_AABUEH taskin_z_Page_013.tif
776f485e0dee8160e6769b428760e00d
ad8b09fe6a50bb4b49ee5f1de13242f297b7299e
118492 F20110216_AABUDT taskin_z_Page_015.jpg
0cf92105053376c6af21a5f2a864c164
0dbfbb831dd827e03455e160d30d67c62c9a6dba
95376 F20110216_AABVHK taskin_z_Page_044.jpg
af8f3fa0583dbbb61b8868c1d939463d
91fa184fddbae0ea6fa0dbad85d443cfcfb1d483
106131 F20110216_AABVGW taskin_z_Page_019.jpg
58c785bb113b3bc26d35217de56b5167
d652576cf563905ddce6961bd126b0109fcb7304
2270 F20110216_AABUEI taskin_z_Page_124.txt
df291fe88b99e06793c08aa04185ee4f
6fe0b834da5d46b5793728c0892d5de1466c2077
2209 F20110216_AABUDU taskin_z_Page_131.txt
65a306bff81a4212c6ccd8890f66eafe
8a039ef84e1c58a0dc8019f87c4482c36220ada7
110747 F20110216_AABVHL taskin_z_Page_045.jpg
99ce212429dca89d5b2d1a5e216227b3
01d434a85e72683284489f67d0c7d979c3cf28af
118074 F20110216_AABVGX taskin_z_Page_020.jpg
47b3bcdb4c55ee7e1b3744c73328d327
7efc6d0d06bd1a0972bed85134884bb103123056
34635 F20110216_AABUEJ taskin_z_Page_127.QC.jpg
bf3b01892ffb6f708c899b040dce5c97
67282937466c2b734c46a97c8c57c614df0bf691
F20110216_AABUDV taskin_z_Page_125.tif
0fb56b9fa0290bbd4a0d2adaa3a84a06
b61bd16595b6846cbbf66f832cfd36580f1c82c4
120997 F20110216_AABVIA taskin_z_Page_075.jpg
c71fee41563f2f3f0cdfc5b18d181176
f766adea5addbf4ef6e5cc706d579f15b3fabee9
110687 F20110216_AABVHM taskin_z_Page_049.jpg
5af4d69c169214a619bf26caad79234f
2780492e08703d7f2b0754759fc8ea1a2fe27ee8
118758 F20110216_AABVGY taskin_z_Page_022.jpg
012fa1fffdfc85a90c2a245c5a04a33f
f9b07cbf13746879d8d8cb95c38be6a9ff42acc9
F20110216_AABUEK taskin_z_Page_154.tif
1582b56db1436db73feb1d9422499b40
35fd1444ec3032491f2c1d1d6599f77311fcd3fb
2415 F20110216_AABUDW taskin_z_Page_051.txt
6c7d543f3faf0c8540ab40832f69b670
6f219f81bbd2181c86621cf103801a768eb7d367
73435 F20110216_AABVIB taskin_z_Page_076.jpg
00ee6b899c3c76cfd486acde0dd54c3c
a2f8b81e10756d5ce28e042fcaab6671d0117694
101111 F20110216_AABVHN taskin_z_Page_050.jpg
4caca8d12dbc96432da7006c0e642cb9
25524fbd73e32e7ba7627d1dabef398de077e4af
78897 F20110216_AABVGZ taskin_z_Page_023.jpg
fae8f3ce7c5eabfceb9150e831b06f55
fd0c574a5c40910166f653b398491e440057d03e
23726 F20110216_AABUEL taskin_z_Page_148.QC.jpg
cb7180a6584324649fa0247735ebe4c3
6cb3301102ca5e3c0b3cf60e63debb6423e76708
6394 F20110216_AABUDX taskin_z_Page_105thm.jpg
075f276f1a4e3c14055bc1fd8873ed29
30d2f611de4e068a6d24079c8c80ba3d4b835f67
73978 F20110216_AABVIC taskin_z_Page_077.jpg
ad93ffe1d0b51ddae11f3f8c7dcae8a4
4a715565e6da6146933be229ab6d3da254d04976
118788 F20110216_AABVHO taskin_z_Page_052.jpg
e6b08fb0161219e9b5c94e94cf89fb0c
00c2674f6ea087d3ad23af2803d9afca3fc6f894
2118 F20110216_AABUEM taskin_z_Page_029.txt
6c342e061f3d594876e0274ada924274
43a5504d9e8775f15849a70c5b4a0ba6b705a807
33005 F20110216_AABUDY taskin_z_Page_049.QC.jpg
b423aeb0ad0daa6250e98ffa014a811d
d6c298cb5fe6d81bda084a376eb870a314cd2889
1037256 F20110216_AABUFA taskin_z_Page_025.jp2
c1b18e36f0a57813466953676cdd3847
4bc5805e88fb3a1f0e48d4bbf487b106935c0450
77681 F20110216_AABVID taskin_z_Page_078.jpg
d7c45d063ae73c2da2c54116c049462c
2196d17842e13725e4bdc3e90e5b7bb415931f3e
102972 F20110216_AABVHP taskin_z_Page_053.jpg
f117819964b678b56cab5aa28b4f7799
9be16fae76bfcab3e1d0a7138f54b0c22a408176
114958 F20110216_AABUEN taskin_z_Page_102.jpg
21a5ef0bef0647e3b01ee1aca52eeba3
bc69861b9f57238c19d1ffa0608d03361b5b0b76
1724 F20110216_AABUDZ taskin_z_Page_098.txt
e3c11df583dd16b3a643f46159a34a5d
b3feb86702f81613212e815e2704799320656f0a
96339 F20110216_AABUFB taskin_z_Page_147.jpg
e29286fda1a24adc336dfecec38f5e37
9cb4e84cdcf9b61d3e3a41118c2ed66711c600ac
86458 F20110216_AABVIE taskin_z_Page_081.jpg
a3689aa294a29e59183998fd481652e8
10dadca44f5938c233cdfef36012c86e48e225be
88556 F20110216_AABVHQ taskin_z_Page_060.jpg
33b1a23926d5d19e0332a41f50e34d75
fa48f534bd11a1cadd1a5dbd04aa751da5ac46d7
1051917 F20110216_AABUEO taskin_z_Page_031.jp2
7d59e86ceabce94026c080c43cb5ce41
b25e4b13faa83a452f4b2c38fb54620eb0394836
2680 F20110216_AABUFC taskin_z_Page_156.txt
098d2b582cf6de19378a5b75b77ed126
0912571b6129483e727e276b2d857eb7b916d255
115854 F20110216_AABVIF taskin_z_Page_082.jpg
9840797caebc13fd6b5190cf6245b198
58787930b6d8d99833d0c531145ad5a05b26ad1a
99664 F20110216_AABVHR taskin_z_Page_061.jpg
0e733888c5966d7f259b502f25e83fb1
00a9eb5e1d8f35fa2296f4a313a6c1507e4d2c7a
7508 F20110216_AABUEP taskin_z_Page_044thm.jpg
edd8d8043ebff3de2dc2787b654134d8
c31300c51a1cd05b4f878b7f3f1c127d2381d90c
2126 F20110216_AABUFD taskin_z_Page_074.txt
043ffd634a0bb5667512a8129ab6fd36
9304b9cc0efe54f0a2809838d313dbb6fa2a4c25
117541 F20110216_AABVIG taskin_z_Page_084.jpg
29a175277c6fc5bbd1fb89818fda2c5a
a548cc8ef58e9021954033a29d331c5cb518c7b8
113965 F20110216_AABVHS taskin_z_Page_063.jpg
6f4e735c7a58fbb4a116a9a453c55f6a
b7d8baf2ec2f920ad08d7584010fb55f9d7b8210
8233 F20110216_AABUEQ taskin_z_Page_126thm.jpg
9ff8b31ae21c3e08c7f7a299d83317e0
3874f48578ed70fcacb5e84dead2b5c8ac2e33cb
135979 F20110216_AABUFE taskin_z_Page_155.jp2
d90032c6db88f46d280f6065e07ef011
5da963d6b54cb73a1f60fdf528a894c9bca82256
122022 F20110216_AABVIH taskin_z_Page_085.jpg
2f891ddc6ea52fddb73e1e5f9d64a9c6
caec8ae09dd17f4fdda17c71a05da2624c8fc58f
111722 F20110216_AABVHT taskin_z_Page_064.jpg
9daf5acee4e3285116db6a654a3fd4e2
753271fcb859fdb35ea151605bbd5cfd559ab2c1
88970 F20110216_AABUER taskin_z_Page_091.jpg
bbfc3db7f06a04e693110eeb2c97b548
8efd2b6ebb9ee6430de32122660bf594b1258a98
1009621 F20110216_AABUFF taskin_z_Page_093.jp2
8dc121b7cf6d75ef8983c1694293fec7
32e2178630f10d7a108455532aa097c65f11927c
114003 F20110216_AABVII taskin_z_Page_086.jpg
baf3c740dd6af72f73f68f519ee7e89c
3afec0b85fb7696b79cb4d76f16785304a130078
80688 F20110216_AABVHU taskin_z_Page_065.jpg
5307fc2e47b294192c1a3faae4ee6125
466aff0a1da0515fde91e400b5efa0e7b6804dab
24886 F20110216_AABUES taskin_z_Page_125.QC.jpg
36a14747acb62513f883e075e4e5f5bf
65a927e4c492af99a05c07a2587c84c2ebfa72f9
F20110216_AABUFG taskin_z_Page_043.tif
8e506cd8604ce86cfca4714f402550ca
dc0aadffb3634aa8df45873c5c73a9fd8a6a1c52
93757 F20110216_AABVIJ taskin_z_Page_088.jpg
4b2b6660950e8c92930d89d30be6a02b
d97cd2541f8792de8920823010917a1969893e5b
119867 F20110216_AABVHV taskin_z_Page_066.jpg
eb82dd6cf92cd2d9f2540dc8b014fe19
8b64d9f7ffc3ca5801d003a60557789ffb42222c
36289 F20110216_AABUET taskin_z_Page_075.QC.jpg
862603e199445fee0bea0ac1e374d936
100035d93e4a5a662e1ad789c8ff356a6badb232
1798 F20110216_AABUFH taskin_z_Page_078.txt
c92380f55259e10a8c8704712743fdea
578cc42b63dce1278d779389e3c69bd647768838
96209 F20110216_AABVIK taskin_z_Page_090.jpg
4cecd392c94c627655f614dacc4b8bb6
64f760bc123accb059f26b0e2c1d6d3681bc2828
85851 F20110216_AABVHW taskin_z_Page_067.jpg
40a57c4486ee32400483479d67f0f3a5
6e15aafe881508f7844ab72295429d9adba0e36b
F20110216_AABUEU taskin_z_Page_117.jp2
13ffb8bf9e2d38074c99a93e51597c5d
308e8cffe07babe228555b8117e8ce6bf1bfcba0
53506 F20110216_AABUFI taskin_z_Page_053.pro
b72fc89032b79706cae2305228e85e5b
9962e9eefa7cda0aad8ba1b3788c34f9d10b643f
91569 F20110216_AABVJA taskin_z_Page_132.jpg
1cb2c9f5aab66559c8a32f1676e9136a
1eaf945bcc24115d90208fb563396fe887011e0f
92212 F20110216_AABVIL taskin_z_Page_093.jpg
dc81995f3af9f3b5220e330ff8e5bf87
f308d1f48fc696ef4b1ba9613a421664366c232b
81771 F20110216_AABVHX taskin_z_Page_068.jpg
6db66166aad7b43ac4d502e181c347ab
2cf8abf5d9f380e93a17e85d9a7bfdc40f73a774
8389 F20110216_AABUEV taskin_z_Page_115thm.jpg
a8163085a7d0d3078896aad39f1bff05
1a2c39036ff9b41407b60711bcb9d1c3ff40520a
90348 F20110216_AABUFJ taskin_z_Page_036.jpg
23e7e11829fd6309e921605daa8804aa
f0ca9a6d92dbadd083379bec9008d6985de10a8d
94384 F20110216_AABVIM taskin_z_Page_094.jpg
820996dee52d09a8a4dd64fa8458fdd4
667e75236419b4333aae0591a502cdbe544c6a16
114426 F20110216_AABVHY taskin_z_Page_069.jpg
2f9475bdcf49460e45d426dee1c53c09
ba400343bfdda9273c2c98f3efd67e246b3fcdf8
8336 F20110216_AABUEW taskin_z_Page_021thm.jpg
a9e015e2d568452d5bfb3cf38417afe9
1a611bffbd11e1ded9347f6ad8792645c131dce4
110779 F20110216_AABUFK taskin_z_Page_070.jpg
ee61cace9d2f98414ab7e7583a51d5f3
6b45175b7c529cfca2fb8797b47e0042f14fef67
58519 F20110216_AABVJB taskin_z_Page_133.jpg
fa0e90e858db8f4c3c875e876f3102b5
d88568dcfa64990ca23883f5c22c48ee3a9fc4b0
82518 F20110216_AABVIN taskin_z_Page_096.jpg
5176591c0e963f090bda19cce1689b2e
9a93242e3d2686eff33d096cc37b76bdacf8a3b5
108009 F20110216_AABVHZ taskin_z_Page_074.jpg
f6377e1ff19429cfc4d72f06bc33470e
4b4bf0ab86b2ab8b708ea2b75b0f4f25fc633741
53460 F20110216_AABUEX taskin_z_Page_141.pro
0986b13410fc4a052b34ce1d32fe4b29
65fb7de8cf284e5b2582679eec130b2f7306fd9a
87437 F20110216_AABUFL taskin_z_Page_109.pro
f14f3a2fa10033d922b7889858aec2b8
392eb2b23ec86e92748ce3ff035496def761eb4d
117591 F20110216_AABVJC taskin_z_Page_135.jpg
20460b56177d2aa9045c614fe96aa012
cfb9adb697db6835cf3ee7a242a066cf17de36d0
82440 F20110216_AABVIO taskin_z_Page_098.jpg
de1adaedbe74b9660bd5b5eab7e45ad8
441142a9f1362e45397b2d8c6b3acc32b204fe46
96649 F20110216_AABUEY taskin_z_Page_029.jpg
94a925488e0b90c81cdb9c813b47262d
6939097f8af0e16434e24fccab712bbc79eb701e
2030 F20110216_AABUGA taskin_z_Page_090.txt
6e1d19958dddeeef3088339721b9b3d0
f11658d63861d1850e826adfab6892f69e47da2e
2594 F20110216_AABUFM taskin_z_Page_153.txt
0a8e1db792b3b5650345c2cf2d83897e
e7349020cec4ae055843897c4bcaf842dfc1e381
114196 F20110216_AABVJD taskin_z_Page_136.jpg
5942d8570fb58acf55bbd50f07904626
e229cab8f0813f4a65c29f8f9fdf7053884543d6
80031 F20110216_AABVIP taskin_z_Page_103.jpg
5f255996893ba8048e8e5dce95cc1b35
e292884d17758c9e29e1ff065c8875de86b44bd4
129052 F20110216_AABUEZ taskin_z_Page_156.jpg
3bb07b417acba07f4ba17450bbc9ab37
91e58421314c7990ffe31e46b90983c6064fec97
124125 F20110216_AABUGB taskin_z_Page_149.jp2
b13ff00270404d3fd6c5221860fb3476
a5c79c7b729f8e1fbe57aa4e78f249d4e6aea780
1051976 F20110216_AABUFN taskin_z_Page_020.jp2
0a2ffa11fbb3eb3c2d13ed952ce50c56
e550ec3be5106e769a58d039bffcd852721b6dbb
117078 F20110216_AABVJE taskin_z_Page_137.jpg
10f23c4b3c95d4bad86a204a835b8285
49a8b3484b2e01264998119789d027c95ceca605
87991 F20110216_AABVIQ taskin_z_Page_104.jpg
f119f7f20858cefdb7a7f54844e40e37
965e31ff73be8d2a422559cb241d4755b08ff901
2307 F20110216_AABUGC taskin_z_Page_138.txt
297e6935540353eee8f220d5220d34a3
86cfeb7456a13be2b13867b7d84ade30670a03f4
8016 F20110216_AABUFO taskin_z_Page_150thm.jpg
fe8b71d07af925337cbcfbf51b9727d0
911b2a1bcf6405b7f067d2c68e3fb1e1f5530c4b
114486 F20110216_AABVJF taskin_z_Page_138.jpg
bc300083882a48bf04cb7a32352a0458
3f044f971603c7185a8df80a1eaf7875099f2047
104422 F20110216_AABVIR taskin_z_Page_106.jpg
05b2d69a584e4e00be9fe90f313bf8c8
7bd0b21910be204ec99ea2c826ad0c9cbf5bf6b1
2017 F20110216_AABUGD taskin_z_Page_094.txt
f6315b9403f4c3470e07d582f7332601
54be23c0a3b22d1160ac1d38258a73cd967a97df
43769 F20110216_AABUFP taskin_z_Page_142.pro
05eabf4fb963696449a38ecba31a5e39
c43660c56e73f9048090872be2a87db9047a26f2
116601 F20110216_AABVJG taskin_z_Page_139.jpg
0d1f2e7a779dec1190d38e754eb264c0
b1a89f1426a64790ce6690969345e923e77cc748
115641 F20110216_AABVIS taskin_z_Page_113.jpg
57194dbd7a654983a1473ec36314f373
e0c5b06ff8d117b4915a82d377988ee2b162ef1c
F20110216_AABUGE taskin_z_Page_100.tif
3123e9cb1f09ad1648a4e0d6b2a27a71
1a04977d18102f0137dca5d0ee3bdfd5841204cd
F20110216_AABUFQ taskin_z_Page_046.tif
db5fa92f6db3f07ac7449ac3c38f4f00
575a2b83509b8fe8a330c2579bdd5b3b0f512eb0
100903 F20110216_AABVJH taskin_z_Page_141.jpg
004536aac0786a95bb2128632966ca5b
e6324bd729ce98a65a2a2e9966f73e5e9c3b3caf
126183 F20110216_AABVIT taskin_z_Page_114.jpg
766d9249bec7adb5fa66bcabe3c5e5ab
cd8120dee1c9091cffd0f02f3af65b0dbd0f07a0
608988 F20110216_AABUGF taskin_z_Page_133.jp2
5d085c12b537ea802ac8b89bb59c6bbe
7f64b9254a6d5069fc5bc54dccdc3e34c1af0ea2
25744 F20110216_AABUFR taskin_z_Page_059.QC.jpg
0a4796966bbcf7829f50d5627b8df3a9
50ee294bcc7c7543d9231f6f4351551ea6456368
84113 F20110216_AABVJI taskin_z_Page_142.jpg
0b37a80a27cef7d14bb7fe6dc7ed1dca
3ebaf90b79302e131e621d06e9f0ab9413be67cc
104171 F20110216_AABVIU taskin_z_Page_117.jpg
efdb297955761d8b938ea95a8e506cb4
9d6e3872a334bf02f480485207ee4829f37cf659
2347 F20110216_AABUGG taskin_z_Page_073.txt
c039bdb00d1a486ef460f7ff73dc711b
36652457936172cefc4e868e24a9db2ab913b650
F20110216_AABUFS taskin_z_Page_016.tif
59a531e06839a3a66c2753dc4d49622c
859c8b091fbe1c1ca98623d593f6ee90c52eda15
10631 F20110216_AABVJJ taskin_z_Page_143.jpg
ef778f29f72eb5dc3472c2f9cbff5a92
1f3d41bb5c62f08fbc2d653e6980d5097bf4183d
90780 F20110216_AABVIV taskin_z_Page_119.jpg
b25af709fe50cd67f22813dace426df4
ec994daea4dcc15580b9af77c092103e154e753b
8274 F20110216_AABUGH taskin_z_Page_014thm.jpg
8d061bcf042b1aaf6436ad5887366db8
b97bfabffe545015b4dfa511baa3a403e899b4e8
75604 F20110216_AABUFT taskin_z_Page_087.jpg
687bb57cbaa8779412d657c0d2be4bec
e032770e03562e311f94d05aa046c5533017c101
104501 F20110216_AABVJK taskin_z_Page_144.jpg
f16056e712cd9755febe9dcd53a59376
5ca942d9770489456aff4984c6563306ddcec30d
113630 F20110216_AABVIW taskin_z_Page_122.jpg
a02e78058eebf3fea4b406498964f298
ca10a21117ca1dfbdd569ce6466b637076870dca
57624 F20110216_AABUGI taskin_z_Page_069.pro
7cbbff82436fcd2d0a29e501f47e792e
9dc07417492d4a0aec313da116778d5b28b98f7f
F20110216_AABUFU taskin_z_Page_070.tif
073068a44a9acc507928ae24c8a32436
d56a5507477f326ef79356117d067c1fe756942a
878095 F20110216_AABVKA taskin_z_Page_023.jp2
0ac6075fffc6833a042135663bd7dcd8
bf1f6b9dd24cb34aa1a8bde314e75c8ec640ab8c
118778 F20110216_AABVJL taskin_z_Page_150.jpg
bb144a25a9a7a69ed5f9ac0ec8d095b4
00391868ad9525aafaca950862fbddf229ce456a
79445 F20110216_AABVIX taskin_z_Page_125.jpg
3db52de743dd61ca9edac10835cb99b0
ece95a8454b0427346863ac54257beffdf9f15e6
F20110216_AABUGJ taskin_z_Page_021.jp2
ea55bdc573c3c466e5653d5ca8e96edf
d82de45a3a9c51d3ab5bd2950dedc6ca900d98cd
54731 F20110216_AABUFV taskin_z_Page_124.pro
de28ecb353d2e6dd2f3ccc667d2a7342
88d6c7cc66707a8a4b4cb29d3250692f133de15d
964300 F20110216_AABVKB taskin_z_Page_024.jp2
daaeca2b5a025bdb4eb830451bc611a1
2479e0f7c5e1d4417f8952f853d0d1a0cb3d04ad
121563 F20110216_AABVJM taskin_z_Page_151.jpg
bfbcfac0609366f9ba39b06cf5f12d24
072effcd5b6148cab50db7ec3ae698b462304c58
111155 F20110216_AABVIY taskin_z_Page_127.jpg
a866864db3cc03d78131c360a033e5af
e253389f7355b27cec82feda4486ea1f631f88ed
1051982 F20110216_AABUGK taskin_z_Page_046.jp2
fe6afc7c7d5af144fdebeccab03b10a0
b95eeff53ccdb4e981c389963d199448c9c25916
35893 F20110216_AABUFW taskin_z_Page_086.QC.jpg
2033dfabe48a7daf6df0f7d0a984fbca
d6537298ba9963eee6b1f941f1d950c5147002f6
119911 F20110216_AABVJN taskin_z_Page_152.jpg
5e1388186d5d163ab354a5cb6d43ab5d
8f0f41870ca5fd448283f8750513b18f0df379ff
92911 F20110216_AABVIZ taskin_z_Page_129.jpg
d13ccc8b1ca96972c3d2188c9d0e4c17
32c607e9d0b324a495576138b183759437a185b4
831927 F20110216_AABUGL taskin_z_Page_105.jp2
25ccbf216de14af481c38379f1105183
bd9c3f1cffd75f1732090601043cbd07337f19fd
2269 F20110216_AABUFX taskin_z_Page_022.txt
ebef39b0d99bf4c1cbe54188c67978e4
ac2a596d48f395cb357d2f05bfe569322368a640
757093 F20110216_AABVKC taskin_z_Page_027.jp2
afd6c461dcd7df29f73450144db6aa1d
408237ad5e1b54cc45820121d5d6c989686575f7
124851 F20110216_AABVJO taskin_z_Page_154.jpg
9b793d20caf9607db766c419f2e178c8
4ec919f022b186df10fa5476c525df379b27bb41
117190 F20110216_AABUHA taskin_z_Page_054.jpg
f1f1cdc538ccde074c75b03da688aefb
e303c698e6596d796f826dc8ebe494d72383f31d
69921 F20110216_AABUGM taskin_z_Page_146.jpg
9b1c79c8082b7c2835985ea65daf1350
a07482f01cbed4f5efb165d5da2de06573ef7511
F20110216_AABUFY taskin_z_Page_088.txt
2929652681a8809f1b3ab0d77bdeffc5
9891a040af6d85e93f04d54c19f4b0147f168f7a
F20110216_AABVKD taskin_z_Page_028.jp2
97e65df242157a14bfec3c870c32fe86
4021db8ac4fc593a5dc6f9bfe9a831ba3a242132
26480 F20110216_AABVJP taskin_z_Page_157.jpg
7fe34cba3bd162aeb47f6349dac130d0
770e5c85d339b339ced14b7c9ce31952252afcf2
F20110216_AABUHB taskin_z_Page_042.tif
ebfaef87b95a2738afdc8a99ab2f65b2
e5aed8881d89b36550367e64c227ba1d49047c10
F20110216_AABUGN taskin_z_Page_066.jp2
36aed4f517543553464cf835832fdcdc
29dd68b8571432829f9a8c877866c1891e97e21b
2084 F20110216_AABUFZ taskin_z_Page_093.txt
7dd222aa905b080f031d5d46d78fff72
cd1983f0187b15765a47293690bd3949fb9dd68a
1051970 F20110216_AABVKE taskin_z_Page_029.jp2
6458fdb39682b6de936ac7d949e9818d
e7c2b6ddbbb7b77f2bbac96de337f5bd3d6be4d7
42889 F20110216_AABVJQ taskin_z_Page_158.jpg
36afc21dd3bb9f1e3aeb85877794ed61
0152c594e2f9638cb926b79b2bbb038315ad942d
F20110216_AABUHC taskin_z_Page_010.tif
a4763760988cb243cbda0c4316fe8875
eab6ac226476c5f53049da40eb0082ab8f0655b0
2114 F20110216_AABUGO taskin_z_Page_026.txt
63bd01ce17eb65a49597d6871c4f6bfe
f8a89c3877c6a64f03d0dc93714953d95fca399e
935855 F20110216_AABVKF taskin_z_Page_030.jp2
61f2ac7805f424fd78eeec3b47651971
e3449455cf9acd946fe1284bfe7dbbfb8a1e2c83
25586 F20110216_AABVJR taskin_z_Page_001.jp2
7dd889cb3ae9c3280dc6a1ac9a9010a3
ecfa5a4f62acc754ad57c588d3e64762e5eb2b93
F20110216_AABUHD taskin_z_Page_150.tif
cd061e0cdc7cb3850ddea3e180fedfac
0d596ed7af13ff19836cc615327e29893aac22a3
2403 F20110216_AABUGP taskin_z_Page_135.txt
2f0c5f3b1425bd96e159022f41f8b12d
0d2c24b8b105786c9ab111255f62c0a637370d1c
906913 F20110216_AABVKG taskin_z_Page_035.jp2
83a18072ca78d48aeaa255b6529b60ee
6c51b12cd75780f97d07ee86daa97d810e61d452
F20110216_AABVJS taskin_z_Page_006.jp2
b883d8a90b98d72f68ee8e59dd3807ff
8fa6c70e45d799fd6dad95788afec4a717adda7e
F20110216_AABUHE taskin_z_Page_109.tif
b3d41b2673787ce36461d9e9fe80c654
d6fc8b1f0c4da6b608e0b62cd2005e6a73708f5c
106824 F20110216_AABUGQ taskin_z_Page_128.jpg
5f8d613b6cdf43b5630a73d2f29ccbd4
4ee09f713fa9309e475e27251fba8b2aac8ddd96
1046032 F20110216_AABVKH taskin_z_Page_036.jp2
df8f324609650212a44d7dd86eb4b82b
57e5d66d13b56a267b877894f3fddb29e520eab2
802746 F20110216_AABVJT taskin_z_Page_007.jp2
ccf153c7bd3cf08fb74b11081459679e
3ecced0dda30e56c7412cf967920069a42ba77d7
61249 F20110216_AABUHF taskin_z_Page_114.pro
6649aeafb65bc588cbe981d9971a40cf
346db1feb2ca1c6af9bb7daf9fc10c9acd2d03f9
101067 F20110216_AABUGR taskin_z_Page_099.jpg
2d681a832f3fef4609f6aa8f2bb4a6cc
e7fb27973a4cdea37c6bb3dde0587e306b14cd2c
98256 F20110216_AABVKI taskin_z_Page_037.jp2
9f0365f12d43aeb7230cd185a5eeb620
e55566c27af810adfbb50d335a91523f41c9c196
F20110216_AABVJU taskin_z_Page_008.jp2
130ac1fbbe7d028d85cd80a6b90b6176
df2ce38a8ee0bc12b0e30ffb3be3f2bf7ed33908
F20110216_AABUHG taskin_z_Page_129.tif
b4f18f827356c05e742dfe2724e321d3
333c46b4613ccf308744fd2ade16885aaf5a349d
F20110216_AABUGS taskin_z_Page_115.jp2
fdbf1c68510db06fda4bb6bfc9c07b5b
cbe460c0262a2d4293d558357a063a5f1d447421
121784 F20110216_AABVKJ taskin_z_Page_038.jp2
f426b453b147cd28b5fb44c6c8fd90e3
db5e65c078b909602a365346aea852f2db59b883
130228 F20110216_AABVJV taskin_z_Page_011.jp2
e4a2ecd29e4dc2dfc8b66987237b1ab5
f45f1cb6f1a21c2b97be4668f6db3e63b9711ddd
56777 F20110216_AABUHH taskin_z_Page_073.pro
e5616d1ea3f38ee97bad25b4267cb77e
ebb5eb71cbb257e30584dfe1e1ed63082db7b3ff
111085 F20110216_AABUGT taskin_z_Page_149.jpg
8364d202d1b966339f918e90988122e3
87a0f253b74ba87c2a25860876f018229b29ebb1
F20110216_AABVKK taskin_z_Page_042.jp2
a0a3c64352fad2ebbf353cf19058bf5d
8eee505c285ea6250380911c61c9c2e3b1a73ad3
58859 F20110216_AABVJW taskin_z_Page_012.jp2
cee1d67a2305d6da65aefe024ccdaf12
ffe0ab477c54702af3489863efc752148abb3eb4
F20110216_AABUHI taskin_z_Page_114.tif
39678797fd69d36f15468ac982c3c0af
c8e49e6d784233e76db5db84aa0d1c25018497b5
2227 F20110216_AABUGU taskin_z_Page_134.txt
fe1044c1c4f71b52fb185c0710a62c68
4f659c1f25c4cdaa709f14cab802c7a383503d2c
F20110216_AABVLA taskin_z_Page_073.jp2
9b05d383a15a6661c36b606b3ed830e8
3274bc128a1bbd1dce69c0c4c7984a67d8b9b2cd
F20110216_AABVKL taskin_z_Page_043.jp2
eeac8ec6e1ae69e5e1e1418c30a89f89
41e8035ba6c30045107b67fa51134dfb419461a7
F20110216_AABVJX taskin_z_Page_013.jp2
cc603554a23a05d8a39371a579abb5c4
467bf1057a3d781e73c28084246312f95343b184
1051960 F20110216_AABUHJ taskin_z_Page_060.jp2
e966d2bf3993aaf6d05a4e80c60b3112
ffbd21b4603741ba605c4c93de01f7b2bbd18f13
2008 F20110216_AABUGV taskin_z_Page_039.txt
c7d80a831dcfe4ea56d8abb778590efe
a0608a1ccc61860f4b6bba427c4165e27872d2a4
F20110216_AABVLB taskin_z_Page_075.jp2
a9ad86591ec61c0f7d79afb3854fb539
a50b3e9d391eddcc1b711c184b77245d73196ff1
1051969 F20110216_AABVKM taskin_z_Page_049.jp2
a3547842ddcf0ec4e0e6d37cba7e7ae0
6a05e856cea2132ef7daec0a23b13faa18edccf0
17671 F20110216_AABVJY taskin_z_Page_018.jp2
5798da177ab1a61da54be991af910275
cd301f2a551a231aa381c03c6025e33973a785b8
1622 F20110216_AABUHK taskin_z_Page_146.txt
90b93bdce27c0e5d20443f7c6fdba302
19500a4f1ee1d10e1bc3f65f6af601d05d27245e
F20110216_AABUGW taskin_z_Page_084.tif
0fcae4044186a3df310a1107f204e873
4e263e4200bb166c1ed29bcd7b4bfcc8d1dd92e6
747292 F20110216_AABVLC taskin_z_Page_076.jp2
76d6e340d19596d7962ba57e718083ab
8839d8a99f689f09d70b5ea1d8d39716ed02d76e
F20110216_AABVKN taskin_z_Page_050.jp2
c91a3890aee24496f13d5d6948cf5b58
cb24401a54d6ce6932e2a9c0528eac530be6b99f
F20110216_AABVJZ taskin_z_Page_022.jp2
02fcb39b7da26bf292910ae427a6bc90
f944d5956bb1a3d725868d7866cf187c4e837a48
97904 F20110216_AABUHL taskin_z_Page_081.jp2
578b6a68be6bf481f6df21d60051930c
33f772633c0e0b9527c1b30ab2a56d5a6cb18fb4
53738 F20110216_AABUGX taskin_z_Page_128.pro
341a5d9f0cd195c59d06a1a537b2b2bc
b6fb16026b12ac6ef5eb06f9cfef7842d980fece
7279 F20110216_AABUIA taskin_z_Page_043thm.jpg
f6b5277e584e1ea7116df3297bbd7eed
9a3607495605876555c665efc12991f3de15334f
F20110216_AABVKO taskin_z_Page_051.jp2
e4c6a46200169883ccdec674a3acd03c
3814db5dd7553877125dd2c02eaa4d945d38a4f9
8083 F20110216_AABUHM taskin_z_Page_066thm.jpg
bcfccfdb857e7b93e75f59163d21116f
ed49cf71ad135d7eeb1dc491a6c2812bdaca4bad
13384 F20110216_AABUGY taskin_z_Page_143.jp2
9c6d640dc3ddf43cea932672c1b2e7fa
9ac18fd578e31e68fd4a6382d6e601e873ac8c89
830564 F20110216_AABVLD taskin_z_Page_079.jp2
6bf147877a29ba87f80441188714f437
bb3b9c493c3e24761a669ae84b3de7f4e9b390d5
1051942 F20110216_AABVKP taskin_z_Page_053.jp2
e6ae3ba7199c631ab65dd6de0f24b1b9
58f892fef304c0ab9e20b007aa0e664cd57c22a6
8544 F20110216_AABUHN taskin_z_Page_156thm.jpg
9c3584547742f7c2375b78ca2055ca78
48ff6e3df011cc05c17a55af8100a5e1c75fe4f6
2089 F20110216_AABUGZ taskin_z_Page_062.txt
5f4d1d6b120b4f529360a4915ceaf3c4
9dba0f2c35cdd0538363ec84db00afa35149e907
2297 F20110216_AABUIB taskin_z_Page_034.txt
774675048002a5cd290844dbe50cde50
d2f5f29334bd5cdd021c7df8415b98f1d1e2480a
F20110216_AABVLE taskin_z_Page_082.jp2
6638cfbfc0688757751cbcaf398cb265
5c45e9ebc1a41a2dfe18af0485561a6dc5e1d314
F20110216_AABVKQ taskin_z_Page_054.jp2
d54c9dbb9ecf78b89d57d119aff2df68
f06e5db97591b900075e21d7b75e97f91897f37f
F20110216_AABUHO taskin_z_Page_057.tif
1fbbf79b74ee9ff2542a777ff4040ce4
7aac851560b148e647fe0f831d4a79333ed4747e
F20110216_AABUIC taskin_z_Page_074.tif
f729e39b2b3715e009da7fd5d25c6c70
7ca099c539c884cf90ff69c6d627626e0bb2e732
750063 F20110216_AABVLF taskin_z_Page_083.jp2
1e72dd3df2b991e71bc7e3d9f8f3a549
ec5c8cb7844cddfecbf230c452c4e403514331f2
F20110216_AABVKR taskin_z_Page_055.jp2
edaddeb380c1a7b1755fe11271e54949
212fcb3976c594f354bdeab74e8f39d9f9dd6795
29070 F20110216_AABUHP taskin_z_Page_060.QC.jpg
175b964b2d246ac3f84de305ed835476
c4b543d8683e733ade0402da1a942deaff479de9
106071 F20110216_AABUID taskin_z_Page_073.jpg
e82c81a4314c3605686e24f81ec34587
bb43b57acbf1382365d7ccb55841712f0535eef7
F20110216_AABVLG taskin_z_Page_084.jp2
a749b9d0f2608b79b617f43135260c72
35e51863b1e39020cd7ba0ab7c6dfab385b0fc9d
F20110216_AABVKS taskin_z_Page_056.jp2
f7668dcf2a6f0a81a5f261f4f974393e
258142f6647605679e0dc67816e76549aad956ff
37708 F20110216_AABUHQ taskin_z_Page_085.QC.jpg
042d2e9416f6121f947fe862f77886c1
04a16ac84f1005c6731fcbeec615533cc7730e8d
8208 F20110216_AABUIE taskin_z_Page_070thm.jpg
27e3410513141362ee96dd8f9e4e5222
d119bc1e024e7b577f760a4feb5a4238f1226e67
F20110216_AABVLH taskin_z_Page_088.jp2
cc7ae2daba01d6115d5a5dd38d05a7dd
ec54c60c06d392cdc991eef92f04b1e3bd38d76b
90213 F20110216_AABVKT taskin_z_Page_057.jp2
99ec002727f22ede843b84c792ac56ae
2fea2cebe13be24c3014386c4600db3aedaaa35f
1051928 F20110216_AABUHR taskin_z_Page_070.jp2
3e5d70cb5837bfafe8e8043c37803862
3f828d7f0f1f96015a117c79f98561edb2e4d785
F20110216_AABUIF taskin_z_Page_104.tif
b1474b71aca534c042a219e3a73b073c
ff5e645492dd22091aa98162b5bae4c74f4b41ca
1005455 F20110216_AABVLI taskin_z_Page_089.jp2
8b2fca90f6ce3d1496f79843296f55c0
934990a7d80abbe74814114cd8af1c6c6be64a2a
882153 F20110216_AABVKU taskin_z_Page_059.jp2
04a675aad0ed15ddfc51e9fa00635f02
7c6cfdfc4a90ec849497b3bbc9eda562b6cbabbf
56446 F20110216_AABUHS taskin_z_Page_113.pro
16d099b44d626021b96fbc694155b99d
3cf6b437e61d17d0caf94079be1e529b7d96b482
F20110216_AABUIG taskin_z_Page_062.tif
31c7d856cbebd9031bf60bf7465399fe
70f8627ae8212a14406b5a2fb08b84a85bc02024
1017150 F20110216_AABVLJ taskin_z_Page_092.jp2
181a68817a87f63405509d8d3eb18de1
f74f6507c196e569a4a63f58b0646859ce25454b
F20110216_AABVKV taskin_z_Page_061.jp2
d055625ae26992b2b236b6af644d4e80
d425d61cd5bb8dcdfd0d6fe04a200b83eb5f1a8f
92 F20110216_AABUHT taskin_z_Page_002.txt
053d5fc0b781230943e5cc87aa240cd5
3a4c7a4c06aa2cf4c049224d35fe6e22227801cd
116456 F20110216_AABUIH taskin_z_Page_130.jpg
184a735de337d7fa16e40f9c47b02016
c3c6c546be6239ba5af9d0b97dee85b4f5cae9b3
F20110216_AABVLK taskin_z_Page_094.jp2
e5b58b2d492d156fa451f5c71d6e3e3f
a3c62af800c6f490c7a6505e59bf642e427cc8c3
1051958 F20110216_AABVKW taskin_z_Page_062.jp2
0c067bb4dae3bb34b585acafe78151d1
441624e94870f47db01f4caecda6b64672b4cecd
6107 F20110216_AABUHU taskin_z_Page_004thm.jpg
4ba3316496c4d2571220bf12ce9a8f43
9550742e9d4786cde8085963e24139a1bd70d093
41365 F20110216_AABUII taskin_z_Page_103.pro
d0ad42a0541a3739588f72c44e8c8c1b
a709b1250398bf2ed760fd86d13e15582d6b7a3d
F20110216_AABVMA taskin_z_Page_126.jp2
0931aa7bf513d6dd1a11754eca2c355f
93de58da61e363d45ee8e1a65ac7946c1602e15e
910856 F20110216_AABVLL taskin_z_Page_096.jp2
d9fd916c16b698d8fb4250681af92ed0
107d3ffa526846c28a7f69bb3c5ca7bb38fad960
F20110216_AABVKX taskin_z_Page_064.jp2
c7587ab5f6eeba5892d6b1efa6a51086
cdbdb189a6e6687ad4aa67d3329b1621acedea68
35910 F20110216_AABUHV taskin_z_Page_065.pro
86f53c5d75ed77605f32d2e11bdd05ed
d62610d11da6f73c870ccd3aa88cff2c7f1f6515
35904 F20110216_AABUIJ taskin_z_Page_046.QC.jpg
672c6de5fc91d25ba8a9b1c8f6c115aa
99caea977e7d6422f6d101e8f155ebdaff3880ff
F20110216_AABVMB taskin_z_Page_127.jp2
c6c788a998ed05a35502702f5cb9e282
eb7a0441d65ad3b91d247dc8cb82cf090fae1724
846513 F20110216_AABVLM taskin_z_Page_098.jp2
c45cdeb04a12a394e15fa5e3783494eb
6bd2a1b02867f864a254881a4aba0acff3c62fe0
910210 F20110216_AABVKY taskin_z_Page_067.jp2
8ecd1d3ce41f842f5b22aab4ca46d4e1
47aad90d5b18dc5f90b5b00ff068b4e381d59589
26385 F20110216_AABUHW taskin_z_Page_096.QC.jpg
ec923687a37e90761b49d4dc117c4a24
d643b7c931615c9cec2397c79a900c3a73bbb9dd
57698 F20110216_AABUIK taskin_z_Page_139.pro
dc34c94c811cbe28d71df91fb14d982c
ed16a1e955db7b05dd6cec285fff1990b23c3643
1051975 F20110216_AABVMC taskin_z_Page_128.jp2
6e0fccc66aed1e8f15cff89000dab376
6b3a4c21c4b2f86504d80eba4aa89dd24727ea27
F20110216_AABVLN taskin_z_Page_100.jp2
9a4d820a73a44cb33fdfb94936aec135
32530c89fa5c1859701439daf860c5c4fda1badf
925893 F20110216_AABVKZ taskin_z_Page_068.jp2
a7b0e95e2ffdd76a5eafd873824e6416
4b4804f7a1da9f63440c6216e61fc491c88c2b9b
8668 F20110216_AABUJA taskin_z_Page_001.QC.jpg
b3bdf174026681267e20d9d815141357
39a4109f702b68923c75621f8a37f50cfbb10621
50242 F20110216_AABUIL taskin_z_Page_090.pro
37e5447ecc2c0e7874f75cb3576e9e45
661dfdd2ecec9322531217d40a88fecfad3c759b
32999 F20110216_AABUHX taskin_z_Page_106.QC.jpg
f402c72ab75e81cdc4db038a46ec844a
360955c90cff21bdb85df9370c438f5feaedf573
102648 F20110216_AABVMD taskin_z_Page_129.jp2
ac934a4a6545dc0098ad250c9901ded8
0594115b1c166944f4c197f1ce66e393a2d70e01
1051941 F20110216_AABVLO taskin_z_Page_102.jp2
35ad832b1157ac0dd386f6272510d7af
9bcd86264fb0c6e35bd9ac242d0a54e162d1f2b4
59838 F20110216_AABUJB taskin_z_Page_017.pro
a27fd5f77d23962e6b7812999a7f66be
0e22b934495e97a5222cd49717467d69c42427ee
2218 F20110216_AABUIM taskin_z_Page_041.txt
f10e00329de1e9b72c98f5d56dcd3884
cff9e8a0ce2091b77fac2558a2f209cdbb86287d
7245 F20110216_AABUHY taskin_z_Page_112.jpg
ce68ef1ccbee176f56c1e4ca7a088c40
7c289758ec852cf9502dc489237ccfa0b831eb57
860014 F20110216_AABVLP taskin_z_Page_103.jp2
355a7a53508b3dc1fa38efb5581adc76
b7e5c95db68b93fa0dfee00b4773916602bf2f20
8133 F20110216_AABUIN taskin_z_Page_072thm.jpg
dcafb40f56df5f2ee32fbf974a32e300
9f53178f56657b962883e91acb2bb3252c020fe0
F20110216_AABUHZ taskin_z_Page_117.tif
2e547f6b77af74700178e655f720525c
df2a79790d4b30908b67cfc2dbf2161b6c2ac74a
966968 F20110216_AABVME taskin_z_Page_131.jp2
4aadb74e27b20c3799812aa1b930ed60
454343fa5beb1d67b04919a12227541fcfc953c3
936724 F20110216_AABVLQ taskin_z_Page_104.jp2
45101a16acb48f8601586ba96370c554
5b8807f73cdcf6df148c2d3f8d7baaa5d4e9877e
73836 F20110216_AABUJC taskin_z_Page_118.jpg
7d8eb23294d5e3c46bf8ca49bb165f74
f146b9ed8f30a4e7911a87e029222e1083f92e9c
31319 F20110216_AABUIO taskin_z_Page_141.QC.jpg
f638be67aa42052e0a99ff4fafb03a81
1f2b1788ed60ea2eb493fc8d77c49d6177acd522
1051932 F20110216_AABVMF taskin_z_Page_135.jp2
d4d4da9efec6e2fd2a3027290b0cc3b9
3b666972f520ede20de479f31da284d0e6f69189
1051924 F20110216_AABVLR taskin_z_Page_111.jp2
759203b368be8cb53fe1d54b58a6bebe
ede3bb1a9852d56611081b5303ca1d8300e0f385
F20110216_AABUJD taskin_z_Page_115.tif
5944b427ef7398cb881cb59da38c69a2
1842b09968e871e48dabba1d522d8213f8a2df1c
65957 F20110216_AABUIP taskin_z_Page_153.pro
c1c8d021e0403c6bfc4ef0628033be09
f52332b6cb4251823ecc588678dcbd731329be2f
1051966 F20110216_AABVMG taskin_z_Page_136.jp2
a2a6f5ae3265bb3ac8af2b11379a2472
5d657dfa0584ceaa95e46202dcedd1f3233a3fae
9637 F20110216_AABVLS taskin_z_Page_112.jp2
08e17cd246bdabca8545d16bb60fe93c
b6152331b8477cf026497f3f4573fbd6c239bd47
57932 F20110216_AABUJE taskin_z_Page_102.pro
685e18a68416d810205122225f18f99d
bcd688bdd2e93d61807ca7315ed246fda2c480a7
62009 F20110216_AABUIQ taskin_z_Page_046.pro
ceb6ffef5bf6aa5e2ee9a4801ea49466
7e2c35fccbc8aabeb5b025cb05db7e3da986387b
125538 F20110216_AABVMH taskin_z_Page_137.jp2
7ea8c21762a4c61d33610daa53a712a1
27a05391aaa4c6509b36373fdedb0def25fdfce0
F20110216_AABVLT taskin_z_Page_114.jp2
ae3c5e6d5233bb7e4ec65f06fab44237
58221ec002335829795706e90972625174bc3d13
52315 F20110216_AABUJF taskin_z_Page_010.pro
fe069c9d74465bee9382cf744b077b18
5e6f6916b8b0f8735503def449a7819578f20446
77570 F20110216_AABUIR taskin_z_Page_105.jpg
37b3b58a5cf75835b7dfe8cf15d8ec14
43f912db28c353220e02379c952dd1f53fa58a17
F20110216_AABVMI taskin_z_Page_144.jp2
930762337304c2e51a420b30adca6330
5da32ba66852e540b3f68c419568d01f7b772a7a
F20110216_AABVLU taskin_z_Page_116.jp2
2f4839dafa4746e8485bb77f2f7c36e2
dd9c29b734e421260288bc1636172e4618ee6622
97802 F20110216_AABUJG taskin_z_Page_033.jpg
e35d035f08d407f95211035d6d266ddd
d82be5b4ab16da30d9d61518607f8711086925ed
36487 F20110216_AABUIS taskin_z_Page_084.QC.jpg
f86d6c45129757277c91de0384891575
cccd093b1a13b0cc40628108d693a133dfa0f5ac
1022542 F20110216_AABVMJ taskin_z_Page_145.jp2
9f1c126c0a0e68aa1be2b63b2a22604a
aa8fe2b8c187a4687d0be4b37188978196db5dbd
724347 F20110216_AABVLV taskin_z_Page_118.jp2
cc0d10640455b1fc37a98a5e127ba987
ee3ca553b47020180456c3524d61b5ba985de1b9
54566 F20110216_AABUJH taskin_z_Page_115.pro
8321ea522e65d00fd58e638efcdd585f
7463042ccc54079740ff7202880cf0ca4112ee98
110357 F20110216_AABUIT taskin_z_Page_031.jpg
285afcfc80af62a963334d5d3c53dd84
2923746d4b8c83e7b1dabcd7c39eaa484d9c0437
132294 F20110216_AABVMK taskin_z_Page_150.jp2
5dda3c85c7490605f7e7d2b68c01db6d
ab110992ffb63ced42ea037c1aa320ba6a3bfd86
1012866 F20110216_AABVLW taskin_z_Page_120.jp2
532b0d0adc1f4e6871b114d30a94161b
fceee405967451917ddac540addc5ed5e87b765c
7012 F20110216_AABUJI taskin_z_Page_006thm.jpg
b9d0be79a37a437c33024d6035b88ac1
43620cadf18621f221b1661a4fc0979bafb4bc2b
F20110216_AABUIU taskin_z_Page_023.txt
29ce061e7f47ce5faef2ae3d879620a9
ec8a4be40c0a1a581687734e389455eabd80c90d
36845 F20110216_AABVNA taskin_z_Page_135.QC.jpg
b3d081506d4d23423ce8e6039c977626
ee7dfd1171d18924bcd10ad4c8c7f339dbb8b0a5
130201 F20110216_AABVML taskin_z_Page_151.jp2
3a6cd5769d52d50cefbcfc13b08deaff
c62a66586fdfc6c3e58aa8cbe93fd78ce2138f51
1033304 F20110216_AABVLX taskin_z_Page_121.jp2
6752a666fbf7f1d78c499e7ff164d2fc
b7f448cff8548f6f6db0b948b2e2d1b960db11e5
F20110216_AABUJJ taskin_z_Page_097.tif
bde5b45bc747bc7ff0239e7bb7b9e53f
723bb21753e8d0d53a92e39119530bfa53abd307
F20110216_AABUIV taskin_z_Page_128.tif
01f59d21074fe45fce10ad550dce51fd
00f38d4ea9d37673c9b8ca168491a04f601fadab
4737 F20110216_AABVNB taskin_z_Page_018.QC.jpg
3d8b65c5764e4db1d657a249e491a30f
59e439323786e60d229a7f2f5d7de2c1ff8e4f77
139153 F20110216_AABVMM taskin_z_Page_153.jp2
f8a8b9dabe6dcb37fe06b1708924b731
7698253a43ca1b3425a350eb1f6ca6db7e10cdc9
1051983 F20110216_AABVLY taskin_z_Page_124.jp2
74ad8ea1e6f02c0bb830209020acc125
6b459dbfbf92bd09b076897cb5960844b4990737
F20110216_AABUJK taskin_z_Page_078.tif
c4bcc39476b2747fdb0ad994cfb5cee0
217e9bb309526d1619a5d3546df742e327a3a3d6
52320 F20110216_AABUIW taskin_z_Page_123.pro
1c5cf99880633aa33517a470206f6a99
b63b4c52edd5cb7f911f1edcc95f42b5c646f69f
478 F20110216_AABVNC taskin_z_Page_002thm.jpg
fa27b69f9489da4ba8f419a66e138f85
46fe012372f9a47b5e01c452090480e836b4e7c4
136919 F20110216_AABVMN taskin_z_Page_154.jp2
44a8648ff88de241f8f6355016dabd67
b9a6ec3088e94b173231bd31b245556c39b90314
857359 F20110216_AABVLZ taskin_z_Page_125.jp2
75771eec769bb1128db5b4e0a6e3cde6
787f7c0a1dcacabadf2a24deebd93fb3d679b52e
F20110216_AABUJL taskin_z_Page_055.tif
280a55b9bfac56fdaea6c753a8e9e67c
a86a7aa2e41fc8bafee572b2e440d065c3384e69
F20110216_AABUIX taskin_z_Page_134.jp2
617d34d2b074a95d70aca9b813a389e2
5c8fc9ef6d617431d25d31b519b481502e2881e3
8106 F20110216_AABUKA taskin_z_Page_082thm.jpg
d17da0385bef957f0b77ab37f7593c00
12079087aaaf8c022d5cddf86b64c25830d3f1e1
34833 F20110216_AABVND taskin_z_Page_041.QC.jpg
4b2bfa1ce45a093441a7ab39487622df
8f69a70bddacc769926ebad42f7529efb3fe1001
28311 F20110216_AABVMO taskin_z_Page_157.jp2
d371f6788bc0d54a84a60ba0899eedc6
920d883014c1006b56ab7087a27f50dd709af209
1469 F20110216_AABUJM taskin_z_Page_027.txt
d8be026d75100e946fb29c84b43ef410
565b13550d71f93368a174dee6a49abaf2efd928
2232 F20110216_AABUIY taskin_z_Page_031.txt
c4c7d84abcdafa33169d88b9acb303f5
bb4543dc64bd7adf4237fac27f6de32ea5eef600
6384 F20110216_AABUKB taskin_z_Page_110thm.jpg
cb11f08c5fa16416ed704404c338e4bd
107da2d8366633f38dfefa08fff6c92128b8c5b5
7974 F20110216_AABVNE taskin_z_Page_137thm.jpg
76d6fa112cdb6ba46abed7d4c6f27784
e2c0139b9d1441ecd5eb9976dad09914cdf1c15d
44596 F20110216_AABVMP taskin_z_Page_158.jp2
7e1bdbffd182804eafc1b3e7bc2a88b9
cc0a446c2c7d90bbe83bbe2c93e6a7c181c70a60
112760 F20110216_AABUJN taskin_z_Page_048.jpg
3d8fd5be2caada35e4a2113d611f27f2
09a41a456f0769f2864943142197c6785734377c
5190 F20110216_AABUIZ taskin_z_Page_002.jp2
c59025ec9489f26bd0fb038bb8c8b94c
ec0d1ef4a368d9c285e49b1afca57c6d0ff660df
1051893 F20110216_AABUKC taskin_z_Page_034.jp2
8698947bd7024affd1c9d989bff20e06
70b1ba660db9f9011f796745fe56d2cfe8165360
2167687 F20110216_AABVMQ taskin_z.pdf
3b2b87fa52ccf50387507158eccf83e3
296564f7cdd7e156846012558048a451ca5833f7
37557 F20110216_AABUJO taskin_z_Page_140.QC.jpg
7b78ef67138b3170e10d8471a90a3315
a37e58a0af68fceb4860f1ef48c6715b1702cdb1
31709 F20110216_AABVNF taskin_z_Page_097.QC.jpg
f59dc80e211fb18c6f25cfbb9e3d8225
51b3ce5dee9ae571d15a4b0e7230d72a0ba82347
6893 F20110216_AABVMR taskin_z_Page_030thm.jpg
2feddf1374cab5f08084b9e0fad7779c
4f78a4e6c0be76062172454163901b3f0f7ebfb1
1051977 F20110216_AABUJP taskin_z_Page_014.jp2
441255f750d082a074ade084806a1fc1
b8230909be63fb55d2d359689c61d35692f864e8
56353 F20110216_AABUKD taskin_z_Page_041.pro
f5f7fcd1813eaef53593705b12cf47d7
0287b9a9320cf4ec9f8922bee2c1b91ca837a080
3077 F20110216_AABVNG taskin_z_Page_158thm.jpg
6121579b9a86bfd473ff64533c0bf10a
26d7dc729e3bcc85ceb10182bd9ef4c0b6db4904
7732 F20110216_AABVMS taskin_z_Page_149thm.jpg
080d1538c3044fb913883fffa23daef0
73d0b971e6db67a195276d52c164747ed1b90a7b
114128 F20110216_AABUJQ taskin_z_Page_013.jpg
8ca1d852eabf457a6f6f86288d082790
3daf8ab48d760faf592c4e0494632e0361b77a71
118887 F20110216_AABUKE taskin_z_Page_071.jpg
e31d700494f6fac7f5c4c69761f4ce83
17db327308a5167e1ef828a1b12adbb89d631c7e
36252 F20110216_AABVNH taskin_z_Page_022.QC.jpg
f04bc42f4ed8f68aea5f447ce53de247
014777877a224f0df3094963701a3e2e8451d551
31994 F20110216_AABVMT taskin_z_Page_099.QC.jpg
332978830808f8292f4456981a943ca8
d1f982625ec73a549b2d9c8f0345e9ad6ddbd640
F20110216_AABUJR taskin_z_Page_098.tif
29c46e1efe2ef1f49315d762b27959e0
1f15382ba1d69c681e29127d9e2725165a1cfc0e
8397 F20110216_AABUKF taskin_z_Page_019thm.jpg
1e45f3abfaaa877f1badc01976985335
b2f4358f72b5e5f60dca5673a7dfcdbc6807ca79
35971 F20110216_AABVNI taskin_z_Page_139.QC.jpg
9e62487553613f807e186cae003ea942
3cf981c8a348e8477d44e3d6298edf5f76b45953
7921 F20110216_AABVMU taskin_z_Page_061thm.jpg
05f27e4e12be0b9291b15fe2c0fc0868
11f4e1821c3e237d088e6a8c8c58f8b6527541f6
50069 F20110216_AABUJS taskin_z_Page_121.pro
f57abf64ab09e2a279d438b4d27aee40
c4e031bde91e4255343af1dbac9e9dca390eb5cb
2823 F20110216_AABUKG taskin_z_Page_112.pro
3633a58487d356ff46501dd787874596
f8cb340893e53aae2c159edee911377811838cb3
36437 F20110216_AABVNJ taskin_z_Page_109.QC.jpg
6940cb44e80126b3956cd7d0dd953db5
eda1f8a5c4c1840765462ec0091ffa704e2104a1
25114 F20110216_AABVMV taskin_z_Page_101.QC.jpg
34388c3803cfd7340179aac90f771e91
b6b16ba1e53ebb60673315d120f13fbd6f002ba6
60423 F20110216_AABUJT taskin_z_Page_150.pro
aa1b44846b72f8176511c0950e87a74b
ebc20b122b459e1d23c80b1295b67053a10745ec
F20110216_AABUKH taskin_z_Page_015.tif
988b9bf96d1f4723e5dfe90c7bf483eb
a6306c4010236362324915c721044274f3540947
28296 F20110216_AABVNK taskin_z_Page_142.QC.jpg
8705e8ab61570ae8d2f364066ce116b5
18cfe151bd7f5db7907e7752fb5293a00bfd01a0
6622 F20110216_AABVMW taskin_z_Page_024thm.jpg
fc8a95652e2e21422d17de7543e18358
96470cd9aa55c6a2b4123f379a87e6e1ab2dd631
F20110216_AABUJU taskin_z_Page_037.tif
ff8dc9694cadbdb057d96b3103c83473
31afcf14daf2aeabffa1b3d41df4e978989fe812
188 F20110216_AABUKI taskin_z_Page_143.txt
19b8c446fe5991574776d80bd1dc31da
8eb246cd3bdcd7668fd4ede7169a269c4c7f096d
37205 F20110216_AABVOA taskin_z_Page_055.QC.jpg
447c43d7b40a8548e47840a468c5cd0d
b629ef1bf8a5196434ee4ad247f6599b64e66e1a
1325 F20110216_AABVNL taskin_z_Page_018thm.jpg
f660e232e81d070dd65e9e74f4f72233
31b7db4badc34de21d9b1de12e366c2330c59356
24175 F20110216_AABVMX taskin_z_Page_065.QC.jpg
47a8284a6eaeb458c326ed314a7a2331
194e33a13cdb036a14d5690271fc0a52da65a0c5
96860 F20110216_AABUJV taskin_z_Page_116.jpg
a926027576422e3c6f95b4e101e9b66f
0a30264aa4d9df21a44b6e8fc86a17db22c41ff7
8296 F20110216_AABUKJ taskin_z_Page_017thm.jpg
996ffa2513384b5708c67f8183842d37
f40282d0a47a73c1245d518ca0e639bd0e88230b
36735 F20110216_AABVOB taskin_z_Page_130.QC.jpg
a921d47cf3a3a582525a4f031294e9d0
5373e9985515b3a108a6c73586d4efad9b511ec9
8268 F20110216_AABVNM taskin_z_Page_055thm.jpg
88d9dc590971aa21cf1a00abb4bdba1e
6c0bfd7ea0e8979c5bf8d0fe0e57517a18cc4c41
6633 F20110216_AABVMY taskin_z_Page_067thm.jpg
eb41e7ce7f87d6d6d0de26e3786950e4
957fa374324065d1ab5c7f20d7a442ebbab720b4
F20110216_AABUJW taskin_z_Page_045.tif
4a1a2d80ba8a02caa89d82efafc489b8
5c21fd830d7128217b3451e1d7cdacc0b4da62a3
55017 F20110216_AABUKK taskin_z_Page_080.pro
83c0fd67188f85de31bd7a087c5cd474
2fced235d46ba2529e872f1a49c6227d404c7851
8158 F20110216_AABVOC taskin_z_Page_046thm.jpg
d217a42600e89d2a63289cdf7f941106
123b6be5b031618dfa8a9e1207fe1cc664460dd3
38406 F20110216_AABVNN taskin_z_Page_107.QC.jpg
659b93c40f618a6548d17ef2218c42ed
d645adb7b9768b9761ce954afe5425d7a1ffa14e
31021 F20110216_AABVMZ taskin_z_Page_029.QC.jpg
64904bbeb20558b424b4a96b436f966e
7d894f49010fc062f8e0d4e0a8d954aa15e4ca5f
F20110216_AABUJX taskin_z_Page_007.tif
59655273a5e7c040b486b8cb609d5675
5cefdfd9445ddcaf7615ab2e71759fd04817a982
F20110216_AABULA taskin_z_Page_109.jp2
86a6e6c6b84b707b04b5458ba64c25aa
64aa20b75ce9e994ee8a47653995f4f9edf2b393
76046 F20110216_AABUKL taskin_z_Page_110.jpg
a2daa947928e83064d3893670e615385
62ad6726d3ecebfb4ca031d85ef8dbdf7d769e28
34251 F20110216_AABVOD taskin_z_Page_150.QC.jpg
1b2fd79673701cc012319c7be82d3803
c0972d42ea9e4da443e75cf4f3b7964a613ec35c
6446 F20110216_AABVNO taskin_z_Page_103thm.jpg
d7a450d76d0b11ff90dd5cae7268c2a3
dc7caec3759484ed44fb8189694a3f8a6be303bf
38558 F20110216_AABUJY taskin_z_Page_098.pro
01fe2f56423dee7e038fb602e29509b5
d9d329605cf479c733b2f814676c40411a096678
F20110216_AABULB taskin_z_Page_082.tif
8020a277cc2aa81a8b3f98ab0465fc05
c223db228c18fcf3becd62d1fa48b18182c1616e
7319 F20110216_AABUKM taskin_z_Page_088thm.jpg
949fb6a9024f45bc23ee11e61f34d804
1191d4d4b6fd0bcd56015032b59de7ed09ab20d2
7610 F20110216_AABVOE taskin_z_Page_092thm.jpg
7cb0a75597da411b742c9783c2216457
087502a0d8543ddfcf319ad4bf9679bb24676d47
7899 F20110216_AABVNP taskin_z_Page_134thm.jpg
488e5c789fc952627043affb653ff4d9
cb53e4d5234af378c098f09a6dbc879e020bcb69
F20110216_AABUJZ taskin_z_Page_069thm.jpg
65e3c90376047e58355a2a0494041eea
387e07743d68d0ce7fddf16e0dfed27c9013ffb3
106623 F20110216_AABULC taskin_z_Page_039.jp2
b44ad7aff2103637cd8a9d2c14a42c1d
8e990407bdd7331551fd684c155be6618647e998
2177 F20110216_AABUKN taskin_z_Page_115.txt
35c7c6eba9925abc7587eb612c18f64b
63bdcd8638bcc775a115a654009774ddf59427b6
30287 F20110216_AABVOF taskin_z_Page_144.QC.jpg
ad9383ebb3b0621f228ab7ceefd88a16
3c854c70e3e24082db325242375b3bdf0a87121e
7958 F20110216_AABVNQ taskin_z_Page_136thm.jpg
c6ea2b3d2e0e077019440649f121566d
4cc03cb7776dad7beccd0627d01ca8393b7b3e53
1683 F20110216_AABULD taskin_z_Page_004.txt
7fc32e828015bc8a755cad24b6ea0554
559fb829fc861433dcf76349ebc1244f494dd365
1667 F20110216_AABUKO taskin_z_Page_076.txt
a580105db4cbc39d2c16dc46fc92ce8f
0871d22fb301f7e197f7143c7c772d6ee96d2dbd
7675 F20110216_AABVNR taskin_z_Page_100thm.jpg
0b8d0c611407a08113776329e968e92e
3932af4e2a7fd7c70d0e8f4f3e407464d25f2563
23892 F20110216_AABUKP taskin_z_Page_007.pro
ae00bff85301efd5c71c669e5b5656e1
0d6146be2c9bcf521720d4d4b140140b584ae8c5
35424 F20110216_AABVOG taskin_z_Page_013.QC.jpg
0166109189e51eb072a46bb4f67289b6
b282c5b66c4b0369aa76be98b85bb3db89d8ecb5
34262 F20110216_AABVNS taskin_z_Page_122.QC.jpg
50d9ba6fa029f328dcd41cdd97f4538f
b0a25ab02d7535c919a786821771b30257f7028c
934900 F20110216_AABULE taskin_z_Page_095.jp2
8867c6d55222ea77a422604cbae53914
a5c9473b9aa610ebd11d97b6f44b83acad6dc475
115441 F20110216_AABUKQ taskin_z_Page_072.jpg
99db5a2a9c97d4ef1dc9f3ed2b2cf31c
8bd5962e18ef4654cc39b23ac93967f0272b4fe9
25484 F20110216_AABVOH taskin_z_Page_023.QC.jpg
9703258782a188fb1523dd690f5412e6
8b59302240247dc58305d20f833e0efb7f3958c0
21930 F20110216_AABVNT taskin_z_Page_009.QC.jpg
a1361a33e16aaf7752825a0db42800bd
e641d64475e7b0133534630b44c13946c4c82069
88476 F20110216_AABULF taskin_z_Page_089.jpg
4b7028cdefd70d4de263f717396ee946
9c63b1db949e8f26f825ffbfcad948c74feba267
2333 F20110216_AABUKR taskin_z_Page_045.txt
f0b469b79665cd63976bae5f6bed77a7
466fbad6241ad8b8b83313a4b9b92bb8c656eacd
32342 F20110216_AABVOI taskin_z_Page_123.QC.jpg
979c734a5636ab6829f6075a673b9588
350942aeae2985dc65fad651b52b58f25b18a903
6347 F20110216_AABVNU taskin_z_Page_065thm.jpg
c8c1f531467174dee876fadd3ed05f2c
ce206174f3e6a08f7071ace1f2e8249110d0fe3e
108423 F20110216_AABULG taskin_z_Page_115.jpg
dab2ca0776634084c2f285d5fc4a3f70
76209880cb4d29c4fe027fbe358bde452ea4d5e4
F20110216_AABUKS taskin_z_Page_052.jp2
5f177c38009c147dd1ddfd3751fddc82
e1462571bcd7329c9e1b2ccd4edc84dfd41956f3
8286 F20110216_AABVOJ taskin_z_Page_045thm.jpg
d20228f0176f7d3afc34707943d518f9
8ebed921718275a9a0f5f0bdab7696560281acf5
5998 F20110216_AABVNV taskin_z_Page_081thm.jpg
e3cfcbe2781dc8c24a3854f1b4c24b9a
9d5c697968cfff8e86cd50fb68216dee0ca3e419
46112 F20110216_AABULH taskin_z_Page_089.pro
d9815f147fb1ebe50a718785dee1327d
9e01677bb7c3f55a700ef3544cb79bc28e45cf76
798 F20110216_AABUKT taskin_z_Page_003thm.jpg
e31e9d20c7db0c47db245e5ac5ac1199
ac44c9b83ed6bb1a34ef87335c5d93882faca695
23250 F20110216_AABVOK taskin_z_Page_083.QC.jpg
62ceccbda90f456f5620e7c0efd25d5c
504fc51ca235b2c00aed3eddb0c47666ce5b2436
8317 F20110216_AABVNW taskin_z_Page_151thm.jpg
a50864290bd49431ec5d15eb057203a6
7585dcad79a67efedb9bb032f21fac0eed1899a6
F20110216_AABULI taskin_z_Page_052.tif
1635ab4b6d86a7e2617ffcc8f6b4ac8f
08cc920cf604108b673035da2825b87c2a89b6fe
3937 F20110216_AABUKU taskin_z_Page_002.jpg
198ae2d736a4f72a4dab92360db65956
d2dce377abf2f837b5ec115e98becc050db009cd
8415 F20110216_AABVPA taskin_z_Page_135thm.jpg
a04b4b209e400da51afd3422dda3c8ef
1d74ee36151afde1f2ce0908d1ca84e724df876c
7825 F20110216_AABVOL taskin_z_Page_080thm.jpg
ba2787ca4f44d58a259a2bc17537a96c
a552c0342791229282f2670527e459df444e3550
34231 F20110216_AABVNX taskin_z_Page_152.QC.jpg
e53d20a6fe42d584f67ebfe6f6f2d744
aa0f14dca2dea8bbb55e13cea2dbd78be6747bb7
F20110216_AABULJ taskin_z_Page_133.tif
1aa4d0bad3a725ebe9085bbca1904d78
929eb5ba8ea00921185a95aaff279db77b996724
33792 F20110216_AABUKV taskin_z_Page_080.QC.jpg
812614c143cf7a49344c92627d916a41
3aa277e714a952bf075352c0781503d1fbab040d
7968 F20110216_AABVPB taskin_z_Page_157.QC.jpg
c163e133b5a8d4a81599f525fbe5ccdb
244ad98777546844f67547507996a3242d057599
8423 F20110216_AABVOM taskin_z_Page_063thm.jpg
28116c936d0b0e99d45b17a05df5ac49
e18434a84d7e31ddda1fd4f7b294f3de4adef707
34476 F20110216_AABVNY taskin_z_Page_031.QC.jpg
49a25e3114fa300d7cd9c098ecef4839
97098c9d3cd85bad284a08c2d30838257d212ae8
1051885 F20110216_AABULK taskin_z_Page_080.jp2
c1dfe272acf9b376040e7aa579b9bfbd
078b5a33a3c3904311a1d923b609d060ec2bf8a6
31149 F20110216_AABUKW taskin_z_Page_027.pro
dac7aa1f07a045e6ed2c737187783520
c7286e50518b8942a508c9dc801a7e812eece14d
32468 F20110216_AABVPC taskin_z_Page_061.QC.jpg
e0c3a056d332c6d7d2a9ee3ce0ea3889
cd9afdc6caea68236c618b0aafb1243fe6c1836c
27200 F20110216_AABVON taskin_z_Page_030.QC.jpg
b9fbca29cd31bd5d238eb75df2d6eb36
e4321841150f6b8916be83bea3fae0973393bbba
F20110216_AABVNZ taskin_z_Page_011thm.jpg
18a0f2c4bfb6f137bc7f41f6263b7c46
75c801053ba643358171298cdf59698d863d0c06
36095 F20110216_AABUMA taskin_z_Page_072.QC.jpg
a84b47dc35783778d4cbbe75186e042f
2ff71d66ab7e4b539f596dec28412ef0f43c9429
F20110216_AABULL taskin_z_Page_108.jp2
7cc13712ba0677b5eeeb7bc92a5eeb1d
523e126a21c18b41afb4108921ccb8096301b680
122400 F20110216_AABUKX taskin_z_Page_008.jpg
abf9b42dc8de0af25ed121e6942f0c60
e145b7905eea710b80a009094142bb32cc2c20ec
7158 F20110216_AABVPD taskin_z_Page_036thm.jpg
28bdcfbf01b3f5ac2c7e41727454c79c
6d99f6fb24298b8a0eca61801691447665bbe4f4
36603 F20110216_AABVOO taskin_z_Page_017.QC.jpg
7faad81fd30aca53ba0fd8b31c7f199a
68b4a1b08459ef6e942aaaa2b27551bf4b2d8caa
70797 F20110216_AABUMB taskin_z_Page_050.pro
f785a6fcd56385bac81d0bcad63aec69
6f2e6a196efa506823551ac4e9213e454d710429
F20110216_AABULM taskin_z_Page_067.tif
170b10f83196ef897ebde6cda7af5317
004f70cca9cd4b7b4ca502f1fb8e82ea5f61afa6
6114 F20110216_AABUKY taskin_z_Page_076thm.jpg
f55651ec0e32cf13aec8c47341b1b816
6a78cf960dd89aa743425341642f6c3cb1c039ef
8178 F20110216_AABVPE taskin_z_Page_127thm.jpg
530437089eebc15e8541d9eedbb0ec4c
e17321434bc1440ead233413d12076823d57be0f
7929 F20110216_AABVOP taskin_z_Page_117thm.jpg
4c67af369168060c53bf72f2156a21bf
59fc968e6a3c65082414b898d3d339d5637f6609
7855 F20110216_AABUMC taskin_z_Page_051thm.jpg
88952002b0c7fdddc55a8ea6d7606bc6
5430dfbc03d2b12622236a3f0abd103fa84c3311
29990 F20110216_AABULN taskin_z_Page_093.QC.jpg
592c3d76d8d4d803a678c18d76b734a6
1030f5f1c5a1f462796e0df51bf9c4866a6e619f
82938 F20110216_AABUKZ taskin_z_Page_058.jp2
6f65f6bc2eb1d931977161ea0c4f1916
03ab061c5322356927e5a953e1f5addb3884866c
26252 F20110216_AABVPF taskin_z_Page_057.QC.jpg
07025b6672128315cf4154e77f08d73c
9895cc62cffb5cd73f306a0efe9aec4b73ead36f
7156 F20110216_AABVOQ taskin_z_Page_095thm.jpg
62d69f3d41ba0f24e4abdcd37dade0bd
06ce1790926cad22a0b99338c8781aef7ae381be
F20110216_AABUMD taskin_z_Page_052.txt
232eea5b71dea82340783699a6fa6d13
08630819f6514aa3306c729199885ac8f6670b2b
2513 F20110216_AABULO taskin_z_Page_155.txt
e57585272dc9cb6f9d7b58ebbd6abce2
96e2229051646e0a30536fb418ad8a322bc48ec3
26625 F20110216_AABVPG taskin_z_Page_081.QC.jpg
b405526eb5c57f3bd92a0d596e626747
d4dc15e3ebcbc29a81198cf7810ab69d61029cf9
35255 F20110216_AABVOR taskin_z_Page_138.QC.jpg
7437a51872814ba45306ef890bc83085
4a52a32c4e9613e80b7afe024d6a92ac88848ee7
34939 F20110216_AABUME taskin_z_Page_154.QC.jpg
c006763757c38bbc970c983ed4dc15b2
f1018c0de3836af4b10e41e9d80446b91ccbbc0d
35466 F20110216_AABULP taskin_z_Page_153.QC.jpg
5f764f8f05a9693e75642818577a0621
c1addbc9d5419da109cdf42edaca3f8ef12d0a58
25636 F20110216_AABVOS taskin_z_Page_103.QC.jpg
121a123f3811e3dd9230fe2b03a7985f
67f768617ed7649c9ca1c6e2c7c5f0e3cd9bb711
56246 F20110216_AABULQ taskin_z_Page_042.pro
b72c73c4bc49d90791f5bfbc6810cb52
18a9030326dec232dfab7ddc01093d8596374ec0
28099 F20110216_AABVPH taskin_z_Page_067.QC.jpg
346d968f4cf4a211760e5b166b45c03d
6f0fdda49df68b4362ea4eeeafd1551a319ee9e3
8395 F20110216_AABVOT taskin_z_Page_111thm.jpg
15bd4885b611aa13f75512ecee1bace8
5f22c30bf1ed43a411ec2be43828f04bff5b106c
F20110216_AABUMF taskin_z_Page_153.tif
752fd85dae97ab5020062827d2361d80
1369c8f16dfd0f425288b91c2fc14e1166a4c6ad
126234 F20110216_AABULR taskin_z_Page_111.jpg
51ed969f73f8d5e07fb3caf284863058
4d6c46e185cc50f214c652bec69449845afecd45
32668 F20110216_AABVPI taskin_z_Page_100.QC.jpg
270a9c9bc03f729b627ea512cd0dc4be
08f976ac673962bc71c022f60fd3ff786d1614cd
2598 F20110216_AABVOU taskin_z_Page_112.QC.jpg
a2b2008a7054b9ae92ec09a66948c22c
4d42256152867d478a2648775dc4d7cbf9481cfb
41152 F20110216_AABUMG taskin_z_Page_023.pro
a6ac967df96e17ed155bdb3e056a511f
4de9a5a6c1e62325125a97488d1bc69698807786
55984 F20110216_AABULS taskin_z_Page_012.jpg
c8df2f645262406be903a9e529d7405d
fd4499db707f0fe6e7696a47f10a844de04b4f53
8534 F20110216_AABVPJ taskin_z_Page_085thm.jpg
fb48d8a2f4b8fd1f0d06cce27f157284
36abe4144692fd14dc2fab6ab7934b462e2ab822
34001 F20110216_AABVOV taskin_z_Page_038.QC.jpg
667e9c0a37807e332660e59d7ba3c7c5
5bbdcae1409aefd78f343102852e3953efce8cf0
8284 F20110216_AABUMH taskin_z_Page_071thm.jpg
2218ffd123db05f54f4179a7f165147b
fcb176c37136a0e9315afdd4c2f838e2bb073ad9
F20110216_AABULT taskin_z_Page_110.tif
58aacb5e144932a4970e93a8c5cca48f
a4d15bc5eef6abcdb1c1a8dd78d68167b3ced41b
8084 F20110216_AABVPK taskin_z_Page_152thm.jpg
000c84d7cc1aa5e088879310307b2973
734ebd059f179a5b4f4c35592bf02690bcf5dd33
36568 F20110216_AABVOW taskin_z_Page_137.QC.jpg
e13d7616378250b06c277e9085c15c0e
8b4bd2bc353bcd7a1ae0e3c589a29477b9191d1e
2204 F20110216_AABUMI taskin_z_Page_117.txt
3ddeba312f2a0b91351c89ceaaf3294e
95274907ca8653f542638b93566ef90b4ef1e1c1
2230 F20110216_AABULU taskin_z_Page_038.txt
360de181fc5bc838e13b10ddc7d26f87
05a2383845638c70461c06609d2ec438b3048ded
34689 F20110216_AABVQA taskin_z_Page_134.QC.jpg
9bb0d50e163125f19d094d2a773f4c12
8ed070fa24d85894f3074ff0b0a69a48ec252c9d
7977 F20110216_AABVPL taskin_z_Page_056thm.jpg
0e9c573f8d0dbcf79de8f82b4f0d6fe0
0bf20efb82f3b0626300cfc636e36567c7da2571
7480 F20110216_AABVOX taskin_z_Page_121thm.jpg
96f989291a742a8fc9f2d44b059aa85b
2e2d663a8cb94b9866c34405a0636c698210a618
2028 F20110216_AABUMJ taskin_z_Page_003.pro
dd88d3308083a428f24b6ab13d37e46f
6c946078b070c431da18cb1862754f7c7ab6e37f
F20110216_AABULV taskin_z_Page_139.tif
2162530313836a1184fd0f29a458503a
e2ad88e4349b4b4c923c5e7e5c4e8bcfca47a213
6322 F20110216_AABVQB taskin_z_Page_023thm.jpg
e66edbae5d899816ffb1ede2bbfe4481
385ab17a29025963febb9527a911cd904993dee1
7335 F20110216_AABVPM taskin_z_Page_097thm.jpg
b377a95b2e2c67d1017a77f99b3bb333
8df6728a3801e552d62d42c0ef5bf82a8fe458ae
6961 F20110216_AABVOY taskin_z_Page_037thm.jpg
80834dbfd85b587c869ac5ac06d383be
e7641ebd884846a9477619f9d874c88254337107
33495 F20110216_AABUMK taskin_z_Page_005.QC.jpg
984dec0a3e13ab60abde4b75979299b4
32370c73cf85dc7b0fac824aebe591b3d873b562
F20110216_AABULW taskin_z_Page_010.txt
d69b57dccb1c9fbde9cc55a7f1d462a4
6e2fc035bd610298075458b15f67bf3c89fe70b9
6753 F20110216_AABVQC taskin_z_Page_035thm.jpg
bf94017f54b6497ea1eb17f158b5d35b
a72f3836712ebf009c769ecde2017f55f92d6cf5
F20110216_AABVPN taskin_z_Page_094thm.jpg
67fc1d5650b7358c619bebe774332989
98cef30658aac34640e89a9e442fb4c40d2c1386
5433 F20110216_AABVOZ taskin_z_Page_148thm.jpg
51a0bc257cef5b1318275b98a790762f
469a9d2de72c0eb2ee504eeba12dd4c3cc5a0f65
2194 F20110216_AABUML taskin_z_Page_141.txt
2dd44e58ff59b3104e2fc52e64c3d277
4fc3492bb836a2f1a7de94ab8515908a62abac77
53782 F20110216_AABULX taskin_z_Page_074.pro
b910190214ce7d744a48f809e1931b6d
16c5721125a4c9f60975d1c9bc33964c5a2612b1
106985 F20110216_AABUNA taskin_z_Page_038.jpg
b5b64f4b7058920720f1f5ef0f4bfd6b
677ec8f33b3d4e92c8c826e0cfcbd05cea8725dc
4966 F20110216_AABVQD taskin_z_Page_009thm.jpg
da76fd6fce9311aa140900f9bb65df76
4d70b55f1a0588e666c5a75ad2db35ccbfe8ab06
8482 F20110216_AABVPO taskin_z_Page_154thm.jpg
ca463511aaac84d34175e9dd8996ad5d
e2c416a548431523695f0bbff15e175be018fce1
F20110216_AABUMM taskin_z_Page_030.tif
9347cf156c427cd7aad0893c7f01a2cb
2fdd7d10da4badad3672fe36d4043e984a9be7f3
124962 F20110216_AABULY taskin_z_Page_108.jpg
c325619a9e9b415ce5a5aed204f20f56
808a86b6b8e0c9fe0f5ab2c9d51789f2888d9737
F20110216_AABUNB taskin_z_Page_019.jp2
de4ac5aa90f1a0ab81c2c1e3fd552d38
6eff95a9628a31950c8f07a453d0cda4819eae3c
36819 F20110216_AABVQE taskin_z_Page_015.QC.jpg
8faf6abfd4259406f6270e11da97578e
646e76169556cc717ce662f80ed23bba4088d2fa
F20110216_AABVPP taskin_z_Page_116thm.jpg
617993095063c3234b2dec04c3d538ed
f162e90e023ef0d5c5530ba79a447f1353d14181
F20110216_AABUMN taskin_z_Page_102thm.jpg
425eb7eeef33eef6288785a37a22c4a8
b102ae65987567d7875d38551aa4f2563c4ba8fd
120033 F20110216_AABULZ taskin_z_Page_021.jpg
724bfe04ce491f9ad7055ef6619ace4e
560bcc548a6740df6474df841486ec529712a94f
1866 F20110216_AABUNC taskin_z_Page_096.txt
f7897829ef104babd6fba9a41fc9f1af
7659d85e78bc5aa47b90a108d1b697bd236aa092
5014 F20110216_AABVQF taskin_z_Page_083thm.jpg
74535c9a012d4ae52318d44f97fff85f
7f3f5132529ebda9c01de88f26984d83cc490e2d
39086 F20110216_AABVPQ taskin_z_Page_114.QC.jpg
45c55fcd9d50f484789f4d73d18b1b6b
23f9e1e6860f01e79e4e87274dd66d41a2d98967
F20110216_AABUMO taskin_z_Page_047.jp2
10c9b2dd7599319911bb634998d13f3e
cc2e2535e3a8331badd6c8c905448d2b1a251c47
789386 F20110216_AABUND taskin_z_Page_078.jp2
169aac3e9f91d87288e719b06c61f25d
dbd2e4522d3d636c653948c1bee0c0f5a3240b83
37179 F20110216_AABVQG taskin_z_Page_052.QC.jpg
87c851418477a01f79c06711bec2cfd7
76c4cad31302f818f8e3fe07beb8169157012007
7261 F20110216_AABVPR taskin_z_Page_039thm.jpg
3da37fe49cf2db6aedb15e1a30145e1f
45e762b2e223c6b9539cc37c3aa2709367ddde25
F20110216_AABUMP taskin_z_Page_053.txt
c1a1d24467dbeb1fd13ca345f6fd366f
86eaac4fb9480603acfd7fbaff8bf9dc702c09c8
F20110216_AABUNE taskin_z_Page_048.jp2
3db70310d00f2bceaacd76c5fe9ff5b6
b9397275e258302f23c130ccd203d829cb24c8ba
26806 F20110216_AABVQH taskin_z_Page_104.QC.jpg
06c84ba39964ea8ce9cf47a621f460bd
d7ee8b72f9bd344c709c0f439ab1f990eb4e6a67
7724 F20110216_AABVPS taskin_z_Page_034thm.jpg
b9feb7a99573e89289a7bb86830f8fa2
7522ff7b0509ef8028f8d91cd1e8d55800ac1074
F20110216_AABUMQ taskin_z_Page_086.tif
eb36679cad33a914d7c57e0c0f9409b7
7e148d87dc18c819d446454ebde340f50c53afa6
48588 F20110216_AABUNF taskin_z_Page_116.pro
88711ec9e32496e0bbbd54bbf0da1db4
6a50999731ab5bcd8587be63ff618c6eaf0fe32c
35222 F20110216_AABVPT taskin_z_Page_136.QC.jpg
68b1d9f116b6f28c12ac5e32748ad81d
3a76cb09fe87031854414c13c50a7cd49938e913
8265 F20110216_AABUMR taskin_z_Page_086thm.jpg
1f3e73e97cb9e73c009aa810415dfcca
c0cfc62a7675a35bfec802d9d317370a2abf5e45
7263 F20110216_AABVQI taskin_z_Page_093thm.jpg
6b5789bf7ff10e9e80d077d5b9f4ced7
9979c2a7dadfb1d66fb8b7d8c7694291e2847b4e
8244 F20110216_AABVPU taskin_z_Page_084thm.jpg
3582905ed30bfacc37c1f6c2b381dd31
b702a4ac87ba981785addea1a447c82e5b879537
49374 F20110216_AABUMS taskin_z_Page_029.pro
157f850410dff98f335547f7117fea59
6c6eb380d5361b801389f1335ac3efd55cdbe672
F20110216_AABUNG taskin_z_Page_023.tif
8a77717ff9da52ca9acb7649240c6d9e
cdc77f5495cd0c92a6aeb31af0b25ed1c5201edb
7949 F20110216_AABVQJ taskin_z_Page_008thm.jpg
8ce7dd255cea301492bca4ee16ba6bc6
b83e417781c9cffcdfb2b3c080259b4919039612
6950 F20110216_AABVPV taskin_z_Page_104thm.jpg
4045cfea6dd01186dae5a65bc0b74398
4d0f94f2a59a26a291da3582fc2c515f33aa1b1d
F20110216_AABUMT taskin_z_Page_138.jp2
b279529132ecfb29c9dbb06bbf173204
1af1319b9c40872ff7e097bfbdfb51fdf5f160b2
32349 F20110216_AABUNH taskin_z_Page_051.QC.jpg
71aa8ddc9f586e0064fea44562221d1e
98c27b1547962caca4c3f2882f02f9ac29214509
7742 F20110216_AABVQK taskin_z_Page_123thm.jpg
9e42446bd46b27f0cfe8afef915a2cb0
13d50e12ce853b28b016f73fb12e7e73cfb74ffb
22822 F20110216_AABVPW taskin_z_Page_118.QC.jpg
66c1c39f5873129bb0d89879a6065838
2f1e46a3cdf59dc26ec6d489a0ccf83a58cf4e1d
F20110216_AABUNI taskin_z_Page_137.tif
78fad86d45c86e33a97cd67bb15000b9
44bd8ccf7eacf8262882e45c774bf179afc50e61
F20110216_AABUMU taskin_z_Page_001.tif
888b80193f26a385447a3ac17a731fc5
d3210daace24844b0040849806b1ceea64aed34f
1406 F20110216_AABVRA taskin_z_Page_002.QC.jpg
bc3218000cd794507e2865d78bc81448
aefcea666d5aca0a14b2bde283d54e73f23d4830
31201 F20110216_AABVQL taskin_z_Page_033.QC.jpg
a34c83f8931d610f9ce86636805683a2
98f8ff5bdbe1f422086a5b5d56f0e7cbd91dadf0
33870 F20110216_AABVPX taskin_z_Page_074.QC.jpg
f328d7cd6313d21d77f7e141608415d5
71f8eb5a648b4c5784b9ba44014c5523d1a3646e
60067 F20110216_AABUNJ taskin_z_Page_021.pro
d20268c9bdc9e36e124019595620c7d4
19f95ac397376dafec8c795f85294c6ef495e4c1
7506 F20110216_AABUMV taskin_z_Page_099thm.jpg
c22b5b2e922d02bd1ba88f9d456f2f72
7d6a8af1fd4f502d5ed5b6c936b5782785290a7c
37847 F20110216_AABVRB taskin_z_Page_011.QC.jpg
7ef09a6ff000985841ba01cd7899eff2
95af54303cc8a40fa2f05335c858e587957559b5
30610 F20110216_AABVQM taskin_z_Page_050.QC.jpg
b4470b843e29a0e30930fe238e1d0d52
a9503c7399c10ea625ebbffe9eb81b842234827b
3669 F20110216_AABVPY taskin_z_Page_143.QC.jpg
ce862323126746b118126dcdaab09987
fe6773d7ba19a88b537bb15793170996cd84c3b7
60827 F20110216_AABUNK taskin_z_Page_135.pro
4781958db94d26e251daa0e25aafc24a
95c40201122f381994f254ade2fc4874fd938962
2398 F20110216_AABUMW taskin_z_Page_114.txt
8f85027385c8ae78abf510ecd8e0b9e3
45944a26484d4883848a1fa677342748db5e07ff
255063 F20110216_AABVRC UFE0024784_00001.xml
61ebd367d2989fba7822936de5662965
19f6c23aeca2af30686ddfd6ccdb9842948001ca
8897 F20110216_AABVQN taskin_z_Page_114thm.jpg
41541567ce51ddf4268cd8f96dc5f5d3
e3f003714cbbbc56b0c0dc15ec78c2787145e43c
33353 F20110216_AABVPZ taskin_z_Page_019.QC.jpg
5bbb9be499b35e781f44dda284f9b7e9
aa61631d1b2c1821ad8e9d28c865958be44dc449
7903 F20110216_AABUOA taskin_z_Page_128thm.jpg
e33867d09a3b9a9bbe43479e6896ed56
af5392173c6a02aafc5277cff5907812b75a2e04
F20110216_AABUNL taskin_z_Page_036.tif
64fbd6ffaf6dace7c54da97beb4f61f4
ecbb410d254aec6cd15345fb1b894104dcb49fe6
F20110216_AABUMX taskin_z_Page_090.jp2
b0f0f0ceba24c1e40c544fb8709e21c8
3ffcbca2e666c036e3693144231fa005849bc401
25833 F20110216_AABVRD taskin_z_Page_004.QC.jpg
ba03ff2d14c89c3a5f605eb7481093ef
44e29e14f283b2c6c735296d672c66b88d422a88
8358 F20110216_AABVQO taskin_z_Page_064thm.jpg
c2066e9e119492413bae18ecbfcc1e04
8828c5e97984d40c9636e340d09c9060ff106a79
79839 F20110216_AABUOB taskin_z_Page_101.jpg
5787cce64f5a31f617196b6b6298963f
d5105ffa677cc939863aad2f75b1f6490ba7bf6a
F20110216_AABUNM taskin_z_Page_091.tif
4b0bfb278c557f4d86bc5b9c340fd5bc
09df7f2214d3c7c6df39d373bea2fbc6bd52a26e
F20110216_AABUMY taskin_z_Page_118.tif
7bb69c5504f30d574a8591944c807434
05549e0e908c9d0bb71ad5b5d8cc1112bce40eb0
16065 F20110216_AABVRE taskin_z_Page_007.QC.jpg
e0b49a44fbef50801e4a50b734cba778
7dd8a2adbf47245a4fc31414024b2d9e5671c3d0
33356 F20110216_AABVQP taskin_z_Page_128.QC.jpg
2719fabc3a2e4709840a31fa1748be7b
2972ee8af8dd4619f6655571b9765ff5d317d760
F20110216_AABUOC taskin_z_Page_130.tif
b2698ff8a12bddd52020d10cd894eadc
8a9b57435c9a38236c84f4c9160cc39e4635fc87
6846 F20110216_AABUNN taskin_z_Page_059thm.jpg
3ff9342925602c5b4637df55282d0756
dee9a149a59ee8aa831786e8dddf7ada9f6e0021
62909 F20110216_AABUMZ taskin_z_Page_011.pro
fe8fd151a6d6394a0230e754cef1f80d
110815e84c5cde3f2b23f8b8bcb049db74d33bd5
17384 F20110216_AABVRF taskin_z_Page_012.QC.jpg
41b44e721e2f6bedac9f3c92269671e1
8f9ac20965feed0a71f538adebfed870c1dc2b80
7737 F20110216_AABVQQ taskin_z_Page_029thm.jpg
aaf895982e522c9cb6343142ac463845
37b7f80789126294aaa31c51c447d7c83a55090a
28999 F20110216_AABUOD taskin_z_Page_036.QC.jpg
3288173bc27bef7a8718d8cd7300472c
e4584212142a148889b54736b1687de252391e84
2210 F20110216_AABUNO taskin_z_Page_116.txt
ff2342778b2e41adecde03c3eb260ae5
502ea0cc34ee3cf7cfb63939d111ad14b17d7405
37452 F20110216_AABVRG taskin_z_Page_014.QC.jpg
a314c5795f4141b90b8c6cb86192edf2
13f6a2818cb5cb53eee9cc54ca24a01406bb44c4
6847 F20110216_AABVQR taskin_z_Page_131thm.jpg
9a2d24581d79c3a1774d3e95f01a5a4d
0b4a5a65085bf2ed3ea36f2a4969d1553bacec73
7987 F20110216_AABUOE taskin_z_Page_141thm.jpg
1c7df36d1289c65cf902d121820a1291
8080b1997b35c1934799a64ace9373617e4fab82
20403 F20110216_AABUNP taskin_z_Page_146.QC.jpg
399f75186cb4ac7112b1c9d22f3d7f03
53f5c9b1a8fac86f54feabe6d83c261fb9ecc137
37450 F20110216_AABVRH taskin_z_Page_016.QC.jpg
31aced7fa8d4aebfbada62a4626e793d
4c4910e644d28aa12b3b55144cf006976ee2e948
36157 F20110216_AABVQS taskin_z_Page_063.QC.jpg
88c0870897ac73ff7df6a17cc70f91bb
5ef272a030e776fad9e98b2e45526d1da9dc57cc
77731 F20110216_AABUOF taskin_z_Page_059.jpg
6b2d0e039f18d9c6c3d08fefb8ac056d
cd1d180f6b3bea3f99bfa8bd94a75080a759e2ea
F20110216_AABUNQ taskin_z_Page_005.jp2
5b97c2882dc9d3816b74daf54fce3250
08bd8bd6fa5939967eb8d335ca62964a5ec5d33b
37300 F20110216_AABVRI taskin_z_Page_021.QC.jpg
91f7f5085563c1c78519dea521e9bbd3
0957e137c8df3cbc300537e142b9b30a9f7212a7
35639 F20110216_AABVQT taskin_z_Page_066.QC.jpg
f998f7541b5a325f9d36468743b3bd10
fc0f9d2be492061d4bd99f09136ab3bbeeca8118
F20110216_AABUOG taskin_z_Page_080.tif
d9cb75634bafabd78d1c1e45dd897a14
4f053e495e66a838f16758609736503cb014db72
2288 F20110216_AABUNR taskin_z_Page_092.txt
765246afab2f314aae8316119477ba62
c084df13dd45f4ebabf615abcb0d1e3e7cf59d37
6490 F20110216_AABVQU taskin_z_Page_118thm.jpg
bb5fdee9569c5be5c333fee959f2e43d
d38acf11a52ba7988ee39eb69df7b72566bc2782
2201 F20110216_AABUNS taskin_z_Page_036.txt
3c0baa36151829aafe46252fea61f7b5
3ce48ca2edb7df53a8405e3777ca4210744401e3
34620 F20110216_AABVRJ taskin_z_Page_028.QC.jpg
c4554b9bea2be45f9245d4dfb6b95a56
f427a01b8551d47fb9383036cf1eee396b45048d
30433 F20110216_AABVQV taskin_z_Page_116.QC.jpg
2aa52bee42ecfb88cb6799e9b9a3ba1e
9e12971fbf40bd756392afbcfb7061f8dda643a0
6469 F20110216_AABUOH taskin_z_Page_096thm.jpg
c19d3c4ca6f3ee032368c24336c22e05
a391fb5afd14aa279d3799b710c9abcc5196a994
F20110216_AABUNT taskin_z_Page_058.tif
8bd487ce9ba13332ba9503b6e7eb7842
8f9a01297087410d761dcbb07e7d890c36de97dd
30876 F20110216_AABVRK taskin_z_Page_039.QC.jpg
39de89eaf70709f24367b00dff196ecb
7923a08da58613eeaaa81d1bff6860568884f487
35956 F20110216_AABVQW taskin_z_Page_113.QC.jpg
ce0317735b5f6adabf8fadf28e41bdad
bcbeef79a22b7d9d007a3a47e1861fd42c3309e5
117 F20110216_AABUOI taskin_z_Page_112.txt
a0054c887a6f8fc8881ac3651b1edb29
5dc69d620c00fc6145a0f142290b2e660474ef14
47067 F20110216_AABUNU taskin_z_Page_036.pro
8fe03aa4758596f744a511692322c9ea
6c6275aa14db13552fed081b87ea1638e6db11e6







ALGORITHMS FOR SOLVING MULTI-LEVEL OPTIMIZATION PROBLEMS WITH
DISCRETE VARIABLES AT MULTIPLE LEVELS



















By

Z. CANER TASKIN


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2009



































2009 Z. Caner Taskm


































To my wife, my parents and my brother; I owe them everything I have









ACKNOWLEDGMENTS

I would like to express my deepest gratitude to Cole Smith for his wise, enlightening

ideas, endless motivation, and patient counseling during the writing of this dissertation.

He has been a great teacher, a mentor and a friend to me in the last four years, and

working with him has been a privilege.

I would like to thank Edwin Romeijn for introducing me to the exciting field of

optimization in health care, and his invaluable contributions to this study. His guidance

and support has been very helpful throughout my graduate education. I would also like

to acknowledge Panos Pardalos and Douglas Dankel for taking part in my dissertation

committee, and their valuable sI-- -i..ii-

My sincere thanks are due to my friends in graduate school. In particular, Chase

has not only been a great colleague, but also a close friend for these four years. I cannot

begin to count the things that I learned from him about the culture, the 11i; -I -le and the

language. My experience in America would not have been nearly as enjo'--,1-'1 without him

and his wife Candace.

I am indebted to my parents and my brother for guiding and supporting me in all

life choices I have made. Finally, I am deeply grateful to my lovely wife, Semra, for her

constant encouragement, support, and understanding. She is the source of my happiness,

and the secret of my success. This dissertation represents the end of my life as a graduate

student. It also represents the beginning of a new stage in my life, every moment of which

I am looking forward to sharing with her.









TABLE OF CONTENTS


page

ACKNOW LEDGMENTS ................................. 4

LIST OF TABLES ....................... ............. 8

LIST OF FIGURES .................................... 9

ABSTRACT . . . . . . . . . . 10

CHAPTER

1 INTRODUCTION ................................ 13

2 STOCHASTIC EDGE-PARTITION PROBLEM ....... ......... 19

2.1 Introduction and Literature Survey ........ .............. 19
2.2 Formulation and Cutting Plane Approach ...... ............ 23
2.3 A Hybrid IP/CP Approach ........ ........... .... 33
2.3.1 First-Stage Problem ................... ..... 35
2.3.2 Second-Stage Problem ......................... 36
2.3.2.1 Foundations ........... ............... 37
2.3.2.2 Domain expansion .................. ..... 38
2.3.2.3 Constraint propagation ............. .. .. 39
2.3.2.4 Forward checking ................ .... .. 40
2.3.2.5 Node selection rule ............. .... .. 41
2.3.2.6 Distribution vector ordering rule . . ..... 41
2.3.3 Third-Stage Problem .................. ..... .. 42
2.3.4 Infeasibility Analysis .................. ..... .. 42
2.3.5 Enhancements for the First-Stage Problem ..... . . 43
2.3.5.1 Valid inequalities ... . . . 43
2.3.5.2 Heuristic for obtaining an initial feasible solution . 44
2.3.5.3 Processing integer solutions .... . . 45
2.4 Computational Results .................. ........ .. .. 45

3 CONSECUTIVE-ONES MATRIX DECOMPOSITION PROBLEM ...... 51

3.1 Introduction and Literature Survey .................. .... 51
3.2 Decomposition Algorithm .................. ........ .. 55
3.2.1 Decomposition Framework .................. .. 56
3.2.2 Master Problem Formulation and Solution Approach . ... 58
3.2.3 Subproblem Analysis and Solution Approach . . ... 61
3.2.4 Valid Inequalities for the Master Problem . . . 66
3.2.4.1 Beam-on-time and number of apertures inequalities . 66
3.2.4.2 Bixel subsequence inequalities . . ...... 68
3.2.5 Constructing a Feasible Matrix Decomposition . . ... 69
3.3 Computational Results and Comparisons ..... . ... 73









3.3.1 Problem Instances ............ . . .. 73
3.3.2 Implementation Issues . . ........ .... 73
3.3.3 Comparison with Langer et al. (2001) Model . . 74
3.3.4 Random Problem Instances .................. .. 75
3.3.5 Clinical Problem Instances ................ .... .. 80

4 RECTANGULAR MATRIX DECOMPOSITION PROBLEM . ... 84

4.1 Introduction and Literature Survey ................. .. 84
4.2 A Mixed-Integer Programming Approach .................. .. 86
4.2.1 Model Development .................. ........ .. 86
4.2.2 Valid Inequalities .................. ....... .. .. 90
4.2.2.1 Adli .. ii rectangles ................. . .. 90
4.2.2.2 Bounding box inequalities ................. .. 90
4.2.2.3 Aggregate intensity inequalities . . ...... 93
4.2.2.4 Special submatrices ................. . .. 94
4.2.2.5 Submatrix inequalities .............. .. .. 96
4.2.3 Partitioning Approach ................ ... .. 97
4.2.3.1 Separable components ............. .. .. 97
4.2.3.2 Independent regions ................ .... .. 98
4.2.3.3 Dependent regions ................ .. .. 100
4.2.3.4 Upper bound calculation ... . . ... 102
4.3 Extensions ................... . . .. 104
4.3.1 Minimize Total Treatment Time ............. .. 104
4.3.2 Optimization with Beam-on-Time Restrictions . . ... 105
4.4 Computational Results ............... ......... 106

5 GRAPH SEARCH PROBLEM .................. ........ .. 113

5.1 Introduction and Literature Review ................ . 113
5.2 Hide-and-Seek Problem ............... ......... .. 115
5.2.1 Mathematical Model ............... .... .. 115
5.2.2 Solution Approach. .................... ........ .. 116
5.2.2.1 Searcher's problem ................ .. .. 117
5.2.2.2 Branch-and-price algorithm ... . . 119
5.3 Pursuit Evasion Problem ............... ........ .. 119
5.3.1 Mathematical Model ............... .... .. 119
5.3.2 Solution Approach. .................... ........ 120
5.3.2.1 Searcher's problem ................ 120
5.3.2.2 Intruder's problem ................ 122
5.3.2.3 Branch-cut-price algorithm. .... . ... 122
5.4 Patrol Problem ............... ............ 123
5.4.1 Problem Description ............... .... .. 123
5.4.2 Mathematical Model ............... .... .. 123
5.4.2.1 Searcher's problem ................ 124
5.4.2.2 Intruder's problem ................ 125









5.4.2.3 Branch-cut-price algorithm. .... . ... 127
5.5 Branching Strategies ............... .......... 127
5.6 Computational Results .................. ........ .. 129

6 CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS . .... 134

6.1 Stochastic Edge-Partition Problem ................... 134
6.2 Matrix Decomposition Problem .................. .. ... 135
6.3 Graph Search Problem . . . ....... . 138
6.4 Master Problem Reformulation in Bi-Level Cutting Plane Optimization
Algorithms .................. ............. .. 139

APPENDIX

A AN IP MODEL FOR C1-MATRIX DECOMPOSITION PROBLEM ...... 144

B COMPLEXITY OF C1-PARTITION .................. ...... 147

REFERENCES .................. ................ .. .. 149

BIOGRAPHICAL SKETCH .................. ............. 158










LIST OF TABLES


Table

2-1 Descriptions of the problem instances used for comparing algorithms .

2-2 Comparison of the algorithms on graphs having edge density =0.2 .

2-3 Comparison of the algorithms on graphs having edge density =0.3 .

2-4 Comparison of the algorithms on graphs having edge density = 0.4 .

2-5 Descriptions of the problem instances used for analyzing three-stage algoi

2-6 Three-Stage algorithm on graphs having edge density 0.2 . .

2-7 Three-Stage algorithm on graphs having edge density =0.3 . .

2-8 Three-Stage algorithm on graphs having edge density = 0.4 . ..

3-1 Dimensions of clinical problem instances . .............

3-2 Comparison of our base algorithm with Langer et al. (2001) model .

3-3 Effect of rotating the MLC head ............... . .....

3-4 Computational results for our base algorithm . .....

3-5 Comparison of heuristic algorithms on clinical data . .

4-1 Effect of valid inequalities and the partitioning strategy . .

4-2 Computational results on model extensions . .....

4-3 Effect of maximum intensity value on solvability . .....


i


page

46


. 47

. 48

. 49

thm 49

. 49

. 50

. 50

. 73

. 75

. 81

. 81

. 83

. 108

. 109

. 111


5-1 Average number of branch-and-bound nodes explored for hide-and-seek problem

5-2 Average number of searchers needed for the hide-and-seek problem .. .....

5-3 Number of instances that are solved within time limit for the pursuit evasion
p rob lem . . . . . . . . . .

5-4 Average number of branch-and-bound nodes explored for the pursuit evasion
p rob lem . . . . . . . . . .

5-5 Average number of searchers needed for the pursuit evasion problem .......

5-6 Number of instances that are solved within time limit for the patrol problem .

5-7 Average number of branch-and-bound nodes explored for the patrol problem .

5-8 Average number of searchers needed for the patrol problem .. ..........









LIST OF FIGURES


Figure page

2-1 (a) An instance of the deterministic edge-partition problem (b) A solution with
K I = 3, 3, b = 20 .. .. ... .. .. .. ... .. .. .. .. .. ..... 19

3-1 (a) A multileaf collimator system (b) The projection of an aperture onto a patient 51

3-2 Comparison of total treatment times on random data . . 76

3-3 Comparison of the number of apertures on random data .......... .77

3-4 Comparison of beam-on-time values on random data . ..... 78

3-5 Comparison of TGI values on random data ............... .... 79

4-1 Example fluence map ............... ............. .. 87

4-2 Example start index ............... ............. .. 91

4-3 Example end index ............... .............. .. 91

4-4 Example bounding box ............... .......... .. .. 92

4-5 Another nondominated bounding box seeded at (6,3) .............. ..93

4-6 Two components of a fluence map ............... ...... 98

4-7 Regions of a connected component ................ ..... 98

4-8 Efficient frontier for number of apertures and beam-on-time . .... 110

5-1 (a) An example graph (b) Time-expanded network for T = 2 . . ... 118









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

ALGORITHMS FOR SOLVING MULTI-LEVEL OPTIMIZATION PROBLEMS WITH
DISCRETE VARIABLES AT MULTIPLE LEVELS

By

Z. Caner Takmn

August 2009

('!C i: J. Cole Smith
Major: Industrial and Systems Engineering

In this dissertation, we investigate a class of multi-level optimization problems, in

which discrete variables are present at multiple stages. Such problems arise in many

practical settings, and they are notoriously difficult to optimize. Benders decomposition,

which is a well-known decomposition method for solving large-scale mixed-integer

programming problems, cannot be utilized for the class of problems that we consider

due to the existence of discrete variables at lower levels. Cutting plane algorithms such

as those proposed by Laporte and Louveaux have been designed for use in bi-level integer

programming problems with binary variables at both levels. However, these are based on

generic cuts, which do not utilize any problem specific structures, and hence often result

in weak convergence. Our goal in this dissertation is to propose novel formulation and

solution strategies for several multi-level optimization problems to solve these problems to

optimality within practical computational limits.

We first consider an edge-partition problem. The motivation for this study is provided

by a Synchronous Optical Network (SONET) design application. In the SONET context,

each edge represents a demand pair between two client nodes, and the weight of each

edge represents the number of communication channels needed between the client

nodes. We consider a stochastic version of the problem, in which the edge weights

are not deterministic, but their underlying probability distribution is known. The

problem is to design a set of SONET iiis" at minimum cost, while ensuring that









the resulting network can handle the random demand with a prespecified level of service.

We first model the problem as a large-scale integer program, and attempt to solve it

via a bi-level decomposition approach, in which both levels contain binary variables.

We then propose a three-level solution approach for the problem, which is based on a

hybrid integer programming/constraint programming decomposition algorithm. We show

computationally that the hybrid algorithm significantly outperforms the other approaches.

Next, we focus on a matrix decomposition problem, which arises in Intensity

Modulated Radiation Therapy (IMRT) treatment planning. The problem input is a matrix

of intensity values that needs to be delivered to a patient, which must be decomposed

into a collection of apertures and corresponding intensities. In a feasible decomposition

the sum of binary shape matrices multiplied by corresponding intensity values is equal to

the original intensity matrix. We consider two variants of the problem: (i) a variant in

which the shape matrices used in the decomposition have to satisfy the "consecutive-ones"

property, and (ii) a variant in which the shape matrices have to be rectangular. For the

first variant, we start by investigating an integer programming model proposed in the

literature, and show how the formulation can be strengthened. We then formulate the

problem as a bi-level optimization problem that has discrete variables at both stages, and

~i-,-. -1 a hybrid integer programming/constraint programming decomposition algorithm

similar to our algorithm for the edge-partition problem. Our tests on data obtained from

patients show that our algorithm is capable of solving problem instances of clinically

relevant dimensions within practical computational limits. We then turn our attention to

the second variant of the matrix decomposition problem. We formulate the problem as a

mixed-integer program, and investigate a decomposition method for solving it. Unlike the

first variant, the second-level problem turns out to be a linear programming problem, and

therefore we are able to derive a Benders decomposition algorithm for solving this variant.

Finally, we investigate a class of graph search problems. In this class of problems, an

intruder is located at an unknown node on the input graph, and a group of searchers needs









to be coordinated to detect the intruder within a limited amount of time. This problem

arises in settings such as search-and-rescue and search-and-capture operations, and patrol

route design. We investigate three variants of the problem: (i) a hide-and-seek version, in

which a stationary intruder is hiding at an unknown node, (ii) a pursuit evasion version,

in which the intruder can move across the edges of the graph to avoid being detected, and

(iii) a patrol problem, in which the searchers are assigned to recurring patrol routes to

protect the graph from intrusion. We model each problem as a large-scale integer program,

and propose a branch-cut-price algorithm to find the minimum number of searchers

needed, and a route for each searcher. In our formulation, both the master problem and

the subproblems corresponding to the searchers and the intruder contain binary variables.

We model the master problem as a set covering problem and propose a solution approach

that is based on dynamic column and row generation.









CHAPTER 1
INTRODUCTION

In most complex decision-making environments, there exist several types of

interdependent decisions that need to be made to optimize some cost or benefit function.

As a simple example, consider a production planning problem. The goal is to determine,

at the very least, the types of products that are to be produced within a time period,

along with the associated production quantities. There might exist individual restrictions

on each type of decision, such as "the total amount of production of products A and B

cannot exceed a," or "if product A is produced, so must product B." There might also

exist restrictions that relate the two types of decisions, such as "if product A is produced,

then the batch size must be at least 3." Modeling an optimization problem involves (i)

defining a decision variable for each individual decision, (ii) formulating the restrictions

on the decisions as constraints, and (iii) defining an objective function to be optimized.

The field of mathematical programming seeks to optimize such models and prove the

optimality of the generated solution, or prove that no feasible solution exists.

Optimization problems in which some variables are restricted to take values from

a discrete set are classified as discrete optimization problems. An important concern

regarding building and solving discrete optimization problems is that the amount of

memory and the computational effort needed to solve such problems grow exponentially

with the number of discrete variables. The traditional approach, which involves making

all decisions simultaneously by solving a monolithic optimization problem, quickly

becomes intractable as the number of discrete variables increases. Multi-level optimization

algorithms, such as Benders decomposition (Benders, 1962), have been developed as

an alternative solution methodology to alleviate this difficulty. Unlike the traditional

approach, these algorithms divide the decision-making process into several stages. For

instance, in Benders decomposition a first-stage master problem is solved for a subset

of variables, and the values of the remaining variables are determined by a second-stage









subproblem given the values of the first-stage variables. In our simplistic production

planning example, the master problem selects a subset of the products to be produced, by

considering only the restrictions such as "if product A is produced, so must product B."

Then, given the set of selected products, a subproblem seeks the production quantities

considering more detailed production restrictions such as "the total amount of production

of products A and B cannot exceed a," and "if a product is produced, then the batch

size must be at least 3." If the subproblem is able to find a solution for the second-stage

variables, then a solution for the overall problem can be obtained by combining the first-

and second-stage decisions. However, the subproblem can also determine that the current

selection of products by the master problem does not yield a feasible solution when the

detailed production constraints are considered. In this case, a constraint that eliminates

the current selection of products from further consideration is added to the master

problem, which is then re-solved. In this manner, the optimization problem is solved via a

series of "solution prop(.- iI- made by the master problem, and i- .-..i- for rejection" by

the subproblem. This iterative algorithm eventually converges to an optimal solution for

the overall problem.

In essence, multi-level optimization algorithms solve a series of small problems

instead of a single large problem. Performing multiple iterations is usually justified due

to the exponentially larger computational resource requirements associated with solving

a larger problem. Furthermore, it is often the case that decisions for several groups of

second-stage variables can be made independently given the first-stage decisions. In such

cases, multi-level optimization algorithms are amenable to parallel implementations. The

advent of efficient parallel computing grids has allowed modern bi-level techniques to

solve problems that were regarded as intractable before (Ntaimo and Sen, 2005). In some

applications, solving problems in multiple stages allows effort to be conserved by avoiding

the explicit solution of problems by mathematical programming, such as the evacuation

network design algorithm of Andreas and Smith (2009). Multi-level optimization has









recently received much attention due to the emerging importance of research in fields

like stochastic programming and network interdiction. Many problems can be formulated

naturally as multi-level optimization problems (\!'gdalas and Pardalos, 1996; Migdalas

et al., 1997), and a wide class of optimization problems can be reformulated as multi-level

optimization problems (Huang and Pardalos, 2002).

Benders decomposition has been particularly successful in solving mixed-integer

linear programming problems arising in a wide variety of applications. In Benders

decomposition, discrete variables of the problem are kept in the master problem, and

continuous variables are moved to the subproblem. In each iteration, given the values

of the discrete variables by the master problem, the subproblem is solved as a linear

program, and a cutting plane to be passed back to the master problem is generated

using linear programming duality. However, this approach cannot be applied directly

when discrete variables also appear in the second stage. The reason is that no dual

information can be extracted from the subproblem if the second-stage problem contains

integer variables. In this case, one can employ cutting planes such as the general-purpose

cuts of Laporte and Louveaux (1993) and combinatorial Benders inequalities (Codato

and Fischetti, 2006). However, these inequalities are often very weak, and result in slow

algorithmic convergence.

Our main line of research is about designing efficient multi-level optimization

algorithms for problems that have discrete variables at multiple stages. We first present

our results on an edge-partition problem arising in a telecommunications network design

context regarding Synchronous Optical Networks (SONET). The edge-partition problem

considers an undirected graph with weighted edges, and simultaneously assigns nodes and

edges to subgraphs such that each edge appears in exactly one subgraph, and such that

no edge is assigned to a subgraph unless both of its incident nodes are also assigned to

that subgraph. Additionally, there are limitations on the number of nodes and on the

sum of edge weights that can be assigned to each subgraph (Goldschmidt et al., 2003).









We consider a stochastic version of the edge-partition problem in C'!i ipter 2, in which we

assign nodes to subgraphs in a first stage, realize a set of edge weights from among a finite

set of alternatives, and then assign edges to subgraphs. We first formulate the problem as

a monolithic integer programming problem, and show that this approach is not tractable

due to the rapidly increasing computational requirements. We then prescribe a bi-level

cutting plane approach having integer variables in both stages and examine computational

difficulties associated both with the generic inequalities by Laporte and Louveaux (1993)

and with our proposed cutting planes. We also prescribe a three-level hybrid integer

programming/constraint programming algorithm having discrete variables at all levels,

and discuss how this hybrid approach resolves some of the difficulties associated with the

bi-level cutting plane approach.

('!C lpters 3 and 4 consider a problem dealing with the efficient delivery of Intensity

Modulated Radiation Therapy (IMRT) to individual patients. In particular, we consider a

matrix decomposition problem that arises at the leaf sequencing stage in IMRT treatment

planning. The problem input is an integer matrix of intensity values that are to be

delivered to a patient from a given beam angle. To deliver this intensity profile to

the patient, we must decompose the input matrix into a collection of apertures and

corresponding intensities. A feasible decomposition is one in which the original desired

intensity profile matrix is equal to the sum of a number of feasible binary matrices

multiplied by corresponding intensity values. To most efficiently treat a patient, we wish

to minimize a measure of total treatment time, which is given as a weighted sum of the

number of apertures and the sum of the aperture intensities used in the decomposition.

In ('!C lpter 3, we describe a version of the problem in which each aperture matrix needs

to satisfy the "consecutive-ones" property, which means that all matrix entries with

value 1 in each row of an aperture matrix must be consecutive. Similar to C'! Ilpter 2,

we prescribe a bi-level hybrid optimization algorithm in which the master problem is an

integer programming problem, and we solve a subproblem for each row of the matrix by









a constraint programming-based backtracking algorithm. ('!I ipter 4 deals with another

variant of the matrix decomposition problem, in which only rectangular apertures can be

used in the decomposition. We develop a Benders decomposition algorithm for solving this

variant. We also propose a scheme for partitioning the problem to obtain simultaneous

upper and lower bounds.

In ('! Ilpter 5, we study a class of graph search problems, where a group of searchers

seek an intruder on a graph. Both the intruder and the searchers are located at some

nodes of the graph, and the searcher can only "see" a subset of the nodes from each

node. At each time period, both the intruder and the searchers can move along an edge

to an .,li i:ent node, or stay at the same node. Our goal is to find the minimum number

of searchers needed to locate the intruder within a given time limit. We investigate

three variants of the graph search problem: (i) a hide-and-seek problem, in which a

stationary intruder "hides" at an unknown node, (ii) a pursuit evasion problem, in

which the intruder moves among the nodes to avoid being detected, and (iii) a patrol

problem, in which no intruder is initially in the graph and each searcher patrols the

graph to protect it from potential intrusion. We formulate these problems as a set

covering problem with an exponential number of variables and constraints, and propose a

branch-cut-price algorithm for solving it. Both our master problem and the subproblems,

which correspond to the intruder and the searchers, have discrete variables. We formulate

the intruder's subproblem as a longest path problem on an auxiliary graph, and the

searcher's subproblems as mixed-integer programming problems.

We conclude our dissertation in ('!C ipter 6, which explores future research directions

regarding multi-level optimization algorithms. We first evaluate the approaches taken

in the edge-partition, matrix decomposition, and graph search problems described in

C'!i ipters 2-5, and discuss future research topics regarding each application. We then

describe our preliminary research on a master problem reformulation technique, which

can be used in a variety of bi-level optimization algorithms. This reformulation technique









can eliminate an exponential number of iterations by adding a quadratic number of

variables to the master problem. Therefore, it has the potential of leading to significant

improvements in solvability of a class of bi-level optimization problems.









CHAPTER 2
STOCHASTIC EDGE-PARTITION PROBLEM

2.1 Introduction and Literature Survey

We begin by describing the edge-partition problem of Goldschmidt et al. (2003), which

is defined on an undirected graph G(N, E). In the deterministic edge-partition problem,

we create a set K of (possibly empty) subgraphs of G such that each edge is contained in

exactly one subgraph, subject to certain restrictions on the composition of the subgraphs.

These restrictions include the stipulations that an edge cannot be assigned to a subgraph

unless both of its incident nodes belong to the subgraph, and that no more than r nodes

can be assigned to any subgraph, for some r E Z+. Additionally, each edge (i,j) E E has

a positive weight of and the sum of edge weights assigned to each subgraph cannot

exceed some given positive number b. The objective of the problem is to minimize the sum

of the number of nodes assigned to each subgraph.

2 3 2 3

3 8 3 8
10 10
14 7 6 1 4 7 6

5 12 5\ 12


(a) (b) 5

Figure 2-1. (a) An instance of the deterministic edge-partition problem (b) A solution
with IK = 3, r 3,b =20


Figure 2-1 illustrates the deterministic edge-partition problem. The graph G and the

corresponding edge weights are shown in Figure 2-la. Figure 2-1b shows a feasible solution

that partitions G into IKI = 3 subgraphs, where the number of nodes in each subgraph

is limited by r = 3, and the total weight assigned to each subgraph is limited by b = 20.

Note that the degree of node 4 is three, which implies that it must be assigned to at least

two subgraphs, or else there would be at least 4 > r nodes in a subgraph. Similarly, node

5 must be assigned to at least two subgraphs. Since nodes 4 and 5 are assigned to two









subgraphs, and every other node is assigned to a single subgraph, the solution represented

by Figure 2-lb is optimal.

Goldschmidt et al. (2003) discuss the edge-partition problem (with deterministic

weights) in the context of designing Synchronous Optical Network (SONET) rings. In

the SONET context, each edge (i,j) E E represents a demand pair between two client

nodes, and the weight ,,' represents the number of communication channels requested

between nodes i and j. All telecommunication traffic is carried over a set of SONET rings,

which are represented by subgraphs in the edge-partition problem. Since each demand

must be carried by exactly one ring, edges must be partitioned among the rings. (Note

that the term iiiig" describes only the physical SONET routing structure, and does

not place any restrictions on topological properties of demand edges included on a ring.

See, e.g., Goldschmidt et al. (2003) for more details.) SONET rings are permitted to

carry communication between nodes only if those nodes have been connected to the ring

by equipment called Add-Drop Multiplexers (ADMs). There are technical limits on the

number of ADMs that can be assigned to each ring (e.g., r), and on the total amount

of channels (e.g., b) that can be assigned to a ring. Since ADMs are quite expensive,

ring networks are preferred that employ as few ADMs as possible, which echoes the

edge-partition problem's objective of minimizing the sum of nodes assigned to each

subgraph.

The primary contribution by Goldschmidt et al. (2003) is an approximation algorithm

for a specific case of the edge-partition problem in which all ,' are equal to one. Sutter

et al. (1998) propose a column-generation algorithm for this problem, and Lee et al.

(2000b) employ a branch-and-cut algorithm on a formulation that we adapt. For the

case in which the weights on the edges can be split among multiple rings, Sherali et al.

(2000) prescribe a mixed-integer programming approach augmented by the use of valid

inequalities, anti-symmetry constraints, and variable branching rules. Smith (2005)

formulates the deterministic version of the edge-partition problem as a constraint program









and examines several issues regarding symmetry and search algorithm design. Specifically,

she shows how adding ..-- regate variables that represent the number of node copies

(similar to our approach in Section 2.3) can improve performance.

We consider a version of the edge-partition problem where the edge weights are

uncertain, and are only realized after the node-to-subgraph assignments have been made.

As we show in Section 2.2, this framework allows us to design more robust solutions than

those in the literature, which are virtually all applied to deterministic data. We seek

a minimum-cardinality set of node-to-subgraph assignments, such that there exists an

assignment of edges to subgraphs satisfying the aforementioned subgraph restrictions with

a pre-specified high probability. Such a probabilistic constraint is extremely hard to deal

with in an optimization framework. One approach, known as scenario approximation (cf.

Calafiore and Campi (2005); Luedtke and Ahmed (2008); Nemirovski and Shapiro (2005))

is to draw independent identically distributed (i.i.d.) realizations of the edge weights

(called scenarios) and require the node-to-subgraph assignments to admit a feasible

edge-to-subgraph assignment in each scenario. It can be shown that, with a sufficiently

large scenario set, a solution to this scenario approximation solution is feasible to the

true probabilistically constrained problem with high confidence. In this study we develop

algorithmic approaches for solving the scenario approximation corresponding to the

discussed probabilistically constrained edge-partition problem. We refer to this scenario

approximation as the stochastic edge-partition problem.

Relatively little work has been done in SONET network design when the edge weights

are uncertain. Smith et al. (2004) consider the SONET ring design problem in which edge

demands can be split among multiple rings and propose a two-stage integer programming

algorithm. The demand splitting relaxation allows the second-stage problems to be

solved as linear programs, and thus standard Benders cuts can easily be derived from the

second-stage recourse problems. However, we have second-stage integer programs from

which dual information cannot be readily obtained.









The edge-partition problem can also be approached using stochastic integer

programming theory. For problems having binary first-stage variables and mixed-integer

second-stage variables, such as our problem, a well-known decomposition approach is

the integer L-shaped method (Laporte and Louveaux, 1993). This method approximates

the second-stage value function by linear "cuts" that are exact at the binary solution

where the cut is generated, and are under-estimates at other binary solutions. Typical

integer programming algorithms progress by solving a sequence of intermediate linear

programming (LP) problems. Using disjunctive programming techniques, it is possible to

derive cuts from the solutions to these intermediate LPs that are valid under-estimators

of the second-stage value function at all binary first-stage solutions (Sen and Higle, 2005;

Sherali and Fraticelli, 2002). This avoids solving difficult integer second-stage problems

to optimality in all iterations of the algorithm, providing significant computational

advantage. Scenario-wise decomposition methods have also been proposed (Car0e and

Schultz, 1999) as an alternative to the above stage-wise decomposition approaches. Here

copies of the first-stage variables are made corresponding to each scenario and are linked

together via non-anticipativity constraints.

Our proposed methodology draws on constraint programming and stochastic integer

programming theory. Hybrid algorithms of this nature have recently been successfully

employ, -1 to solve notoriously difficult problems. Jain and Grossmann (2001) and

Boclhii 'vr and Pisaruk (2006) devise hybrid integer programming/constraint programming

algorithms for solving machine scheduling problems. Thorsteinsson (2001) proposes a

framework for integrating integer programming and constraint programming approaches.

Hooker and Ottosson (2003) extend the Benders decomposition framework so that

constraint logic programs can be used as subproblems to generate cuts that are added to

a mixed-integer linear master problem. A recent work by Hooker (2007) uses logic-based

Benders decomposition to solve several planning and scheduling problems.









The remainder of this chapter is organized as follows. In Section 2.2, we develop a

mixed-integer programming formulation for the stochastic edge-partition problem, and

provide cutting planes that can be used within a two-stage decomposition algorithm.

In Section 2.3, we prescribe an alternative three-stage algorithm to overcome the

computational difficulties associated with the weakness of the proposed cutting planes.

Finally, we compare the efficacy of these algorithms in Section 2.4 on a set of randomly

generated test instances.

2.2 Formulation and Cutting Plane Approach

Let us introduce binary decision variables xik = 1 if node i is assigned to subgraph k

and 0 otherwise, Vi E N, k E K. For this formulation, we specify a value of IKI that is

sufficiently large to ensure that a feasible solution exists to the problem (as discussed in

Section 2.4). We denote the vector of node-to-subgraph assignments by x. Let w denote

the random vector of edge weights with known distribution, and w denote a realization

with components ,, .. We define binary decision variables yijk = 1 if edge (i,j) is

assigned to subgraph k. Given an allowed violation probability c E (0, 1) the probabilistic

edge-partition problem can be formulated as follows:


Minimize xaik (2-1)
iEN kEK

subject to Y ik iEN

Xik E {0, 1} Vi c N, k E K (2-3)

Pr {G(x,w) < b > 1 -, (2-4)


where


G(x, w) =Minimize z (2-5)

subject to Yijk 1= V(i,j) E, (2-6)
kEK









S.,, < z Vk e K (2-7)
(i,j)EE

ijk i xk, ijk < k V(i,j) E E, k e K (2-8)

Yijk {0O, 1} V(i,j) E, kE K. (2-9)

The objective (2-1) minimizes the total number of nodes assigned to subgraphs.

Constraints (2-2) limit the number of nodes assigned to each subgraph. Constraints (2-6)

require that the edges be partitioned among the subgraphs. Constraints (2-7) compute the

maximum assigned weight over all subgraphs. Constraints (2-8) require that no edge can

be assigned to a subgraph unless both of its incident nodes are assigned to that subgraph,

and (2-3) and (2-9) state logical restrictions on the variables. By convention, the optimal

value G(x, w) of the integer program (2-5)-(2-9) is +oc if the problem is infeasible. Given

a node-to-subgraph assignment vector x and edge weight vector w there exists a feasible

edge-to-subgraph assignment if and only if G(x, w) < b, i.e., the weight assigned to any

subgraph does not exceed b. Thus the probabilistic edge-partition problem (2-1)-(2-4)

seeks a minimum cost node-to-subgraph assignment such that the probability that there

will be a feasible edge-to-subgraph assignment when the edge weights are realized is

sufficiently high.

To build a scenario approximation of the probabilistic edge-partition problem

(2-1)-(2-4), we generate an i.i.d. sample of w denoted by {wq}q), (we call each

realization a scenario). The scenario approximation is then:


Minimize xik (2-10)
iEN kEK

subject to ik < r Vk e K (2-11)
iEN

xik E {0, 1} Vi e N, k e K (2-12)

G(x, wq)







where the probabilistic constraint is replaced by the deterministic requirement that there
must be a feasible edge-to-subgraph assignment for each scenario. As mentioned before we
refer to the above problem as the stochastic edge-partition problem. The following result,
which follows from the general results in Luedtke and Ahmed (2008), provides justification
for considering the scenario approximation.
Proposition 1. Let a desired corfi,. ,..' level 6 E (0, 1) be given. If the sample size IQI



|Q1 > INIKI ln2 ln] (214)

then i,.;, feasible solution to the stochastic edge-partition problem (2-10)-(2-13) is feasible
to the probabilistic edge-partition problem (2-1)-(2-4) with i, ..1,,i./l./:./;/ at least 1 6.

Proof. Let X denote the set of solutions satisfying the deterministic constraints (2-2) and
(2-3), let X' denote the set of feasible solutions to the probabilistic edge-partition problem
(satisfying (2-2)-(2-4)), and let XQ denote the set of feasible solutions to the stochastic
edge-partition problem corresponding to a sample Q (satisfying (2-11)-(2-13)). We want
to bound Q\ such that Pr{XQ C X} > 1 6.
Consider a solution x E X \ X', i.e., Pr{G(x,w) < b} < 1 c. Then x E XQ if
and only if G(x, wq) < b for all q e Q. Since the wq for q E Q are i.i.d. it follows that
Pr{x c XQ} < ( e)I I. Now

Pr{XQ X'} = Pr{3 x XQ s.t. Pr{G(x, w) < b} < 1- e}

< Excx\x, Pr{x E XQ} < IX \ X'(1 c < IX|(1 -e)Q.

Thus Pr{XQ C X'} > 1 X|(1 e)IQI. To guarantee that Pr{XQ C X'} > 1 6 we need
IX(1 c)I'I < 6 or equivalently

|Q|> cIn X -In6/ In a(- .)

The claimed bound then follows by noting that IX| < 2 J11^ and ln(l/(l e)) > e. O









The above result -i'-i-'- -1 that we can obtain feasible solutions to the probabilistically

constrained edge-partition problem by solving the stochastic edge-partition problem with a

"not too 1 i;,. number of scenarios. Key to this sampling-based approach is the ability to

efficiently solve stochastic edge-partition instances having a modest number of scenarios,

which is the motivation of this chapter.

Next we describe an extensive form model of the stochastic edge-partition problem.

Let E' be the set of edges with non-zero weights under scenario q. We define binary

decision variables y 1k 1 if edge (i,j) is assigned to subgraph k in scenario q and 0

otherwise, Vq E Q, (i,j) E Eq, and k c K. The stochastic edge-partition problem can then

be formulated as follows:

Minimize X ik (2-15)
iEN kEK
subject to Y k = Vq e Q, (i,j) e Eq (2-16)
kEK

Y ik iEN
S". ti' k < b Vq Q, k K (2-18)
(i,j)EEq

yik < Xik, jk < Xjk Vq e Q, (i,j) e Et, k e K (2-19)

xik E {0,1} Vi E N, k E K (2-20)

k {0, 1} Vq e Q, (i, j) cE E, k e K. (2-21)

Observe that if one were to solve the above extensive form problem given by (2-15)-(2-21),

integrality restrictions need only be imposed on the y-variables, which would in turn

enforce the integrality of the x-variables at optimality. Note also that given a fixed set

of x-values, this problem decomposes into IQI separable integer programs, where the

subproblem corresponding to scenario q E Q is given by:

Sq(x) = Maximize 0 (2-22)










subject to (2-16), (2-18), (2-19), and (2-21).


Under the foregoing model, it is useful to define ., = min{xik, jk} as a part of the

first-stage decision variables, V(i,j) E E, k E K. The presence of these variables allow

us to formulate stronger cutting planes than would be possible with just x-variables (see

also Smith et al. (2004)). Assuming that UqEQE = E, the extensive form problem is now

equivalent to:


Minimize Y Yxik
iEN kEK
subject to YXik iEN
X Xik, < Xjk

>1 V(ij)
kEK
xik e {0, 1} Vi e N

F(v) < b Vq E


F(v) Minimize maxf{ ', jk }
(i,j)EEq
subject to Yijk V(i,j) E Eq
kEK
Yijk-< ., V(i,j) E, k cK

yjk {0, 1} V(i, j) E, k cK.


The valid inequalities (2-27) require that for each edge (i,j) E E, both i and j must

be assigned to some common subgraph, and are useful in improving the computational

efficacy of the decomposition algorithm that we propose. Note that an optimal solution

exists in which = min{xik,xjk} V(i,j) E E, k E K, without enforcing integrality

restrictions or lower bounds on the v-variables.


(2-24)


V(i,j) e k K

EE

, keK

Q,


where


(2-30)


(2-23)


(









There can exist up to IKI! 1 alternative optimal solutions to this problem by simply

reindexing the subgraph indices. These symmetric solutions are known to impede the

performance of branch-and-bound algorithms (Sherali and Smith, 2001; Sherali et al.,

2000). To reduce model symmetry we can rewrite the cardinality constraints (2-25) (or

(2-17) for the extensive form problem) by using the following inequalities:


r> x,1 > x .> > Y xi|K. (2-34)
iEN iEN iEN

For a scenario q and a given vector 9, the problem (2-30)-(2-33) is essentially an

identical parallel machine scheduling problem to minimize makespan (P/ /Cmx) (with

some assignment restrictions). In particular, there would be IKI machines and IEqI jobs,

whose processing times are given by ,,'., V(i,j) E Eq. Each job must be assigned to

exactly one machine, and the v-variables impose some restrictions on the assignments. The

integer programming scheme developed in Smith (2004) is tailored for a similar problem in

which the (weighted) number of demands that cannot be placed on one of these subgraphs

is minimized (i.e., minimum weighted number of tardy jobs). This is not equivalent to

solving a minimum makespan problem; however, the optimal solution of Fq(Q) is no more

than b if and only if the minimum number of tardy jobs is equal to 0. If a positive lower

bound to the problem of minimizing the number of tardy jobs is established, one can

terminate the subproblem algorithm and conclude infeasibility.

We now present a cutting plane algorithm for solving (2-24)-(2-29). The scheme

relaxes constraints (2-29) and adds cutting planes as necessary to enforce feasibility to the

subproblems. Let us call the problem (2-24)-(2-28) the master problem ( IP).

1. Solve MP. If MP is infeasible then STOP; the problem is infeasible. Otherwise let 9
be an optimal solution of MP.

2. For q e Q, compute Fq(9). If Fq(9) < b for all q, then STOP; the current solution is
optimal. Otherwise, continue to Step 3.

3. Update MP by adding a cutting plane of the form (2-37) as presented in Remark 1,
and return to Step 1.









After a finite number of steps, the cutting plane algorithm terminates with an optimal

solution or detects infeasibility.

Remark 1. Suppose F4(9) > b for some scenario q and a solution vector 9 to MP.

Let L be a g/1/,al lower bound on F(v), i.e., L4 < F4(v) for all v. Also 1. it,., I(v)

{(ijk) : ., = 1} and O(v) = {(ijk) : ., = 0}. The integer oi//'.:,,r/,'h; cut proposed by
Laporte and Louveaux (1993) for this class of problems is given by


(F (q)- L) < + < (F( L)- L)(|I(I)|))- L. (2-35)
(ijk)EI(v) (ijk)EO(v)

Since L4 < b for ,:;, feasible instance, and since F4(9) > b by assumption, we can 'il
(Clu',,ll ii,,,.J':i,,1 to (2-35) by dividing both sides by (F4(9) L4) and rounding down to

obtain

S + (1 -)>1 (2-36)
(ijk)EO(V) (ijk)eI(V)
However, the following .,:,. ,q,' l;.'; is also a valid cutting plane that dominates (2-36):


S1 .1 > 1. (2-37)
(ijk)EO(v)

To see that (2-37) is valid, consider a solution v' that does not .li; fy the above .,:,. ,;,;,;;.:,

i.e., vik = 0 for all (ijk) E O((). Therefore, vik < I ., for all (ijk). Then FV(v') >

Fq(i) > b, and v' is not feasible. I,. ,!;,;l.:/;i (2-37) dominates (2-36) since the left-hand-

side of (2-37) is not more than that of (2-36), and the right-hand-sides are both equal to

1. Thus, (2-37) serves as a cutting plane that can be used in Step 3 of the above il',' .:thm.

Another reformulation of our subproblem might admit stronger cutting planes than

the ones of the form (2-37). In the parlance of machine scheduling, instead of trying to

minimize the maximum makespan, we may wish to minimize the total sum of tardiness.

Let Ck, Vk c K be a nonnegative variable that denotes the amount of capacity deficit in

subgraph k. Then, the problem of minimizing the total capacity deficit can be formulated









as problem MT (v) below:


MT(v) : T(v) = Minimize Yck (2-38)
kEK
subject to y = 1 V(ij) Eq (2-39)
kEK
y < V1(i,j) E E, k K (2-40)

Ck > q "b./k-b Vk c K (2-41)
(i,j)EEq

k > 0 Vk K (242)

yJ{k E 0, 1} V(i,j) E k K. (2-43)

Clearly, Fq(v) < b if and only if Tq(v) = 0, and so we can replace master problem
constraints (2-29) with the restrictions that Tq(v) = 0 for all scenarios q E Q. If
subproblems Tq(v) are used in lieu of Fq(v), we would obtain (2-36) (directly, this time)
from Laporte and Louveaux's integer feasibility cut. However, we can state a stronger
cutting plane for a solution vector 9 having TV(r) > 0 for some scenario q, by requiring
that the total amount of additional capacity that must be allocated to the collection of
subgraphs is at least TV(r). This inequality is formally stated in the following proposition.
Proposition 2. Suppose for some solution vector v and for some scenario q E Q, we
obtain a lower bound LB4(*) > 0 for TV(r). Then the following .:,., ..,l.:/;o; is a valid
cutting plane for problem MP, and is at least as strong as (2-37):


S min{'"', LB(i)} ., > LB4(v). (2-44)
(ijk)EO(v)

Proof. Suppose by contradiction that there exists a binary vector v* such that T4(v*) = 0,

but E(ijk)eO(v) i"' 6jk < LB(i-). Then there exists a solution (y*,c*) to MTV(v*) having
c* = 0 Vk e K. We will show that the existence of such a v* contradicts the assumption
that LB4(9) is a valid lower bound on TV(9). We now build a solution (y, d) to MTV(9).
First, we construct ? as follows:








1.For (i,j) E E, if y*k = 1 and ., 1, then set ijk = 1 as well.

2.For (i,j) E E4, if Yj k 1 and ., 0, then set yj = 1 for any k E K for which
(ijk) e I(ir). (Note that (ijk) e 0(() since ;., 0.)

3.Set all other ijk = 0.
In other words, y is constructed in two phases. In the first phase, we ensure that if edge

(i,j) was assigned to subgraph k in solution y*, then (i,j) is assigned to k in y as well,
unless = 0 (prohibiting this assignment). In the second phase, if yjk = 1 but = 0,
then we assign (i,j) to any k such that vj k 1. Note that this assignment results in a
solution feasible to (2-39), (2-40), and (2-43). Next, let us construct d. Observe that in
the first phase of assigning edges to subgraphs based on (ijk) E I(i) for which y*jk 1 no
subgraph capacities are violated since c* 0, Vk E K, and so we initialize ck = 0, Vk E K.
In the second phase, we guarantee feasibility to (2-41) (and maintain feasibility to (2-42))
by increasing ck by 'ii,. Thus (y, d) is a feasible solution to MTq(r).
At the end of the second phase of assignments, we have kEK Ckk (ijk)EO(,) ".i' lijk
since YZkK Ck is increased by i, only when both vjk 1 and (ijk) E O(9).
However, by assumption, we have that Z(ijk)EO(,) II" 'Vjk < LB4((). Since >ZkE k k

Y(ijk)EO(v) "' ,ijk, we have that Y:kK Ck < LB'(1), which contradicts the fact that
LB(9)) < TV(r). Therefore, all feasible solutions must obey the inequality


,, ., > LB4( ),
(ijk)EO(v)
from which (2-44) is readily derived. Finally, by dividing both sides of (2-44) by LB(91),
we see that (2-44) implies (2-37). E

Remark 2. In cutting plane implementations based on (2-37), once iw; scenario q is
found such that the current v vector is proven to be infeasible with respect to scenario q,
a cutting plane is generated and the master problem is re-solved. No further scenarios are
tested, since an identical cut would be generated for each infeasible scenario. However, a
cutting plane implementation based on problem (2-38) (2-43) above with cutting planes









(2-44) might '. i". from deriving multiple cuts for each infeasible scenario, since these

cuts could be distinct.

Remark 3. Smith et al. (2004) explore the inclusion of ',iir.,,,:'1 constraints" in the

master problem, which enforce simple necessary conditions for f, ... -,;T/i to SONET

problems. Denote the degree of node i E N by deg(i), and the set of nodes adjacent to i by

A(i). Lee et al. (:'iiiiii) show that node i must be i--',gr./ to at least subgraphs,

since otherwise, more than r nodes would be assigned to some '-l'ci.',l Similarly, for

scenario q E Q, the total weight associated with node i E N is given by EjeA(i) "'.. Since

the total weight that can be assigned to a -b;l'ai'.i, is limited by b, CA(i) is a lower

bound on the number of copies of node i. We can then compute

f [deg(i) EjA(i) 1 -
Smax r- max b (245)

and impose the following valid inequalities in the master problem:


Y xik >i Vi N. (2-46)
kEK
Let i denote a node having the 1r/, -/ lower bound, so that f? > f Vi E N. Node i can

be assigned ar.:l,,I'/.:l; to -i1','l,,l, 1,... ,C, and we fix xil = 2 = x = 1

accordingly. Note that the symmetry-br'.rl. .:, constraints (2-34) need to be adjusted so that

they are enforced separately for I.'l'g.'li'- 1,... ,Cf, and f+ 1,..., IK1. Sherali et al. (2000)

show comp,llI.:..a,,ll;, that such a variable-f i:,:.j scheme improves -...1.tl /ii, of problem

instances.

Smith et al. (2004) note that a node i cannot be assigned to a -;1'.g.Jli k in an

optimal solution unless an adjacent node is also assigned to the same -;'1li.g,', Therefore,

we also include the following constraints in MP:


Xik< j xjk Vi e N, k e K. (2-47)
jEA(i)









Smith (2005) describes valid inequalities that can be derived by ,n,,l;, ...:,i the './''.*'.* it

of the pgil'r/, First, consider an edge (i,j) E E such that ,i = j 1. Let A(i,j)

A(i) U A(j) {i,j} denote the set of distinct nodes that are adjacent to i or j. If

|A(i,j)| > r 1, then i or j must be assigned to at least two '-l';'g,.'l,- Similarly, we
, I,,,: W(i, j) = EkA(i,j)('q + t 't,) + ', and note that if Wq(i,j) > b for some q E Q,

then we cannot f/. .il;;, assign nodes i and j to a single .-,l'q',.l, If A(i,j) > r 1 or

Wq(i, j) > b, then we state the following valid .:,, ,;,,;.l/;;,


S Xik + xjk > 3. (2-48)
kEK kEK

Second, for each edge (i,j) E E, suppose deg(i) > r, deg(j) < r, and IA(i,j)l > 2r 3.

Smith (2005) shows that nodes i and j collectively need to be assigned to at least four

'1,''1qrl''- which we state as:

5 Xik + xk > 4. (2-49)
kEK kEK

2.3 A Hybrid IP/CP Approach

The cutting plane algorithms presented in Section 2.2 are preferable to solving

stochastic edge-partition instances by the extensive form problem given by (2-15)-(2-21),

as we show in Section 2.4. However, the two-stage cutting plane algorithms still suffer

from several computational difficulties. First, the master problem, MP, contains IN |IK

binary variables, IEIIKI continuous variables, and O(IEIIKI) constraints, which results in

large integer programs. Second, the linear programming relaxation of MP is quite weak

for i rn 'i: problem instances. Furthermore, the lower bound improves slowly as cuts of the

type (2-37) or (2-44) are added to MP in each iteration. The main reason for this slow
convergence is the existence of symmetry in MP. Inequalities (2-34) reduce, but do not

completely eliminate, symmetric solutions in MP. Therefore, when a solution of MP is

found to be infeasible to a subproblem, MP often simply switches to a symmetric solution









having the same objective function value. On the other hand, stronger anti-symmetry

constraints tend to make MP very difficult to solve.

In this section we develop a new decomposition framework to remedy these

difficulties. We combat symmetry due to reshuffling of subgraphs by representing

subgraphs as configurations. A configuration c is identified by a subgraph node set N1 (we

allow N = 0) and a positive integer ac, which gives the number of subgraphs having node

set N1. A solution is represented by a configuration multiset C whose elements are pairs

(N,, ac). We eliminate symmetry by ensuring that no isomorphic configuration multisets

(i.e., those that are identical after reindexing configuration indices) are encountered in our

search.

A configuration multiset C satisfies the following necessary feasibility conditions.

Fl: Ecc ac= IKI (partitions E into IKI subgraphs)

F2: I|N, < r, V c C (no subgraph contains more than r nodes)

F3: V(i,j) E E, 3c E C such that i E N1, j E N1 (for each edge (i,j), there is at least
one subgraph to which (i,j) can be assigned)

A multiset C that satisfies Fl, F2, and F3 represents a feasible solution if all edges

can be partitioned on the set of subgraphs corresponding to C without violating the

weight restrictions for any scenario. Note that the number of distinct configurations in C,

which we denote by | C, is dynamically determined in our algorithm.

We now provide an overview of our three-stage hybrid algorithm.

1. The first-stage problem determines (via optimal solution of a mixed-integer program)
the number of times we assign each node to the configurations in C. For instance,
in the example given in Figure 2-la, we could specify that we must use two copies of
nodes 4 and 5, and one copy of the other nodes.

2. In the second stage, we seek a multiset C that uses exactly the number of node
assignments specified in the first phase and satisfies Fl, F2, and F3. In the example
mentioned above, a multiset C having configurations {1, 2, 4}, {3, 4, 5}, and {5, 6}
(each with multiplicity one) could be generated based on the first-stage solution.









3. Finally, in the third stage, we determine whether C is feasible. If C is feasible then
we stop with an optimal solution. Else, we return to the second stage, and generate
a different multiset meeting the stated criteria. If no such multiset exists, a cut is
added to the first-stage problem, which is then re-solved. For the example given
above, the multiset yields a feasible solution (see Figure 2-lb).

2.3.1 First-Stage Problem

For all i E N, let zi be an integer variable that represents the number of copies of

node i to be used in forming configurations. We v- that an INI-dimensional vector z

induces a multiset C if C contains exactly zi copies of node i, Vi E N. The first-stage

problem can succinctly be written as:


Minimize > zi (2-50)
iEN

subject to z induces a feasible multiset (2-51)

S< i < IKI Vi N (2-52)

zi integer, (2-53)


where i is a lower bound on the number of copies required for node i, as given in (2-45).

To formulate the first-stage problem as an integer program, we rewrite (2-51) as an

exponential set of linear inequalities by considering the z-vectors that violate it. We first

need to introduce auxiliary binary variables tik, Vi E N, k = Li,..., K so that tik 1 if

zi = k. Then, given a vector 2 that does not induce a feasible multiset, we note that no z

such that zi < zi, Vi E N, induces a feasible multiset. Hence, at least one component of z

must be increased, and so
IKI
tik > 1 (2-54)
iEN k= i+1
is a valid inequality. Our first-stage problem can now be expressed as the following integer

program:


Minimize > zi (2-55)
iEN









IKI
subject to i > ktik Vi e N (2-56)
k =4
IKI
tik 1 Vi N (2-57)
k ~=

5 tik > 1 V2 eZ (2-58)
iEN k= i+1

tik binary Vi N, k ..., K, (2-59)


where Z is the set of all z-vectors that do not induce a feasible multiset. (The z-variables

are in fact unnecessary in this formulation, but we keep them for ease of exposition.) In

our algorithm we relax constraints (2-58) in the first-stage problem, and add them in a

cutting plane fashion. In every iteration we solve the first-stage problem to find 2, and

solve the second- and third-stage problems to seek a feasible multiset induced by 2. If a

feasible multiset is found, then 2 induces an optimal solution and we stop. Otherwise, we

add a cut of type (2-58) and re-solve the first-stage problem.

2.3.2 Second-Stage Problem

Our second-stage problem seeks a multiset induced by 2 that satisfies Fl, F2, and

F3, using a constraint programming search. Given a set of constraints, a set of variables,

and the domain of each variable (i.e., the set of values that each variable can take),

constraint programming seeks a value assignment to each variable that satisfies all

constraints. Constraints are propagated to reduce variable domains, which in turn

trigger new constraint propagations. When no more domain reductions are possible,

the algorithm searches for a solution by fixing a variable to a value in its domain, then

recursively propagating constraints and reducing variable domains. If the domain of a

variable becomes empty during constraint propagation, then the algorithm backtracks. We

refer the reader to Smith (1995), Lustig and Puget (2001), and Rossi et al. (2006) for a

thorough discussion of constraint programming techniques.









2.3.2.1 Foundations

In our second-stage algorithm, a solution corresponds to a multiset C induced by

2 that meets conditions Fl, F2, and F3. In a solution each node i has a corresponding

|C -dimensional distribution vector i, which represents the number of copies of node i to

be allocated to each existing configuration in C. Note that < cannot exceed ac, and that

ecc pc = i. The domain of a node i E N is the set of possible p'-vectors that i can take.
We -ic that a node i is processed if we have selected its distribution vector P3. A partial

multiset is constructed by processing a subset of the nodes in N.

For instance, consider a five-node graph, and let the z-vector obtained by the

first-stage problem be 2 = (2, 3, 1, 4, 3). Suppose that nodes 1, 2, and 3 have been

processed, and the following partial multiset with CI = 3 has been obtained:

* N = 0, ac = 5,

* N= {1,2}, a2= 2, and

SN3 {2, 3}, a3 -1.

Suppose that we process node 4 by choosing its distribution vector as 4 = (2, 1,1).

Adding node 4 to two of the five copies of N1 creates a new configuration N2 whose node

set consists only of node 4 (with multiplicity two) and reduces the multiplicity of N1 by

two. After similarly adding one copy of node 4 to N2 and one copy of node 4 to N3, we

obtain the following partial multiset with IC' = 5:

* N1 0, a1 = 3 (reduced ca),

* NI {4}, a' = 2 (generated from configuration 1 by adding node 4 to N1),

* N = {1,2}, a2= 1 (reduced a2),

* N2 {1, 2, 4}, a' = 1 (generated from configuration 2 by adding node 4 to N2), and

* N3 {2, 3, 4}, a3 = 1 (added node 4 to N3).

In general, when we process node i by choosing a distribution vector /', we update the

partial multiset C as follows. For each configuration c E C if 3 =- 0, then no changes









are made to c (since no copies of node i are added to c). If 3 = ac, then we update

configuration c by setting Nc = N, U {i}. Finally, if 0 < Pf < a,, then we create a new

configuration c' having N, = N, U {i}, = P, and update configuration c by setting

aOc = c O3.

Remark 4. Recall that the configurations in a partial multiset C can be ordered in |C|!

-,lii,,I ii.,,,- Our l1'j.,,:hm avoids this symmetry by generating only one such ordering

after processing a node. Furthermore, the conr fii ,il..-'i multisets that we compute by

processing node i according to the u'-vectors in its domain must be pairwise nonisomor-

phic, since the f'-values in the domain of node i are distinct. Hence, we never encounter

isomorphic configuration multisets in the second-stage search.

2.3.2.2 Domain expansion

Processing a node modifies the current partial multiset, and therefore distribution

vectors of the remaining unprocessed nodes need to be updated. Domains of nodes are

reduced by constraint propagation as we describe in the next section, but must also

be expanded as new configurations are generated. We describe the initialization and

expansion of node domains below.

In the beginning of the second stage we initialize our multiset C with a single

configuration having N1 = 0 and ac = |K. Each node can only be added to the lone

configuration, and so the domain for node i is initially the single one-dimensional vector

fi = (z). Our algorithm next processes some node i E N and updates the existing set

of configurations: N1 = 0, a1 = IKI zi and N2 = {i, a2 = i. Next, the domains

of all unprocessed nodes are updated to reflect the changes in C. For each unprocessed

node j, we enumerate all possible v--v of partitioning zj copies into node sets N1 and N2.

This logic is repeated at all future steps as well. For instance, in the example given above,

suppose that 35 = (2, 0, 1) was the only vector in the domain of node 5 before processing

node 4. Since processing node 4 modifies the first configuration by reducing ac and

generates a new configuration (N', a'), we expand the domain of node 5 by enumerating









all possible v--v of assigning 3 = 2 copies of node 5 to configurations (N1, al) and

(NI, a'c). On the other hand, since 35 does not assign node 5 to the second configuration,

the distribution vectors in the expanded domain do not add node 5 to (N2, a2) or (2N, 2).

Finally, since processing node 4 does not generate any new configurations from the third

configuration, all distribution vectors in the expanded domain of node 5 assign a single

copy of node 5 to (N3, a3). After processing node 4 and updating the configurations as

described above, the domain of node 5 is expanded to:


{(2,0,0,0, 1), (1, 1,0,0, 1), (0,2,0,0, 1)}.

2.3.2.3 Constraint propagation

The constraints we impose in the second-stage problem limit the number of nodes in

each configuration (F2) and require that each edge has both its end points in at least one

configuration (F3). Condition Fl (requiring IKI total configurations) is implicitly satisfied.

We apply constraint propagation algorithms to remove distribution vectors inconsistent

with F2 or F3 from the expanded node domains. Let i E N be the last processed node,

and let Ci C C represent the subset of configurations to which node i has been added. We

only need to execute constraint propagation for configurations c E CQ, since these are the

only newly modified configurations.

To enforce F2, the propagation algorithm identifies all configurations to which r nodes

have been assigned. For each such configuration c, we remove all distribution vectors 3P

having P3 > 0 from the domains of all unprocessed nodes j e N. To enforce F3, the

propagation algorithm iterates over the domains of the unprocessed nodes j .-lIi i.:ent

to i, and removes all distribution vectors that do not add at least one copy of j to any

configuration in C. Otherwise, the configurations containing node i would be disjoint from

those containing node j, which violates F3.









2.3.2.4 Forward checking

After all constraints are propagated, we first check whether the domain of any

unprocessed node is empty; if so, then we backtrack. Else, we further analyze the current

partial multiset before resuming the search with the next unprocessed node. This step

identifies whether the current partial multiset can eventually yield a feasible multiset as

early as possible to avoid performing unnecessary backtracking steps (van Beek, 2006).

We call one such test implied node assignment i,.il,-.: Suppose that we identify a

processed node i such that zi 1, and the configuration c to which i has been assigned.

By condition F3 it follows that all unprocessed nodes j .,.i i,:ent to i must also be assigned

to configuration c. We use this analysis to augment partial configurations with implied

node assignments, and then check whether any augmented configuration contains more

than r nodes, and hence violates F2.

We also perform an implied edge i/:I..:w, l, i,.rl, i.:j by finding all edges that can

only be assigned to a single configuration. For each (i, j) E E, if both nodes i and j have

been processed, then we check whether both i and j are in a single configuration c for

which ac 1. In this case edge (i,j) can only be assigned to configuration c. On the other

hand if (without loss of generality) node i has been processed but node j has not yet been

processed, and Zi 1, then edge (i,j) can only be assigned to the configuration to which

i has been assigned. After finding all implied edge assignments, we check whether F3 is

violated for any scenario.

Finally, we consider a singleton node wi,';l,..:> in which we ensure that each node is

.,1i ,i:ent to at least one other node in each configuration. For each processed node i, and

for all configurations c E Ci, we seek a node j .Ii] i,:ent to i so that either j E Nc (if j also

has been processed), or f3 > 0 for some distribution vector in the domain of j (if j has

not been processed). If no such j can be found for a configuration c E Ci, then the current

partial solution cannot lead to an optimal solution; node i can ultimately be removed









from configuration c without affecting feasibility conditions, leading to a reduction in the

objective function value.

2.3.2.5 Node selection rule

The order in which variables are processed can significantly affect the performance

of constraint programming algorithms (Lustig and Puget, 2001; Smith, 1995). Especially

for infeasible second-stage problem instances, processing the 1'l I.' i,, i, c" nodes first

can quickly lead to the detection of infeasibility and can result in significant savings in

computational time. We employ a dynamic node selection rule in which the order of

nodes considered can vary in different sections of the search tree. In accordance with the

"fail-first" principle widely used in constraint programming algorithms (Haralick and

Elliott, 1980; van Beek, 2006), our node selection rule first picks an unprocessed node that

1. has the fewest number of distribution vectors in its domain,

2. has the fewest number of copies to be partitioned, and

3. has the largest number of unprocessed .,.i i:ent nodes,

breaking ties in the given order. In this manner, we can quickly enumerate all possible

distribution vectors of a few key nodes, allowing constraint propagation to quickly reduce

the size of the remaining search space.

2.3.2.6 Distribution vector ordering rule

Once the next node to be processed has been identified, all distribution vectors in

its domain need to be tried one-by-one to see if any of them leads to a feasible multiset.

For an infeasible second-stage problem instance, the order in which these vectors are

instantiated does not matter, because all vectors must be enumerated before infeasibility

can be concluded. However, for feasible problem instances it is important to find a vector

that leads to a feasible multiset as soon as possible to curtail our search. Our ordering

rule attempts to sort the distribution vectors in nonincreasing order of the likelihood that

the vector leads to a feasible multiset. We calculate the feasibility likelihood score of a

distribution vector /P in the domain of an unprocessed node i with respect to a partial









multiset C as:


FL(i, C, ') = 3 {j N : (ij) e E}. (2-60)
cEC
FL(i, C,/3 ) measures the total number of .,li i,:ent node pairs (i, j) that would be added

across all configurations if fO is selected to be the distribution vector for node i. Our

vector ordering rule sorts vectors in the domain of the chosen node in nondecreasing order

of their FL-scores. By allowing for a higher degree of flexibility in assigning edges, we

increase the likelihood that a feasible partition of edges to subgraphs can be found.

2.3.3 Third-Stage Problem

Given a solution of the second-stage problem that consists of a configuration multiset

C satisfying Fl, F2, and F3, the third-stage problem must verify whether C is feasible.

We first generate the set of subgraphs from the multiset C {cl, c2,..., clcl} by assigning

the nodes in Nc, to the first ac, subgraphs, then assigning the nodes in N12 to the next ac2

subgraphs, and so on. Since we have enforced cecc a= K this transformation creates

exactly IKI subgraphs, some of which can be empty. Then we iterate over all subgraphs

and set ., = 1 if nodes i and j are in subgraph k, and ., = 0 otherwise. We then use

formulation (2-38)-(2-43) to solve the third-stage problem.

Note that this transformation re-introduces symmetry into the third-stage problem.

However, the solution of the third-stage problems does not constitute a bottleneck in the

algorithm, and symmetry-breaking constraints appended to the transformed subproblem

will not impact the computational efficacy of the overall algorithm.

2.3.4 Infeasibility Analysis

If a z-vector is found not to induce a feasible multiset, we add a constraint to the

first-stage problem so that the same z-vector is not generated in subsequent iterations.

Constraints (2-58) state that the number of copies of some node must be increased,

but they do not contain any information about which nodes need to be added. We

observe that the progress of our second-stage algorithm can be analyzed to identify a

11 ''I" in I 1c" subset of nodes whose corresponding z-values cause infeasibility regardless









of other variable values. Given a vector 2 for which no feasible multiset exists, if a node

i E N has not been processed, or has not been identified as the reason of infeasibility

in any step of the backtracking algorithm, then 2 will not induce a feasible multiset for

any value of zi. Let P C N denote the set of nodes that have been processed, or whose

domains have become empty due to constraint propagation in the second-stage algorithm,

possibly during different backtracking steps. The following is a valid inequality:

IKI
S ik t> 1. (2-61)
iEP k =i+1

Constraints (2-61) clearly dominate (2-58) for any P C N, and get stronger as |P|

decreases. Based on this observation, we update our node selection rule by giving

preference to selecting nodes that have already been added to P. Our revised node

selection rule first picks a node that

0. has been added to P in a previous backtracking step,

1. has the fewest number of distribution vectors in its domain,

2. has the fewest number of copies to be partitioned, and

3. has the largest number of unprocessed .,.1] i:ent nodes,

again breaking ties in the stated order.

2.3.5 Enhancements for the First-Stage Problem

Our computational studies revealed that the first-stage integer programming

model solution represents the bottleneck operation of our algorithm. To decrease the

computational time spent by the first-stage problem, we investigate several strategies.

2.3.5.1 Valid inequalities

The valid inequalities that we discuss in Remark 3 can be adapted to the first-stage

problem to eliminate the z-vectors that violate the corresponding necessary feasibility

conditions. In particular, constraints (2-46) translate to simple lower bounds (2-52) on

the z-variables. Constraints (2-48), which are written for node pairs that satisfy the









conditions discussed in Remark 3, can be written as:


z, + zj > 3. (2-62)


Similarly, each constraint of type (2-49) can be equivalently represented as following:


z, + zj > 4. (2-63)


Smith (2005) discusses an additional valid inequality, which cannot be represented using

the x-variables in our two-stage algorithm, but can be written in terms of the z- and

t-variables in the first-stage problem of our hybrid algorithm. For nodes i E N and j E N,

if (i,j) E, deg(i) < r 1,deg(j) < r 1, IA(i,j)| > r 1, and there exists a common

neighbor k E N so that k c A(i), k c A(j), deg(k) > r, and if i,j, k have more than

2r 4 distinct neighbors in total, then zi = 1, zy = 1 implies zk > 3. This condition can

be written as:

zk > -1 + 2(tl + tji), (2-64)

which reduces to Zk > 3 for zi j = 1, and is redundant otherwise.

2.3.5.2 Heuristic for obtaining an initial feasible solution

The existence of a good initial feasible solution can help improve the performance

of the first-stage problem because it provides a good upper bound, and allows the solver

to apply strategies such as reduced cost fixing. We first solve the first-stage model

enhanced with valid inequalities (2-62)-(2-64) to obtain an initial solution z, and execute

the second- and third-stage algorithms to seek a feasible multiset. If one is found, we

terminate with an optimal solution. Otherwise, we investigate the set of processed nodes

P C N, and pick a node c E P having the fewest number of copies (breaking ties by

picking a node having the largest degree). We then set z, = z, + 1 and re-invoke the

second- and third-stage algorithms. This algorithm eventually finds a feasible multiset or

concludes that the entire problem is infeasible after generating the solution zi = KI, Vi E









N. We also generate a cut of type (2-61) for each 2 generated before a feasible multiset is

found, which we add to the first-stage problem to improve the lower bound.

2.3.5.3 Processing integer solutions

We can interrupt the branch-and-bound solution process of the first-stage problem

each time the solver finds an integer solution 2, and check whether 2 induces a feasible

multiset by solving the second- and third-stage problems. If a feasible multiset exists, we

accept 2 as the new incumbent and resume solving the first-stage problem. Otherwise,

we reject 2, generate a constraint of type (2-61), and again resume the solution process.

The same idea is also applicable to the master problem (lI'P) of the two-stage algorithm

discussed in Section 2.2.

In our tests, this approach turned out to be more effective than solving the first-stage

problem to optimality in each iteration, adding a cut, and re-solving it. The reason is that

the problem is solved using a single branch-and-bound tree, which we tighten by adding

cuts as necessary on integral nodes, instead of repeatedly generating a branch-and-bound

tree in each iteration. It also allows us to obtain good feasible solutions for problem

instances that are too difficult to solve to optimality.

We note that this approach requires a minor modification to the second-stage

algorithm. All constraint propagation (Section 2.3.2.3) and forward checking rules (Section

2.3.2.4) except for singleton node analysis are based on necessary conditions for feasibility

of configurations, and therefore they are valid for any integral 2. However, singleton node

analysis is based on an optimality condition and hence can only be used if 2 is a candidate

optimal solution to the first-stage problem.

2.4 Computational Results

We implemented the algorithms discussed in the previous sections using CPLEX 11.1

running on a Windows XP PC with a 3.4 GHz CPU and 2 GB RAM. Our base set of

test problem instances consists of 225 randomly generated problem instances for which

the expected edge density of the graph (measured as i l-) takes values 0.2, 0.3, and
NJ.xQNJ 1)1 e aus0.,03 n










0.4, the number of nodes ranges from 5 to 15, and the number of scenarios is between 1

(corresponding to the deterministic edge-partition problem) and 100. There is no practical

limit on the number of subgraphs (IKI), but a limit needs to be specified to model the

problem (see Goldschmidt et al. (2003); Sherali et al. (2000); Smith (2005)). C'! .. -ig IKI

too small may make the problem infeasible, and large values of IKI increase difficulty of

the problem. In our tests, we chose IKI sufficiently large to yield a feasible edge partition

in each problem instance. In generating instances we first picked a random subset of edges

to have a positive weight, and then we assigned a weight uniformly distributed between

1 and 10 to each edge in each scenario. We generated five problem instances for each

problem size, which is determined by the expected edge density, the number of nodes, and

the number of scenarios. The data set names and details used in our experiments are given

in Table 2-1.

Table 2-1. Descriptions of the problem instances used for comparing algorithms
Name INI IKI IQI r b Name INI IKI IQI r b
5-1 5 5 1 4 20 12-1 12 10 1 5 50
5-30 5 5 30 4 20 12-30 12 10 30 5 50
5-100 5 5 100 4 20 12-100 12 10 100 5 50
8-1 8 7 1 4 35 15-1 15 10 1 8 70
8-30 8 7 30 4 35 15-30 15 10 30 8 70
8-100 8 7 100 4 35 15-100 15 10 100 8 70
10-1 10 8 1 5 40
10-30 10 8 30 5 40
10-100 10 8 100 5 40


We used the default options of CPLEX for solving the extensive form problems.

Preliminary computational experience on our two-stage algorithm indicated that the

best implementation includes the valid inequalities (2-27), (2-46)-(2-49), and the

symmetry-breaking constraints (2-34), and uses the model given by (2-38)-(2-43)

for the subproblem, which is the formulation that minimizes the total tardiness. In

our base setting for the three-stage algorithm, we used our heuristic to find an initial

feasible solution, generated valid inequalities (2-62)-(2-64), and (similar to the two-stage

algorithm) we used formulation (2-38)-(2-43) for the third-stage problem. We used

callback functions of CPLEX to generate a single branch-and-bound tree for both

two-stage and three-stage algorithms as discussed in Section 2.3.5. We imposed a half-hour










(1800 seconds) time limit past which we halted the execution of an algorithm in all our

experiments.

Our first experiment compares the performance of the extensive form, two-stage, and

three-stage algorithms. Table 2-2 summarizes the results of these three algorithms on

low density graphs having expected edge density 0.2. For each problem size, we report

the following statistics calculated over five random instances: (i) the number of problems

solved to optimality ("Solvh 'i ), (ii) the average optimality gap obtained at the root node

("Root C p1 ), (iii) the average final optimality gap for instances that could not be solved

within the allowed time limit ("Final C(; ), (iv) the average amount of time spent by

each algorithm on the instances that were solved to optimality ("Time"). Out of the

75 instances in this data set, CPLEX could solve the extensive form to optimality for

61 instances, while both two-stage and three-stage algorithms solved all 75 instances to

optimality within a few seconds. The results reveal that the performance of the extensive

form formulation deteriorates rapidly as the number of scenarios increases, but the effect

of the number of scenarios is mitigated for the two-stage and three-stage algorithms. We

observe that the average optimality gap obtained by the three-stage algorithm at the

root node is 1..!' which is significantly less than the initial gaps obtained using other

approaches.

Table 2-2. Comparison of the algorithms on graphs having edge density = 0.2
Extensive Form Two-Stage Three-Stage
Root Final Root Final Root Final
Name Solved Gap Gap Time Solved Gap Gap Time Solved Gap Gap Time
5-1 5 0.00% 0.1 5 5.00% 0.1 5 0.00% 0.1
5-30 5 18.33% 6.6 5 4.00% 0.2 5 0.00% 0.1
5-100 5 12.38% 5.4 5 11.00% 0.6 5 0.00% 0.3
8-1 5 25.90% 0.4 5 6.67% 0.1 5 0.00% 0.1
8-30 5 12.89% 4.0 5 3.64% 0.2 5 0.00% 0.1
8-100 5 37.61% 223.1 5 14.84% 1.2 5 0.00% 0.3
10-1 5 19.58% 0.5 5 17.80% 0.4 5 0.00% 0.1
10-30 5 57.01% 147.3 5 10.71% 0.8 5 0.00% 0.2
10-100 4 30.35% 7.14% 684.1 5 13.94% 2.0 5 0.00% 0.4
12-1 5 47.25% 8.1 5 24.66% 2.2 5 0.00% 0.1
12-30 4 55.09% 25.00% 507.3 5 17.99% 4.3 5 3.08% 0.3
12-100 2 62.21% 24.88% 713.1 5 36.28% 4.7 5 2.11% 0.8
15-1 5 31.85% 33.0 5 64.38% 16.4 5 4.56% 0.2
15-30 1 65.29% 21.65% 864.6 5 39.13% 27.1 5 7.29% 0.6
15-100 0 57.33% 28.47% 5 24.49% 20.4 5 4.86% 1.2










Tables 2-3 and 2-4 compare the three approaches on denser graphs having edge

density 0.3 (medium density) and 0.4 (high density), respectively. We observe that

performances of all three algorithms deteriorate as the edge density increases, which is

not surprising due to the nature of the edge-partition problem. The number of instances

Table 2-3. Comparison of the algorithms on graphs having edge density = 0.3
Extensive Form Two-Stage Three-Stage
Root Final Root Final Root Final
Name Solved Gap Gap Time Solved Gap Gap Time Solved Gap Gap Time
5-1 5 0.00% 0.1 5 2.86% 0.1 5 0.00% 0.1
5-30 5 25.76% 10.4 5 6.15% 0.5 5 0.00% 0.1
5-100 5 10.00% 3.1 5 10.77% 0.4 5 0.00% 0.2
8-1 5 31.30% 0.5 5 11.20% 0.1 5 0.00% 0.1
8-30 5 42.57% 18.0 5 7.48% 0.4 5 0.00% 0.2
8-100 4 39.37% 7.14% 110.0 5 16.19% 1.3 5 1.43% 0.3
10-1 5 32.42% 3.7 5 16.27% 0.6 5 1.18% 0.1
10-30 4 51.33% 21.05% 953.0 5 40.82% 8.2 5 5.83% 0.3
10-100 2 61.24% 29.05% 382.6 5 35.43% 302.7 5 8.89% 0.5
12-1 5 53.85% 312.0 5 39.49% 16.8 5 4.65% 0.2
12-30 0 63.41% 27.06% 5 46.98% 120.5 5 9.31% 0.8
12-100 0 84.24% 65.50% 4 42.78% 4.35% 89.0 5 11.99% 1.4
15-1 4 46.88% 11.54% 460.4 5 72.86% 250.5 5 12.93% 0.9
15-30 0 66.01% 42.41% 2 72.20% 16.02% 30.4 5 13.51% 3.5
15-100 0 80.76% 74.48% 0 53.05% 13.21% 5 16.31% 4.1



that can be solved by the extensive form decreases from 61 for low density graphs to 49 for

medium density graphs, and finally to 36 for high-density graphs. The two-stage algorithm

also exhibits a similar behavior; it can solve 75, 66, and 61 instances for low, medium, and

high-density graphs, respectively. On the other hand, the three-stage algorithm is able

to solve almost all instances, failing to solve two instances in the high-density 15-30 and

15-100 data sets to optimality within the allowed time limit. Table 2-4 clearly shows that

the three-stage algorithm dominates the other approaches, and the two-stage algorithm

provides better results than directly solving the extensive formulation. Our analysis of

optimal solutions obtained for the problem instances shown in Tables 2-2-2-4 showed

that the average objective function value for the deterministic (single-scenario) problem

instances is 14.8. This value is smaller than the average objective function value for

30-scenario and 100-scenario instances (15.52 and 15.6, respectively). We also observe that

several subgraphs can be empty in an optimal solution.











Table 2-4. Comparison of the algorithms on graphs having edge density = 0.4
Extensive Form Two-Stage Three-Stage
Root Final Root Final Root Final
Name Solved Gap Gap Time Solved Gap Gap Time Solved Gap Gap Time
5-1 5 5.00% 0.1 5 0.00% 0.1 5 0.00% 0.1
5-30 5 24.67% 2.6 5 19.79% 0.3 5 0.00% 0.1
5-100 5 12.38% 5.6 5 8.31% 0.6 5 0.00% 0.2
8-1 5 41.32% 2.0 5 3.33% 0.1 5 0.00% 0.1
8-30 5 48.89% 140.9 5 17.68% 1.1 5 1.43% 0.1
8-100 3 47.23% 22.50% 113.0 5 21.08% 8.7 5 2.50% 0.4
10-1 5 45.08% 48.3 5 32.36% 3.5 5 2.16% 0.1
10-30 0 61.52% 20.64% 5 56.55% 39.5 5 8.45% 0.4
10-100 0 64.47% 50.91% 3 54.82% 7.50% 151.7 5 12.73% 1.5
12-1 1 67.13% 14.30% 33.2 5 40.60% 327.3 5 7.86% 0.5
12-30 0 88.61% 46.74% 5 42.93% 160.9 5 3.16% 0.8
12-100 0 84.37% 68.24% 5 51.54% 583.7 5 13.91% 1.7
15-1 2 60.11% 11.21% 369.6 3 53.01% 5.56% 410.0 5 11.57% 0.9
15-30 0 85.29% 65.29% 0 66.72% 22.66% 3 18.03% 4.74% 120.2
15-100 0 96.00% 86.92% 0 62.62% 24.58% 3 19.98% 6.45% 173.8


. Descriptions of the problem instances
| IKI r b
5 4 20
7 4 35
8 5 40
10 5 50
10 8 70
10 8 100
10 10 120
10 10 140


used for analyzing three-stage algorithm


Our next experiment analyzes the performance of our three-stage algorithm for larger

instances. For this experiment, we generated additional random problem instances using

the parameter settings given in Table 2-5. Similar to our previous experiments, we

Table 2-6. Three-Stage algorithm on graphs having edge density = 0.2
6 = 0.05 = 0.01
Root Final Heuristic Root Final Heuristic
Name |Q| Solved Gap Gap Time Gap |Q| Solved Gap Gap Time Gap
5 407 5 0.00% 0.8 2.86% 2194 5 0.00% 3.5 0.00%
8 837 5 0.00% 2.2 1.54% 4343 5 0.00% 12.7 3.33%
10 1169 5 10.88% 5.6 5.09% 6006 5 1.33% 19.6 1.33%
12 1724 5 1.11% 13.2 4.19% 8779 5 3.00% 58.5 3.33%
15 2140 5 7.24% 22.1 2.74% 10858 5 12.41% 170.9 4.37%
17 2417 5 10.19% 41.0 4.78% 12245 5 9.82% 211.1 8.01%
20 2833 5 16.55% 79.8 7.01% 14324 5 12.40% 403.7 4.51%
22 3110 5 14.77% 128.2 6.49% 15710 5 15.78% 699.9 6.46%



generated problem instances for which the expected edge density of the graph takes values

0.2, 0.3, and 0.4. For each data set, we calculated the number of scenarios corresponding

to e, 6 = 0.05 and e, 6 = 0.01 using Proposition 1. Hence, inequality (2-14) ensures

that we can be 95'. (9'-'., respectively) certain that all demands can be satisfied 95'

(9I'' respectively) of the time. We generated five random instances for each data set,


Table
Name
5
8
10
12
15
17
20
22










Table 2-7. Three-Stage algorithm on graphs having edge density = 0.3
c, = 0.05 = 0.01
Root Final Heuristic Root Final Heuristic
Name |Q| Solved Gap Gap Time Gap |Q| Solved Gap Gap Time Gap
5 407 5 0.00% 0.8 2.86% 2194 5 0.00% 3.5 0.00%
8 837 5 0.00% 2.6 2.86% 4343 5 3.33% 13.0 2.86%
10 1169 5 6.58% 7.5 8.99% 6006 5 11.86% 27.4 4.80%
12 1724 5 8.61% 16.1 4.51% 8779 5 8.22% 71.7 4.31%
15 2140 5 15.45% 45.1 3.05% 10858 4 18.53% 3.45% 176.4 4.25%
17 2417 5 13.63% 42.4 3.43% 12245 5 9.93% 189.3 2.68%
20 2833 5 17.47% 362.5 3.24% 14324 5 18.13% 639.1 3.32%
22 3110 4 16.18% 4.76% 159.3 5.82% 15710 5 15.86% 738.5 3.45%


resulting in 240 instances in total. In addition to the columns given in Table 2-2, Tables

2-6, 2-7, and 2-8 show the relative gap between the quality of the solution found by our

initial heuristic (Section 2.3.5.2) and the best lower bound obtained ("Heuristic C; Ip ).

Our algorithm can solve 206 instances out of 240 to optimality, and provides an average

Table 2-8. Three-Stage algorithm on graphs having edge density = 0.4
c, = 0.05 e, = 0.01
Root Final Heuristic Root Final Heuristic
Name |Q| Solved Gap Gap Time Gap |Q| Solved Gap Gap Time Gap
5 407 5 0.00% 0.8 2.22% 2194 5 0.00% 3.2 0.00%
8 837 5 7.71% 2.7 2.43% 4343 5 7.25% 17.3 5.33%
10 1169 5 16.38% 9.4 5.71% 6006 5 15.84% 52.1 9.73%
12 1724 5 16.71% 63.2 5.41% 8779 5 14.13% 118.4 5.45%
15 2417 4 16.24% 2.86% 338.8 4.07% 12245 2 16.55% 4.71% 549.8 6.86%
17 2140 1 23.46% 8.46% 993.7 9.23% 10858 1 24.77% 11.44% 1515.5 13.07%
20 2833 0 18.83% 9.46% 9.46% 14324 0 20.01% 11.30% 11.81%
22 3110 0 18.36% 11.05% 11.47% 15710 0 17.83% 9.90% 10.29%



optimality gap of 9.21 for the 34 instances that it cannot solve to optimality. The

maximum optimality gap obtained for the entire data set is 21.22''. The results also

-,--i., -I that our heuristic for finding an initial feasible solution is quite effective: the

average optimality gap for our heuristic is 4.97'. and the maximum optimality gap is

22.72''. Since these calculations are based on the lower bounds obtained for the problem

instances that could not be solved to optimality, our reported gaps possibly overestimate

the true gap between heuristic and optimal objective values.










CHAPTER 3
CONSECUTIVE-ONES MATRIX DECOMPOSITION PROBLEM

3.1 Introduction and Literature Survey

Cancer is one of the leading causes of death throughout the world. In the last

century, external beam radiation therapy has emerged as a very important and powerful

modality for treating many forms of cancer, either in primary form or in conjunction with

other treatment modalities such as surgery, chemotherapy, or medication. In the United

States .-1 I,-iv, approximately two-thirds of all newly diagnosed cancer patients receive

radiation therapy for treatment. Since the radiation beams employ, ,1 in radiation therapy

damages all cells traversed by the beams, both in targeted areas in the patient that

contain cancerous cells as well as any cells in I. i111!:' organs and tissues, the treatment

must be carefully designed. This can partially be achieved by delivering radiation from

several different directions, also called beam orientations. Therefore, patients receiving

radiation therapy are typically treated on a clinical radiation-delivery device that can

rotate around the patient. The most common device is called a linear accelerator and is

typically equipped with a so-called multileaf collimator (_llC) system which can be used

to judiciously shape the beams by forming apertures, thereby providing a high degree of

control over the dose distribution that is received by a patient (see Figure 3-1 1 ). This

Multileaf Photon
collimator It therapy
system beam





Left and right
leaves form
aperture,
creating an I ..
irregularly
shaped beam '.
(a) (b)

Figure 3-1. (a) A multileaf collimator system (b) The projection of an aperture onto a
patient









technique has been named :i/. -.i ih modulated radiation therapy (IMRT).

Since the mid 1990's, large-scale optimization of the fluence applied from a number of

beam orientations around a patient has been used to design treatments from MLC-equipped

linear accelerators. A typical approach to IMRT treatment planning is to first select the

number and orientations of the beams to use as well as an intensity profile or fluence map

for each of these beams, where the fluence map takes the form of a matrix of intensities.

This problem has been studied extensively and can be solved satisfactorily, in particular

when (as is common in clinical practice) the beam orientations are selected manually

by the physician or clinician based on their insight and expertise regarding treatment

planning. For optimization approaches to the fluence map optimization problem with fixed

beam orientations we refer to the review paper by Shepard et al. (1999). More recently,

Romeijn et al. (2006) proposed new convex programming models, and Hamacher and

Kiifer (2002) and Kiifer et al. (2003) considered a multi-criteria approach to the problem.

Lee et al. (2000a, 2003) studied mixed-integer programming approaches to the extension

of the fluence map optimization problem that also optimizes the number and orientations

of the beams to be used. However, to enable delivery of the optimal fluence maps by

the MLC system, they need to be decomposed into a collection of deliverable apertures.

(For examples of integrated approaches to fluence map optimization, also referred to as

aperture modulation, we refer to Shepard et al. (2002), Preciado-Walters et al. (2004), and

Romeijn et al. (2005).)

The vast i i ii iy of MLC systems contain a collection of leaves that can be moved

in parallel, thereby blocking part of the radiation beam. This architecture implies that

we can view each beam as a matrix of beamlets or bixels (the smallest deliverable square

beam that can be created by the MLC), so that each aperture can be represented by

a collection of rows (or, by rotating the MLC head, columns) of bixels, each of which



1 Varian Medical Systems; http://www.varian.com/orad/prd056.html.









should be convex. In other words, each fluence map should be decomposed into either

constant-intensity row-convex apertures or constant-intensity column-convex apertures.

Due to the time required for setup and verification, clinical practice prohibits using both

types of apertures for a given fluence map, so that without loss of generality we focus on

row-convex apertures only. Note that while some manufacturers of MLC systems impose

additional constraints on the apertures, we assume that all row-convex apertures are

deliverable. As an example, consider the fluence map given by the following 2 x 3 matrix

of bixel intensities (see Baatar et al. (2005)):


3 6 4

2 1 5

If we represent an aperture by a binary matrix in which an element is equal to one if and

only if the associated bixel is exposed (i.e., not blocked by either the left or right leaf of

the MLC system), row-convexity corresponds to the property that, in each row of the

corresponding matrix, the elements that are equal to one are consecutive (often referred to

as the consecutive-ones 1y, '. i/;i). Now observe that this fluence map can be decomposed

into three apertures with corresponding intensities:

3 6 4 1 0 0 1 1 0 0 1 1
1x +2x +4x
2 1 5 0 1 1 1 0 0 0 0 1

Since, in general, there are many v- -- of decomposing a given fluence map into

row-convex apertures, it is desirable to select the decomposition that can be delivered

most efficiently. The two main efficiency criteria that pl li a role are the total beam-on-

time, i.e., the total amount of time that the patient is being irradiated, and the total

setup time, i.e., the total amount of time that is spent shaping the apertures. The former

metric is proportional to the sum of intensities used in the decomposition, while the

latter is approximately proportional to the number of matrices used in the decomposition.

Although closely related, these two efficiency criteria are not equivalent. The example









given above shows the unique decomposition using only three apertures and with a

beam-on-time of 7. However, the minimum beam-on-time for this fluence map is 6, which

can be realized by four apertures using the following decomposition:

3 6 4 1 1 00 1 0 111 0 11
1x +lx +2x +2x
2 1 5 1 0 0 1 1 1 0 0 1 0 0 1

The problem of decomposing a fluence map while minimizing beam-on-time is

polynomially solvable and has been widely studied, leading to several solution approaches

for this problem. Bortfeld et al. (1994) proposed the sweep method, which Almi and

Hamacher (2005) (who derived an equivalent method) showed to indeed yield an optimal

solution; other exact algorithms were proposed by Kamath et al. (2003), and Siochi

(1999). In addition, Baatar et al. (2005), Boland et al. (2004), Kalinowski (2005a),

Kamath et al. (2004a,b,c,d), Lenzen (2000), and Siochi (1999) studied the problem of

minimizing beam-on-time under additional hardware constraints, while Kalinowski (2005b)

studied the benefits of allowing rotation of the MLC head.

Although the time required by the MLC system to transition between apertures

formally depends on the apertures themselves, the fact that these times are similar and

the presence of significant (aperture-independent) verification and recording overhead

times justifies the use of the total number of setups (or, equivalently, the total number

of apertures) to measure the total setup time. In addition, delivering IMRT with a

small number of apertures provides the additional benefits of less wear-and-tear on the

collimators (less stopping and starting) and a less error-prone delivery as IMRT delivery

errors are known to be proportional to the number of apertures (see Stell et al. (2004)).

The problem of decomposing a fluence map into the minimum number of row-convex

apertures has been shown to be strongly NP-hard (see Baatar et al. (2005)), leading to the

development of a large number of heuristics for solving this problem. Notable examples

are the heuristics proposed by Baatar et al. (2005) (who also identify some polynomially









solvable special cases), Agazaryan and Solberg (2003), Dai and Zhu (2001), Que (1999),

Que et al. (2004), Siochi (1999, 2004, 2007), Van Santvoort and Heijmen (1996), Xia and

Verhey (1998). In addition, Engel (2005), Kalinowski (2005a), and Lim and Choi (2007)

developed heuristics to minimize the number of apertures while constraining the total

beam-on-time to be minimal. Finally, Langer et al. (2001) developed a mixed-integer

programming formulation of the problem, while Kalinowski (2004) proposed an exact

dynamic programming approach for the related problem of minimizing the number of

apertures that yields the minimum beam-on-time. Baatar et al. (2007) described integer

programming and constraint programming models for the same problem, and Ernst

et al. (2009) proposed a constraint programming approach for minimizing the number of

apertures. However, computational studies reported in Baatar et al. (2007); Ernst et al.

(2009); Kalinowski (2004); Langer et al. (2001) reveal that these approaches can only be

used to efficiently solve small problem instances to optimality. Our primary contribution

is that we develop the first algorithm capable of solving clinical problem instances to

optimality (or to provably near-optimality) within clinically acceptable computational time

limits.

In this chapter, our focus is on the problem of finding a decomposition of a fluence

map into row-convex apertures that minimizes total treatment time, as measured by

the sum of the total setup time and beam-on-time. In Section 3.2 we develop our

decomposition-based solution approach. In Section 3.3 we discuss the application of

our algorithm on a collection of clinical and randomly generated test data, and compare

its performance with alternative exact and heuristic techniques.

3.2 Decomposition Algorithm

Throughout this chapter, we denote the fluence map to be delivered by a matrix

B E -1' where the element at row i and column j, (i,j), corresponds to a bixel with

required intensity bi. Let w, be the time required by the MLC to form an aperture

and w2 denote the time required for the delivery of one unit of fluence. We refer to the









problem of minimizing the total treatment time, i.e., the sum of the aperture transition

times and the total delivery time, as the optimal leaf sequencing problem.

We start this section by describing a decomposition framework for the optimal leaf

sequencing problem in Section 3.2.1 and use this to formulate our master problem in

Section 3.2.2. We introduce our subproblem in Section 3.2.3, prove its complexity, and

provide a combinatorial search algorithm for its solution. We then enhance the empirical

performance of our decomposition algorithm by introducing classes of valid inequalities to

the master problem in Section 3.2.4, and finally describe an algorithm for constructing a

feasible solution with medically desired properties in Section 3.2.5.

3.2.1 Decomposition Framework

To establish motivation for our approach, observe that if the objective is to minimize

beam-on-time, the optimal leaf sequencing problem is decomposable by the rows of

the fluence map. In particular, if the beam-on-time is minimized for each bixel row,

the maximum of the corresponding beam-on-time values is equal to the minimum

beam-on-time for the overall fluence map (see, e.g., Ehrgott et al. (2008)). However, this

approach is not directly applicable when the objective is to minimize the total treatment

time.

Even though the optimal leaf sequencing problem is not directly decomposable by

rows, the fact that leaves corresponding to different rows can be positioned independently

can still be exploited. Denote a particular positioning of left and right leaves for a row as

a leaf position; an aperture is composed of a leaf position for each row of B. Our main

observation is that given a collection of intensities, which can be used in apertures that

collectively cover the fluence map, the rows are independent of one another. That is, we

can determine the leaf positions to be used for covering each row independently, and then

form apertures for covering the entire fluence map by combining individual leaf positions

for each row that are assigned to the same intensity.









We define an allowable I':/. i,-/l multiset to be a collection of (potentially non-unique)

intensity values, each of which can be assigned to a single aperture in our solution. We

.iv that an allowable intensity multiset is compatible with a row if there exists a feasible

decomposition of the row into leaf positions using a subset of that allowable intensity

multiset. If an allowable intensity multiset is compatible with all rows, then it corresponds

to a feasible decomposition of the fluence map and we call it a feasible I'/. ui-.i, multiset.

As an example, consider the fluence map given by the following 3 x 3 matrix:

148
B= 3 8 5 (31)

453
4 5 3

Consider the allowable intensity multiset {1, 3, 5}. Assigning each of these values to at

most one leaf position, the first row can be decomposed as


[1 4 8] x [1 1 0] + 3 x [0 1 1] + 5 x [0 0 1], (3-2)


so that the allowable intensity multiset is compatible with the first row. Similarly, the

second row can be decomposed as


[3 8 5] 3 x [1 1 0] + 5 x [0 1 1]. (3-3)

However, the first bixel in the third row must be covered by two leaf positions assigned

to intensities 1 and 3, and the second bixel must be covered by a single leaf position

assigned to intensity 5. Therefore, all allowable intensities must be used to cover the first

two bixels, and the third bixel with required intensity 3 cannot be covered. Hence, the

allowable intensity multiset is not compatible with the third row. Alternatively, consider

an allowable intensity multiset that contains the values 1, 3, and 4 for the same fluence

map. The rows can be decomposed as


[1 4 8] x [1 1 1] + 3 x [0 1 1] + 4 x [0 0 1] ,










[3 8 5] 1 x [0 1 1] + 3 x [1 1 0] + 4 x [0 1 1] and (3-4)

[4 5 3] x [0 1 0] + 3 x [0 0 1] + 4 x [1 1 0].


Since the allowable intensity multiset is compatible with all rows, it is a feasible intensity

multiset having three leaf positions and a beam-on-time of 8. Furthermore, observe that

the intensity requirements of the bixels in the first row strictly increase from left to right,

implying that a leaf position must start at each bixel. Thus, any feasible decomposition

of the first row uses at least three leaf positions, which yields a lower bound on the

number of apertures. Also, the largest element of B is 8, which yields a lower bound on

the beam-on-time. Since the given decomposition achieves the lower bounds on both

objectives, we have an optimal solution to the optimal leaf sequencing problem.

3.2.2 Master Problem Formulation and Solution Approach

We represent an allowable intensity multiset by an integer vector x = (x,..., XL),

where L = maxi= 1... m;j=1,...,nb bij is the maximum intensity value in the fluence map, and

where xr is the number of times that intensity value E occurs in the allowable intensity

multiset. It is easy to see that, assuming all allowable intensity values are used, the

number of apertures and the beam-on-time are, respectively, equal to

L L
xi and EUx. (3-5)
e=1 e= 1

The master problem can therefore succinctly be written as

L L
minimize wi E + w2 (36)
e=1 e= 1

subject to


x is compatible with row i V i 1,...,m (3-7)

r integer V 1,... ,L. (3-8)










Clearly, our model contains the problem of minimizing the number of apertures as a

special case by setting wl = 1 and w2 = 0. Moreover, if we wish to minimize the number of

apertures required while limiting the beam-on-time to no more than T, we simply add the

following constraint to the model:
L
fx, e=1
where of course T cannot be less than the minimum achievable beam-on-time z (which can

be found in polynomial time using the algorithms mentioned in Section 3.1).

To formulate our master problem as an integer programming problem, we introduce

binary variables ye,, V 1,..., L, r = 1,..., Re, where ye = 1 if and only if xa = r, and

Re is an upper bound on the number of apertures having intensity used in an optimal

solution. (We can compute Re by computing an initial upper bound on the optimal

objective function value via any of the heuristics mentioned in Section 3.1, and then

setting Rf to the largest value such that wAlR + W2Re is no more than this bound.) Using

these decision variables, we can reformulate the master problem (M!lP) as follows:

L L
minimize w, xy + w2 x (3-10)
e=1 e= 1

subject to

RE
Y.ry Vxf -1,...,L (3-11)
r=i
RE
Ye < 1 V f 1,...,L (312)
r=i
x is compatible with row i V i 1,..., m (3-13)

X' integer V 1,...,L (3-14)

yNr binary V = 1,...,L, r 1,...,R. (3-15)


We next formulate (3-13) as a set of linear inequalities by deriving valid inequalities that

cut off precisely those vectors x that violate (3-13). To this end, consider a particular










allowable intensity multiset represented by i that is incompatible with at least one row.

It is then clear that we should only consider vectors x that are different from 5 in at least

one component. We can achieve this by imposing the following constraint:

L Re
yi > 1. (3-16)
=1 r=l 1

Since all integer solutions except for i satisfy (3-16), it is indeed a valid inequality.

Constraint (3-16) can be tightened by observing that if the solution i is incompatible

with row i, then any solution x such that xL < Vx, V 1,..., L, is also incompatible with

row i. Therefore, we require that x contain at least one component that is larger than its

corresponding component in i, which yields the stronger valid inequality

L Re
y E r>l (3-17)
=1 r= e+l

Constraint (3-17) can, in turn, be tightened further by explicitly considering the rows

for which x is incompatible. Let Li = maxj, 1,.., bij be the maximum intensity in

the fluence map for row i. By the same argument as above, if the current solution i is

incompatible with row i, then any solution x such that xL < VX, V = 1,..., Li, is also

incompatible with row i, since no leaf positions with intensity greater than Li can be

used in decomposing row i. Therefore, we require that x is larger than i in at least one

component 1,... Li:

Li Re
SZ y&r > 1 V rows i incompatible with k. (3-18)
=1 r= e+1

Since (3-18) is stronger than (3-16) or (3-17), we use the latter inequalities in our model.

Note also that (3-18) stated for row i1 dominates a cut generated for row i2 if Li, < Li,.

Thus, we consider the bixel rows in nondecreasing order of their Li-values, halt when an

infeasible row is detected, and add a single inequality of the form (3-18). This sequence

also tends to minimize subproblem execution time, since rows having a small maximum









intensity are easier to solve by the nature of the backtracking algorithm discussed in

Section 3.2.3.

Since the collection (3-18) contains an exponential number of valid inequalities,

we add them only as needed in a cutting plane fashion. In particular, this means that

we relax (3-18), solve the relaxation of (l IP) and generate an x-solution representing a

candidate allowable intensity multiset. We then solve a subproblem for each bixel row to

determine if the allowable intensity multiset is incompatible with that row. If not, we have

found an optimal solution to (\! I). Otherwise, we add a constraint of the form (3-18) to

(\! P) that cuts off that solution.

3.2.3 Subproblem Analysis and Solution Approach

In this section, we consider the subproblem of checking whether a given intensity

multiset x is compatible with a particular bixel row. For convenience and wherever the

interpretation is clear from the context, we suppress the index i of the bixel row and

denote a typical row of the fluence map B by b = (bi,..., b,).

We represent a feasible decomposition as a collection of n-dimensional binary vectors

v&r (e = 1,... L; r = 1,... r). The values of ve, that equal 1 correspond to the

(consecutive) exposed bixels in the rth aperture having intensity For example, the

decomposition in equation (3-2) corresponds to vi = (1, 0), v31 = (0, 1, 1), V51

(0, 0, 1), and ve = 0 for other r. (Note that this decomposition would be feasible as long

as xl, X3, Xs > 1.) The subproblem can then formally be presented as follows:

CI-PARTITION

INSTANCE: An n-dimensional vector of nonnegative integers b and an integer vector

x= (Xi,... ,XL).

QUESTION: Do there exist n-dimensional binary vectors v&r (e 1, ..., L; r

1,..., xr) that satisfy the consecutive-ones property such that C L1 1 1V r = b?


Proposition 3. CI-PARTITION is strongly NP-complete.









Proof. See Appendix B.


In principle, the CI-PARTITION problem can be formulated and solved as an integer

program. However, we have developed a computationally more effective backtracking

algorithm that focuses on partitioning intensity requirements individually for each bixel.

An integer vector p = (pJ,... ,p ) provides a bixel decomposition of bixel j E {1,..., n} in

row b if and only if bj = I1 p". We then attempt to form a collection of leaf positions

that realizes the individual bixel partitions. We call such a collection of leaf positions a

leaf decomposition of b.

To more effectively conduct our subproblem searches, we describe a property that

holds in some leaf decomposition (if one exists) that satisfies the given collection of bixel

decompositions.

Lemma 1. Consider candidate bixel decompositions for bixels j and j + 1, for some

j E {1,..., n 1}, and suppose that these have a common decomposition ;/,l.. '-./i value f,

i.e., ,p jp > 0. Then, if a leaf decomposition exists, one exists in which a leaf position

having ",'1. i, ii exposes both bixels j and j + 1.

Proof. Assume that there exists a leaf decomposition V in which bixels j and j + 1 are

exposed by two separate leaf positions, vi and v2, respectively, each having intensity f.

Now consider the leaf position V3 = V1+v2 having intensity f. Then V' = {v3}UV\{v, v2}

is also a leaf decomposition that realizes the given bixel decomposition. E

We next derive a necessary condition that any feasible bixel decomposition must

satisfy so that the corresponding set of leaf positions is compatible with a given allowable

intensity multiset x. Similar to the idea behind Lemma 1, if p > p+l, then p p+l leaf

positions having intensity f must expose bixel j but not j + 1. Lemma 2 formalizes this

idea.

Lemma 2. Let x represent an allowable ';/,.I. ,'li multiset, and pJ7' denote candidate bixel

decompositions for bixels j,, Vr 1, ..., n' such that 1 < ji < .. < jn, < n. The following









set of conditions must be -/.:-/7. in I ,; feasible solution.


n/
Z max{Op"-p + p"

Proof. If p& > p", at least p"- p leaf positions having intensity must expose bixel

Ji- but not j,. Also, at least pj"' leaf positions having intensity must expose bixel j,'.

Since all leaf positions listed above are necessarily disjoint, the lemma holds. E

We next describe our backtracking algorithm. In this algorithm, we first enumerate

all possible v--i- of decomposing the bixel intensities in b using a subset of the allowable

intensity multiset given by x. We denote the set of all candidate bixel decompositions for

bixel j by Pj, where for each p E Uj I'j, we must have pe < xz, V = 1,..., L.

The backtracking algorithm for solving the subproblem is stated formally in

Algorithm 1. We begin by enumerating each possible element of Pj, V j = 1,..., n.

We denote the set of processed bixels by F (for which a candidate Il, bixel

decomposition has been established), and the set of unprocessed bixels by R. In each

iteration, we check to see if the set of candidate bixel decompositions Pj for any j E R

is empty. If so, the current active bixel decompositions do not yield a feasible solution,

and the algorithm backtracks. Otherwise, we consider an unprocessed bixel j E 7, and

choose an untried bixel decomposition p' E CPy to be active for bixel j. Next, we move j

from R to F, creating updated sets R' and F', and invoke Lemma 2 to update the set of

bixel decompositions for the bixels in R'. Specifically, for each j E R' and pJ E 'j, we

calculate the number of leaf positions that would be required due to selecting pJ as the

active bixel decomposition for bixel j, in addition to those already selected for bixels in

F'. We eliminate pJ if a condition of type (3-19) is violated. We then recursively call the

procedure to continue with a new bixel j' E R'.

We stop either when we find a feasible bixel decomposition for all bixels, or when we

exhaust all bixel decompositions without finding a feasible solution. In the former case, a

leaf decomposition that realizes the bixel decompositions for bixels j e {1,..., n} can be









found by invoking Algorithm 3, which is based on the repeated application of Lemma 1.

To see that Algorithm 3 recovers a feasible leaf decomposition, note that Algorithms 1 and

2 provide bixel decompositions that satisfy Lemma 2, and in particular, the condition


max{0,pF 1 + j=2

Algorithm 3 recovers a feasible leaf decomposition if, in the outer while-loop corresponding

to each 1,..., L, the counter r is never incremented more than xr times. Note that

r is incremented each time the inner while-loop terminates, which occurs either when

J > n (a total of p7 times), or when p = 0 (p-1 pj times) for = 2,..., n. The total

number of times that r is incremented in the outer while-loop for 1,..., L is thus the

left-hand-side of (3-20), which is no more than xe, as required.

If we exhaust all bixel decompositions without finding a feasible solution, we conclude

that the current allowable intensity multiset is incompatible with the current row.

Algorithm 1 C1-PARTITION(b, x)
Input: b {n-dimensional vector representing bixel intensity requirements}
Input: x {L-dimensional vector representing an allowable intensity multiset}
{This algorithm finds whether there exists a CI-PARTITION of b compatible with x}
S<-- 0 {F is the set of processed bixels}
R <-- {1,..., n} {7R is the set of unprocessed bixels}
for all j E {1,...,n} do
Pj -- Enumerate all bixel decompositions compatible with x for bixel j
<-- {P ,.... ,P }
return C1-PARTITIONRECURSIVE(b, x, F, 7, P)


Since Algorithm 1 is a backtracking algorithm, and therefore in the worst case

investigates all possible bixel decompositions, it is of exponential time complexity (as

expected, due to Proposition 3). However, the empirical running time of the algorithm can

be reduced using the following observations:

(i) If two .,.li i.:ent bixels in a row have the same required intensity value, there must
exist an optimal solution in which they are exposed by the same leaf positions.
This result can be proven in a similar way as Lemma 1, and is therefore omitted
for brevity. This observation implies that we can preprocess the data by merging











Algorithm 2 C1-PARTITIONRECURSIVE(b, x, F, 7, 7P)
if 7R 0 then
return true {all bixels have been processed, P represents a feasible solution}
else
if 3j ER : Pj = 0 then
return false {there is no remaining way of decomposing bixel j}
else
j argminj,-'Pj {j is a bixel having the smallest number of bixel decompositions}

for all p e c P do
p7' '- p, 'P <-- {pJ} {p is now the active decomposition for bixel j}
F' Fu { j}, R' R\ {j}
for all j R' do
-- Pj\ {all elements eliminated by Lemma 2, given the active decompositions
p3 for jE C '}
if C1-PARTITIONRECURSIVE(b, x, F', R', 7') then
return true {a feasible solution that uses pi to decompose bixel j is found}
return false {all bixel decompositions of bixel j have been exhausted}





Algorithm 3 RECOVERLEAFDECOMPOSITION(b, x, P)
Require: Pj {pJ} Vj e {1,..., n} {all bixels have been processed}
Output: vr ( = 1,..., L; r = 1,... ,xf)
{ve, is an n-dimensional binary vector that represents a leaf position}
for all c {1,..., L}, r {1,..., x} do
vr <-- 0
for all f {1,..., L} do
r <-- 1, j <-- 1
while j < n do
if pi > 0 then
j <-- j {a new leaf position must start at bixel j}
while j < n and p' > 0 do
{expand the new leaf position as much as possible}
+ -- 1
-- pi 1,J J+1
r r+ 1
else
j -- j + 1 {all leaf positions that start at bixel j have been recovered}









all .,ili i, ent bixels in a bixel row having the same intensity requirement, thereby
reducing the dimensionality of the problem instance.

(ii) In choosing the next bixel to be processed, we pick a bixel j E R having the smallest
number of remaining candidate bixel decompositions. In this manner, we can quickly
enumerate all possible bixel decompositions for a few key bixels and eliminate a
significant portion of bixel decompositions for the remaining bixels without wasting
effort by unnecessary backtracking steps.

(iii) In choosing the next candidate bixel decomposition pJ E Pj for a chosen bixel
j E 7?, we select an untried bixel decomposition having the fewest number of
intensity values. Since each intensity value used in decomposing a bixel needs to
be assigned to a different aperture, this rule favors a bixel decomposition using the
fewest number of apertures to decompose the chosen bixel. Therefore, it tends to
retain the availability of more elements of the allowable intensity multiset (and hence
apertures) for the remaining bixels, making it easier to find a feasible solution (if one
exists).

3.2.4 Valid Inequalities for the Master Problem

The initial optimal solution to the relaxation of ( lP) in which none of the inequalities

(3-18) have yet been added to the model will set all variables equal to zero, which is

clearly incompatible with all rows. In this section, we derive some characteristics of all

feasible solutions and use these to define valid inequalities for (\! I). In this way, we

attempt to improve the convergence rate of the decomposition algorithm by eliminating

some clearly infeasible solutions before the initial execution of the master problem.

3.2.4.1 Beam-on-time and number of apertures inequalities

Our first observation uses and generalizes the fact that the beam-on-time, number of

apertures, and total treatment time required for the decomposition of any single row into

leaf positions provide lower bounds on the minimum beam-on-time, number of apertures,

and total treatment time, respectively, needed to deliver the entire fluence map. More

generally, consider any collection of nonnegative objective weights w' and w' in place of

w1 and w2, and let Ti(w', w') be the minimum value of the objective with respect to these

weights over all decompositions for row i only. Then the following are valid inequalities for









(\!P):
L L
w' x+w' x > T, w,w ) V i 1,...,m. (3-21)
e=1 e=1
We formulate an integer programming model to determine T (w w%) for a given row i.

First, denote the set of possible leaf positions for that row by /C, and define n-dimensional

binary vectors vk for k E /C (where |1C = O(n2)), such that = 1 if and only if bixel j

is exposed by leaf position k. In addition to decision variables xe as in ( \ P), define binary

decision variables zke, V k E /C, f 1,..., Lk such that zk = 1 if and only if leaf position

k is used with intensity (where Lk = min,: .v=1 b is an upper bound on the intensity of

leaf position k.) Then T (w', w') is the optimal objective function value of the following

optimization problem, (SR):

L L
minimize wu4 Y x' + w' x' (3-22)
e=1 e= 1

subject to


Svk i =z b Vj = 1,...,n (3-23)

Lk
5 Zke < 1 V k C (3-24)
e= 1
zkU V f 1,...,L (3-25)
kEIC:Lk>_
zke {0,1} V k cIC, = 1,...,Lk (3-26)

xe > 0 and integer V 1,...,L. (3-27)


Constraints (3-23) ensure that each bixel receives exactly its required amount of dose

while constraints (3-24) guarantee that each leaf position is either not used or is assigned

to a single intensity value. Finally, constraints (3-25) relate the x- and z-variables.

A practical difficulty in implementing the valid inequalities of the form (3-21) is

that we must determine appropriate values for the weights w' and w'. However, Baatar

(2005) shows that, when decomposing a single bixel row, there exists a set of leaf positions










that simultaneously minimizes both beam-on-time and the number of apertures. If we let

Ni Ti(l, 0) represent the minimum number of apertures for row i, and i = Ti(0, 1)

represent the minimum beam-on-time for row i, this implies that Ti(w, w') = wNi+ + w'i,

so that we can replace (3-21) by

L L
w xe+w' x>wwNi + ,,'_ V 1, ...,m. (328)
1 1

It is easy to see that we can capture all of these valid inequalities by restricting ourselves

to the coefficient pairs (w, w) = (1, 0) and (0, 1) only:

L
x > max {NJ} (3-29)
=1
L
x, > max {Js}. (3-30)
=1

We can generalize this idea as follows. Let R() denote the set of rows for which

the maximum intensity requirement is bounded by L for some L E {1,..., L}, i.e.,

R() = {i {1,... ,m} : Li < }. Since intensity values greater than L cannot be used in

decomposing the rows in R(), a similar approach to the one above can be used to derive

the following family of valid inequalities


xe > max NJ} V 1,..., L (3-31)
1iER(L)
e 1

> x > max {zJ} V 1, ..., L. (3-32)
1iER(L)
= 1

Finally, note that the values of Ni and zi can be found by solving (SR) with w' = 1, w

1 or by using the method of Kalinowski (2004), since there exists a solution that minimizes

both beam-on-time and the number of apertures (Baatar, 2005).

3.2.4.2 Bixel subsequence inequalities

Recall that (3 16) (3 18) represent necessary conditions for feasibility of an allowable

intensity multiset with respect to a particular row. It is possible to develop stronger









necessary conditions if we examine subsequences of a row, i.e., a subset of the required

intensity values in a row that preserves their order in the fluence map. First, Lemma 3

shows that, if a given allowable intensity multiset is incompatible with a subsequence s of

row i, then it also must be incompatible with row i.

Lemma 3. Consider an allowable <;I/. <,;iu- multiset x, an n-dimensional vector b that

represents the :<,l. -,i1/; requirements of the bixels in some row of B, and an n'-dimensional

vector s = (bl,... ba,) where 1 < jl < i < jn < n. If x is not compatible with s, then

it is also not compatible with b.

Proof. We prove the equivalent statement that if x is compatible with b, then it is also

compatible with s. Assume that x is compatible with b. By definition, there exists a bixel

decomposition for each bixel j 1,..., n so that the resulting set of leaf positions is

compatible with x. The bixel decompositions corresponding to only the bixels in s are also

compatible with x, since the order of the bixels in s is the same as that in b. O

Note that we can invoke Lemma 3 to associate a subproblem with each of the 0(2')

subsequences of a bixel row b. Each of these subproblems can then be used to generate

cutting planes of the form (3-18), as well as valid inequalities of the form (3-31) and

(3-32). However, since the strength of (3-18), (3-31) and (3-32) depend on the largest

intensity value in a bixel row, we form subsequences of each bixel row by, for L 1,... L,

considering only those bixels having required intensity less than or equal to L. The valid

inequalities generated by the O(min(n, L)) subsequences generated in this fashion imply

all 0(2") valid inequalities associated with all possible subsequences.

3.2.5 Constructing a Feasible Matrix Decomposition

Our algorithm finds an optimal allowable intensity multiset and a bixel decomposition

for each bixel row. To construct a corresponding matrix decomposition, we need to apply

Algorithm 3 to find a leaf decomposition for each row. We can then generate aperture

matrices by arbitrarily combining leaf positions using the same intensity values in different









rows. We have found empirically that this simple approach yields a feasible matrix

decomposition very quickly.

Since any pair of leaf positions assigned to the same intensity value in different rows

can be combined, there are up to (nH ,(X!) aperture matrices that can be constructed

from a given feasible leaf decomposition for each row. Even though each such choice

represents an alternative optimal solution to the optimal leaf sequencing problem, some

matrix decompositions may clinically be preferable to others based on their structural

properties. Perhaps the most challenging structural consideration pertains to the so-called

"tongue-and-giuui, effect observed in MLCs. We refer the reader to the works of Deng

et al. (2001) and Que et al. (2004) for technical details of the tongue-and-groove effect in

dynamic MLC dose delivery. For the purposes of this study, it is sufficient to understand

that leaves in .,.i ,i:ent rows often interlock with a tongue on the bottom of one row sliding

along a groove in the top of another row. Tongue-and-groove underdosage occurs since a

leaf's tongue blocks dosage intended for cells beneath it. Therefore, it is desirable to limit

such underdosages.

To measure the amount of tongue-and-groove effect in a treatment plan, Que et al.

(2004) note that it is generally not desirable to deliver one aperture in which some bixel

(i,j) is blocked by a leaf while bixel (i + 1,j) is not blocked, if another aperture is

being delivered where (i,j) is not blocked by a leaf while (i + 1,j) is blocked. Based on

this observation, Que et al. (2004) derive the following tongue-ai,:l-,roove index (TGI).

Suppose a treatment plan consists of K apertures described by binary values vjk, where

v =k 0 if cell (i,j) is blocked by a leaf in aperture k and vk = 1 otherwise, for each

i = 1,..., j = 1,..., n, k = ,..., K. Let Ik be the intensity delivered in aperture

k = 1,..., K. Then the TGI of a matrix decomposition is defined as:

m-1 n K-1 K
min{Ik, I ik (t V+1,k) t )U i+1,
i 1 j=1 k=1 =k+1

+ (1 Vk) (kvj -1 -+) (3-33)
i v (t









We thus can calculate the TGI component induced by rows 1 and 2 (of all aperture

pairs), then rows 2 and 3, and so on, down to rows m 1 and m. This observation allows

us to focus on pairs of rows instead of pairs of entire aperture matrices while reducing

TGI, allowing us to design an efficient algorithm for TGI reduction given a set of bixel

decompositions for each row.

Given a pair of .,l1i ient rows, we attempt to match individual leaf positions in the

two rows to minimize the TGI induced by the .,li ,i:ent row pair. To limit computational

overhead in this phase of our algorithm, we reduce TGI indirectly by the following scheme.

Let us denote a leaf position for row i by a binary n-vector v', where vj = 1 if the leaf

position exposes bixel j in row i. We measure the overlap between two leaf positions

having the same intensity value in consecutive rows by counting the number of columns

that both leaf positions expose simultaneously. Formally, we define the overlap between

leaf positions vi and vi+1 as 0(vi, vi+1) = E, vjv1'. Our approach is to heuristically

minimize TGI by maximizing the total overlap between all leaf position pairs, which can

efficiently be solved as an assignment problem. The efficiency of the assignment problems

can be further improved by noting that the problem decomposes over the intensity values

Se {1,..., L}, since only leaf positions having the same intensity value can be combined.

Therefore, we can generate a matrix decomposition by finding a leaf decomposition for

each row, and then matching leaf positions in .,l-i i,:ent rows having the same intensity

value by solving an assignment problem so that the total overlap is maximized.

The TGI minimization step described in the previous paragraph can be improved as

follows. Typically, multiple bixel decompositions exist for each row that are compatible

with a given feasible intensity multiset. Algorithm 2 can be modified in a straightforward

manner so that it finds all leaf decompositions of a row, instead of stopping once the first

feasible bixel decomposition for all bixels is found. Since different bixel decompositions

for a bixel row correspond to different leaf decompositions, considering alternative bixel

decompositions can lead to a matrix decomposition having a smaller TGI.









Given alternative leaf decompositions for each row, the problem of minimizing TGI

can be formulated as a shortest path problem as follows. We create al 1,. I network in

which each l? -r corresponds to a bixel row i E {1,..., m}, and node Nid represents the

dth leaf decomposition of row i. We add a directed arc from each node Nid to all nodes

N(1i+l)d, for all i = ,... m 1. The cost of the arc from node Nid to N(i+l)dl is given by

the TGI value resulting from the assignment solution corresponding to the candidate leaf

decompositions represented by d for row i, and d' for row i + 1. Finally, we add a start

node S and a finish node F. We create zero-cost arcs from S to all nodes in the first 1~.l-r,

and from all nodes in the last l?--r to F. A shortest S-F path in this graph represents a

matrix decomposition having a minimum TGI from among the provided options. Since the

graph is .,. i-, the shortest path problem can be solved in O(|A|) time, where A is the

set of all arcs.

Remark 5. The shortest path approach to i,,.,:.:i,, .t:, TGI can be difficult to solve

!.':. 1./ when bixel rows have a ,,',, number of alternative leaf decompositions, since an

arc joins each pair of nodes corresponding to adjacent bixel rows. To y., ;/.:ill,/ overcome

this difficulty, we limit the number of bixel decompositions found by Algorithm 2 by

terminating once 250 feasible bixel decompositions have been .,J. ,.,1/; Next, note that a

straightforward '. ;/. /.:'. shortest path implementation processes 7.';,.. one at a time, and

does not generate a feasible S-F path before processing the last 17;, ,. Since being able to

I'... :fy a time limit is a desired feature in a practical setting, we use a i/;,l, .:i .r'i' .:hm for

solving the shortest path problem. Our i1'.,i,.:hm starts by processing 7.;1,.. one-by-one,

i.,l.':,.,j node labels as usual. If a shortest path is not found when a given initial time

limit expires, our il'.>rithm switches to a depth-first-search (DFS) procedure, which we

terminate after a given final time limit. We start DFS from an unprocessed node Nid

having a smallest label, select a minimum-cost arc (Nid, N(i+l)d,) exiting that node, and

update the label of N(i+l)di if we have found a new shortest S-N(i+l)d, path. Else, the










i1,.-' ,:thm backtracks and seeks another arc from Nid. We then return the shortest S-F

path found by this procedure when the final time limit is reached.

3.3 Computational Results and Comparisons

3.3.1 Problem Instances

In our experiments we have used two classes of problem instances. Our base set of

test problem instances consists of 25 clinical problem instances that were obtained from

treatment plans for five head-and-neck cancer patients treated using five beam angles

each. Table 3-1 reports the problem characteristics for these problem instances in terms

of the matrix dimensions m and n. The maximum intensity value is L = 20 for all these

instances. In addition, to allow comparison of our results with published results on other

approaches to the problem, we generated 100 random problem instances of dimensions

20 x 20 having maximum intensity value L = 10.

However, since these problem instances are generally too large to be solvable by the

integer programming model from Langer et al. (2001) and its modification described in

Appendix A, we also randomly generated eight instances ( i. -I -.:;. ", ..., "test6x7b")

to demonstrate the computational limitations of the latter approaches. Unless otherwise

specified, we used wl = 7 and w2 = 1 as the objective weights for the number of apertures

and beam-on-time, respectively.

Table 3-1. Dimensions of clinical problem instances
Name m n Name m n Name m n Name m n Name m n
clbl 15 14 c2bl 18 20 c3bl 22 17 c4bl 19 22 c5bl 15 16
clb2 11 15 c2b2 17 19 c3b2 15 19 c4b2 13 24 c5b2 13 17
clb3 15 15 c2b3 18 18 c3b3 20 17 c4b3 18 23 c5b3 14 16
clb4 15 15 c2b4 18 18 c3b4 19 17 c4b4 17 23 c5b4 14 16
clb5 11 15 c2b5 17 18 c3b5 15 19 c4b5 12 24 c5b5 12 17


3.3.2 Implementation Issues

We have implemented our decomposition algorithm using CPLEX 11.0 running on

a Windows XP PC with a 3.4 GHz CPU and 2 GB RAM. We use callback functions of

CPLEX to generate a single branch-and-bound tree in which we solve the subproblems

corresponding to each integer solution found in the tree, and add cuts to tighten the









master problem as necessary. This implementation turned out to be consistently faster

than one which re-solves the master problem each time a cutting plane is added to the

model. Furthermore, in our base algorithm, we use the subsequence inequalities (3-31)

and (3-32) described in Section 3.2.4.2. We also use Engel's heuristic (Engel, 2005), which

executes in well under one CPU second for each instance and generates a solution having

minimum beam-on-time, to (i) obtain an initial upper bound and (ii) compute the upper

bounds Re (f 1,..., L).

3.3.3 Comparison with Langer et al. (2001) Model

Our first experiment compares our base algorithm that minimizes the total treatment

time to that of Langer et al. (2001) and to the modification of their model as described

in Appendix A. We choose randomly generated test instances of various dimensions

to identify the problem sizes that can be solved by each algorithm, as well as four of

the smallest clinical instances to compare the effectiveness of the algorithms on clinical

instances. We imposed a one-hour time limit past which we halted the execution of an

algorithm. For these experiments we disabled the use of Engel's heuristic as an initial

heuristic to test the ability of these models to efficiently find good-quality upper bounds.

Table 3-2 summarizes the results of these three algorithms in terms of the execution

time, the best upper and lower bounds found within the time limit, and the optimality

gap (calculated as the difference between the upper and lower bound as a percentage of

the upper bound). Our decomposition algorithm can solve all 15 instances in this data

set within a few seconds, whereas only six instances can be solved to optimality within

an hour by either integer programming formulation. We conclude that, even though the

integer programming formulation given in (Langer et al., 2001) can solve small instances

to optimality, it cannot be used to solve clinical problem instances to optimality within

practical computation time limits.










Table 3-2. Comparison of our base algorithm with Langer et al. (2001) model
Two stage Langer Modified Langer
Name m n L CPU Optimal CPU UB LB Gap CPU UB LB Gap
test3x3 3 3 8 0.1 29 1.6 29 29.00 0.0% 0.9 29 29.00 0.0%
test3x4 3 4 8 0.1 37 5.2 37 37.00 0.0% 1.6 37 37.00 0.0%
test4x4 4 4 8 0.1 36 30.4 36 36.00 0.0% 10.7 36 36.00 0.0%
test5x5a 5 5 10 0.2 45 2069.6 45 45.00 0.0% 86.4 45 45.00 0.0%
test5x5b 5 5 15 0.2 50 198.2 50 50.00 0.0% 92.6 50 50.00 0.0%
test5x6a 5 6 10 0.2 55 3600 61 33.53 45.0% 3600 55 40.95 25.5%
test5x6b 5 6 18 0.4 71 3600 84 51.58 38.6% 3600 77 58.67 23.8%
test6x6a 6 6 13 0.3 55 3600 55 45.63 17.0% 3600 55 48.00 12.7%
test6x6b 6 6 13 0.3 52 3600 57 43.82 23.1% 3600 57 50.00 12.3%
test6x7a 7 6 10 0.2 45 690.0 45 45.00 0.0% 435.1 45 45.00 0.0%
test6x7b 6 7 15 0.4 74 3600 94 35.69 62.0% 3600 80 47.88 40.1%
clbl 15 14 20 1.3 111 3600 336 48.58 85.5% 3600 273 42.00 84.6%
clb2 11 15 20 0.8 104 3600 280 38.26 86.3% 3600 132 39.55 70.0%
clb5 11 15 20 3.1 104 3600 280 46.20 83.5% 3600 140 49.29 64.8%
c5b4 14 16 20 2.5 124 3600 360 34.00 90.6% 3600 360 39.11 89.1%


3.3.4 Random Problem Instances

For our next experiment, we first solved each of the 20 x 20 random problem

instances in our data set to optimality for the problems of (i) minimizing total treatment

time ("Total Time"), (ii) minimizing the number of apertures while constraining the

beam-on-time to be minimal ("Lexicographic"), and (iii) minimizing the number of

apertures ("# Ap. i I i. ). We also implemented three heuristic algorithms proposed

by Siochi (2007), Engel (2005), and Xia and Verhey (1998), which we executed on the

same data set. (The results we present from Siochi (2007) refer to the Variable Depth

Recursion (VDR) algorithm without tongue-and-groove constraints, using the parameters

recommended in the paper. We discuss the effect of including tongue-and-groove

considerations in the algorithm below.)

Figure 3-2 summarizes the total treatment times associated with the solutions

generated by the six algorithms we tested. Each algorithm is represented by a curve

that depicts quality of the solutions obtained by the corresponding algorithm. For each

value T of total treatment time on the horizontal axis, each curve plots the number of

problem instances for which the corresponding algorithm was able to find a solution

having total treatment time no more than T. For instance, Figure 3-2 shows that Siochi's

heuristic found a solution with a total treatment time of at most 175 time units in 5'

of the problem instances, while an optimal solution (represented by "Total Time") has
























C-)
0
I 50-
LL0

40-


30-


20 / p Xia-Verhey
-/ Siochi-VDR
Engel
10- ----# Apertures
SLexicographic
STotal Time
0
150 200 250 300
Total Treatment Time

Figure 3-2. Comparison of total treatment times on random data


the same quality level in 97'. of the problem instances. We observe that all three exact

algorithms find solutions having similar treatment times. Solution qualities generated

by the Engel and Siochi heuristics are similar, with the Siochi heuristic being slightly

better. A comparison of the heuristic solutions with optimal solutions reveals that average

optimality gaps for Siochi, Engel and Xia-Verhey heuristics are 10.1 12.0' and 51.5' ,

respectively.

Figure 3-3 compares the algorithms with respect to the number of apertures used

in their respective solutions. We note that our algorithm that minimizes total treatment

time ("Total Time") finds a solution that also minimizes the number of apertures for most

problem instances. As expected, lexicographic minimization of the two objective functions
























C-)
= 50-
L0

40-


30-


20 Xia-Verhey
/ --Siochi-VDR
/ -- Engel
10- // # Apertures
Lexicographic
e -- Total Time
0-
14 16 18 20 22 24 26 28 30 32
Number of Apertures

Figure 3-3. Comparison of the number of apertures on random data


results in an increased number of apertures. For this objective the "# Ap. i Ii i -

algorithm finds optimal solutions. Average optimality gaps for the heuristics of Siochi,

Engel, and Xia-\V-rl- v are 15.1,'. 18.9' and 62.;:', respectively.

We analyze the beam-on-time values of the solutions generated by each algorithm in

Figure 3-4. Since both Engel's heuristic and our "Lexicographic" algorithm find optimal

solutions having minimum beam-on-time, their curves overlap. We observe that the

Siochi heuristic and our "Total Time" algorithm tend to generate solutions having small

beam-on-time values, but the solutions generated by our "# Ap. iIi algorithm, and

by the Xia-Verhey heuristic have higher beam-on-time values. We calculated the average

optimality gaps for the latter two algorithms as 12.1,'. and 32.1 respectively.

























C--
0
50-
LL
40-


30-


20 Xia-Verhey
-- Siochi-VDR
-- Engel
10 / # Apertures
Lexicographic
STotal Time
0- r r r T I -
45 50 55 60 65 70 75 80 85
Beam-on-Time


Figure 3-4. Comparison of beam-on-time values on random data


The final measure of solution quality that we consider is TGI, which is a measure

of the tongue-and-groove effect given by (3 33). Figure 3-5 reveals that the solutions

obtained by all three variants of our decomposition algorithm have significantly lower

TGI values than the heuristic procedures. This result implies that, even though our

TGI-reduction algorithm described in Section 3.2.5 does not guarantee a minimum TGI,

it is highly effective in finding solutions with TGI values superior to the other heuristic

approaches. To estimate optimality gaps for the heuristics we compare heuristic solutions

with the solutions generated by our "Lexicographic" algorithm, which provides the best

TGI among all methods mentioned above. We note that average gaps for Siochi, Engel

and Xia-Verhey heuristics are 162.1 164. !' and 205.!' respectively. We also note













90


80


70


60
0
50


40


30


20 Xia-Verhey
-- Siochi-VDR
7Engel
10 # Apertures
Lexicographic
STotal Time

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400
Tongue-and-Groove Index (TGI)

Figure 3-5. Comparison of TGI values on random data


that these heuristics do not attempt to minimize TGI, and it might be possible to modify

them to obtain solutions with lower TGI values. It is interesting to note that a variant of

Siochi's algorithm (Siochi, 2007) is capable of completely eliminating TGI at the expense

of creating additional apertures. This variant is reported to increase the number of

apertures by 10('. to 3 :' i relative to the variant that does not remove TGI (Siochi, 2007).

Finally, the Engel and Xia-Verhey heuristics took less than one second of CPU time

in all instances we tested. The average CPU time for Siochi's heuristic, "Total Time"

algorithm, "# Ap. ii wi. algorithm, and "Lexicographic" algorithm were 31.5, 963.1,

414.8, and 421.4 seconds, respectively. We note that all variants of our two-stage algorithm

showed a "heavy-tail" behavior, where about 1I' '. of the problem instances were solved to









optimality in less than the average solution time. For instance, using the "# Ap. 11 i -

algorithm, we were able to solve 40 instances within one minute, 58 within two minutes, 81

within 414.8 seconds (the average solution time for this algorithm), 90 within 15 minutes,

and all but three instances were solved within an hour. The remaining three instances

were solved within three hours.

3.3.5 Clinical Problem Instances

Recall from Section 3.1 that in clinical practice, we can deliver each fluence map

using a decomposition into either row-convex or column-convex apertures, where the latter

requires rotation of the MLC head. Our final set of experiments compares the algorithms

on clinical problem instances in our data set, allowing for MLC head rotation.

We first show the results of applying our decomposition algorithm to decompose each

of the 25 clinical fluence maps into row-convex apertures, and column-convex apertures,

where the latter is achieved by applying our algorithm to the transpose of each fluence

map. Table 3-3 reports the performance of our algorithm when the objective function is

set to minimize total treatment time, and di-p'1 v the number of apertures ("nAper"),

beam-on-time ("BOT"), total treatment time ("Time"), tongue-and-groove index ("TGI"),

and CPU time used ("CPU") for the algorithm.

Our algorithm finds an optimal solution to several instances within a few seconds

while four instances take more than 10 minutes of CPU time to be solved to optimality.

Comparing the solutions obtained for row-convex and column-convex decompositions, we

observe that rotating the MLC head is most beneficial (in terms of treatment time) for

instances in which the number of rows is much smaller than the number of columns. These

benefits are most apparent on instances c4b2 and c4b5, where rotating the MLC head

can result in more than 50'. reduction in total treatment time. We also note that several

problem instances require much less computational time to solve for a column-convex

decomposition compared to a row-convex decomposition.















Table 3-3. Effect of rotating the MLC head
Row-Convex Column-Convex
Name nAper BOT Time TGI CPU nAper BOT Time TGI CPU
clbl 10 41 111 102 1.1 11 38 115 50 5.5
clb2 10 34 104 80 0.8 8 23 79 14 0.7
clb3 11 31 108 97 11.4 9 28 91 59 1.0
clb4 11 33 110 74 37.0 11 37 114 146 7.0
clb5 10 34 104 133 4.3 8 32 88 49 1.2
c2bl 14 34 132 134 26.5 12 30 114 187 11.5
c2b2 13 41 132 159 20.1 11 33 110 192 8.0
c2b3 13 49 140 245 14.7 11 28 105 151 3.1
c2b4 14 51 149 316 87.3 12 34 118 148 8.3
c2b5 13 41 132 217 395.6 10 27 97 120 2.0
c3bl 13 41 132 323 310.0 14 40 138 254 23.0
c3b2 14 46 144 320 4759.8 8 23 79 86 1.1
c3b3 13 49 140 533 10373.9 12 40 124 360 18.6
c3b4 12 44 128 481 524.9 12 40 124 327 428.2
c3b5 13 34 125 133 3.3 9 27 90 75 2.6
c4bl 16 40 152 216 34.9 12 46 130 244 10.6
c4b2 16 69 181 450 20901.0 9 27 90 149 15.8
c4b3 14 41 139 130 44.7 10 32 102 129 3.3
c4b4 14 44 142 246 164.3 10 27 97 163 8.0
c4b5 17 76 195 470 14511.4 9 24 87 48 4.0
c5bl 10 26 96 68 0.5 10 35 105 41 0.5
c5b2 12 41 125 59 14.3 8 25 81 27 0.6
c5b3 10 34 104 155 3.1 9 23 86 42 1.0
c5b4 12 40 124 105 2.2 10 32 102 87 4.3
c5b5 12 46 130 151 51.9 8 31 87 17 0.8


Table 3-4. Computational results for our base algorithm
Total Time # Apertures Lexicographic
Name nAper BOT TGI CPU nAper BOT TGI CPU nAper BOT TGI CPU
clbl 10 41 102 4.7 10 41 102 2.3 11 38 50 4.8
clb2 8 23 14 1.1 8 23 14 1.1 8 23 14 1.1
clb3 9 28 59 3.0 9 28 59 4.5 9 28 59 4.5
clb4 11 33 74 41.2 11 37 128 27.1 11 33 74 12.2
clb5 8 32 49 2.1 8 34 56 1.3 9 26 9 1.9
c2bl 12 30 187 15.6 12 30 187 14.9 12 30 187 14.4
c2b2 11 33 192 10.8 11 38 161 6.9 11 33 146 7.8
c2b3 11 28 149 8.9 11 28 113 9.9 11 28 197 10.8
c2b4 12 34 148 16.8 12 34 148 16.8 12 34 148 17.1
c2b5 10 27 120 6.1 10 31 155 6.2 10 27 120 6.2
c3bl 13 41 323 315.0 12 51 521 62.1 13 41 325 31.4
c3b2 8 23 86 4.4 8 26 87 4.5 8 23 62 5.6
c3b3 12 40 360 27.4 12 40 360 894.7 12 40 365 20.1
c3b4 12 40 327 442.2 12 46 284 548.8 13 38 928 55.1
c3b5 9 27 75 5.6 9 27 75 5.4 9 27 75 5.7
c4bl 12 46 244 16.8 12 46 227 10.6 12 46 227 11.3
c4b2 9 27 149 45.5 9 32 150 56.2 9 27 135 35.0
c4b3 10 32 129 15.7 10 34 108 14.9 10 32 129 15.6
c4b4 10 27 163 32.0 10 28 112 32.6 11 26 72 29.9
c4b5 9 24 48 27.8 9 24 48 27.7 9 24 48 27.0
c5bl 10 26 68 1.2 10 26 68 1.2 10 26 68 1.2
c5b2 8 25 27 1.1 8 25 27 1.0 9 23 8 1.1
c5b3 9 23 42 3.6 9 24 45 3.2 9 23 83 3.1
c5b4 10 32 87 5.8 10 41 101 2.7 10 32 87 2.8
c5b5 8 31 17 1.4 8 33 16 1.2 8 31 71 1.1









Motivated by this observation, we modify our algorithm to directly solve for the

best orientation by using obtained upper and lower bounds to quickly prove whether

rotating the MLC head is beneficial. Assume that we have lower and upper bounds for

the row-convex and column-convex problems, and suppose that the lower bound of the

row-convex problem is greater than the upper bound of the column-convex problem. In

this case, we can conclude that an optimal solution minimizing total treatment time for

the given fluence map must be a column-convex decomposition. We use this argument to

solve one of the problems, and then use the bound information to avoid having to solve

the other one to optimality. We pick the first problem to solve by selecting one having

the least initial lower bound, breaking ties if applicable by choosing the problem for which

n < m, since the subproblems tend to solve faster for smaller values of n. Table 3-4 shows

the nAper, BOT, TGI, and CPU metrics obtained from our algorithm enhanced with

the above bounding scheme, corresponding to the "Total Time," "# Apertures," and

"Lexicographic" objectives. Observe that all 25 instances, under any metric, terminate

in under 15 minutes of CPU time with a solution that is optimal with respect to the

corresponding objective, and all instances are solved to optimality within a minute using

the "Lexicographic" algorithm.

Recall that the "BOT" column in "Lexicographic" reports the minimum achievable

beam-on-time, and the "nAper" column under the objective "# Ap. 'i I,. reports the

minimum number of apertures needed to decompose each instance. Perhaps surprisingly,

in comparing these values with the results of "Total Time," we observe that there exists a

solution that minimizes both the number of shapes and the beam-on-time simultaneously

in 19 of the 25 instances.

Finally, we analyze performance of the three heuristics on clinical data, where we

execute each heuristic on each problem instance and its transpose (corresponding to

row-convex and column-convex decompositions), and pick the solution yielding the

smallest treatment time. Table 3-5 shows the number of apertures, beam-on-time, TGI












Table 3-5. Comparison of heuristic algorithms on clinical data
Siochi Engel Xia-Verhey
Name nAper BOT TGI CPU nAper BOT TGI CPU nAper BOT TGI CPU
clbl 11 38 245 14.0 12 38 261 < 1 13 40 219 < 1
clb2 8 23 109 3.0 8 23 127 < 1 10 32 133 < 1
clb3 9 28 213 4.0 10 28 192 < 1 12 34 198 < 1
clb4 12 34 306 9.5 11 37 398 < 1 14 42 355 < 1
clb5 9 26 103 3.6 9 26 175 < 1 12 35 124 < 1
c2bl 12 30 652 11.2 12 30 738 < 1 15 45 635 < 1
c2b2 12 33 395 17.8 12 33 464 < 1 15 45 460 < 1
c2b3 12 28 625 34.8 12 28 429 < 1 15 43 459 < 1
c2b4 12 34 628 43.2 12 34 723 < 1 18 56 417 < 1
c2b5 11 27 463 15.3 11 27 465 < 1 14 41 375 < 1
c3bl 14 43 828 36.3 15 40 1054 < 1 17 55 765 < 1
c3b2 9 23 143 11.1 9 23 127 < 1 12 36 289 < 1
c3b3 14 40 1316 40.8 14 40 869 < 1 19 60 1038 < 1
c3b4 13 48 678 33.3 14 38 765 < 1 17 55 553 < 1
c3b5 9 28 263 7.0 9 27 325 < 1 13 45 261 < 1
c4bl 13 46 617 29.4 14 46 625 < 1 18 62 531 < 1
c4b2 10 29 295 73.8 10 27 466 < 1 14 44 350 < 1
c4b3 11 32 339 19.4 11 32 365 < 1 14 48 428 < 1
c4b4 11 26 489 13.5 11 26 540 < 1 15 46 424 < 1
c4b5 9 24 236 89.8 9 24 328 < 1 15 44 328 < 1
c5bl 11 26 188 4.6 12 26 176 < 1 12 38 185 < 1
c5b2 9 23 129 6.9 9 23 100 < 1 10 33 145 < 1
c5b3 9 26 201 5.1 10 23 293 < 1 12 32 189 < 1
c5b4 11 32 218 11.2 11 32 322 < 1 13 46 243 < 1
c5b5 8 32 217 7.2 9 31 211 < 1 11 35 138 < 1



metrics for each solution as well as the CPU time spent by each heuristic. Comparison

with the "Total Time" columns in Table 3-4 reveals that even though the heuristics

consistently generated high-quality solutions, the Siochi and Engel heuristics were able to

find an optimal solution in only five problem instances, and Xia-Verhey heuristic could not

find an optimal solution to any instance.









CHAPTER 4
RECTANGULAR MATRIX DECOMPOSITION PROBLEM

4.1 Introduction and Literature Survey

Over the past decade, Intensity Modulated Radiation Therapy (IMRT) has developed

into the most successful external-beam radiation therapy delivery technique for many

forms of cancer. This is due to its ability to deliver highly complex dose distributions to

cancer patients that enable the eradication of cancerous cells while limiting damage to

nearby L. il' 1!: organs and tissues. Patients treated with IMRT therefore often experience

a higher chance of cure, suffer from fewer side effects of the treatment, or both. In

this chapter, we study an optimization problem that is related to the efficient clinical

implementation of IMRT using a simpler technology than currently used, which, if

successful, will reduce the cost as well as the complexity of delivering IMRT and thereby

make such superior treatments accessible to significantly more patients worldwide.

External-beam radiation therapy is delivered from multiple angles by a device that

can rotate around a patient. The use of multiple (typically 3-9) angles is one of the tools

that allow for the treatment of deep-seated tumors while limiting the radiation dose to

surrounding functioning organs. Conventional conformal radiation therapy then further

uses blocks and wedges to shape the beams (see, e.g., Lim (2002) and Lim et al. (2004,

2007)). IMRT is a more powerful therapy that instead modulates beam intensity. The

most common technique for achieving this modulation is to dynamically shape beams

with the help of a multileaf collimator (! IC) system. Such systems can dynamically

form many complex apertures by independently moving leaf pairs that block part of

the radiation beam. Unfortunately, MLC systems are very costly and technologically

advanced, and are therefore difficult and expensive to operate and maintain. Moreover,

MLC systems are currently only available for use with a so-called linear accelerator that

generates high-energy photon beams for treatment. However, the use of radioactive 60Co

(Cobalt) sources for radiation therapy is still ubiquitous in many parts of the world and is









poised to experience a revival in the United States and Europe through the RenaissanceTM

device that is under development by ViewRay, Inc. based in Cleveland, Ohio. Without

a MLC, IMRT delivery may be achieved through the use of compensators: high-density

blocks that control the intensity profile of a radiation beam. Such blocks are custom-made

for each individual patient, which makes compensator-based IMRT not only labor and

storage space intensive, but it also makes the actual treatment very time-consuming

due to the fact that therapists must enter the treatment room to place each individual

compensator. In addition, compensators have several undesirable properties that make it

difficult to perform accurate dose calculations, thereby reducing the advantages of IMRT

(see, e.g., Earl et al. (2007)). Recently, researchers have begun to explore the clinical

feasibility of delivering IMRT using conventional jaws that are already integrated into

radiation delivery devices and can create apertures that are rectangular in shape (see,

e.g., Earl et al. (2007), Kim et al. (2007), and Men et al. (2007)). Successful application of

this much simpler delivery technique depends critically on the ability to eff;. : /ibi deliver

high-quality treatment plans. We therefore develop and test new optimization approaches

to minimize the treatment time required for a particular treatment plan using rectangular

apertures only.

Solving a so-called fluence map optimization problem yields an optimal IMRT

treatment plan that resolves different, and conflicting, clinical measures of treatment

plan quality related to tumor control and side effects (see, e.g., Shepard et al. (1999) for

a review; Lee et al. (2000a, 2003) for mixed-integer programming approaches; Romeijn

et al. (2006) for convex programming models; and Hamacher and Kiifer (2002) and Kiifer

et al. (2003) for a multicriteria approach). A treatment plan then consists of a collection

of nonnegative intensity matrices, often referred to as fluence maps, one corresponding

to each beam angle. To limit treatment time, each of these matrices is then expressed

as a multiple of an integral fluence map in which the maximum element is on the order

of 10-20. To allow delivery of the treatment plan, each of these fluence maps should be









decomposed into a number of apertures and corresponding intensities, where the collection

of apertures that may be used depends on the delivery equipment. For MLC delivery this

problem is called the leaf sequencing problem and is very widely studied; for examples, we

refer to Al!mi and Hamacher (2005), Boland et al. (2004), Kamath et al. (2003), Engel

(2005), Kalinowski (2005a), and Taskm et al. (2009b). (Note that integrated approaches to

fluence map optimization, also referred to as aperture modulation, have been proposed as

well; we refer to, e.g., Preciado-Walters et al. (2004), Romeijn et al. (2005), and Men et al.

(2007).)

The problem that we study is the decomposition of an integral fluence map into

rectangular apertures and corresponding intensities. While Dai and Hu (1999) proposed a

straightforward heuristic for a variant of this decomposition problem, we develop the first

computationally viable optimization approach to this problem. In Section 4.2 we consider

the core problem of decomposing an (integral) fluence map while minimizing the number

of rectangular apertures. In Section 4.3 we then extend our models to the problems of (i)

minimizing total treatment time (as measured by the sum of the required aperture setup

times and the beam-on-time, i.e., the actual time that radiation is being delivered); and

(ii) minimizing the number of apertures subject to beam-on-time being minimal. Finally,

Section 4.4 discusses our computational results on a collection of clinical fluence maps.

4.2 A Mixed-Integer Programming Approach

We begin in Section 4.2.1 by formally describing the optimization model under

investigation and modeling it with a mixed-integer programming formulation. We next

describe several classes of valid inequalities in Section 4.2.2. Finally, we discuss methods

for partitioning the input matrix in Section 4.2.3, which leads to effective lower and upper

bounding techniques.

4.2.1 Model Development

In this section, we discuss an integer programming approach to decomposing a fluence

map into a minimum number of rectangular apertures and corresponding intensities. We









denote the fluence map to be delivered by a matrix B E -1' where the element at

row i and column j, (i,j), corresponds to a bixel with required intensity bi. We call a

bixel having an intensity requirement of zero a zero-bixel. We also define a nonzero-bixel

analogously. Figure 4-1 shows an example fluence map, which we use throughout this

chapter.

2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 14 10 9 0 3
5 8 20 7 1 0 4
5 9 5 4 0 0 3


Figure 4-1. Example fluence map


Let R be the set of all O(n2r2) possible rectangular apertures (i.e., submatrices of B

having contiguous rows and columns) that can be used to decompose B, excluding those

that contain a zero-bixel. For each rectangle r E R we define a continuous variable x, that

represents the intensity assigned to rectangle r, and a binary variable y, that equals 1 if

rectangle r is used in decomposing B (i.e., if xr > 0), and equals 0 otherwise. Let Cr be

the set of bixels that is exposed by rectangle r. We define Mr = min(ij)ec,{bij} to be the

minimum intensity requirement among the bixels covered by rectangle r. Furthermore, we

denote the set of rectangles that cover bixel (i,j) by R(i,j). Given these definitions, we

can formulate the problem as follows:


IPR: Minimize Syr (4-1)
rER

subject to: Y z xr bij Vi 1,...,m, j 1,...,n (4-2)
rER(i,j)

Xr
xr > 0, yr binary Vr e R. (4-4)









The objective function (4-1) minimizes the number of rectangles used in the decomposition.

Constraints (4-2) guarantee that each bixel receives exactly the required dose. Constraints

(4-3) enforce the condition that x, cannot be positive unless yr = 1. Finally, (4-4) states

bounds and logical restrictions on the variables. Note that the objective (4-1) guarantees

that y, = 0 when x, = 0 in any optimal solution of IPR.

Formulation IPR contains two variables and a constraint for each rectangle, resulting

in a large-scale mixed-integer program for problem instances of clinically relevant sizes.

Furthermore, the Mr-terms in constraints (4-3) lead to a weak linear programming

relaxation; with no valid inequalities or branching yet performed on the problem, we have

that y, = x,/Mr at optimality to the linear programming relaxation of IPR. An alternative

formulation that does not require Mr-terms employs a decomposition method. Recall that

we investigated the problem of decomposing an integer matrix into "consecutive-ones"

matrices in Chapter 3, where in each decomposed matrix all nonzero values take the same

value and appear consecutively on each row. Our computational results showed that

solvability of the problem is significantly improved by applying a bi-level optimization

algorithm. A similar approach for the problem we consider in this chapter would formulate

a master problem as:


MP: Minimize Y yr (4-5)
rER

subject to: y corresponds to a feasible decomposition (4-6)

yr binary Vr E R, (4-7)


where we address the form of (4-6) in the sequel. Given a vector y, we can check whether

constraint (4-6) is satisfied by solving the following linear program:


SP(f): Minimize 0 (4-8)

subject to: x xr bij Vi 1,...,m, j 1,...,n (4-9)
rER(i,j)









x, < 31 i^ Vr E R


x, > 0 Vr e R. (411)

Associating variables ac with (4-9), and /3 with (4-10), we obtain the dual formulation:


DSP(y): Maximize bijaij + If (4 12)
i=1 j=1 rER

subject to: Y c + f3 > 0 Vr ER (4-13)
(ij)ecC,-

yij unrestricted Vi 1,...,m, j 1,...,n (4-14)

0,< 0 Vr R. (415)


Our Benders decomposition strategy first solves MP, which yields 9. If SP(S) is feasible,

then 9 corresponds to a feasible decomposition and is optimal. Else, DSP(f) is unbounded

(since the trivial all-zero solution guarantees its feasibility). Let (&, 3) be an extreme dual

ray of DSP(9) such that 1 1 bijij + TrER 1i 1 > 0. Then, all y-vectors that are

feasible with respect to (4-6) must satisfy


Z- bZ j + -(M,,r)yr < 0. (4-16)
i=1 j=1 rER

We add (4-16) in a cutting plane fashion as necessary.

Remark 6. Even though the number of r. in,.J, that can be used in ,';,1.:1.:;/.,':.i; the

input matrix B is O(n2m2), we observe that optimal solutions 'I/,'./ .l'/; use only a small

percentage of the total number of r. /, ,..,jl, This observation suggests that another

way to overcome the dimensional ,. m,;pl. :~ ii associated with solving IPR is to apply a

column generation approach. In this approach, we start with a feasible set of columns

and rows corresponding to a subset of r,. I,-gl, and generate additional columns and

rows as necessary within a branch-cut-price (BCP) il,'.-, :thm. Even though this approach

requires the solution of much smaller linear p.ji 'i':r,",:,J' relaxations, several features

of the branch-and-cut ili,-rithm such as preprocessing and automatic cutting plane


(4-10)









generation are not applicable. As a result, our implementation of the BCP approach was

not ,,i*,*l/,/'L, *,', ll; competitive with the other il' ', :thms we presented, and further details

are therefore omitted.

4.2.2 Valid Inequalities

In this section we discuss several valid inequalities and optimality conditions for our

problem. All inequalities that we describe in this section are applicable to both the integer

programming formulation and the master problem of the Benders decomposition approach

we described in Section 4.2.1.

4.2.2.1 Adjacent rectangles

We call two non-overlapping rectangles rl and r2 adjacent if either of the following

conditions is satisfied:

(a) rl and r2 cover an identical range of columns, with rl having bottom row i and r2
having top row i + 1, or

(b) rl and r2 cover an identical range of rows, with ri having right-most column j and r2
having left-most column j + 1.

We observe that there exists an optimal solution in which no two .,i.i i:ent rectangles are

used in the decomposition. To see this, assume that .,.i i:ent rectangles rl and r2 have

intensities x,1 and x,2, respectively, where x,, < x,2 without loss of generality. In this case,

an alternative optimal solution can be constructed by extending ri into r2. Specifically, let

r' be the rectangle for which C,, = C,, U C2. An alternative optimal solution that does not

contain any .,.i i,:ent rectangles uses r2 having intensity x,2 x,,, and r' having intensity

xr1. This dominance criterion can be written as:


y~r + y,2 < 1 V .,l.i ,.ent rectangles r, r2, (4 17)

which states that no pair of .,I.i ient rectangles can be selected in an optimal solution.

4.2.2.2 Bounding box inequalities

We first observe that intensity requirements of .,Ii i.ent bixels can be used to derive

certain necessary conditions that any feasible decomposition of a matrix needs to satisfy.









We z- that a rectangle starts at bixel (i, j) if the upper-left corner of the rectangle is

located at (i,j). Consider the bixel (5, 3) marked with dark gray in Figure 4-2. Since

b43 = 2, the total intensity delivered to (5, 3) by all rectangles that start in rows i

1,..., 4 cannot exceed 2. However, b53 = 14 > 2, and hence at least one rectangle

that starts in row 5 is required to cover bixel (5, 3). Similarly, b53 > b52 implies that at

least one rectangle that starts in column 3 is required to cover the same bixel. These

results can be strengthened by considering both (4, 3) and (5, 2) simultaneously. Since

b53 > b43 + b52, we conclude that at least one rectangle that starts at bixel (5, 3) is required

in any feasible decomposition of the fluence map. In general, a rectangle must start at

(i,j) if bij > b(_i-)j + bi(j-_) is satisfied. Figure 4-3 illustrates a similar idea, where we

2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 10 9 0 3
5 8 20 7 1 0 4
5 9 5 4 0 0 3

Figure 4-2. Example start index

compare the intensity requirement of bixel (6, 4) with the bixel below it, and the one on its

right. Using arguments similar to the ones regarding starting indices, we conclude that a

rectangle must end (i.e., have a lower-right corner) at (6, 4) since b64 > b74 + b65.

2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 14 10 9 0 3
5 8 20 1 0 4
5 9 5 4 0 0 3

Figure 4-3. Example end index


Starting and ending index conditions can be generalized further as follows. Assume

that there exist integers u E [0, i 1], d E [i + 1, m + 1], 1 E [0,j 1], and re [j + 1, n + 1]









so that bij > bi + bj + bi, + bdj, where we define bio = boj = bm+l,j = bi,+l = 0 for

i E {0,... ,m + 1},j E {0,..., n + 1}. In this case, we i- that (1,u, r,d) is a bounding

box for bixel (i,j). Figure 4-4 illustrates a bounding box for bixel (6, 3) (marked in dark

gray), which corresponds to (1, u, r, d) = (2,4,5, 7). The four bixels that represent the
borders of a bounding box are marked in light gray. We note that any rectangle that

2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 14 10 9 0 3
5 8 7 1 0 4
5 9 5 4 0 0 3

Figure 4-4. Example bounding box


contains bixel (6,3), and does not start inside the bounding box (at (5,3) or (6,3)) or

end inside the bounding box (at (6,3) or (6,4)), has to contain at least one of the four

bixels on the border. Therefore, the sum of intensities of those rectangles is bounded

by the total required intensity of the bixels in light gray. Since the intensity of the dark

gray bixel cannot be satisfied by those rectangles alone, it follows that at least one

rectangle contained within the bounding box must be used to cover bixel (6, 3). Let BByj

represent the interior of a bounding box for bixel (i, j), i.e., given (1, u, r, d) all bixels at

the intersection of rows u + 1,..., d 1 and columns 1 + 1,... ,r 1. We denote the

set of rectangles in R(i,j) that are contained within BBBy by R(BBBy). In this case, the

following inequality is valid:


yr > 1. (4-18)
rER(BBij)

Note that (0, 0, n+ l, m+ 1), which corresponds to the input matrix, is a bounding box

for any bixel. Therefore there can be multiple bounding boxes associated with each bixel.

Let BBij and BBj be two bounding boxes for bixel (i,j). We that BBBy dominates

BBI if R(BBij) c R(BB'B). Since the inequality (4-18) that corresponds to a dominated









bounding box is implied by the inequality that is associated with the corresponding

dominating bounding box, we are only interested in generating nondominated bounding

boxes. Figure 4-5 di- pl'-1 another nondominated bounding box for the bixel considered in

Figure 4-4.

2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 14 10 9 0 3
5 8 7 1 0 4
5 9 5 4 0 0 3

Figure 4-5. Another nondominated bounding box seeded at (6,3)


To generate nondominated bounding boxes, we first make the following observation.

A nondominated bounding box for bixel (i, j) is minimal in the sense that none of its

edges can be shifted closer to (i,j) without violating the bounding box intensity property.

We use this observation to design an algorithm that finds several nondominated bounding

boxes associated with a given bixel. In our algorithm, we start at a bixel (i,j), and

first move in a vertical or horizontal direction until we encounter a bixel (i',j') having

bi/j, < bij. We mark (i',j') as an edge of the bounding box, reduce bij by bi//, and return

to (i,j). We then move in the remaining directions one-by-one, updating bij after each

step, to find the remaining edges of the bounding box. We repeat the same procedure

for all 4! permutations of the directions, and obtain a nondominated bounding box in

each iteration. Finally, we eliminate duplicates to obtain a set of nondominated bounding

boxes, and we generate a constraint of type (4-18) for each bounding box.

4.2.2.3 Aggregate intensity inequalities

We derive a simple class of valid inequalities by observing that the total intensity that

can be delivered to each bixel needs to be greater than or equal to its required intensity.









Formally,


> Mryr, > bi Vi 1,... m, ,j 1,... n. (4-19)
rER(i,j)

We note that inequalities (4-19) are implied by (4-2) and (4-3) in IPR. However, (4-19)

can be used to tighten the master problem of the Benders decomposition approach

discussed in Section 4.2.1. Furthermore, various tightening procedures can be applied

to (4-19) for use in either the direct solution of IPR or in the Benders master problem.

In our implementation, we apply a C'!Orv; 1 -Gomory rounding procedure (see, e.g.,

Nemhauser and Wolsey (1988)) in which we divide both sides of the inequality by the

smallest Mr coefficient on the left-hand-side (unless bij is divisible by that number), and

round up coefficients on both sides of the inequality. If bij is divisible by the smallest

Mr-coefficient on the left-hand-side of (4-19), then the rounding procedure yields an

inequality implied by (4-19), and hence we do not generate it.

4.2.2.4 Special submatrices

An alternative strategy to the one described in Section 4.2.2.3 divides both sides

of (4-19) by bij 1, provided that bij > 2, and then rounds up all coefficients and the

right-hand-side. Noting that all coefficients on the left-hand-side are bounded from above

by bij, this process yields:


yr + 2 > 2 Vi 1,.,m, j 1,.,n. (420)
rER(i,j): rER(i,j):
Mr
Equations (4-20) imply that bixel (i,j) can either be covered by a single rectangle having

a maximum intensity of bij, or otherwise needs to be covered by at least two rectangles.

The idea behind (4-20) can be extended to other special cases. For instance, consider the

following lemma.

Lemma 4. Consider ':1; 1 x 2 or 2 x 1 submatrix of B in which both elements equal a

common nonzero value, q. D. fi;,. A1 as the set of r,. la,.jl, that cover il. /11 one of the









two bixels, and have a maximum ':i,. I,.l'i of q. Let A' be the set of all r,. In,.l/l. that

cover, i. i//:l one of the two elements, and have a maximum ':,/ :,il/i less than q. D, Fi,

Ay and Af i",i., i.'..-li for r.. ,l.i, that cover both elements. The following .,, 8.;,/.:/u is

valid:


4 y + 2 y, + 2 y, + y, > 4. (4-21)
rEA2 rEA2 rEA1 rEA1

Proof. Consider any feasible solution, and let vector v denote how many rectangles exist

in the solution belonging to Ay, Af, A1, and A<, respectively. We claim (without proof,

for brevity) that the following vectors Vl,..., v6 are minimal, in the sense that v > vi

for at least one i = 1,..., 6, for every feasible v: v = (1,0, 0, 0), 2 = (0,1,0, 2),V3

(0, 2,0, 0), V4 (0, 0, 1, 2),V5 = (0,0,2, 0), V6 (0, 0, 0, 4). Note that each solution

represented by vi satisfies (4-21), and thus all v corresponding to a feasible solution must

also satisfy (4-21). E

Similarly, consider submatrices of the form


Lr quR

or its transpose, where we assume 0 < qL < qR without loss of generality. We define AL

and A< to be the sets of rectangles that cover qL, but not qR, with maximum intensity

qL, and less than qL, respectively. Let AR and A< be defined for rectangles that cover qR

but not qL, with a maximum intensity greater than or equal to (qR qL) and less than

(qR qL), respectively. We define Ay and A< as before, with a maximum intensity of qL,

and less than qL, respectively. A similar analysis as in proof of Lemma 4 reveals that the

following inequality is valid:


2 y + 2 yr + 2 yr + yr + yr + r > 4. (4-22)
rEAL rEAr rEA2 rEA< rEA< rEA2









The last special case that we consider is a nonzero submatrix of the form:

q q

q q

We define Ai to be the sets of rectangles having maximum intensity equal to q, and

covering exactly i elements of the 2 x 2 submatrix, for i 1, 2, and 4. Similarly, define A<

to be the sets of rectangles having maximum intensity less than q, and covering exactly i

elements of the submatrix. Given these definitions, we obtain:


8 y + 4 y y + 4 y + 2 + 2 yr + 2 > 8. (4-23)
rEAT rEA A^ EAA reA2 reA rEAA


4.2.2.5 Submatrix inequalities

It is possible to generate valid inequalities using arguments similar to the ones

discussed in Section 4.2.2.4 for other submatrices as well. However, this process is very

tedious, and there is a large number of possible submatrix combinations. In this section

we describe a similar set of inequalities, which are weaker than those described in the

previous section, but are easier to generate. We first observe that the formulation IPR

can be solved quickly for small input matrices. Let S denote a submatrix of the input

matrix, and R(S) represent the set of rectangles that cover at least one bixel in S. Let

LB(S) be a lower bound on the number of rectangles required to decompose S. Since

LB(S) constitutes a lower bound on the total number of rectangles required, the following

inequality is valid for any submatrix S:


Syr > LB(S)]. (4-24)
reR(S)

We can obtain LB(S) by formulating an auxiliary integer programming problem of type

IPR for S, and setting a limit on the maximum solution time.









4.2.3 Partitioning Approach

In this section, we propose a partitioning approach for our problem. We first propose

an algorithm for detecting completely separable regions of the input matrix, which can be

solved independently. Next, we explore methods for partitioning the large components, to

obtain simultaneous upper and lower bounds, which we use to improve the solvability of

our formulation.

4.2.3.1 Separable components

Our observations on clinical data sets sI-.-, -1 that input matrices can usually be

decomposed into several small components, and one or two large components. The small

components can usually be solved to optimality by formulation IPR enhanced with the

valid inequalities discussed in Section 4.2.2.

We observe on clinical data that several regions of the input matrix are completely

surrounded by zero-bixels. Since no rectangle can cover a zero-bixel, each of these regions

can be solved independently. A connected subset of the input matrix obeys the property

that a rectilinear path exists between any two nonzero-bixels of the subset, such that each

bixel in the path is also a nonzero-bixel that belongs to the subset. We call a connected

set of nonzero-bixels a component of the input matrix if it is .,i1i ient to zero-bixels across

all of its boundaries (i.e., if the subset is not contained within a larger connected subset).

To identify the components of the input matrix, we generate a graph G in which

each nonzero-bixel has a corresponding node. We add an arc between a pair of nodes if

and only if the corresponding bixels are .,11] went in the input matrix. We then identify

connected components on G by running a standard depth-first-search algorithm. Each

connected component on G corresponds to a component of the input matrix, which can

be solved independently of other components. Figure 4-6 depicts the components of the

fluence map given in Figure 4-1.


















Figure 4-6. Two components of a fluence map


4.2.3.2 Independent regions

After finding separable components of the input matrix, we attempt to further

partition each component into smaller regions. We -iv that distinct regions of a

component are independent if no rectangle intersects two bixels belonging to different

regions without also intersecting a zero-bixel. In Figure 4-7, the regions with light and

dark gray background are independent. If we solve IPR separately over all independent

regions, the sum of rectangles required to decompose each independent region yields a

lower bound on the objective function for the corresponding component.

23 30


S 14 10 9 061
5 8 0 7 1 0
5 9 5 4 0 0

Figure 4-7. Regions of a connected component


In general, there are multiple v-iva of partitioning a component into independent

regions, with each yielding possibly different lower bounds. The problem of finding a

partition that yields the best lower bound can be thought of as a "dual" of finding the

minimum number of rectangles to decompose a component. To solve this dual problem, we

need to balance two conflicting criteria:

The number of bixels assigned to each independent region needs to be small enough
so that each region can be solved quickly.


2 3 0 8 2 4 2
2 1 0 5 1 2 1
3 0 0 5 0 0 3
5 0 2 8 6 0 3
0 8 14 10 9 0 3
5 8 207 1 0 4
5 9 5 4 0 0 3
595411003









* The number of bixels not assigned to any independent regions needs to be as small
as possible to obtain a good lower bound.

We use a heuristic procedure to partition a component into independent regions, which

employs an auxiliary objective of maximizing the number of component bixels covered by

an independent region. Each bixel (i,j) is called "committed" if it either belongs to an

independent region, or if (i, j) is contained within some rectangle in R that also covers

bixels in an independent region (and hence, (i,j) cannot belong to another independent

region). All other bixels are called "uncommitted." We select our independent regions one

at a time, until no more uncommitted bixels remain. The procedure's details are described

as follows.

Initialization. Labels all nonzero-bixels as "uncommitted."

Step 1. Each candidate independent region (or just "candidate") is seeded from a

rectangle r E R such that rectangle r contains only uncommitted bixels, and such that the

number of bixels in the rectangle is no more than some limit L. For each such rectangle r,

define to be the (initial) candidate region.

Step 2. For each candidate r, if f, covers exactly L bixels, then go to Step 4. Else,

continue to Step 3.

Step 3. For each candidate r, determine if there exists an uncommitted bixel (i,j)

.,-i i,:ent to r (i.e., a bixel (i,j) } r such that either (i 1,j), (i + 1,j), (i,j 1), or

(i,j + 1) belongs to e,), such that for every r' E R(i,j), all bixels in r' either belong to r,,

or would already become committed due to the selection of 4r as an independent region.

That is, adding (i,j) to 4, would not increase the number of bixels committed by selecting

fr as a new independent region. If such a bixel exists, then add (i,j) to -r, and return to

Step 2. Else, continue to Step 4.

Step 4. For each candidate r,, compute Kr = the number of bixels in ,, and K, = the

number of uncommitted bixels (i,j) such that some rectangle in R includes both (i,j)









and a bixel in 4,. If any candidates exist such that ,K = 0, then choose e, to be any such

candidate. Else, choose to be any candidate that maximizes K /0. Go to Step 5.

Step 5. Create an independent region corresponding to f*. For each bixel (i,j) that can

be covered by a rectangle in R intersecting at least one bixel in e", change the status of

(i,j) to "committed." (This includes all bixels in e4 itself.) If all bixels are committed,

terminate the procedure; else, return to Step 1.

In our algorithm for solving a component, we execute the foregoing heuristic to

find a set of independent regions. We formulate IPR for each region, with a limit on

the maximum solution time. We then use the lower bound obtained for each region to

generate an inequality of type (4-24). (It is often prudent to skip this step if only one

region is computed for a component.)

4.2.3.3 Dependent regions

In this section, we attempt to improve the lower bound obtained using independent

regions by focusing on those bixels not included in the union of independent regions.

We define a dependent region to be a connected set of bixels in a component that does

not overlap with any of the independent regions in that component. In our example, the

region with black background in Figure 4-7 is a dependent region. Let D represent the set

of bixels in a dependent region, and let R(D) represent the set of rectangles that cover

only a subset of the bixels in D.

To improve our lower bound, we wish to compute the minimum number of rectangles

required to cover D; however, we wish to avoid double-counting those rectangles used

to cover bixels in independent regions. Accordingly, we seek the minimum number of

rectangles in R(D), perhaps in concert with rectangles outside R(D), required to cover

the bixels in D. Using the x- and y-variables as before, we formulate the following

variation of IPR to find the minimum number of rectangles in R(D) required to partition












DPR: Minimize > y, (4-25)
rER(D)

subject to: Xr bij V(i,j) E D (4-26)
rER(ij)

S< V- ,..., n, j) D (4 27)
rER(i,j)

x, <3 1, Vr E R(D) (428)

x, > 0 Vr e R, yr binary Vr e R(D) (4-29)

Objective (4-25) minimizes the number of rectangles in R(D) used in the solution.

Constraints (4-26) ensure that the bixels in D get partitioned exactly, where (4-27) limit

the intensity delivered to the remaining bixels. Constraints (4-28) relate the x- and y-

variables as done in IPR, and finally (4-29) define variable types. As before, we set a time

limit for the solution of DPR, and obtain a lower bound on the objective function value,

which we denote by LB(D). Given this value, the following inequality is valid:


Syr > [LB(D)]. (4-30)
rER(D)

In our example, the optimal value of DPR for the black (dependent) region is 1 since

the intensity requirement of bixel (1, 4) cannot be satisfied completely by rectangles that

cover bixels in the gray (independent) regions (in fact, this result can also be seen due

to the bounding box constraint implying that one rectangle representing the singleton

bixel (1,4) must appear in any feasible solution). We note that the rectangles in R(D),

by definition, do not intersect any other (dependent or independent) regions. Therefore,

the lower bounds obtained for all regions can be summed to obtain a lower bound on the

minimum number of rectangles required to decompose a component.









4.2.3.4 Upper bound calculation

In this section, we discuss how a related approach leads to a heuristic algorithm

to obtain a feasible decomposition of a component. We first note that a feasible

decomposition of a component can be obtained by combining feasible solutions obtained

for individual regions within a component. Feasible solutions for independent regions

are readily available from the integer programming problems solved for obtaining lower

bounds on those regions, as discussed in Section 4.2.3.2. Feasible solutions for dependent

regions can be extracted from solutions of the formulation given by DPR. However, since

DPR minimizes the number of rectangles that are contained within a dependent region,

and not necessarily the total number of rectangles required to decompose a dependent

region, the solutions obtained from DPR potentially use an unnecessarily large number of

rectangles not contained in R(D).

A better way of obtaining feasible solutions for dependent regions is to formulate

the problem IPR for each dependent region. Since IPR explicitly minimizes the total

number of rectangles required, we expect this approach to result in feasible solutions

of higher quality. However, this approach does not consider the fact that some of the

rectangles that are already used for decomposing independent regions can be extended into

dependent regions without increasing the total number of rectangles. To permit the use of

rectangles that intersect independent and dependent regions, we require a revised integer

programming formulation.

In our approach, we solve the integer programming formulations for decomposing

the independent regions first, and store the best feasible solutions found within the

allowed time limit. Let x, represent the intensity assigned to rectangle r for decomposing

independent regions. Next, we generate a feasible solution for each dependent region,

one at a time, as follows. We first find the set of rectangles that can be extended into

the current dependent region, and determine how those rectangles can be extended. Let

E(D, r) represent the set of rectangles in R that extend rectangle r into dependent region









D. We also define the parameter I(r)e equal to one if bixel (i,j) E D is covered by

extension e of rectangle r, and zero otherwise. Let ze be a binary variable that equals 1

if and only if extension e of rectangle r is used in the solution. We define the x- and y-

variables as before, and formulate the following problem:

EPR: Minimize > Y, (4-31)
rER(D)

subject to: x, b, (xI(r) ,) z V(i,j) e D (4-32)
rER(i,j) rER eEE(D,r)

SZr < 1 Vr R (433)
eEE(D,r)

x, < 3 Vr E RZ(D) (4-34)

x, > 0, yr binary Vr e R(D) (4-35)

zr, binary Vr e R, e E (D, r). (4-36)

We generate a feasible solution by combining three types of rectangles: (i) rectangles

used to decompose independent regions that are not extended by EPR; (ii) rectangles

obtained by extending rectangles from independent regions into dependent regions by

EPR; and (iii) rectangles in R(D) used by EPR.

Note that the optimal value of EPR for the dependent region given in Figure 4-7

is 1. This can be seen by observing that the rectangle(s) that cover bixel (3, 4) can be

extended up to fully satisfy the intensity requirement of bixel (2, 4) without any penalty

on the objective function of EPR formulated for the dependent region. Therefore, a single

rectangle contained in the dependent region solves EPR optimally. Since the optimal value

of DPR for the dependent region is also 1, our partition solves the problem of finding the

minimum number of rectangles to optimality.









4.3 Extensions

In this section, we briefly discuss how our model can be adjusted to tackle the

problems of minimizing total treatment time, and lexicographically minimizing beam-on-time

and number of apertures.

4.3.1 Minimize Total Treatment Time

The total time spent delivering a given fluence map is composed of (i) time required

to move the jaws to form the next rectangular aperture (setup time), and (ii) time during

which radiation is delivered (beam-on-time). Even though the setup time required for

switching from one rectangular aperture to the next one depends on the jaw settings

corresponding to these apertures, and hence is sequence-dependent, we make the common

assumption that total setup time is proportional to the total number of apertures used.

With this assumption, our model can easily be adjusted to explicitly minimize the total

treatment time by changing the objective function of IPR to


Minimize w Y r + .r, (4-37)
rER rER

where w is a parameter that represents the average setup time per aperture relative to the

time required to deliver a unit of intensity.

The Benders decomposition procedure discussed in Section 4.2.1 also needs to be

adjusted accordingly. We first add a continuous variable t to MP, which "predicts" the

minimum beam-on-time that can be obtained by the set of rectangles chosen by MP. The

updated master problem can be written as follows.


MPTT: Minimize w y 2 + t (4-38)
rER

subject to: y corresponds to a feasible decomposition (4-39)

t > minimum beam-on-time corresponding to y (4-40)

yr binary Vr E R. (4-41)









Given a vector 9, we can find the minimum beam-on-time for the corresponding

decomposition, if one exists, by solving:


SPTT(f): Minimize x, (4-42)
rER

subject to: x= bij Vi 1,...,m, j 1,...,n (4-43)
rER(i,j)

xr < i, Vr R (444)

xr > 0 Vr e R. (445)


Note that SPTT is obtained by simply changing the objective function of SP. If SPTT(y)

is infeasible, then we add a Benders feasibility cut of type (4-16) as before, and re-solve

MPTT. Otherwise, let the value of t in MPTT be t, and the optimal objective function

value of SPTT be t*. If t = t*, then (y, t) is an optimal solution of MPTT that minimizes

the total treatment time. However, if t > t*, then we add the following Benders optimality

cut
mn n
t ^bipiij + -),y-, -(446)
i=1 j=1 rER

where &yi and /3 are optimal dual multipliers associated with constraints (4-43) and

(4-44), respectively.

4.3.2 Optimization with Beam-on-Time Restrictions

Another related problem that we consider is finding the minimum number of

rectangles that yields the minimum beam-on-time. Note that the minimum beam-on-time

required to decompose a fluence map can be found (in polynomial time) by solving SPTT,

which is a linear program, for the vector yr 1, Vr E R. Let T* denote the optimal

objective function value of SPTT(1), where 1 is the vector of all 1's. Given this value, it is

sufficient to add


x, < T* (4-47)
rER









to minimize the number of rectangles while limiting beam-on-time to T*.

The modifications required for the Benders decomposition algorithm are also

straightforward. To enforce the minimum beam-on-time restriction, we add (4-47) to

SP, which checks whether a given set of rectangles can decompose the fluence map. The

updated feasibility cut is given by


> 1 &(M + ,M)r)y, + TtO < 0, (448)
i= j=1 rER

where 0 is the dual variable associated with (4-47) in SP. Finally, we need to check

whether the solution generated by our heuristic discussed in Section 4.2.3.4 satisfies

constraint (4-47); if so, then it can be used as an initial upper bound.

4.4 Computational Results

We have implemented our algorithms using CPLEX 11 running on a Windows XP PC

with a 3.4 GHz CPU and 2 GB RAM. Our base set of test problem instances consists of

25 clinical problem instances ("clbl", ..., "c5b5"). These instances were obtained from

treatment plans for five patients treated using five beam angles each. We report problem

characteristics in terms of the number of rows m, the number of columns n, and the

maximum intensity value L. We imposed a time limit of 1800 seconds (30 minutes) in all

of our tests. For problem instances that were not solved to optimality within the imposed

time limit, we report the best upper and lower bounds obtained, where we round lower

bounds up for the cases in which the objective function is guaranteed to have an integral

value.

Our preliminary computational tests showed that the naive implementation of our

Benders decomposition approach, in which we add a cut and re-solve the master problem

in each iteration, was not computationally competitive with solving the explicit integer

programming formulation. This is due to the fact that repetitively solving the master

problem, which is an integer programming problem, is computationally very expensive.

We instead used callback functions of CPLEX to generate a single branch-and-bound









tree in which we solve SP(9) (or (SPTT(9)) corresponding to each integer solution found

in the branch-and-bound tree, and add cuts to tighten the master problem as necessary.

While this approach produced better results than the naive implementation, it still yielded

inferior bounds than those obtained from the explicit formulation. Therefore, we omit

further Benders-based computational results.

Our first experiment quantifies the effects of the valid inequalities discussed in Section

4.2.2, and the partitioning approach discussed in Section 4.2.3 on solution quality and

execution time. In Table 4-1, the set of columns labeled "Default CPLEX" shows the

results we obtained by solving the formulation IPR on each problem instance using default

CPLEX options. The "+ Valid inequ I!i i. columns represent the IPR formulation

enhanced with the .,i1] i:ent rectangle inequalities (4-17), bounding box inequalities (4-18),

strengthened .i,--regate intensity inequalities (4-19) and (4-20), and 1 x 2 submatrix

inequalities (4-21) and (4-22). (Additional computational results showed that the 2 x 2

submatrix inequalities (4-23) and the arbitrary submatrix inequalities (4-24) did not

improve the solvability of the model.) The set of columns labeled "+ Partitions" shows

the results we obtained by partitioning the problem into separable components (Section

4.2.3.1), further partitioning each component into independent and dependent regions

(Sections 4.2.3.2 and 4.2.3.3), and using our upper bounding heuristic (Section 4.2.3.4)

in addition to the valid inequalities used for the tests in the previous set of columns. We

refer to the latter settings as our base il/,.>rithm in the remaining computational tests.

Each set of columns in Table 4-1 di- pl1iv the time spent for each problem instance

("CPU"), and upper bound ("UB"), lower bound ("LB"), and optimality gap ("GAP")

obtained. We also report the average and maximum gaps over all problem instances. We

observe that none of the problem instances were solved to optimality using the default

CPLEX options, whereas clb2 and c5b2 were solved to optimality after adding the

valid inequalities of Section 4.2.2. An additional instance (c5b5) was solved using the

partitioning strategy described in Section 4.2.3. We note that even though our approach











Table 4-1. Effect of valid inequalities and the partitioning strategy
Default CPLEX + Valid inequalities + Partitions
Name m n L CPU UB LB Gap CPU UB LB Gap CPU UB LB Gap
clbl 15 14 20 1800 66 60 0.09 1800 63 62 0.02 1800 64 62 0.03
clb2 11 15 20 1800 48 47 0.02 138.2 48 48 0 1009.9 48 48 0
clb3 15 15 20 1800 57 54 0.05 1800 57 54 0.05 1800 58 54 0.07
clb4 15 15 20 1800 61 52 0.15 1800 61 53 0.13 1800 59 55 0.07
clb5 11 15 20 1800 47 45 0.04 1800 46 45 0.02 1800 47 45 0.04
c2bl 18 20 20 1800 114 79 0.31 1800 119 85 0.29 1800 103 87 0.16
c2b2 17 19 20 1800 95 69 0.27 1800 96 81 0.16 1800 94 82 0.13
c2b3 18 18 20 1800 98 73 0.26 1800 103 77 0.25 1800 94 77 0.18
c2b4 18 18 20 1800 114 80 0.3 1800 115 84 0.27 1800 105 88 0.16
c2b5 17 18 20 1800 94 64 0.32 1800 98 72 0.27 1800 91 72 0.21
c3bl 22 17 20 1800 121 69 0.43 1800 134 79 0.41 1800 119 79 0.34
c3b2 15 19 20 1800 73 46 0.37 1800 71 52 0.27 1800 70 52 0.26
c3b3 20 17 20 1800 119 69 0.42 1800 119 75 0.37 1800 107 77 0.28
c3b4 19 17 20 1800 103 69 0.33 1800 106 73 0.31 1800 99 78 0.21
c3b5 15 19 20 1800 73 55 0.25 1800 71 58 0.18 1800 73 58 0.21
c4bl 19 22 20 1800 106 79 0.25 1800 107 89 0.17 1800 109 89 0.18
c4b2 13 24 20 1800 88 54 0.39 1800 99 58 0.41 1800 91 58 0.36
c4b3 18 23 20 1800 95 71 0.25 1800 99 75 0.24 1800 93 77 0.17
c4b4 17 23 20 1800 103 78 0.24 1800 102 81 0.21 1800 98 83 0.15
c4b5 18 24 20 1800 93 62 0.33 1800 93 66 0.29 1800 87 67 0.23
c5bl 15 16 20 1800 66 64 0.03 1800 66 65 0.02 1800 66 65 0.02
c5b2 13 17 20 1800 58 57 0.02 102.1 58 58 0 213.6 58 58 0
c5b3 14 16 20 1800 63 54 0.14 1800 68 56 0.18 1800 65 57 0.12
c5b4 14 16 20 1800 63 57 0.1 1800 64 59 0.08 1800 62 59 0.05
c5b5 12 17 20 1800 53 47 0.11 1800 51 48 0.06 36.2 49 49 0


was not able to provide provably optimal solutions for most instances, it significantly

improved both lower and upper bounds for several instances.

Our next experiment tests our base algorithm under the extensions discussed in

Section 4.3. The set of columns labeled as "Total Time" in Table 4-2 presents the

extension in which the objective function is defined as a linear combination of the

beam-on-time and the number of rectangles. The actual value of w depends on the

particular treatment delivery equipment used in the clinic, where values of w in the

range 1-10 are typical (see, e.g., Dai and Hu (1999), and Takm net al. (2009b)). In

our experiments, we therefore used w = 7 as a representative value. The next set of

columns ("Lexicographic") is dedicated to the extension in which we first minimize

beam-on-time, T*, and then find the minimum number of rectangles that yields the

minimum beam-on-time. The column "BOT" represents the value of T*, and "Total

Time" represents the total treatment time associated with the solution found, where we

again use w = 7 as the average setup time per rectangle. We observe that our algorithm

could solve more problem instances to optimality for both extensions compared to the











Table 4-2. Computational results on model extensions
Total Time Lexicographic
Name m n L CPU UB LB Gap CPU UB LB Gap BOT Total Time
clbl 15 14 20 255.9 621 621 0 36.5 66 66 0 176 638
clb2 11 15 20 330.3 459 459 0 132.6 50 50 0 121 471
clb3 15 15 20 1800 548 542.72 0.01 130.4 62 62 0 147 581
clb4 15 15 20 1800 557 542.49 0.03 186.9 62 62 0 136 570
clb5 11 15 20 1800 451 443.63 0.02 30.9 53 53 0 115 486
c2bl 18 20 20 1800 962 814.24 0.15 1800 107 104 0.03 194 943
c2b2 17 19 20 1800 883 797.74 0.1 1800 96 92 0.04 207 879
c2b3 18 18 20 1800 918 797.6 0.13 1800 96 88 0.08 237 909
c2b4 18 18 20 1800 1028 889.36 0.13 1800 111 106 0.05 258 1035
c2b5 17 18 20 1800 890 721.13 0.19 1800 92 83 0.1 207 851
c3bl 22 17 20 1800 1161 858.9 0.26 1800 116 103 0.11 266 1078
c3b2 15 19 20 1800 668 533.24 0.2 1800 70 64 0.09 151 641
c3b3 20 17 20 1800 1066 847.09 0.21 1800 111 95 0.14 278 1055
c3b4 19 17 20 1800 1023 857.91 0.16 1800 103 95 0.08 287 1008
c3b5 15 19 20 1800 722 610.01 0.16 204.4 76 76 0 182 714
c4bl 19 22 20 1800 1044 918.57 0.12 1800 108 105 0.03 275 1031
c4b2 13 24 20 1800 895 656.15 0.27 1800 95 76 0.2 232 897
c4b3 18 23 20 1800 858 743.62 0.13 1800 92 89 0.03 189 833
c4b4 17 23 20 1800 943 834.32 0.12 1800 101 96 0.05 235 942
c4b5 18 24 20 1800 913 740.19 0.19 1800 86 77 0.1 260 862
c5bl 15 16 20 271.4 626 626 0 5.5 71 71 0 158 655
c5b2 13 17 20 33.4 597 597 0 19.9 63 63 0 156 597
c5b3 14 16 20 1800 623 597.96 0.04 869.2 68 68 0 180 656
c5b4 14 16 20 1800 584 571.15 0.02 192.4 66 66 0 145 607
c5b5 12 17 20 90.4 503 503 0 37.2 57 57 0 147 546


problem of finding the minimum number of rectangles. To understand why this is the

case, we first note that the difficulty of the matrix decomposition problem varies greatly

based on the objective function used. On one hand, minimizing the number of rectangles

is strongly NP-hard, even for fluence maps having a single row (see Baatar et al. (2005)).

On the other hand, minimizing the beam-on-time is a polynomially solvable problem (see

Section 4.3). Therefore, we expect that the problem should become easier as the weight of

the beam-on-time term in the objective function increases. The reason the lexicographic

minimization problem is easier to solve than the other two variations is because the

additional beam-on-time constraint considerably shrinks the feasible solution space.

Another way of looking at the problem of balancing the number of apertures and

the beam-on-time is to view the problem as a multicriteria optimization problem. In this

setting, we are interested in constructing the Pareto efficient frontier of solutions with the

property that neither of the two criteria can be improved without deteriorating the other.

Note that the lexicographic approach that we considered above determines a particular

Pareto optimal solution to the multicriteria problem. To generate other non-dominated











solutions for the multicriteria version of the problem, we sequentially impose different

upper bounds on the number of apertures allowed, -, 7, and find the corresponding

minimum beam-on-time for these values of 7. As an example, we considered the problem

instance c5b5. For this instance, we note that the minimum number of apertures is 49 (see

Table 4-1) with a corresponding beam-on-time of 160, while the minimum beam-on-time

for this problem instance is 147 (see Table 4-2) which requires 57 apertures. Figure 4-8

then depicts (i) the non-dominated solutions; (ii) the Pareto efficient frontier for values of

7 E [49, 57], and (iii) the (boundary of the) convex hull of the Pareto set. The solutions on

the latter are the optimal solutions to the problem of minimizing total treatment time that

can be obtained with different values of w.

162


160
,.\ [ ----

158


156 \


0154
E

152

150
o 15 2


148
~ -

146
48 49 50 51 52 53 54 55 56 57 58
Number of apertures


Figure 4-8. Efficient frontier for number of apertures and beam-on-time



Our final experiment analyzes the effect of the maximum intensity value L. Usually

fluence maps are obtained by solving a nonlinear optimization problem for each beam

angle to determine an intensity profile for each beam angle, which is represented by a











Table 4-3. Effect of maximum intensity value on solvability
L 5 L 10 L 15
Name m n CPU UB LB Gap CPU UB LB Gap CPU UB LB Gap
clbl 15 14 4.4 305 305 0 22.9 441 441 0 1800 539 536.99 0
clb2 11 15 1.4 238 238 0 4.4 320 320 0 498.5 394 394 0
clb3 15 15 9.6 287 287 0 228.6 377 377 0 1800 495 487.67 0.01
clb4 15 15 5.8 269 269 0 1800 393 377.42 0.04 1800 513 493.13 0.04
clb5 11 15 2.4 216 216 0 28.9 326 326 0 316.7 411 411 0
c2bl 18 20 31.3 440 440 0 1800 648 635.1 0.02 1800 826 732.68 0.11
c2b2 17 19 51.5 448 448 0 1800 625 599.84 0.04 1800 759 690.67 0.09
c2b3 18 18 144.4 428 428 0 1800 645 615.64 0.05 1800 800 704.61 0.12
c2b4 18 18 1593.1 487 487 0 1800 755 678.49 0.1 1800 919 796.13 0.13
c2b5 17 18 197.9 429 429 0 1800 606 538.62 0.11 1800 728 621.9 0.15
c3bl 22 17 1359.4 480 480 0 1800 747 662.47 0.11 1800 1080 779.16 0.28
c3b2 15 19 1800 280 274.97 0.02 1800 414 376.01 0.09 1800 532 461.6 0.13
c3b3 20 17 1800 461 446.77 0.03 1800 731 641.2 0.12 1800 908 755.25 0.17
c3b4 19 17 1800 463 456.42 0.01 1800 713 634.26 0.11 1800 900 762.02 0.15
c3b5 15 19 1800 332 325.87 0.02 1800 481 466.6 0.03 1800 582 532.2 0.09
c4bl 19 22 39.9 529 529 0 1800 758 719.53 0.05 1800 899 827.73 0.08
c4b2 13 24 1800 422 408.14 0.03 1800 595 503.92 0.15 1800 764 582.69 0.24
c4b3 18 23 126.3 409 409 0 1800 579 564.5 0.03 1800 695 666.23 0.04
c4b4 17 23 321.4 444 444 0 1800 662 649.1 0.02 1800 815 742.81 0.09
c4b5 18 24 1194.7 414 414 0 1800 636 573.02 0.1 1800 794 662.74 0.17
c5bl 15 16 5.9 342 342 0 5.6 442 442 0 28.9 584 584 0
c5b2 13 17 4.3 289 289 0 5.5 439 439 0 49.3 516 516 0
c5b3 14 16 1574.4 294 294 0 1800 453 441.44 0.03 1800 566 538.22 0.05
c5b4 14 16 3.6 239 239 0 1800 473 468.49 0.01 1800 534 524.31 0.02
c5b5 12 17 2.1 252 252 0 9.3 354 354 0 63.9 441 441 0


matrix of real numbers. Later, these matrices are rounded to integer matrices to limit

the delivery time. To analyze the trade-off between round-off errors and treatment time,

we started from clinical treatment plans, and applied rounding with different levels of

granularity. Specifically, we generated problem instances from the same fluence maps

with L {5, 10, 15, 20}, and used our algorithm to find the minimum total treatment

time required as a measure of delivery efficiency. Table 4-3 shows the results of our

experiments. We observe that our algorithm produces smaller optimality gaps as L

decreases, which is not surprising since IPR becomes tighter as the Mr-coefficients (which

are bounded by L) decrease. Furthermore, delivery efficiency is also higher for small values

of L. The average treatment time (calculated over the lower bounds) for all problem

instances increases from 366.05 for L = 5 to 513.79 for L = 10, 609.79 for L = 15, and

684.96 for L = 20, which is calculated using the set of columns labeled "Total Time"

in Table 4-2. Our results show that the choice of granularity chosen for rounding has a

significant effect on the treatment time. For each individual patient, the risks associated

with the deterioration in treatment plan quality due to the rounding of intensities needs









to be weighed against the disadvantages of a longer treatment time by the physician or

clinician.









CHAPTER 5
GRAPH SEARCH PROBLEM

5.1 Introduction and Literature Review

In this chapter we consider several variants of a search problem on graphs, which can

be seen as a game between an intruder and a group of searchers. Consider an undirected

graph G = (N, E). The intruder and searchers occupy some nodes of the graph. At each

time period, a pl liv r (intruder or searcher) located at node i e N can move along an

edge to an .,I.i ient node, or stay at their current position. A searcher located at node

i E N can detect an intruder if it is located at some node in S(i) C N. Searchers can

be deploy, .1 at any node, but have no information about the location of the intruder.

Therefore, they have to systematically search the graph to be able to detect the intruder,

even if the intruder has perfect information about the searchers' plan, and utilizes this

knowledge to evade detection for as long as possible. The problem is to find the minimum

number of searchers needed and a routing plan for each searcher that guarantees detection

of the intruder within a given time limit.

The graph search problem was initially defined by Parsons (1978) in the context

of seeking a person lost in a cave. The cave is represented as a graph, where tunnels

of the cave correspond to edges of the graph. Searchers have to sweep edges of the

graph to locate the missing person, who is assumed to be wandering unpredictably or is

purposefully trying to evade searchers. The search number s(G) of a graph G is defined to

be the minimum number of searchers needed so that the missing person can be found even

if he could move infinitely fast along any path not occupied by searchers (Parsons, 1978).

Computing s(G) is NP-hard for general graphs (Bienstock and Seymour, 1991; LaPaugh,

1993; Megiddo et al., 1988), but it can be computed in linear time for trees (Alspach,

2004; Megiddo et al., 1988; Peng et al., 2000). The search number of a graph has been

shown to be related to other important parameters such as tree-width, path-width, and

vertex separation (Dendris et al., 1997; Ellis et al., 1994; Seymour and Thomas, 1993).









Several variants of the graph search problem have been investigated in the literature.

In decontamination problems, edges or nodes of a graph are infected by a contaminant

such as a computer virus or a chemical agent, which spreads across the graph (Flocchini

et al., 2008; LaPaugh, 1993; Penuel and Smith, 2009). In rendezvous problems different

p1 i, rs, who are not aware of the location of others, try to meet at a common node as

quickly as possible (Alpern, 1995; Alpern and Gal, 2003; Kikuta and Ruckle, 2007).

Hide-and-seek problems consider an intruder that "hides" in a stationary location, while

the searchers try to locate the intruder in minimum time (Alpern, 2008; Jotshi and Batta,

2008). Such problems also arise in search-and-rescue settings (Benkoski et al., 1991).

Pursuit evasion (or "cops-and-robber") games model an intruder that tries to avoid being

captured by searchers (Aigner and Fromme, 1984; Alspach et al., 2008; Hahn, 2007;

Isler and Karnad, 2008). In some applications nodes of a graph need to be patrolled for

protection or supervision (C'!, i. I!. yre et al., 2004; Sak et al., 2008). In particular, one

interesting application coordinates automated software searchers so that they patrol the

Internet to find web sites that exploit browser vulnerabilities (Wang et al., 2005). We

refer the reader to Alpern and Gal (2003); Alspach (2004); Fomin and Thilikos (2008) for

detailed surveys of the literature on search problems and applications in various practical

settings.

Most of the previous research on graph search problems has focused on theoretical

aspects of the problems (e.g. C'!, i,!yre et al. (2004); Dendris et al. (1997); Ellis et al.

(1994); Goldstein and Reingold (1995); Seymour and Thomas (1993)) or designing

algorithms for solving the problems on special graph structures (e.g. Alpern (2008);

Flocchini et al. (2008); Kikuta and Ruckle (2007); Peng et al. (2000)). Our contribution

is an exact optimization algorithm for solving several variants of the search problem on

general graphs (see also Penuel and Smith (2009) for a decontamination problem in which

the intruder location has been determined). In particular, we consider three specific graph

search problems: (i) a hide-and-seek problem, (ii) a pursuit evasion problem, and (iii) a









patrol problem. We model these problems as large-scale integer programs, and propose

a branch-cut-price algorithm that is capable of solving all three problems with some

modifications.

A variant of the branch-and-bound algorithm, which adds cutting planes to linear

programming relaxations to tighten dual bounds is called branch-and-cut, and is

emploiv '1 in most commercial solvers for solving integer programs ( r'chand et al.,

2002; Nemhauser and Wolsey, 1988; Wolsey, 1998). An effective method for solving

integer programs having a large number of variables is branch-and-price, which is based

on dynamic column generation (Barnhart et al., 1998). Branch-cut-price is essentially an

algorithm that combines dynamic column generation with dynamic row generation (Jiinger

and Thienel, 2000).

The remainder of this chapter is organized as follows. In Section 5.2 we describe a

hide-and-seek problem and propose a column generation algorithm for solving its linear

programming relaxation. Similarly, Sections 5.3 and 5.4 analyze the pursuit evasion and

patrol problems, respectively. We describe some branching rules that can be used in all

three algorithms to obtain an optimal solution to these problems in Section 5.5. Finally,

we give computational results in Section 5.6.

5.2 Hide-and-Seek Problem

We start our discussion with a hide-and-seek problem, in which a group of searchers

seek to locate a stationary intruder within T time steps. While this problem is relatively

easy to model and analyze, it serves as the basis for our more complex search algorithms.

5.2.1 Mathematical Model

We denote the set of nodes .,.i i,'ent to node i e N by A(i). Since we allow the

intruder and searchers to stay at their current location, we assume that i E A(i), Vi E N.

Recall that a walk on a graph G = (N, E) is a sequence i1,..., ir of nodes such that for all

1 < k < r 1, ik+l E A(ik) (Almi I et al., 1993). We define the length of a walk as the

number of traversed edges on the walk. Let P(T) denote the set of all possible walks of









length at most T that can be taken by a searcher. We assume that each node i e N can

be observed from some node, i.e., there exists a j E N such that j E S(i). Let dpi be a

parameter whose value is 1 if a searcher following walk p can detect an intruder located at

node i, and 0 otherwise. Let Ap be a binary variable that equals 1 if a searcher is assigned

to follow walk p, and 0 otherwise. Given these definitions, our hide-and-seek problem can

be formulated as the following set covering problem.


HS: minimize Ap (5 1)
pEP(T)

subject to dpi, > 1 Vi E N (5-2)
pEP(T)

A, {0, } Vp P(T) (5-3)

The objective function (5-1) minimizes the number of selected searchers, and constraints

(5-2) guarantee that each node is covered by at least one searcher within the allowed

time frame. We note that a special case of this problem for which a searcher located at

node i e N can observe node i and its neighbors, and we need to guarantee immediate

detection of the intruder (i.e., when S(i) = A(i), Vi E N and T = 0), is equivalent to the

minimum dominating set problem, which is known to be NP-hard (Garey and Johnson,

1979). Therefore, the hide-and-seek problem that we consider is NP-hard.

5.2.2 Solution Approach

In principle, all walks of length at most T can be enumerated, and HS can be solved

directly. This approach might be practical for small values of T, but in general there is

an exponential number of such walks, which corresponds to an exponential number of

variables. We instead propose a column generation approach to solve HS. Given a subset

of walks P'(T) C P(T), we can construct a limited hide-and-seek (LHS) problem identical

to HS, with P(T) replaced by P'(T). The linear programming relaxation of LHS is:


LHSLP: minimize A (5-4)
pEP'(T)









subject to d > 1 Vi c N (5-5)
pEP'(T)

A > 0 Vp P'(T), (5-6)

where upper bounds on the A-variables are not necessary at optimality. Given an optimal

dual vector y, the reduced cost of Ap, which we denote by cp, can be calculated as 1 -

CiN idpi. Since y is an optimal dual vector, c, > 0 for all p P'(T). We can conclude
that the current solution of LHSLP is also optimal for the linear programming relaxation

of HS if cp > 0 for all p P(T). On the other hand, if cp < 0 for some p E P(T) \ P'(T),

then adding p to P'(T) can potentially decrease the value of the objective function (5-4).

We discuss our pricing problem, which seeks such a p, in the next section.

5.2.2.1 Searcher's problem

Let <7 be the dual variable associated with the constraint of type (5-5) corresponding

to node i e N. Also, let yi be a decision variable that equals 1 if node i is "seen"

by a searcher following a walk that we generate, and 0 otherwise, Vi e N. Given an

optimal dual vector y, we solve the following pricing problem to seek a A-variable having

a negative reduced cost: max ECEN ^ /, subject to the restriction that (/i,..., yINI)

corresponds to a set of nodes observed by a walk of length no more than T.

The pricing problem can be formulated as a mixed-integer programming problem on a

time-expanded network consisting of T + 1 stages. In particular, we create a node it for

each i E N, t = 0,..., T. We create an arc from node i, Vi E N, t = 0, ., T 1 to

nodes Nj(t+l) for all j E A(i). For this problem it is easy to see that an optimal solution

exists in which all searchers move at each time period. Therefore, we omit arcs between

nodes it and i(t+l) for each i E N. Figure 5-1 di-tpli- a simple example graph, and the

corresponding time-expanded network for T = 2.

To formulate the pricing problem as a mixed-integer program, we introduce binary

variables xt = 1 if the searcher is at node i at time t. Then, an integer programming
























(a) K > (b) K u

Figure 5-1. (a) An example graph (b) Time-expanded network for T = 2


formulation of the problem can be given as:


maximize ^ I
iEN


subject to


x> t 1 Vt ,...,T
iEN


(5-7)


(5-8)


(5-9)


(5-10)


x < t-1 Vi N,t =,...,T
jeA(i)
T
y, < xt Vi EN
t=0 j:ies(j)

0
xJ e {0, 1} Vi e N,t= 0,...,T.


Constraints (5-8) represent the fact that the searcher can visit only one node at a time.

Constraints (5-9) ensure that node i can be visited at time t only if one of its neighbors

has been visited at time t 1. Constraints (5-10) force the value of yi to zero unless node

i can be observed by the searcher at some time period. Note that the y-variables will

take on binary values in an optimal solution, and therefore we relax them as continuous

variables. If the optimal objective function value of (5-7)-(5-12) is greater than 1, then









we have found a variable that has a negative reduced cost, and we therefore add the

generated column to LHSLP.

5.2.2.2 Branch-and-price algorithm

The hide-and-seek problem can be solved using the following branch-and-price

algorithm.

* Step 0: C'!.... a feasible set of initial search walks, in which every node is seen by
at least one searcher.

Step 1: Solve LHSLP, and generate columns by solving the searcher's problem until
LHSLP has been solved to optimality. If the optimal solution is fractional, then
branch, and go back to Step 1 for the subproblems. Else, stop processing the current
subproblem with an integral solution.

We initialize our algorithm by generating a stationary searcher that stays at node i for

T periods, for each i E N. Even though these elementary searchers are not likely to

be selected in an optimal solution, they guarantee the feasibility of LHSLP. We discuss

several branching strategies that can be used for Step 1 in Section 5.5.

5.3 Pursuit Evasion Problem

In this section, we consider a pursuit evasion variant of the search problem. Unlike

the hide-and-seek problem, the intruder is also mobile in this variant, and tries to

avoid pursuit by the searchers. We note that this problem also reduces to the minimum

dominating set problem for S(i) = A(i), Vi E N and T = 0, and hence it is NP-hard.

5.3.1 Mathematical Model

Similar to the hide-and-seek problem, we define P(T) to be the set of all possible

walks of length at most T that can be taken by searchers. Similarly, let R(T) denote the

set of all possible walks of length T + 1 that can be taken by the intruder, thus potentially

evading the searchers for more than T time periods. Let dpr be a parameter whose value is

1 if a searcher following walk p detects an intruder following walk r, and 0 otherwise. This









problem can be formulated as a set covering problem, which we denote by PE:


PE: minimize A, (5-13)
pEP(T)

subject to dpA > 1 Vr e R(T) (5-14)
pEP(T)

A, {0,1} Vp P(T), (5-15)

where Ap again equals 1 if and only if a searcher is assigned to follow walk p. The

objective function (5-13) minimizes the number of searchers. Constraints (5-14) ensure

that for each possible intruder walk of length T + 1, at least one searcher is selected to

detect it.

5.3.2 Solution Approach

Once again, rather than enumerating all possible search patterns of length at most

T, and all evasion patterns of length T + 1, we propose a dynamic column and row

generation algorithm to solve the problem. We start with a subset of search patterns

P'(T) C P(T) and evasion patterns R'(T) C R(T), and solve a limited pursuit evasion

(LPE) problem, given P'(T) and R'(T). We next describe how we generate new search

and evasion patterns as needed.

5.3.2.1 Searcher's problem

Given a subset R'(T) of intruder walks, the searcher's problem is similar to the

pricing problem in the hide-and-seek problem. Let 7, be the dual variable associated with

the constraint of type (5-14) corresponding to intruder walk r E R'(T). Also, let y, be

a decision variable that equals 1 if an intruder following walk r is detected by a searcher

following a walk that we generate, and 0 otherwise, Vr E R'(T). Given an optimal dual

solution y of the linear programming relaxation to LPE, we solve the following pricing

problem to seek a A-variable having a negative reduced cost: max reiR'(T) I/ subject

to the restriction that (yl,..., YR'(T)r) corresponds to a set of intruder walks detected by a

searcher walk of length no more than T.









Similar to the previous pricing problem, this pricing problem can also be formulated

as a mixed-integer programming problem on a time-expanded network consisting of

T + 1 stages. In particular, we create a node Nit for each i E N, t = 0,..., T. Unlike

the previous problem, searchers do not necessarily have to move in each time period.

Therefore, we connect each node to its neighbors and the (< li of itself in the next stage.

We define a binary variable xt for all i e N, t = 0,..., T, which equals 1 if the searcher

is located at node i at time t, and 0 otherwise. We also define a parameter dt = 1 if a

searcher located at node i at time t can detect an intruder following walk r. An integer

programming formulation of the pricing problem can be given as follows:


maximize ^ 1' (5-16)
rER'(T)

subject to x 1 Vt 0,...,T (5-17)
iEN

x< > t- 1 Vi N,t=1,...,T (5-18)
jEA(i)
T
yr,< Y t Vr R'(T) (5-19)
t=0 iEN

0 < r < 1 Vr e R'(T) (520)

x e {0,1} V1 e N,t= 1,..., T. (5-21)


Constraints (5-17) ensure that the searcher cannot be located at multiple nodes

simultaneously. Constraints (5-18) model the fact that the searcher can either stay at

the same node, or can move to an .,.li i:ent node at each period. Constraints (5-19)

represent the condition that the searcher detects intruder r E R'(T) only if it moves to

a node where it can detect the intruder during the pursuit. We note that the y-variables

can be relaxed as continuous variables in this case, too. This property allows the number

of binary variables in the searcher's problem to stay constant as new evasion paths for the

intruder are discovered. As before, if the optimal objective function value of (5-16)-(5-21)

is greater than 1, then we have found a variable whose reduced cost is negative.









5.3.2.2 Intruder's problem

Given an integer feasible solution A of LPE, in which a subset of the searchers

has been selected, we need to solve a subproblem for the intruder to seek a walk that

evades the searchers for more than T time units. To solve this problem, we generate a

time-expanded network consisting of T + 1 stages, which contains a node Nit for each

i E N, t = 0, ,T. We add a dummy start node s, and connect s to all nodes in the first

stage, which corresponds to the initial intrusion at t = 0. Similar to the network generated

for the searcher's problem, we connect each node to its neighbors and the copy of itself

in the next stage. Finally, we connect all nodes in the last stage to a dummy node q. We

then trace each selected searcher's walk, and eliminate the nodes (and the corresponding

arcs) from the time-expanded network that would lead to the detection of the intruder.

After constructing the time-expanded network as described, we seek a feasible s-q

path on the network by a standard breadth-first-search algorithm, which works in O(N2T)

time in the worst case if G is dense. If such a path exists, then it corresponds to a walk r

that the intruder can take to avoid detection for T + 1 time units. In this case, we add r

to R'(T), and generate the associated constraint of type (5-14). On the other hand, if no

such path exists, then A is a feasible solution of PA.

5.3.2.3 Branch-cut-price algorithm

The pursuit evasion problem can be solved using the following branch-cut-price

algorithm.

* Step 0: C'!....-. a set of initial search and evasion walks so that every node is seen by
at least one searcher.

Step 1: Solve the linear programming relaxation of LPE, and generate columns by
solving the searcher's problem until linear programming relaxation of LPE has been
solved to optimality. If all columns are integer-valued, go to Step 2. Else, branch,
and go back to Step 1 for the subproblems.

Step 2: Evade by solving the intruder's problem. If the intruder can evade the
searchers, then add the evasion walk as a cutting plane of type (5-14), and go back
to Step 1. Else, stop processing the current subproblem with an integral solution.









We initialize our algorithm by generating a stationary searcher that stays at node i for T

periods, and a stationary evader that stays at node i for T + 1 periods, for each i E N. We

discuss several branching strategies that can be used for Step 1 in Section 5.5.

5.4 Patrol Problem

5.4.1 Problem Description

The setting for the patrol problem that we consider in this section is as follows.

The searchers are assigned to repeated patrol circuits, which they follow indefinitely. We

assume that the period of a patrol circuit is bounded from above by a parameter K, where

K > 1. Such a restriction might be due to a capacity or range limit of the searchers, or

due to desired frequency of visits to individual nodes. Initially there is no intruder in the

system. The intruder observes the searchers for a duration of time that is long enough

to identify search patterns, and then picks a node and time to enter the system. It then

tries to stay in the system as long as possible without being detected. The goal is to find

the minimum number of searchers needed, along with the corresponding patrol routes, to

ensure that the intruder is detected within T time periods after the intrusion. We note

that this problem also reduces to the minimum dominating set problem (for K = 1,

S(i) = A(i), Vi E N and T = 0), and hence is also NP-hard.

5.4.2 Mathematical Model

Let PC(K) denote the set of all possible circuits of period no more than K that can

be taken by searchers. Similarly, let R(T) denote the set of all possible walks of length

T + 1 that can be taken by the intruder. Let dp, be a parameter whose value is 1 if a

searcher following circuit p detects an intruder following walk r, and 0 otherwise. Let us

define a binary variable Ap for all p E PC(K), which equals 1 if a searcher is assigned to

follow circuit p, and 0 otherwise. The patrol problem can be formulated as a set covering

problem as follows.


PP: minimize Ap (5-22)
pEPC(K)









subject to d,,, > 1 Vr R(T) (5-23)
pEPC(K)

Ap {0,1} Vp P(K) (5-24)

We propose a branch-cut-price algorithm similar to the pursuit evasion problem for solving

this problem. We start with a subset of patrol routes P'`(K) and evasion walks R'(T),

and solve the resulting limited patrol problem (LPP). We generate new patrol routes and

evasion walks as needed.

5.4.2.1 Searcher's problem

The searcher's problem is similar to the pricing problems discussed before. Let 7, be

the dual variable associated with the constraint of type (5-23) corresponding to intruder

walk r c R'(T). We define y, to be a decision variable that equals 1 if an intruder

following walk r is detected by a searcher following a patrol circuit that we generate, and

0 otherwise, Vr e R'(T). Given an optimal dual solution 7 of the linear programming

relaxation to LPP, we solve the following pricing problem to seek a A-variable having a

negative reduced cost: max rER'(T) ^ I/ subject to the restriction that (yi,... YR'(T)I)

corresponds to a set of intruder walks detected by a searcher following a patrol circuit of

period no more than K.

We can solve the searcher's problem by solving a series of integer programs as follows.

Let T denote the length of the current circuit under consideration. By considering different

values of T E {1,..., K} we can find a circuit that optimizes the searcher's problem. Note

that some values of T may not correspond to any circuits in G. For each value of -, we

generate a time-expanded network containing T + 1 levels, where the first level corresponds

to the initial deployment of the searcher, and the last lw-r is a dummy li.-r that we use

to model the recurring patrol patterns. We connect each node to its neighbors in the next

stage. As before, we define a binary variable x for all i C N, t = 0,..., T, which equals

1 if the searcher is located at node i at time t. We also define a parameter dr, = 1 if a

searcher located at node i at time t can detect an intruder following walk r E R'(T). We









solve the following integer program to seek an optimal searcher circuit visiting exactly 'r

nodes.


maximize ^ n (5-25)
rER'(T)

subject to x 1 Vt 0,..., (5-26)
iEN
t < x -1 Vi e N, t 1, (5-27)
jEA(i),j i

x x Vi EN (5-28)
T-1
yr < Vr e R'(T) (5-29)
t=0 iEN

0 < yr < 1 Vr R'(T) (5-30)

x~e {0,1} Vi e N, t 0,...,7. (5-31)


Constraints (5-26) and (5-27) ensure that each feasible solution corresponds to a walk.

Constraints (5-28) guarantee that the first and the last nodes visited by the searcher

are the same, and hence the searcher's walk forms a circuit. Finally, Constraints (5-29)

relate the x- and y-variables, where we can once again relax integrality restrictions on the

y-variables.

Integer programs corresponding to different values of r can be solved in any sequence.

We note that a good solution obtained by solving the searcher problem for a particular

value of r can be used to prune problems to be solved later for different values of 7 by

bound. Therefore, we can start by solving a searcher problem for the largest value of r,

since a searcher following a longer circuit is more likely to detect more intruder walks.

Also, we skip any r if we determine that no circuit of length r exists in G.

5.4.2.2 Intruder's problem

Given a selected subset of the searcher circuits, the intruder seeks a way of -I ii-; in

the system as long as possible without being detected. Our main observation is that the









state of the system with respect to the location of the searchers is cyclic, and its period is

equal to the least common multiple of the periods of searchers' circuits, which we denote

by L. Therefore, the intruder can stay in the system indefinitely if it can identify a walk

that allows it to return to its initial location at the end of a multiple of L steps.

To solve the intruder's problem, we can generate a time-expanded network consisting

of L stages similar to the pursuit evasion problem. We connect each node to the copy

of itself and its neighbors in the next stage by a directed arc having length 1. We also

connect the nodes corresponding to stage L to the nodes to their neighbors in the first

stage with directed arc having length 1 (modeling the fact that the overall search pattern

repeats after L periods). We add a dummy start node s and a dummy end node q. We

connect s to all nodes by a directed arc having length 0 (reflecting our assumption that

the intruder can enter the system at any time and location), and connect all nodes to q

by a directed arc having length 0. Finally, we trace each selected searcher's circuit, and

remove nodes and arcs from the intruder's network that would lead to the detection of the

intruder by the searcher.

We can solve the intruder's problem on the generated graph by seeking a longest s-q

path. We first seek a topological ordering of the nodes using a standard depth-first-search

algorithm, whose complexity is O(N2L) for a dense G. Since a directed graph is .,. iv lic

if and only if it is has a topological order, this step identifies whether the graph is cyclic.

If there is a cycle in this graph, then the intruder can stay in the system forever without

being detected by the searchers. In this case, we generate a cut of type (5-23), and stop.

Otherwise, the graph is .,. iI. ii and given a topological order of the nodes, a longest path

can be found in polynomial time by a dynamic programming algorithm whose complexity

is O(N2L) (Am!li et al., 1993). If the length of a longest path is greater than T, then the

intruder can successfully evade the searchers. We can use such a longest walk to generate

a cut of type (5-23).









5.4.2.3 Branch-cut-price algorithm

The patrol problem can be solved using a branch-cut-price algorithm similar to the

algorithm we propose for the pursuit evasion problem.

* Step 0: C'! .. ..-- an initial set of patrol circuits and evasion walks so that so that
every node is seen by at least one searcher.

Step 1: Solve the linear programming relaxation of LPP, and generate columns
by solving the searcher's problem until the linear programming relaxation of LPP
has been solved to optimality. If all columns are integer-valued, go to Step 2. Else,
branch, and go back to Step 1 for the subproblems.

Step 2: Evade by solving the intruder's problem. If the intruder can evade the
searchers for more than T time periods, then add the evasion walk as a cutting plane
of type (5-23), and go back to Step 1. Else, stop processing the current subproblem
with an integer solution.

As before, we initialize our algorithm by generating a stationary searcher that r l- at

node i, and a stationary evader that li- ,- in node i for T + 1 periods, for each i E N. We

discuss several branching strategies that can be used in Step 1 in the next section.

5.5 Branching Strategies

If an optimal solution of the linear programming relaxation of the master problem

is integer feasible after all necessary variables and constraints have been added in the

column- and row-generation phases, then we have found an optimal solution. Else, if

some A-variable is fractional, then we need to branch. C'! ....-ig a branching rule that

does not make the pricing problem too difficult to solve is essential in column generation

algorithms. Finding such a branching rule is relatively easy if the master problem is a set

packing or set partitioning problem, since in this case it is possible to branch by choosing

a subset of variables X and setting x = 0, Vx E X in one branch, and x = 0, Vx X in

the other branch. These branching constraints can be added implicitly by removing the

corresponding variables from the problem, and adjusting the pricing problem so that those

variables are never generated in future iterations. Since no constraints are added explicitly,

no new dual variables are created. Therefore, the structure of the pricing problem Il li-









the same throughout all nodes of the branch-and-bound tree (Barnhart et al., 1998; Jiinger

and Thienel, 2000).

However, our master problem is a set cover problem, and the approach described

above is not directly applicable. Any constraint on a group of variables for our set cover

problem would necessitate adding a branching constraint like :.xe x < A in one branch,

and Yxex x > A + 1 (for some suitable A) in the other branch. Since these constraints

cannot be added implicitly, we need to handle a new dual variable for each branching

constraint in the subproblem.

As an alternative, we propose a multi-tiered branching rule. Given a fractional

solution A, we first evaluate the value of each constraint expression vr EpEp(T) dpr~p.

If there exists a fractional v, value for some r E R'(T), then we branch on the constraint

(5-23) corresponding to r as follows. In the up-branch, we simply change the right-hand-side

value of the constraint as [vr]. On the down-branch, we set the upper bound of the

constraint expression to Lvr] (and convert it to an equality constraint if Lvr] = 1).

Note that branching on a constraint in this manner does not introduce any new dual

variables or constraints that need to be considered in any of the pricing problems. Another

benefit of this branching scheme is the following. If we branch down on a constraint and

obtain ,,pc d,,Ap 1, we can then use set-partition type of branching schemes on the

corresponding subproblem. This allows us to branch on a group of A-variables in the

subsequent branches without d. -I i vi-.; the pricing problem structure.

Note that sign of the dual variable 7, can change after branching down on the

constraint (5-23) corresponding to r, which makes y/ = 0 optimal regardless of the values

of x-variables. Therefore, we need to add constraints that force y, to 1 if the searcher's

chosen path detects intruder r. Hence, we add constraints


y > x Vi e N, j S(i), t 0,...,T (5-32)

yr > dx Vr e R'(T), i N, t 0,...,T (5-33)









yr > d',x4 Vr e R'(T), ie N, t ,...,-r 1


to the searcher's pricing problem for the hide-and-seek, pursuit evasion, and patrol

problems, respectively.

In general it is possible to have a fractional solution A for which all v-values are

integral, and therefore the branching rule described above cannot be applied. In such

cases we can apply a simple variable-based branching rule. If there is a variable Ap whose

value is fractional, we can simply create two branches with Ap = 0 and Ap = 1. In the

down-branch, we simply eliminate the column corresponding to Ap from the set covering

formulation. In the up-branch we need to adjust the right-hand-side vector of our set

covering problem before eliminating Ap. In either case, we need to adjust the pricing

problems so that the same variable cannot be regenerated. Recall that the solution

methods we propose for all three pricing problems are based on a time-expanded network

formulation. Let us denote by W(p) the set of time-expanded node indices corresponding

to a searcher walk (or circuit) p. We can enforce the condition that a walk p is not

generated again by adding the following constraint to the corresponding searcher's pricing

problem:


x < |W(p)|-l. (5-35)
(i,t)EW(p)

Branching on a single variable is likely to be quite weak on the down-branch since

only one particular searcher walk (or circuit) is avoided. Hence, we only apply this rule if

our constraint-based branching rule fails. Also note that the difficulty of solving the set

covering problem does not increase under either branching rule, since no constraints are

added explicitly while branching.

5.6 Computational Results

We implemented the algorithms discussed in the previous sections in C++ on a

Windows XP PC with a 3.4 GHz CPU and 2 GB RAM. We used CPLEX 11.2 to solve the


(5-34)









pricing problems and the linear programming relaxations of our set covering formulations.

We implemented our algorithms for solving the intruder's problem using Boost Graph

Library version 1.36 (Siek et al., 2001). Our base set of test problem instances consists

of 150 randomly generated problem instances for which the expected edge density of the

graph (measured as iN Vi where we do not consider self-loop edges in calculating

edge density) is 10' the number of nodes N ranges from 5 to 25, and the maximum

allowed time to detection T ranges from 0 to 5. In generating instances we first picked

a random subset of edges so that the edge density is 10' and if necessary added the

minimum number of edges needed to make the graph connected (see Siek et al. (2001)).

We then added self-loop edges, and we generated five problem instances for each problem

size, which is determined by the number of nodes. Finally, we solved each problem

instance with different values of T E {0,..., 5} for the hide-and-seek, pursuit evasion

and patrol problems. In each case, we assume that a searcher located at node i E N can

observe node i and all nodes .,.1] i:ent to it, and hence we set S(i) = A(i), for all i E N.

We imposed a 1200-second time limit past which we stopped the execution of an algorithm

in all our experiments.

Recall that all problems that we consider in this chapter reduce to the minimum

dominating set problem for T = 0. We use this property to calculate an initial upper

bound by solving our LHS formulation, where we initialize R'(0) by adding a searcher

located at each node i E N.

Our first experiment focuses on the hide-and-seek problem. For this experiment,

we executed our branch-and-price algorithm described in Section 5.2.2.2 on our base

data set. All 150 problem instances in our data set were solved to optimality within 12.8

seconds. Table 5-1 di-tpl1i- the average number of branch-and-bound nodes evaluated

in our branch-and-price algorithm for different values of N and T. Each value represents

the average of the results obtained on five randomly generated graphs. We observe that

the number of branch-and-bound nodes that are explored increases as N increases, which











Table 5-1. Average number of branch-and-bound nodes explored for hide-and-seek problem
N T=0 T=1 T=2 T=3 T=4 T=5
5 1 1 1 1 1 1
10 1 1.4 1.8 1.4 1.4 1
15 1 1 1 1 1 1
20 1 2.6 3.4 1.8 1.8 1
25 1 7.4 2.2 2.6 1.4 1


is expected since the difficulty of the problem increases with increasing N. Table 5-2

Table 5-2. Average number of searchers needed for the hide-and-seek problem
N T=0 T=1 T=2 T=3 T 4 T=5
5 2 1.8 1 1 1 1
10 3.4 2.8 2 1.8 1.6 1
15 4.6 3.2 2.2 2 2 1.6
20 4.6 3.6 2.6 2 2 1.8
25 5.8 3.8 2.8 2 2 2



shows the average number of searchers needed for different values of N and T, where once

again each value is an average of five problem instances. We note that the T = 0 column

corresponds to the cardinality of the minimum dominating set. Table 5-2 reveals that

the number of searchers needed increases as the graph gets larger, and decreases as the

maximum allowed time to detect the intruder increases.

We analyze the performance of our branch-cut-price algorithm described in Section

5.3.2.3 for the pursuit evasion problem. Table 5-3 shows that our algorithm was able

to solve 128 out of the 150 instances within the prescribed time limit. As expected,

the difficulty of the problem increases as N and T increase, since this setting allows for

more evasion routes for the intruder, and hence requires the searchers to develop more

sophisticated routes.

Table 5-3. Number of instances that are solved within time limit for the pursuit evasion
problem
N T=0 T=1 T=2 T=3 T 4 T=5
5 5 5 5 5 5 5
10 5 5 5 5 5 5
15 5 5 5 5 5 5
20 5 5 5 3 1 2
25 5 5 4 3 0 0



Table 5-4 dip-~'i the average number of branch-and-bound nodes evaluated in

our branch-cut-price algorithm for the pursuit evasion problem. We note that for this











Table 5-4. Average number of branch-and-bound nodes explored for the pursuit evasion
problem
N T=0 T=1 T=2 T=3 T=4 T=5
5 1 1 1 1 1 1
10 1 1.8 3 1.8 1.8 1
15 1 5 3.4 6.2 4.2 3
20 1 28.2 109 334.6 101 14.2
25 1 23.4 869 324.6 41.4 11.4


problem, processing each node takes longer than the hide-and-seek problem, and therefore

our algorithm can process fewer nodes within the time limit. Finally, Table 5-5 di-ipl-,

the average number of searchers needed for this problem, where we use the best known

solutions for instances that were not solved to optimality within the time limit. A

Table 5-5. Average number of searchers needed for the pursuit evasion problem
N T=0 T=1 T=2 T=3 T=4 T=5
5 2 1.8 1 1 1 1
10 3.4 3 2.2 2 2 2
15 4.6 3.4 2.6 2 2 2
20 4.6 4 3 2.8 2.8 2.6
25 5.8 4 3.4 3.2 3.2 3



comparison of Tables 5-2 and 5-5 reveals that the number of searchers needed for the

hide-and-seek problem is less than that for the pursuit evasion problem. This result is

not surprising since the intruder is stationary in the former problem, while it can move to

avoid the searchers in the latter problem.

Table 5-6. Number of instances that are solved within time limit for the patrol problem
N T=0 T=1 T=2 T=3 T=4 T=5
5 5 5 5 5 5 5
10 5 5 5 5 5 5
15 5 5 5 5 5 5
20 5 1 1 1 1 0
25 5 0 0 0 0 0



Our last experiment evaluates our branch-cut-price algorithm described in Section

5.4.2.3 for the patrol problem. Table 5-6 reveals that our algorithm for the patrol problem

can solve fewer instances in our data set within the time limit compared to our algorithms

for the other problems. This can be explained by (i) our assumption that the intruder can

pick a time to enter the system, and (ii) our solution algorithm for the searcher's problem,

which requires solving multiple mixed-integer programs. The first factor makes it easier for











the intruder to evade the searchers, while the second factor makes the searcher's problem

more difficult to solve. The combined effect of these factors is that processing each node

takes longer than the other problems, which can also be seen by considering Table 5-7.

Table 5-7. Average number of branch-and-bound nodes explored for the patrol problem
N T=0 T=l T=2 T=3 T 4 T=5
5 1 1 1 1 1 1
10 1 1 1.4 1.4 1.4 1.4
15 1 44.6 20.2 6.2 11.4 3
20 1 12.2 5.4 4.2 3.4 1
25 1 5 3.2 2.1 1.2 1



Finally, we report our results on the number of searchers needed for the patrol

problem in Table 5-8. As before, our calculations are based on the best known solutions

and do not necessarily correspond to optimal solutions for the instances that were not

solved within the time limit. However, we observe that the number of searchers needed for

the patrol problem is larger than the other two problems as expected.

Table 5-8. Average number of searchers needed for the patrol problem
N T=0 T=1 T=2 T=3 T 4 T=5
5 2 1.8 1.6 1.2 1 1
10 3.4 3 2.8 2.6 2.6 2.6
15 4.6 3.4 3 2.8 2.8 2.6
20 4.6 4.2 3.8 3.1 3 2.8
25 5.8 5.3 4.7 3.8 3.6 3.2









CHAPTER 6
CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS

In this chapter, we focus on future research areas regarding the problems we have

described in the previous chapters. We analyze the techniques we employ. .1 identify

associated weaknesses, and -,-.-.i -1 improvements. We also discuss a research topic

that is based on a technique for reformulating the master problem for a class of bi-level

optimization problems.

6.1 Stochastic Edge-Partition Problem

For solving the stochastic edge-partition problem (C!i lpter 2, see also Taskm et al.

(2009a)) we first developed an integer programming formulation of the problem, and

prescribed a bi-level reformulation of the problem that has integer variables in both

stages. Our computational tests showed that both the direct solution of the integer

programming formulation and our integer programming-based cutting plane algorithm for

the bi-level formulation are capable of solving only small problem instances to optimality.

We then designed a hybrid integer programming/constraint programming algorithm to

overcome the computational difficulties encountered by the first two approaches. Our

hybrid approach first allocates node copies that are to be distributed across configurations

using an integer programming formulation, and then assigns nodes to subgraphs using

a constraint programming algorithm. After assigning nodes to subgraphs, it partitions

edges to subgraphs for each scenario in a third stage, using another integer programming

formulation. Our computational experiments show that the hybrid approach significantly

outperforms the other approaches.

In our study, we have assumed that the number of subgraphs, IKI, is a part of the

problem input. In SONET network design application there is no practical limit on the

number of subgraphs, but a limit is specified to model the problem (Goldschmidt et al.,

2003; Sherali et al., 2000; Smith, 2005). ('!i ..-! ig IKI too small may make the problem

infeasible or suboptimal, and choosing IKI too large increases the difficulty of the problem









as we discussed in Section 2.4. In our experiments, we manually chose IKI sufficiently

large to yield a feasible solution in each problem instance. An area for future research

is to treat IKI not as a parameter but as a variable, and to find the smallest possible

value of IKI that guarantees the existence of an optimal edge partition that minimizes

the number of node copies. This problem appears to be very difficult in general, but some

upper bounds can be derived for our problem. We first note that choosing IKI = IE

is guaranteed to yield a feasible edge partition (and not eliminate any feasible edge

partitions from consideration), since each edge can be partitioned into a unique subgraph.

Furthermore, if a feasible edge partition having an objective function value z is known,

then [L/2J is an upper bound on IKI. This bound follows from the fact that there exists

an optimal solution that contains at least two nodes in each non-empty subgraph. Such

a z can be calculated by a simple improvement heuristic. We start with IKI = IE,

and initially assign each edge to a unique subgraph. We then seek two subgraphs that

can be merged without violating any constraints in any scenario, while also improving

the objective function value. If we find such subgraphs, we merge them and repeat this

process until no more subgraphs can be merged. Since this algorithm starts with a feasible

solution, and retains feasibility in each iteration, it yields a feasible solution. However,

our preliminary analysis -ii--.- -1- that the bound on IKI that we get using this approach

is quite weak. Improving bounds on IKI and finding the smallest possible number of

subgraphs that yields an optimal edge partition is a future research area.

6.2 Matrix Decomposition Problem

In C'! Ilpter 3 we have described exact decomposition algorithms for solving leaf

sequencing problems arising in IMRT treatment planning (also see Taskm et al. (2009b)).

Our solution algorithm for the matrix decomposition into apertures satisfying the

consecutive-ones property is based on an integer programming model for finding a set

of intensity values to be assigned to apertures, and a backtracking algorithm that forms

apertures by finding compatible leaf positions for each row. Our computational results









show that an optimal solution to many clinical problem instances can be found within

a few minutes. As such, not only can this algorithm reasonably be used in real clinical

settings, but also the bounds obtained from our algorithm can serve as benchmark criteria

to compare the performance of several heuristic methods prescribed for this problem.

Our solution technique for the consecutive-ones matrix decomposition problem and

our three-stage approach to the stochastic edge-partition problem are based on a similar

idea. In both problems, we add ,.-'-regate vy iil I! to our formulations, which describe

important structural properties of solutions, but are not enough by themselves to encode

complete solutions. In each case we represent the optimization problem in terms of our

. I:-. regate variables in a master problem, and provide a subproblem that seeks a complete

feasible solution corresponding to the values of the .,. .--regate variables chosen by the

master problem. In both applications, our master problems are discrete optimization

problems, which we solve using integer programming methods, and our subproblems are

discrete feasibility problems, which we solve using constraint programming methods.

Separating critical optimization decisions from feasibility decisions, and utilizing strengths

of integer and constraint programming techniques in a hybrid algorithm has allowed us

to obtain significantly better results than other methods. A i, r i theme in our future

research is going to be on generalizing this hybrid approach to handle a broader class of

problems.

In C'!i lter 4 we studied a different variant of the matrix decomposition problem,

in which the aperture matrices need to be rectangular in shape (see also Takmn et al.

(2008)). Rectangular apertures can be formed by using conventional jaws already

integrated into IMRT treatment devices, and do not need an advanced MLC system,

which is costly to manufacture and operate. In Chapter 4, we proposed an exact

optimization algorithm that can be used to i", i1. whether a jaws-only treatment system

can deliver fluence maps efficiently. Our algorithm is based on an integer programming









formulation, which we enhance using several valid inequalities and by partitioning the

problem into simpler problems.

We also derived a bi-level Benders decomposition algorithm for this problem.

The master problem of our Benders decomposition approach chooses a subset of

the rectangular shapes that can be used in decomposing the input matrix. Later, a

subproblem checks whether the selected subset of rectangles can completely decompose

the input matrix. Unfortunately, our computational tests showed that our Benders

decomposition algorithm is computationally inferior to the integer programming approach.

The main reason of slow convergence is the weakness of cuts generated in each iteration.

Specifically, given an infeasible subset of rectangles chosen by the master problem, our

subproblem (which is a linear programming problem) detects infeasibility, and returns

a cut, which is generated based on a dual extreme ray. However, this extreme ray is a

mathematical proof of infeasibility, but does not necessarily identify the underlying rea-

son of infeasibility. In other words, it does not identify which bixels in the input matrix

cannot be partitioned with the selected subset of rectangles. Furthermore, there are

typically many extreme dual rays for a single infeasible master problem solution, from

which several non-dominated cuts can be derived. One way of improving the convergence

can be applying a two-dimensional binary search algorithm on the input matrix to find

out which region of the input matrix cannot be partitioned with the selected rectangles.

That is, if the entire matrix cannot be partitioned, we try to partition the left-hand-side

and the right-hand-side halves of the matrix independently. If one of these submatrices

cannot be partitioned, we immediately have a more specific reason for infeasibility (and

hence a stronger cut), because this result implies that the infeasibility in a submatrix

needs to be fixed using a subset of the rectangles, which cover that part of the matrix.

This idea can be applied recursively to find possibly multiple infeasible regions of the

matrix. The cuts associated with these infeasible submatrices are much stronger than a

single cut derived based on the entire matrix. Furthermore, the information regarding the









description of infeasibility generated by analyzing submatrices in a single iteration can

only be retrieved after several iterations of the original decomposition algorithm, which

is based on solving the entire matrix only. This idea of obtaining a better description of

infeasibility by analyzing subsets of variables can be generalized to the general Benders

decomposition algorithm, and has the potential of improving the convergence properties in

many applications.

In our study, we have developed solution techniques for two versions of the matrix

decomposition problem, which apply to most available IMRT treatment machinery.

However, there are other types of machinery that have different aperture shape restrictions,

such as interdigitation or connectedness constraints (see e.g., Lim (2002)). In a related line

of research, we are planning to design exact optimization algorithms to solve other variants

of the matrix decomposition problem to optimality. Quantifying the effect of several shape

constraints enforced by different types of machinery on radiation delivery efficiency would

be a valuable contribution to the medical physics field.

6.3 Graph Search Problem

In C'!h pter 5 we considered three variants of a graph search problem: (i) a hide-and-seek

problem, where a set of searchers seek a stationary intruder, (ii) a pursuit evasion

problem, where the intruder moves to avoid detection, and (iii) a patrol problem, where

searchers follow recurring patrol routes. The aim in each problem is to find the minimum

number of searchers needed so that the intruder cannot stay in the system without being

detected for longer than a prespecified duration of time. We proposed a branch-cut-price

algorithm, which can be adapted to all three problems with certain modifications. Our

main contribution is that we do not make any assumptions on the topology of the input

graph, and our algorithms work on general graphs.

Even though all three problems that we consider are NP-hard, our computational

tests show that the hide-and-seek and pursuit evasion problems can be solved to

optimality for modest problems sizes within reasonable computational time. However,









our algorithm for the patrol problem can only solve small problem instances to optimality

within the limits that we enforced. Our future research will initially focus on improving

the performance of our algorithm. In particular, we are planning to (i) add heuristics for

the searcher's problem to seek variables having a negative reduced cost before switching

to our mixed-integer programming models, (ii) investigate flow-based formulations for the

searcher's problem, which might have a tighter linear programming relaxation than the

node-based formulations that we proposed, and (iii) seek valid inequalities to improve

dual bounds. We are also planning to extend our models and solution algorithms to

handle different types of searchers with different capabilities and costs. For instance,

some searchers might be able to detect the intruder from a longer distance, while some

others might be stationary but cheaper to operate. A related problem that can be

investigated is a setting in which we are given a fixed number of searchers, and which

must be coordinated so that the amount of time an intruder can stay in the system is

minimized.

Our set covering formulation, which is based on the idea that at least one searcher

must be chosen for each route the intruder can take, can be generalized to other variants

of the graph search problem. In particular, a problem that has been widely studied in the

literature assumes that the intruder can reside at the edges of the graph. This problem

has been investigated from a theoretical perspective (see, e.g., Dendris et al. (1997); Ellis

et al. (1994); Seymour and Thomas (1993)); however, to the best of our knowledge no

exact optimization method that works on general graphs has been proposed. We are

planning to extend our algorithm so that it can also solve this problem.

6.4 Master Problem Reformulation in Bi-Level Cutting Plane Optimization
Algorithms

In this line of research, we are planning to explore v--iv of reformulating master

problems to improve the coordination between the master and subproblems in bi-level

Benders optimization algorithms, resulting in faster convergence to an optimal solution.









The key concept behind this line of research is a surprising property of decomposition

algorithms, namely the existence of several alternative optimal solutions (or extreme rays)

to the dual of the second-stage problems, each resulting in a different Benders inequality.

There can be an exponential number of these inequalities, each of which is nondominated.

Worse, it is possible that each of these inequalities may need to be generated one at a

time after each iteration of the master problem. However, it turns out that it is possible

to reformulate the master problem to avoid this ::p. ii!' li I cut" difficulty in some cases.

In a stochastic SONET design problem (Smith et al., 2004), and in a product introduction

and interdiction game, Smith et al. (2008) consider the addition of a quadratic-order set

of variables in the master problem. These new variables are passed to the subproblem,

and a Benders cut is generated in terms of the new variables that implies all of the

(exponentially-many) Benders cuts that could have been generated in the original variable

space. This master problem reformulation technique has the potential to dramatically

reduce the number of iterations required by Benders decomposition to converge to an

optimal solution, with only a modest increase in the size of the formulation. In both cases

mentioned above, the trade-off of increasing model size to improve the strength of Benders

cuts was computationally beneficial.

Let us describe the idea in more detail in the context of a SONET network design

problem (Smith et al., 2004), which is similar to the edge-partition problem discussed

in C'i Ilpter 2. There exist a set of demand pairs (i,j) E E that may be satisfied on a

single communications network (all demand pairs have to be satisfied in our edge-partition

problem). The communications network is composed of i i-" k = 1,...,K. If both

clients i and j have been linked to ring k, then we may choose to satisfy the demand

request between i and j on ring k. Define y k to be a continuous variable that represents

the fraction of the demand between i and j that is satisfied on ring k in scenario q (these

variables are defined to be binary in our edge-partition problem since we assume that each

demand pair needs to be satisfied on a single ring). Let Eq be the set of demand pairs









(i,j) in scenario q. We also define xik as a binary variable equal to one if and only if client

i is assigned to ring k. Setting aside other issues in this problem such as ring capacity, a

subset of scenario restrictions is:


k < Xik Vq e Q, (i,j) e Et, k e K (6-1)

yk
where the existence of scenarios q E Q is due to uncertain demands between clients i and

j. Using a straightforward decomposition approach, the problem can be decomposed so

that the x-variables are parts of the master problem formulation, while the yq-variables

are determined in subproblems corresponding to scenarios q E Q. Constraints (6-1) and

(6-2) essentially state that in order for the communication demand between customers

i and j to be assigned to ring k, both customers i and j have to be assigned to ring k.

Unfortunately, cuts enforcing this relationship cannot be represented in the original space

of x-variables. Smith et al. (2004) show that there can exist an exponential number of

alternative dual solutions associated with a master problem solution represented by k:,

each leading to a non-dominated Benders cut. Then they reformulate the master problem

by adding uijk variables, which represent the minimum of xik and Xjk. In other words,

uijk = 1 if both customers i and j are assigned to ring k. Given the u-variables, the
constraints (6-1) and (6-2) can be replaced by


qk < uijk Vq e Q, (i,j) e Eq, k e K. (6-3)

Smith et al. (2004) show that a single Benders cut based on the u-variables dominates

an exponential number of cuts based on the original x-variables. In other words, adding

a quadratic number of variables to the master problem can save an exponential number

of iterations of the Benders decomposition algorithm. In this particular problem, the

authors recognized that the y-variables in the subproblem are dependent on min(xik, xjk),

and used this nonlinear relationship between the x-variables to reformulate the master









problem. This kind of relationship is quite common in bi-level optimization algorithms.

Our goal in this line of research will be to generalize this approach to other types

relationships between variables so that structures that can be exploited by master problem

reformulation are identified automatically. We observe that such nonlinear relationships

can be induced using suitable relaxations of the subproblem. For instance, assume that the

subproblem contains two constraints like


ayI + 0 2 + + al < I (6 4)

alyl + ,II_. + + a2yn < X2, (6-5)

where x1 and x2 are variables of the master problem. Since both (6-4) and (6-5) are of <

type, we can take the component-wise minimum of the two constraints to obtain

min(al, a )yi + min(at, a y2 + min(a, a )yn < xi (6 6)

min(al, a2)yi + min(a, a),y2 + + min(al, a )yn 2 x. (6 7)


Since the left-hand-sides of (6-6) and (6-7) are the same, we can combine the two

constraints into

min(a, a l)yi + min(a), ay2 + min(aa )Yn < min(xi,2) (6 8)


Constraint (6-8) describes a nonlinear relationship between the y- and x-variables. At this

point, the master problem can be reformulated by adding a variable v2 = min(xl, x2), and

the subproblem can be reformulated by using this variable as

min(a, a + min(a, a minl(a,) v 2 1 (69)
a1)yl 1 al)Y2 +''" + n ) n <_ 12"

Even though (6-9) is weaker than (6-4) and (6-5), this reformulation might improve

the convergence of the algorithm due to the ::II" "i i i-cut" behavior, especially in

the beginning iterations of the Benders process. In our future research we are planning









to formalize and generalize this idea, and determine how to apply this reformulation

approach to improve computational performance in bi-level cutting plane algorithms.









APPENDIX A
AN IP MODEL FOR C1-MATRIX DECOMPOSITION PROBLEM

In this appendix we discuss an integer programming approach to decomposing a

fluence map into a number of apertures and corresponding intensities that is based on

a model proposed by Langer et al. (2001). Given a maximum number of unit-intensity

apertures, v- T, this formulation determines the positions of the left and right leaves in

each row of each of these apertures. We develop the model by separately studying four

components:

* Fluence map requirements. Define, for each aperture t = 1,..., T and each bixel
(i,j) E {1,..., m} x {1,..., n}, a binary variable dt that is equal to one if and
only if bixel (i,j) is exposed, i.e., not covered by a left leaf or a right leaf. Since each
aperture has unit intensity, the following constraints then ensure that the desired
fluence map is delivered:
T
d bi V Il, ,...,m, j=l,...,n. (A-l)
t=1

Aperture del.i i.b.iil.hi constraints. Define, for each aperture t = 1,... ,T and each
bixel (i,j) E {1,..., m} x {1,..., n}, binary variables p and li that are equal to
one if and only if bixel (i,j) is covered by the right leaf or the left leaf in row i of
aperture t, respectively. The following set of constraints then ensure that each of the
T apertures is deliverable:

pt + I+dt 1 Vi 1,...,m, j 1,...,n, t 1,...,T (A 2)
Pij < pij+l V i- 1,...,m, j,= 1,..., n- 1, t= 1,...,T (A-3)
it < i, V i 1,...,m, j=2,...,n, t= 1,...,T. (A-4)

In particular, constraints (A-2) state that each bixel is either covered by a
right-hand leaf, covered by a left-hand leaf, or uncovered (where the d-variables
are included only for convenience and can be substituted out of the formulation).
Constraints (A-3) and (A-4) state that if any bixel (i,j) is covered by a right-hand
leaf (resp. left-hand leaf), then bixel (i,j + 1) (resp. (i,j- 1)) should be covered by a
right-hand leaf (resp. left-hand leaf) as well.

Beam-on-time. We associate a binary variable zt with each aperture t = 1,...,T
that is equal to one if there are uncovered bixels in aperture t and zero otherwise, so









that the beam-on-time is simply given by


T
Szt. (A5)
t= 1

While Langer et al. (2001) impose the following constraints to ensure that these
variables have (at least) their desired value:
m n
Z di< (mn)zt Vt ,...,IT, (A-6)
i=1 j 1

we note that the following stronger formulation, which would actually not require
enforcing the z-variables to be binary, can be obtained by di-.i._r:egating (A-6).

dt < z V i= 1,..., j 1,...,n, t= 1,...,T. (A-6')

Note that this model allows zt to be equal to one even if in aperture t no bixels are
exposed, so that formally speaking (A-5) is an upper bound on the beam-on-time.
The objective function ensures that the z-variables take on their minimum possible
value.

Number of apertures. We associate a binary variable gt with each aperture t =
1,..., T 1 that is equal to one if aperture t is different from aperture t + 1 and zero
otherwise. The number of setups is then given by

T
Egt. (A-7)
t=1

(If any aperture is used more than once but separated by another one, we consider
the second occurrence of the aperture to be a new setup. However, when minimizing
total treatment time there alv--,i exists an optimal solution in which identical
apertures are delivered sequentially.) Now let c and uj be auxiliary binary
variables such that the former is equal to one if bixel (i,j) is exposed in aperture t
but not in aperture t + 1 and zero otherwise, and the latter is equal to one if bixel
(i,j) is covered in aperture t but not in aperture t + 1. This relationship is stated by
-cj dij -d^
Langer et al. (2001) then use the following constraints to ensure that the variables gt
have (at least) their desired value:


i (c + U) < (mrn)g Vt ,...,T -. (A-9)
i= 1 j= 1









However, note that again a stronger set of inequalities (that permit g to be
equivalently relaxed as continuous variables) is obtained using di-,.--_egation:

c + U < g Vi -1,...,m, j 1, ... ,n, t = 1, ... ,T 1. (A-9')

Similar to the case of the beam-on-time, this model allows gt to be equal to one even
if apertures t and t + 1 are identical, although our objective function ensures that the
g-variables are chosen sufficiently small.

Langer et al. (2001) then study the problem of minimizing the number of setups

(A-7) subject to the constraints (A-2)-(A-4), (A-6), (A-9), the constraint that the

beam-on-time is minimal:
T
z < (A-10)
t=1
and binary constraints on the variables, where we recommend determining z via one of the

polynomial-time procedures mentioned in Section 3.1. We note that an equivalent model is

obtained by simply setting T = 5, which reduces the problem dimension, and hence should

be more efficient than adding a beam-on-time constraint.

We wish to minimize the total treatment time as measured by

T T
Wi gt + W2 Z (A-11)
t=1 t=1

subject to constraints (A-2)-(A-4), (A-6'), (A-9'), and binary constraints on the

appropriate variables (and hence we do not impose (A-10)).









APPENDIX B
COMPLEXITY OF C1-PARTITION

Proposition 3. CI-PARTITION is strongly NP-complete.


Proof. Let ( be the subset of {1,..., L} such that E c if and only if x > 0. Formally

speaking, the problem size is given by log2(L), n, and | (since the zero entries of x need

not be encoded).

Let KC denote the set of all O(n2) n-dimensional binary vectors whose ones appear

consecutively, where vk is the binary vector corresponding to k E /C. Consider a guessed

solution that consists of Il -dimensional nonnegative integer vectors dk, Vk E /C, where

dke denotes the number of times leaf position vk, k E /C, is assigned to intensity E c.

Since all dkW < L in some feasible solution, we restrict the guessed d-vectors as such.

The size of the guessed vectors is thus O(r2ll log2(L)). We can verify whether or not

CkEkC Z dk3vk = b in 0(n2| |) additions. Therefore, CI-PARTITION is in NP.
To show that CI-PARTITION is NP-complete, we reduce 3-PARTITION to it. 3-

PARTITION is a strongly NP-complete problem and seeks whether a given multiset of

integers can be partitioned into triplets having the same sum. Formally, it can be defined

as follows (see Garey and Johnson (1979)):

3-PARTITION

INSTANCE: A multiset A of 3v positive integers al,..., a3, and a positive integer B

such that B/4 < ai < B/2 for i = 1,..., 3v and such that E3, a, = vB.

QUESTION: Can A be partitioned into v disjoint multisets A1,..., A, such that

jA aj = B for i =,..., v?

Given an arbitrary instance of 3-PARTITION, we construct an instance for CI-PARTITION

as follows. First, we define x to be an integer vector whose fth component, x^, is equal to

the number of indices i for which ai = Furthermore, we let b be a (2v 1)-dimensional









vector of the form [B 0 B 0 ... 0 B]. We construct a feasible solution to C1-

PARTITION that employs only the odd-indexed unit vectors of IC. Denote these vectors as

el, e3,..., e2~-l, and index their associated d-vectors as dl, d3,..., d2,-l.

Assume that the 3-PARTITION instance is a yes-instance, and hence there exist

multisets A1,..., A, such that EA aj = B. In this case, a feasible solution of the

CI-PARTITION instance lets d2j-l,L be the number of elements of intensity e in Aj, for

each j = 1,..., v, and assigns dk = 0 for all other k. Similarly, suppose that the CI-

PARTITION instance is a yes-instance. Since all positive values in b are ..1i i,:ent to 0, in

any feasible solution to the instance of CI-PARTITION, we may only use leaf positions that

expose a single odd-index bixel. Also, since B/4 < ai < B/2, V i = 1,..., 3v, vector dk

must be used to deliver exactly three intensity values, for k = 1, 3,..., 2v 1. Then a

feasible solution of the 3-PARTITION instance is given as multisets A,,..., A, recovered

from di,, d,..., d2v-1 as described above. Therefore, an arbitrary 3-PARTITION instance

is a yes-instance if and only if the corresponding transformed CI-PARTITION instance is

a yes-instance. Since 3-PARTITION is strongly NP-complete, and since the transformation

provided is polynomial in terms of the size of the problem and the instance data, it follows

that CI-PARTITION is also strongly NP-complete. E









REFERENCES


Agazaryan, N., T. D. Solberg. 2003. Segmental and dynamic intensity-modulated
radiotherapy delivery techniques for micro-multileaf collimator. Medical Ph;,i. 30(7)
1758-1765.

Al!li R. K., H. W. Hamacher. 2005. A network flow algorithm to minimize beam-on-time
for unconstrained multileaf collimator problems in cancer radiation therapy. Networks
45(1) 36-41.

Alli R. K., T. L. Magnanti, J. B. Orlin. 1993. Network Flows: Th(.. ,; Algorithms, and
Applications. Prentice Hall, Upper Saddle River, NJ.

Aigner, M., M. Fromme. 1984. A game of cops and robbers. Discrete Applied Mathematics
8(1) 1-12.

Alpern, S. 1995. The rendezvous search problem. SIAM Journal on Control and Optimiza-
tion 33(3) 673-683.

Alpern, S. 2008. Hide-and-seek games on a tree to which Eulerian networks are attached.
Networks 52(3) 162-166.

Alpern, S., S. Gal. 2003. The Ti, **i of Search Games and Rendezvous, International
Series in Operations Research & Mi r,,ir,. i,, ,. Science, vol. 55. Kluwer Academic
Publishers, Dordrecht, The Netherlands.

Alspach, B. 2004. Searching and sweeping graphs: a brief survey. Le Matematiche 59
5-37.

Alspach, B., D. Dyer, D. Hanson, B. Yang. 2008. Time constrained graph searching.
Theoretical Computer Science 399(3) 158-168.

Andreas, A. K., J. C. Smith. 2009. Decomposition algorithms for the design of a
nonsimultaneous capacitated evacuation tree network. Networks 53(2) 91-103.

Baatar, D. 2005. Matrix decomposition with time and cardinality objectives: Theory,
algorithms, and application to multileaf collimator sequencing. Ph.D. thesis,
Department of Mathematics, Technische Universitat Kaiserslautern, Kaiserslautern,
Germany.

Baatar, D., N. Boland, S. Brand, P. J. Stuckey. 2007. Minimum cardinality matrix
decomposition into consecutive-ones matrices: CP and IP approaches. Integration of
AI and OR Techniques in Constraint P,. Problems. Berlin, 1-15.

Baatar, D., H. W. Hamacher, M. Ehrgott, G. J. Woeginger. 2005. Decomposition of
integer matrices and multileaf collimator sequencing. Discrete Applied Mathematics
152(1) 6-34.









Barnhart, C., E. L. Johnson, G. L. Nemhauser., M. W. P. Savelsbergh, P. H. Vance. 1998.
Branch-and-price: Column generation for solving huge integer programs. Operations
Research 46(3) 316-329.

Benders, J. F. 1962. Partitioning procedures for solving mixed variables programming
problems. Numerische Mathematik 4 238-252.

Benkoski, S. J., M.G. Monticino, J. R. Weisinger. 1991. A survey of the search theory
literature. Naval Research Logistics 38 469-494.

Bienstock, D., P. Seymour. 1991. Monotonicity in graph searching. Journal of Algorithms
12(2) 239-245.

Boc(.i v-r, A., N. Pisaruk. 2006. Detecting infeasibility and generating cuts for mixed
integer programming using constraint programming. Computers & Operations Research
33(10) 2777-2786.

Boland, N., H. W. Hamacher, F. Lenzen. 2004. Minimizing beam-on time in cancer
radiation treatment using multileaf collimators. Networks 43(4) 226-240.

Bortfeld, T. R., D. L. Kahler, T. J. Waldron, A. L. Bo.-r 1994. X-ray field compensation
with multileaf collimators. International Journal of Radiation O, ..'.1 ';, B:. /..i; Ph. -.
28(3) 723-730.

Calafiore, G. C., M. C. Campi. 2005. Uncertain convex programs: Randomized solutions
and confidence levels. Mathematical P,.;'i ,i,, i,:'il 102 25-46.

Care, C. C., R. Schultz. 1999. Dual decomposition in stochastic integer programming.
Operations Research Letters 24 37-45.

C'!. i!. yre, Y., F. Sempe, G. Ramalho. 2004. A theoretical analysis of multi-agent
patrolling strategies. AAMAS '04: Proceedings of the Third International Joint
Conference on Autonomous Agents and MI/ll'.r,. I,./ S';,-/.;l -. IEEE Computer Society,
Washington, DC, USA, 1524-1525.

Codato, G., M. Fischetti. 2006. Combinatorial Benders' cuts for mixed-integer linear
programming. Operations Research 54(4) 756-766.

Dai, J., Y. Hu. 1999. Intensity-modulation radiotherapy using independent collimators: an
algorithm study. Medical Ph;-.i- 26(12) 2562-2570.

Dai, J., Y. Zhu. 2001. Minimizing the number of segments in a delivery sequence for
intensity-modulated radiation therapy with a multileaf collimator. Medical Ph ;-.
28(10) 2113-2120.

Dendris, N. D., L. M. Kirousis, D. M. Thilikos. 1997. Fugitive-search games on graphs and
related parameters. Theoretical Computer Science 172 233-254.









Deng, J., T. Pawlicki, Y. C'! i. J. Li, S. B. Jiang, C. M. Ma. 2001. The MLC
tongue-and-groove effect on IMRT dose distributions. Ph;,-. in Medicine and Bi-
ology 46(4) 1039-1060.

Earl, M. A., M. K. N. Afghan, C. X. Yu, Z. Jiang, D. M. Shepard. 2007. Jaws-only IMRT
using direct aperture optimization. Medical Ph;,'. 34(1) 307-314.

Ehrgott, M., H. W. Hamacher, M. Nu3baum. 2008. Decomposition of matrices and static
multileaf collimators: A survey. C. J. S. Alves, P. M. Pardalos, L. N. Vicente, eds.,
Optimization in Medicine. Springer, New York, NY, 25-46.

Ellis, J. A., I. H. Sudborough, J. S. Turner. 1994. The vertex separation and search
number of a graph. Information and Computation 113(1) 50-79.

Engel, K. 2005. A new algorithm for optimal multileaf collimator field segmentation.
Discrete Applied Mathematics 152(1) 35-51.

Ernst, A. T., V. H. Mak, L. A. Mason. 2009. An exact method for the minimum
cardinality problem in the planning of IMRT. INFORMS Journal on Computing
Forthcoming.

Flocchini, P., M. J. Huang, F. L. Luccio. 2008. Decontamination of hypercubes by mobile
agents. Networks 52(3) 167-178.

Fomin, F. V., D. M. Thilikos. 2008. An annotated bibliography on guaranteed graph
searching. T7, .,., I.. ,,l Computer Science 399(3) 236-245.

Garey, M. R., D. S. Johnson. 1979. Computers and Intractability; A Guide to the Theory
of NP-Completeness. W. H. Freeman & Co., San Francisco, CA.

Goldschmidt, O., D. S. Hochbaum, A. Levin, E. V. Olinick. 2003. The SONET
edge-partition problem. Networks 41 13-23.

Goldstein, A. S., E. M. Reingold. 1995. The complexity of pursuit on a graph. Theoretical
Computer Science 143(1) 93-112.

Hahn, G. 2007. Cops, robbers and graphs. Tatra Mountains Mathematical Publications 36
1-14.

Hamacher, H. W., K.-H. Kilfer. 2002. Inverse radiation therapy planning -a multiple
objective optimization approach. Discrete Applied Mathematics 118(1) 145-161.

Haralick, R. M., G. L. Elliott. 1980. Increasing tree search efficiency for constraint
satisfaction problems. Arl.:i, .:,. Intelligence 14(3) 263-313.

Hooker, J. N. 2007. Planning and scheduling by logic-based Benders decomposition.
Operations Research 55 588-602.

Hooker, J. N., G. Ottosson. 2003. Logic-based Benders decomposition. Mathematical
P, i..i, .,,i i ,, 96 33-60.









Huang, H. X., P. M. Pardalos. 2002. A multivariate partition approach to optimization
problems. C;il ,, .. / 1. and S.,'1 m- A,.l;.-.:- 38(2) 265-275.

Isler, V., N. Karnad. 2008. The role of information in the cop-robber game. Theoretical
Computer Science 399(3) 179-190.

Jain, V., I. E. Grossmann. 2001. Algorithms for hybrid MILP/CP models for a class of
optimization problems. INFORMS Journal on Conr,;,/.:,ij 13 258-276.

Jotshi, A., R. Batta. 2008. Search for an immobile entity on a network. European Journal
of Operational Research 191(2) 347-359.

Jiinger, M., S. Thienel. 2000. The ABACUS system for branch-and-cut-and-price
algorithms in integer programming and combinatorial optimization. Software Prac-
tice and Experience 30(11) 1325-1352.

Kalinowski, T. 2004. The algorithmic complexity of the minimization of the number of
segments in multileaf collimator field segmentation. Tech. rep., Fachbereich Mathematik,
Universitat Rostock, Rostock, Germany.

Kalinowski, T. 2005a. A duality based algorithm for multileaf collimator field
segmentation with interleaf collision constraint. Discrete Applied Mathematics 152(1)
52 88.

Kalinowski, T. 2005b. Reducing the number of monitor units in multileaf collimator field
segmentation. Ph;-i. in Medicine and B':/. ',1;, 50(6) 1147-1161.

Kamath, S., S. Sahni, J. Li, J. Palta, S. Ranka. 2003. Leaf sequencing algorithms for
segmented multileaf collimation. Ph, /. in Medicine and B:.:.. .. ,. 48(3) 307-324.

Kamath, S., S. Sahni, J. Palta, S. Ranka. 2004a. Algorithms for optimal sequencing of
dynamic multileaf collimators. Ph,;-. in Medicine and B.:.. /.. ,; 49(1) 33-54.

Kamath, S., S. Sahni, J. Palta, S. Ranka, J. Li. 2004b. Optimal leaf sequencing with
elimination of tongue-and-groove underdosage. Ph,-i. in Medicine and B'.. /. 'i; 49(3)
N7-N19.

Kamath, S., S. Sahni, S. Ranka, J. Li, J. Palta. 2004c. A comparison of step-and-shoot
leaf sequencing algorithms that eliminate tongue-and-groove effects. PhI- in Medicine
and B:../ ..,/i, 49(14) 3137-3143.

Kamath, S., S. Sahni, S. Ranka, J. Li, J. Palta. 2004d. Optimal field splitting for large
intensity-modulated fields. Medical Ph;,. i- 31(12) 3314-3323.

Kikuta, K., W. H. Ruckle. 2007. Rendezvous search on a star graph with examination
costs. European Journal of Operational Research 181(1) 298-304.









Kim, Y., L. J. Verhey, P. Xia. 2007. A feasibility study of using conventional jaws to
deliver IMRT plans in the treatment of prostate cancer. Ph,.':. in Medicine and
Bf.. ,, 52(8) 2147-2156.

Kilfer, K.-H., M. Monz, A. Scherrer, H. L. Trinkaus, T. Bortfeld, C. Thieke. 2003.
Intensity modulated radiotherapy -a large scale multi-criteria programming problem.
OR Spektrum 25 223-249.

Langer, M., V. Thai, L. Papiez. 2001. Improved leaf sequencing reduces segments or
monitor units needed to deliver IMRT using multileaf collimators. Medical Ph -: -
28(12) 2450-2458.

LaPaugh, A. S. 1993. Recontamination does not help to search a graph. Journal of the
ACI 40(2) 224-245.

Laporte, G., F. V. Louveaux. 1993. The integer L-shaped method for stochastic integer
programs with complete recourse. Operations Research Letters 13 133-142.

Lee, E. K., T. Fox, I. Crocker. 2000a. Optimization of radiosurgery treatment planning via
mixed integer programming. Medical Ph;,' 27(5) 995-1004.

Lee, E. K., T. Fox, I. Crocker. 2003. Integer programming applied to intensity-modulated
radiation treatment planning. Annals of Operations Research 119 165-181.

Lee, Y., H. D. Sherali, J. Han, S. Kim. 2000b. A branch-and-cut algorithm for solving an
intra-ring synchronous optical network design problem. Networks 35(3) 223-232.

Lenzen, F. 2000. An integer programming approach to the multileaf collimator
problem. Master's thesis, Department of Mathematics, University of Kaiserslautern,
Kaiserslautern, Germany.

Lim, G. J. 2002. Optimization in radiation treatment planning. Ph.D. thesis, Computer
Sciences Department, University of Wisconsin-Madison, Madison, WI.

Lim, G. J., J. Choi. 2007. A two-stage integer programming approach for optimizing leaf
sequence in IMRT. Tech. rep., Department of Industrial Engineering, University of
Houston, Houston, TX.

Lim, G. J., M. C. Ferris, D. M. Shepard. 2004. Optimization tools for radiation treatment
planning in Matlab. M. L. Brandeau, F. Sainfort, W. P. Pierskalla, eds., Operations
Research and Health Care: A Handbook of Methods and Applications. Kluwer Academic
Publishers, Boston, MA, 775-806.

Lim, G. J., M. C. Ferris, S. J. Wright, D. M. Shepard, M. A. Earl. 2007. An optimization
framework for conformal radiation treatment planning. INFORMS Journal on Comput-
ing 19(3) 366-380.

Luedtke, J., S. Ahmed. 2008. A sample approximation approach for optimization with
probabilistic constraints. SIAM Journal on Optimization 19 674-699.









Lustig, I. J., J.-F. Puget. 2001. Program does not equal program: Constraint programming
and its relationship to mathematical programming. Ir,/ 'f i., 31(6) 29-53.

Marchand, H., A. Martin, R. Weismantel, L. Wolsey. 2002. Cutting planes in integer and
mixed integer programming. Discrete Applied Mathematics 123 397-446.

Megiddo, N., S. L. Hakimi, M. R. Garey, D. S. Johnson, C. H. Papadimitriou. 1988. The
complexity of searching a graph. Journal of the AC(I[ 35(1) 18-44.

Men, C., H. E. Romeijn, Z. C. Taskm, J. F. Dempsey. 2007. An exact approach to direct
aperture optimization in IMRT treatment planning. P/;,--. in Medicine and B.: .,,I
52(24) 7333-7352.

Migdalas, A., P. M. Pardalos. 1996. Hierarchical and bilevel programming. Journal of
Global Optimization 8(3).

Migdalas, A., P.M. Pardalos, P. Varbrand, eds. 1997. Multilevel Optimization: Algorithms
and Applications. Kluwer Academic Publishers.

Nemhauser, G. L., L. A. Wolsey. 1988. Integer and Combinatorial Optimization. John
Wiley & Sons, New York, NY.

Nemirovski, A., A. Shapiro. 2005. Scenario approximation of chance constraints.
G. Calafiore, F. Dabbene, eds., Probabilistic and Randomized Methods for Design
Under Unceri',,ii,/. Springer, London, 3-48.

Ntaimo, L., S. Sen. 2005. The million-variable i1, i ch" for stochastic combinatorial
optimization. Journal of Global Optimization 32(3) 385-400.

Parsons, T. D. 1978. Pursuit-evasion in a graph. Theory and Applications of Graphs.
Springer, Berlin, 426-441.

Peng, S., C. Ho, T. Hsu, M. Ko, C. Y. Tang. 2000. Edge and node searching problems on
trees. Theoretical Computer Science 240(2) 429-446.

Penuel, J., J. C. Smith. 2009. Models and complexity analysis for the graph
decontamination problem with mobile agents. Tech. rep., Department of Industrial
and Systems Engineering, University of Florida, Gainesville, FL.

Preciado-Walters, F., R. Rardin, M. Langer, V. Thai. 2004. A coupled column generation,
mixed integer approach to optimal planning of intensity modulated radiation therapy for
cancer. Mathematical Pir;,,',,,,,:,.j 101(2) 319-338.

Que, W. 1999. Comparison of algorithms for multileaf collimator field segmentation.
Medical Ph;,; 26(11) 2390-2396.

Que, W., J. Kung, J. Dai. 2004. "Tongue-and-giuulo effect in intensity modulated
radiotherapy with static multileaf collimator fields. Ph/ i, in Medicine and B.:' .4 .,;
49(3) 399-405.









Romeijn, H. E., R. K. Ali I J. F. Dempsey, A. Kumar. 2005. A column generation
approach to radiation therapy treatment planning using aperture modulation. SIAM
Journal on Optimization 15(3) 8 :- -,.2

Romeijn, H. E., R. K. Al i I J. F. Dempsey, A. Kumar. 2006. A new linear programming
approach to radiation therapy treatment planning problems. Operations Research 54(2)
201-216.

Rossi, F., P. van Beek, T. Walsh, eds. 2006. Handbook of Constraint P,.g,.irl,'.:,j.
Elsevier.

Sak, T., J. W., S. K. Goldenstein. 2008. Probabilistic multiagent patrolling. Advances in
Ari.:?l .:.'1 Intelligence SBIA o'w,''. Springer, Berlin, 124-133.

Sen, S., J. L. Higle. 2005. The C3 theorem and a D2 algorithm for large scale stochastic
integer programming: Set convexification. Mathematical PI,..' ',.ir ,,.: j 104 1-20.

Seymour, P., R. Thomas. 1993. Graph searching and a min-max theorem for tree-width.
Jounal of Combinatorial The(.',; Series B 58(1) 22 33.

Shepard, D. M., M. A. Earl, X. A. Li, S. Naqvi, C. Yu. 2002. Direct aperture optimization:
a turnkey solution for step-and-shoot IMRT. Medical PhA,-. 29(6) 1007-1018.

Shepard, D. M., M. C. Ferris, G. H. Olivera, T. R. Mackie. 1999. Optimizing the delivery
of radiation therapy to cancer patients. SIAM Review 41(4) 721-744.

Sherali, H. D., B. M. P. Fraticelli. 2002. A modification of Benders' decomposition
algorithm for discrete subproblems: an approach for stochastic programs with integer
recourse. Journal of Global Optimization 22 319-342.

Sherali, H. D., J. C. Smith. 2001. Improving zero-one model representations via symmetry
considerations. M i,.ir, iI ,t. Science 47(10) 1396-1407.

Sherali, H. D., J. C. Smith, Y. Lee. 2000. Enhanced model representations for an
intra-ring synchronous optical network design problem allowing demand splitting.
INFORMS Journal on CornTJ;'.:Il 12(4) 284-298.

Siek, J. G., L. Lee, A. Lumsdaine. 2001. The Boost Graph Library: User Guide and
Reference Manual. Addison-Wesley.

Siochi, R. A. C. 1999. Minimizing static intensity modulation delivery time using an
intensity solid paradigm. International Journal of Radiation O,.,, .. iI; B'.:.i ,,; Ph,' -
43(3) 671-680.

Siochi, R. A. C. 2004. Modifications to the IMFAST leaf sequencing optimization
algorithm. Medical PhA-.. 31(12) 3267-3278.

Siochi, R. A. C. 2007. Variable depth recursion algorithm for leaf sequencing. Medical
Ph;,.. 34(2) 664-672.









Smith, B. M. 1995. A tutorial on constraint programming. Tech. rep., School of
Computing, University of Leeds, Leeds, UK.

Smith, B. M. 2005. Symmetry and search in a network design problem. Roman Bartik,
Michela Milano, eds., Integration of AI and OR Techniques in Constraint P,..oi,,,,,,:,,j
for Combinatorial Optimization Problems. Springer, Berlin, 336-350.

Smith, J. C. 2004. Algorithms for distributing telecommunication traffic on a multiple-ring
SONET-based network. European Journal of Operational Research 154(3) 659-672.

Smith, J. C., C. Lim, A. Alptekinoglu. 2008. Optimal mixed-integer programming and
heuristic methods for a bilevel Stackelberg product introduction game. Tech. rep.,
Department of Industrial and Systems E:iir.. -i 1i. University of Florida, Gainesville,
FL.

Smith, J. C., A. J. Schaefer, J. W. Yen. 2004. A stochastic integer programming approach
to solving a synchronous optical network ring design problem. Networks 44(1) 12-26.

Stell, A. M., O. A. Zeidan, J. G. Li, J. F. Dempsey. 2004. An extensive log-file analysis
of step-and-shoot IMRT subfield delivery errors. International Journal of Radiation
Orn., ]..,'; B ':.., Ph,.-.. 31(6) 1593-1602.

Sutter, A., F. Vanderbeck, L. A. Wolsey. 1998. Optimal placement of add/drop
multiplexers: Heuristic and exact algorithms. Operations Research 46(5) 719-728.

Takmn, Z. C., J. C. Smith, S. Ahmed, A. J. Schaefer. 2009a. Cutting plane algorithms for
solving a stochastic edge-partition problem. Discrete Optimization Forthcoming.

Taskm, Z. C., J. C. Smith, H. E. Romeijn. 2008. Mixed-integer programming techniques
for decomposing IMRT fluence maps using rectangular apertures. Tech. rep.,
Department of Industrial and Systems Eir-,iii. liir. University of Florida, Gainesville,
FL.

Takmn, Z. C., J. C. Smith, H. E. Romeijn, J. F. Dempsey. 2009b. Optimal multileaf
collimator leaf sequencing in IMRT treatment planning. Operations Research
Forthcoming.

Thorsteinsson, E. S. 2001. Branch-and-check: A hybrid framework integrating mixed
integer programming and constraint logic programming. Toby Walsh, ed., CP '01:
Proceedings of the 7th International Conference on Principles and Practice of Constraint
P,.g ,ri,,i,,:, : Springer, 16-30.

van Beek, P. 2006. Backtracking search algorithms. F. Rossi, P. van Beek, T. Walsh, eds.,
Handbook of Constraint P,..q,,,i,,,:,,,j. Elsevier, 85-134.

Van Santvoort, J. P. C., B. J. M. Heijmen. 1996. Dynamic multileaf collimation without
"tongue-and-giuu,t underdosage effects. PhA;-., in Medicine and B '...it,.; 41(10)
2091-2105.









W\i. Y., D. Beck, X. Jiang, R. Roussev. 2005. Automated web patrol with strider
honeymonkeys: Finding web sites that exploit browser vulnerabilities. Tech. rep.,
Microsoft Research, Redmond, WA.

Wolsey, L A. 1998. Integer P,..,r,,,,,:,uj. Wiley-Interscience, New York, NY.

Xia, P., L. J. Verhey. 1998. Multileaf collimator leaf sequencing algorithm for intensity
modulated beams with multiple static segments. Medical PhI,'. 25(8) 1424-1434.









BIOGRAPHICAL SKETCH

Z. Caner Takmn was born in Bahkesir, Turkey on September 8, 1981. He earned his

B.S. and M.S. degrees in Industrial Engineering from Bogazigi University, Istanbul in

2003 and 2005, respectively. Before starting his doctorate study, he worked for ICRON

Technologies as a product consultant, where he took role in advanced planning and

scheduling projects for customers in several industries including steel, automotive,

electronics and glass manufacturing industries. He will finish his Ph.D. degree in

Industrial and Systems Engineering at the University of Florida in August 2009. Following

graduation, he will join Department of Industrial Engineering at Bogazici University as a

faculty member.





PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

IwouldliketoexpressmydeepestgratitudetoColeSmithforhiswise,enlighteningideas,endlessmotivation,andpatientcounselingduringthewritingofthisdissertation.Hehasbeenagreatteacher,amentorandafriendtomeinthelastfouryears,andworkingwithhimhasbeenaprivilege.IwouldliketothankEdwinRomeijnforintroducingmetotheexcitingeldofoptimizationinhealthcare,andhisinvaluablecontributionstothisstudy.Hisguidanceandsupporthasbeenveryhelpfulthroughoutmygraduateeducation.IwouldalsoliketoacknowledgePanosPardalosandDouglasDankelfortakingpartinmydissertationcommittee,andtheirvaluablesuggestions.Mysincerethanksareduetomyfriendsingraduateschool.Inparticular,Chasehasnotonlybeenagreatcolleague,butalsoaclosefriendforthesefouryears.IcannotbegintocountthethingsthatIlearnedfromhimabouttheculture,thelifestyleandthelanguage.MyexperienceinAmericawouldnothavebeennearlyasenjoyablewithouthimandhiswifeCandace.IamindebtedtomyparentsandmybrotherforguidingandsupportingmeinalllifechoicesIhavemade.Finally,Iamdeeplygratefultomylovelywife,Semra,forherconstantencouragement,support,andunderstanding.Sheisthesourceofmyhappiness,andthesecretofmysuccess.Thisdissertationrepresentstheendofmylifeasagraduatestudent.Italsorepresentsthebeginningofanewstageinmylife,everymomentofwhichIamlookingforwardtosharingwithher. 4

PAGE 5

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 8 LISTOFFIGURES .................................... 9 ABSTRACT ........................................ 10 CHAPTER 1INTRODUCTION .................................. 13 2STOCHASTICEDGE-PARTITIONPROBLEM .................. 19 2.1IntroductionandLiteratureSurvey ...................... 19 2.2FormulationandCuttingPlaneApproach .................. 23 2.3AHybridIP/CPApproach .......................... 33 2.3.1First-StageProblem ........................... 35 2.3.2Second-StageProblem ......................... 36 2.3.2.1Foundations .......................... 37 2.3.2.2Domainexpansion ...................... 38 2.3.2.3Constraintpropagation .................... 39 2.3.2.4Forwardchecking ....................... 40 2.3.2.5Nodeselectionrule ...................... 41 2.3.2.6Distributionvectororderingrule .............. 41 2.3.3Third-StageProblem .......................... 42 2.3.4InfeasibilityAnalysis .......................... 42 2.3.5EnhancementsfortheFirst-StageProblem .............. 43 2.3.5.1Validinequalities ....................... 43 2.3.5.2Heuristicforobtaininganinitialfeasiblesolution ..... 44 2.3.5.3Processingintegersolutions ................. 45 2.4ComputationalResults ............................. 45 3CONSECUTIVE-ONESMATRIXDECOMPOSITIONPROBLEM ....... 51 3.1IntroductionandLiteratureSurvey ...................... 51 3.2DecompositionAlgorithm ........................... 55 3.2.1DecompositionFramework ....................... 56 3.2.2MasterProblemFormulationandSolutionApproach ......... 58 3.2.3SubproblemAnalysisandSolutionApproach ............. 61 3.2.4ValidInequalitiesfortheMasterProblem ............... 66 3.2.4.1Beam-on-timeandnumberofaperturesinequalities .... 66 3.2.4.2Bixelsubsequenceinequalities ................ 68 3.2.5ConstructingaFeasibleMatrixDecomposition ............ 69 3.3ComputationalResultsandComparisons ................... 73 5

PAGE 6

............................ 73 3.3.2ImplementationIssues ......................... 73 3.3.3ComparisonwithLangeretal.(2001)Model ............. 74 3.3.4RandomProblemInstances ....................... 75 3.3.5ClinicalProblemInstances ....................... 80 4RECTANGULARMATRIXDECOMPOSITIONPROBLEM .......... 84 4.1IntroductionandLiteratureSurvey ...................... 84 4.2AMixed-IntegerProgrammingApproach ................... 86 4.2.1ModelDevelopment ........................... 86 4.2.2ValidInequalities ............................ 90 4.2.2.1Adjacentrectangles ...................... 90 4.2.2.2Boundingboxinequalities .................. 90 4.2.2.3Aggregateintensityinequalities ............... 93 4.2.2.4Specialsubmatrices ...................... 94 4.2.2.5Submatrixinequalities .................... 96 4.2.3PartitioningApproach ......................... 97 4.2.3.1Separablecomponents .................... 97 4.2.3.2Independentregions ..................... 98 4.2.3.3Dependentregions ...................... 100 4.2.3.4Upperboundcalculation ................... 102 4.3Extensions .................................... 104 4.3.1MinimizeTotalTreatmentTime .................... 104 4.3.2OptimizationwithBeam-on-TimeRestrictions ............ 105 4.4ComputationalResults ............................. 106 5GRAPHSEARCHPROBLEM ........................... 113 5.1IntroductionandLiteratureReview ...................... 113 5.2Hide-and-SeekProblem ............................. 115 5.2.1MathematicalModel .......................... 115 5.2.2SolutionApproach ............................ 116 5.2.2.1Searcher'sproblem ...................... 117 5.2.2.2Branch-and-pricealgorithm ................. 119 5.3PursuitEvasionProblem ............................ 119 5.3.1MathematicalModel .......................... 119 5.3.2SolutionApproach ............................ 120 5.3.2.1Searcher'sproblem ...................... 120 5.3.2.2Intruder'sproblem ...................... 122 5.3.2.3Branch-cut-pricealgorithm .................. 122 5.4PatrolProblem ................................. 123 5.4.1ProblemDescription .......................... 123 5.4.2MathematicalModel .......................... 123 5.4.2.1Searcher'sproblem ...................... 124 5.4.2.2Intruder'sproblem ...................... 125 6

PAGE 7

.................. 127 5.5BranchingStrategies .............................. 127 5.6ComputationalResults ............................. 129 6CONCLUSIONSANDFUTURERESEARCHDIRECTIONS .......... 134 6.1StochasticEdge-PartitionProblem ...................... 134 6.2MatrixDecompositionProblem ........................ 135 6.3GraphSearchProblem ............................. 138 6.4MasterProblemReformulationinBi-LevelCuttingPlaneOptimizationAlgorithms ................................... 139 APPENDIX AANIPMODELFORC1-MATRIXDECOMPOSITIONPROBLEM ...... 144 BCOMPLEXITYOFC1-PARTITION ........................ 147 REFERENCES ....................................... 149 BIOGRAPHICALSKETCH ................................ 158 7

PAGE 8

Table page 2-1Descriptionsoftheprobleminstancesusedforcomparingalgorithms ...... 46 2-2Comparisonofthealgorithmsongraphshavingedgedensity=0:2 ....... 47 2-3Comparisonofthealgorithmsongraphshavingedgedensity=0:3 ....... 48 2-4Comparisonofthealgorithmsongraphshavingedgedensity=0:4 ....... 49 2-5Descriptionsoftheprobleminstancesusedforanalyzingthree-stagealgorithm 49 2-6Three-Stagealgorithmongraphshavingedgedensity=0:2 ........... 49 2-7Three-Stagealgorithmongraphshavingedgedensity=0:3 ........... 50 2-8Three-Stagealgorithmongraphshavingedgedensity=0:4 ........... 50 3-1Dimensionsofclinicalprobleminstances ...................... 73 3-2Comparisonofourbasealgorithmwith Langeretal. ( 2001 )model ........ 75 3-3EectofrotatingtheMLChead ........................... 81 3-4Computationalresultsforourbasealgorithm ................... 81 3-5Comparisonofheuristicalgorithmsonclinicaldata ................ 83 4-1Eectofvalidinequalitiesandthepartitioningstrategy .............. 108 4-2Computationalresultsonmodelextensions ..................... 109 4-3Eectofmaximumintensityvalueonsolvability .................. 111 5-1Averagenumberofbranch-and-boundnodesexploredforhide-and-seekproblem 131 5-2Averagenumberofsearchersneededforthehide-and-seekproblem ....... 131 5-3Numberofinstancesthataresolvedwithintimelimitforthepursuitevasionproblem ........................................ 131 5-4Averagenumberofbranch-and-boundnodesexploredforthepursuitevasionproblem ........................................ 132 5-5Averagenumberofsearchersneededforthepursuitevasionproblem ....... 132 5-6Numberofinstancesthataresolvedwithintimelimitforthepatrolproblem .. 132 5-7Averagenumberofbranch-and-boundnodesexploredforthepatrolproblem .. 133 5-8Averagenumberofsearchersneededforthepatrolproblem ............ 133 8

PAGE 9

Figure page 2-1(a)Aninstanceofthedeterministicedge-partitionproblem(b)AsolutionwithjKj=3;r=3;b=20 ................................ 19 3-1(a)Amultileafcollimatorsystem(b)Theprojectionofanapertureontoapatient 51 3-2Comparisonoftotaltreatmenttimesonrandomdata ............... 76 3-3Comparisonofthenumberofaperturesonrandomdata ............. 77 3-4Comparisonofbeam-on-timevaluesonrandomdata ............... 78 3-5ComparisonofTGIvaluesonrandomdata ..................... 79 4-1Exampleuencemap ................................. 87 4-2Examplestartindex ................................. 91 4-3Exampleendindex .................................. 91 4-4Exampleboundingbox ................................ 92 4-5Anothernondominatedboundingboxseededat(6,3) ............... 93 4-6Twocomponentsofauencemap .......................... 98 4-7Regionsofaconnectedcomponent ......................... 98 4-8Ecientfrontierfornumberofaperturesandbeam-on-time ........... 110 5-1(a)Anexamplegraph(b)Time-expandednetworkforT=2 ........... 118 9

PAGE 10

Inthisdissertation,weinvestigateaclassofmulti-leveloptimizationproblems,inwhichdiscretevariablesarepresentatmultiplestages.Suchproblemsariseinmanypracticalsettings,andtheyarenotoriouslydiculttooptimize.Bendersdecomposition,whichisawell-knowndecompositionmethodforsolvinglarge-scalemixed-integerprogrammingproblems,cannotbeutilizedfortheclassofproblemsthatweconsiderduetotheexistenceofdiscretevariablesatlowerlevels.CuttingplanealgorithmssuchasthoseproposedbyLaporteandLouveauxhavebeendesignedforuseinbi-levelintegerprogrammingproblemswithbinaryvariablesatbothlevels.However,thesearebasedongenericcuts,whichdonotutilizeanyproblemspecicstructures,andhenceoftenresultinweakconvergence.Ourgoalinthisdissertationistoproposenovelformulationandsolutionstrategiesforseveralmulti-leveloptimizationproblemstosolvetheseproblemstooptimalitywithinpracticalcomputationallimits. Werstconsideranedge-partitionproblem.ThemotivationforthisstudyisprovidedbyaSynchronousOpticalNetwork(SONET)designapplication.IntheSONETcontext,eachedgerepresentsademandpairbetweentwoclientnodes,andtheweightofeachedgerepresentsthenumberofcommunicationchannelsneededbetweentheclientnodes.Weconsiderastochasticversionoftheproblem,inwhichtheedgeweightsarenotdeterministic,buttheirunderlyingprobabilitydistributionisknown.TheproblemistodesignasetofSONET\rings"atminimumcost,whileensuringthat 10

PAGE 11

Next,wefocusonamatrixdecompositionproblem,whicharisesinIntensityModulatedRadiationTherapy(IMRT)treatmentplanning.Theprobleminputisamatrixofintensityvaluesthatneedstobedeliveredtoapatient,whichmustbedecomposedintoacollectionofaperturesandcorrespondingintensities.Inafeasibledecompositionthesumofbinaryshapematricesmultipliedbycorrespondingintensityvaluesisequaltotheoriginalintensitymatrix.Weconsidertwovariantsoftheproblem:(i)avariantinwhichtheshapematricesusedinthedecompositionhavetosatisfythe\consecutive-ones"property,and(ii)avariantinwhichtheshapematriceshavetoberectangular.Fortherstvariant,westartbyinvestigatinganintegerprogrammingmodelproposedintheliterature,andshowhowtheformulationcanbestrengthened.Wethenformulatetheproblemasabi-leveloptimizationproblemthathasdiscretevariablesatbothstages,andsuggestahybridintegerprogramming/constraintprogrammingdecompositionalgorithmsimilartoouralgorithmfortheedge-partitionproblem.Ourtestsondataobtainedfrompatientsshowthatouralgorithmiscapableofsolvingprobleminstancesofclinicallyrelevantdimensionswithinpracticalcomputationallimits.Wethenturnourattentiontothesecondvariantofthematrixdecompositionproblem.Weformulatetheproblemasamixed-integerprogram,andinvestigateadecompositionmethodforsolvingit.Unliketherstvariant,thesecond-levelproblemturnsouttobealinearprogrammingproblem,andthereforeweareabletoderiveaBendersdecompositionalgorithmforsolvingthisvariant. Finally,weinvestigateaclassofgraphsearchproblems.Inthisclassofproblems,anintruderislocatedatanunknownnodeontheinputgraph,andagroupofsearchersneeds 11

PAGE 12

12

PAGE 13

Inmostcomplexdecision-makingenvironments,thereexistseveraltypesofinterdependentdecisionsthatneedtobemadetooptimizesomecostorbenetfunction.Asasimpleexample,consideraproductionplanningproblem.Thegoalistodetermine,attheveryleast,thetypesofproductsthataretobeproducedwithinatimeperiod,alongwiththeassociatedproductionquantities.Theremightexistindividualrestrictionsoneachtypeofdecision,suchas\thetotalamountofproductionofproductsAandBcannotexceed,"or\ifproductAisproduced,somustproductB."Theremightalsoexistrestrictionsthatrelatethetwotypesofdecisions,suchas\ifproductAisproduced,thenthebatchsizemustbeatleast."Modelinganoptimizationprobleminvolves(i)deningadecisionvariableforeachindividualdecision,(ii)formulatingtherestrictionsonthedecisionsasconstraints,and(iii)deninganobjectivefunctiontobeoptimized.Theeldofmathematicalprogrammingseekstooptimizesuchmodelsandprovetheoptimalityofthegeneratedsolution,orprovethatnofeasiblesolutionexists. Optimizationproblemsinwhichsomevariablesarerestrictedtotakevaluesfromadiscretesetareclassiedasdiscreteoptimizationproblems.Animportantconcernregardingbuildingandsolvingdiscreteoptimizationproblemsisthattheamountofmemoryandthecomputationaleortneededtosolvesuchproblemsgrowexponentiallywiththenumberofdiscretevariables.Thetraditionalapproach,whichinvolvesmakingalldecisionssimultaneouslybysolvingamonolithicoptimizationproblem,quicklybecomesintractableasthenumberofdiscretevariablesincreases.Multi-leveloptimizationalgorithms,suchasBendersdecomposition( Benders 1962 ),havebeendevelopedasanalternativesolutionmethodologytoalleviatethisdiculty.Unlikethetraditionalapproach,thesealgorithmsdividethedecision-makingprocessintoseveralstages.Forinstance,inBendersdecompositionarst-stagemasterproblemissolvedforasubsetofvariables,andthevaluesoftheremainingvariablesaredeterminedbyasecond-stage 13

PAGE 14

Inessence,multi-leveloptimizationalgorithmssolveaseriesofsmallproblemsinsteadofasinglelargeproblem.Performingmultipleiterationsisusuallyjustiedduetotheexponentiallylargercomputationalresourcerequirementsassociatedwithsolvingalargerproblem.Furthermore,itisoftenthecasethatdecisionsforseveralgroupsofsecond-stagevariablescanbemadeindependentlygiventherst-stagedecisions.Insuchcases,multi-leveloptimizationalgorithmsareamenabletoparallelimplementations.Theadventofecientparallelcomputinggridshasallowedmodernbi-leveltechniquestosolveproblemsthatwereregardedasintractablebefore( NtaimoandSen 2005 ).Insomeapplications,solvingproblemsinmultiplestagesallowseorttobeconservedbyavoidingtheexplicitsolutionofproblemsbymathematicalprogramming,suchastheevacuationnetworkdesignalgorithmof AndreasandSmith ( 2009 ).Multi-leveloptimizationhas 14

PAGE 15

MigdalasandPardalos 1996 ; Migdalasetal. 1997 ),andawideclassofoptimizationproblemscanbereformulatedasmulti-leveloptimizationproblems( HuangandPardalos 2002 ). Bendersdecompositionhasbeenparticularlysuccessfulinsolvingmixed-integerlinearprogrammingproblemsarisinginawidevarietyofapplications.InBendersdecomposition,discretevariablesoftheproblemarekeptinthemasterproblem,andcontinuousvariablesaremovedtothesubproblem.Ineachiteration,giventhevaluesofthediscretevariablesbythemasterproblem,thesubproblemissolvedasalinearprogram,andacuttingplanetobepassedbacktothemasterproblemisgeneratedusinglinearprogrammingduality.However,thisapproachcannotbeapplieddirectlywhendiscretevariablesalsoappearinthesecondstage.Thereasonisthatnodualinformationcanbeextractedfromthesubproblemifthesecond-stageproblemcontainsintegervariables.Inthiscase,onecanemploycuttingplanessuchasthegeneral-purposecutsof LaporteandLouveaux ( 1993 )andcombinatorialBendersinequalities( CodatoandFischetti 2006 ).However,theseinequalitiesareoftenveryweak,andresultinslowalgorithmicconvergence. Ourmainlineofresearchisaboutdesigningecientmulti-leveloptimizationalgorithmsforproblemsthathavediscretevariablesatmultiplestages.Werstpresentourresultsonanedge-partitionproblemarisinginatelecommunicationsnetworkdesigncontextregardingSynchronousOpticalNetworks(SONET).Theedge-partitionproblemconsidersanundirectedgraphwithweightededges,andsimultaneouslyassignsnodesandedgestosubgraphssuchthateachedgeappearsinexactlyonesubgraph,andsuchthatnoedgeisassignedtoasubgraphunlessbothofitsincidentnodesarealsoassignedtothatsubgraph.Additionally,therearelimitationsonthenumberofnodesandonthesumofedgeweightsthatcanbeassignedtoeachsubgraph( Goldschmidtetal. 2003 ). 15

PAGE 16

2 ,inwhichweassignnodestosubgraphsinarststage,realizeasetofedgeweightsfromamonganitesetofalternatives,andthenassignedgestosubgraphs.Werstformulatetheproblemasamonolithicintegerprogrammingproblem,andshowthatthisapproachisnottractableduetotherapidlyincreasingcomputationalrequirements.Wethenprescribeabi-levelcuttingplaneapproachhavingintegervariablesinbothstagesandexaminecomputationaldicultiesassociatedbothwiththegenericinequalitiesby LaporteandLouveaux ( 1993 )andwithourproposedcuttingplanes.Wealsoprescribeathree-levelhybridintegerprogramming/constraintprogrammingalgorithmhavingdiscretevariablesatalllevels,anddiscusshowthishybridapproachresolvessomeofthedicultiesassociatedwiththebi-levelcuttingplaneapproach. Chapters 3 and 4 consideraproblemdealingwiththeecientdeliveryofIntensityModulatedRadiationTherapy(IMRT)toindividualpatients.Inparticular,weconsideramatrixdecompositionproblemthatarisesattheleafsequencingstageinIMRTtreatmentplanning.Theprobleminputisanintegermatrixofintensityvaluesthataretobedeliveredtoapatientfromagivenbeamangle.Todeliverthisintensityproletothepatient,wemustdecomposetheinputmatrixintoacollectionofaperturesandcorrespondingintensities.Afeasibledecompositionisoneinwhichtheoriginaldesiredintensityprolematrixisequaltothesumofanumberoffeasiblebinarymatricesmultipliedbycorrespondingintensityvalues.Tomostecientlytreatapatient,wewishtominimizeameasureoftotaltreatmenttime,whichisgivenasaweightedsumofthenumberofaperturesandthesumoftheapertureintensitiesusedinthedecomposition.InChapter 3 ,wedescribeaversionoftheprobleminwhicheachaperturematrixneedstosatisfythe\consecutive-ones"property,whichmeansthatallmatrixentrieswithvalue1ineachrowofanaperturematrixmustbeconsecutive.SimilartoChapter 2 ,weprescribeabi-levelhybridoptimizationalgorithminwhichthemasterproblemisanintegerprogrammingproblem,andwesolveasubproblemforeachrowofthematrixby 16

PAGE 17

4 dealswithanothervariantofthematrixdecompositionproblem,inwhichonlyrectangularaperturescanbeusedinthedecomposition.WedevelopaBendersdecompositionalgorithmforsolvingthisvariant.Wealsoproposeaschemeforpartitioningtheproblemtoobtainsimultaneousupperandlowerbounds. InChapter 5 ,westudyaclassofgraphsearchproblems,whereagroupofsearchersseekanintruderonagraph.Boththeintruderandthesearchersarelocatedatsomenodesofthegraph,andthesearchercanonly\see"asubsetofthenodesfromeachnode.Ateachtimeperiod,boththeintruderandthesearcherscanmovealonganedgetoanadjacentnode,orstayatthesamenode.Ourgoalistondtheminimumnumberofsearchersneededtolocatetheintruderwithinagiventimelimit.Weinvestigatethreevariantsofthegraphsearchproblem:(i)ahide-and-seekproblem,inwhichastationaryintruder\hides"atanunknownnode,(ii)apursuitevasionproblem,inwhichtheintrudermovesamongthenodestoavoidbeingdetected,and(iii)apatrolproblem,inwhichnointruderisinitiallyinthegraphandeachsearcherpatrolsthegraphtoprotectitfrompotentialintrusion.Weformulatetheseproblemsasasetcoveringproblemwithanexponentialnumberofvariablesandconstraints,andproposeabranch-cut-pricealgorithmforsolvingit.Bothourmasterproblemandthesubproblems,whichcorrespondtotheintruderandthesearchers,havediscretevariables.Weformulatetheintruder'ssubproblemasalongestpathproblemonanauxiliarygraph,andthesearcher'ssubproblemsasmixed-integerprogrammingproblems. WeconcludeourdissertationinChapter 6 ,whichexploresfutureresearchdirectionsregardingmulti-leveloptimizationalgorithms.Werstevaluatetheapproachestakenintheedge-partition,matrixdecomposition,andgraphsearchproblemsdescribedinChapters 2 { 5 ,anddiscussfutureresearchtopicsregardingeachapplication.Wethendescribeourpreliminaryresearchonamasterproblemreformulationtechnique,whichcanbeusedinavarietyofbi-leveloptimizationalgorithms.Thisreformulationtechnique 17

PAGE 18

18

PAGE 19

Goldschmidtetal. ( 2003 ),whichisdenedonanundirectedgraphG(N;E).Inthedeterministicedge-partitionproblem,wecreateasetKof(possiblyempty)subgraphsofGsuchthateachedgeiscontainedinexactlyonesubgraph,subjecttocertainrestrictionsonthecompositionofthesubgraphs.Theserestrictionsincludethestipulationsthatanedgecannotbeassignedtoasubgraphunlessbothofitsincidentnodesbelongtothesubgraph,andthatnomorethanrnodescanbeassignedtoanysubgraph,forsomer2Z+.Additionally,eachedge(i;j)2Ehasapositiveweightofwij,andthesumofedgeweightsassignedtoeachsubgraphcannotexceedsomegivenpositivenumberb.Theobjectiveoftheproblemistominimizethesumofthenumberofnodesassignedtoeachsubgraph. (a) (b) Figure2-1. (a)Aninstanceofthedeterministicedge-partitionproblem(b)AsolutionwithjKj=3;r=3;b=20 Figure 2-1 illustratesthedeterministicedge-partitionproblem.ThegraphGandthecorrespondingedgeweightsareshowninFigure 2-1 a.Figure 2-1 bshowsafeasiblesolutionthatpartitionsGintojKj=3subgraphs,wherethenumberofnodesineachsubgraphislimitedbyr=3,andthetotalweightassignedtoeachsubgraphislimitedbyb=20.Notethatthedegreeofnode4isthree,whichimpliesthatitmustbeassignedtoatleasttwosubgraphs,orelsetherewouldbeatleast4>rnodesinasubgraph.Similarly,node5mustbeassignedtoatleasttwosubgraphs.Sincenodes4and5areassignedtotwo 19

PAGE 20

2-1 bisoptimal. Goldschmidtetal. ( 2003 )discusstheedge-partitionproblem(withdeterministicweights)inthecontextofdesigningSynchronousOpticalNetwork(SONET)rings.IntheSONETcontext,eachedge(i;j)2Erepresentsademandpairbetweentwoclientnodes,andtheweightwijrepresentsthenumberofcommunicationchannelsrequestedbetweennodesiandj.AlltelecommunicationtraciscarriedoverasetofSONETrings,whicharerepresentedbysubgraphsintheedge-partitionproblem.Sinceeachdemandmustbecarriedbyexactlyonering,edgesmustbepartitionedamongtherings.(Notethattheterm\ring"describesonlythephysicalSONETroutingstructure,anddoesnotplaceanyrestrictionsontopologicalpropertiesofdemandedgesincludedonaring.See,e.g., Goldschmidtetal. ( 2003 )formoredetails.)SONETringsarepermittedtocarrycommunicationbetweennodesonlyifthosenodeshavebeenconnectedtotheringbyequipmentcalledAdd-DropMultiplexers(ADMs).TherearetechnicallimitsonthenumberofADMsthatcanbeassignedtoeachring(e.g.,r),andonthetotalamountofchannels(e.g.,b)thatcanbeassignedtoaring.SinceADMsarequiteexpensive,ringnetworksarepreferredthatemployasfewADMsaspossible,whichechoestheedge-partitionproblem'sobjectiveofminimizingthesumofnodesassignedtoeachsubgraph. Theprimarycontributionby Goldschmidtetal. ( 2003 )isanapproximationalgorithmforaspeciccaseoftheedge-partitionprobleminwhichallwijareequaltoone. Sutteretal. ( 1998 )proposeacolumn-generationalgorithmforthisproblem,and Leeetal. ( 2000b )employabranch-and-cutalgorithmonaformulationthatweadapt.Forthecaseinwhichtheweightsontheedgescanbesplitamongmultiplerings, Sheralietal. ( 2000 )prescribeamixed-integerprogrammingapproachaugmentedbytheuseofvalidinequalities,anti-symmetryconstraints,andvariablebranchingrules. Smith ( 2005 )formulatesthedeterministicversionoftheedge-partitionproblemasaconstraintprogram 20

PAGE 21

2.3 )canimproveperformance. Weconsideraversionoftheedge-partitionproblemwheretheedgeweightsareuncertain,andareonlyrealizedafterthenode-to-subgraphassignmentshavebeenmade.AsweshowinSection 2.2 ,thisframeworkallowsustodesignmorerobustsolutionsthanthoseintheliterature,whicharevirtuallyallappliedtodeterministicdata.Weseekaminimum-cardinalitysetofnode-to-subgraphassignments,suchthatthereexistsanassignmentofedgestosubgraphssatisfyingtheaforementionedsubgraphrestrictionswithapre-speciedhighprobability.Suchaprobabilisticconstraintisextremelyhardtodealwithinanoptimizationframework.Oneapproach,knownasscenarioapproximation(cf. CalaoreandCampi ( 2005 ); LuedtkeandAhmed ( 2008 ); NemirovskiandShapiro ( 2005 ))istodrawindependentidenticallydistributed(i.i.d.)realizationsoftheedgeweights(calledscenarios)andrequirethenode-to-subgraphassignmentstoadmitafeasibleedge-to-subgraphassignmentineachscenario.Itcanbeshownthat,withasucientlylargescenarioset,asolutiontothisscenarioapproximationsolutionisfeasibletothetrueprobabilisticallyconstrainedproblemwithhighcondence.Inthisstudywedevelopalgorithmicapproachesforsolvingthescenarioapproximationcorrespondingtothediscussedprobabilisticallyconstrainededge-partitionproblem.Werefertothisscenarioapproximationasthestochasticedge-partitionproblem. RelativelylittleworkhasbeendoneinSONETnetworkdesignwhentheedgeweightsareuncertain. Smithetal. ( 2004 )considertheSONETringdesignprobleminwhichedgedemandscanbesplitamongmultipleringsandproposeatwo-stageintegerprogrammingalgorithm.Thedemandsplittingrelaxationallowsthesecond-stageproblemstobesolvedaslinearprograms,andthusstandardBenderscutscaneasilybederivedfromthesecond-stagerecourseproblems.However,wehavesecond-stageintegerprogramsfromwhichdualinformationcannotbereadilyobtained. 21

PAGE 22

LaporteandLouveaux 1993 ).Thismethodapproximatesthesecond-stagevaluefunctionbylinear\cuts"thatareexactatthebinarysolutionwherethecutisgenerated,andareunder-estimatesatotherbinarysolutions.Typicalintegerprogrammingalgorithmsprogressbysolvingasequenceofintermediatelinearprogramming(LP)problems.Usingdisjunctiveprogrammingtechniques,itispossibletoderivecutsfromthesolutionstotheseintermediateLPsthatarevalidunder-estimatorsofthesecond-stagevaluefunctionatallbinaryrst-stagesolutions( SenandHigle 2005 ; SheraliandFraticelli 2002 ).Thisavoidssolvingdicultintegersecond-stageproblemstooptimalityinalliterationsofthealgorithm,providingsignicantcomputationaladvantage.Scenario-wisedecompositionmethodshavealsobeenproposed( CareandSchultz 1999 )asanalternativetotheabovestage-wisedecompositionapproaches.Herecopiesoftherst-stagevariablesaremadecorrespondingtoeachscenarioandarelinkedtogethervianon-anticipativityconstraints. Ourproposedmethodologydrawsonconstraintprogrammingandstochasticintegerprogrammingtheory.Hybridalgorithmsofthisnaturehaverecentlybeensuccessfullyemployedtosolvenotoriouslydicultproblems. JainandGrossmann ( 2001 )and BockmayrandPisaruk ( 2006 )devisehybridintegerprogramming/constraintprogrammingalgorithmsforsolvingmachineschedulingproblems. Thorsteinsson ( 2001 )proposesaframeworkforintegratingintegerprogrammingandconstraintprogrammingapproaches. HookerandOttosson ( 2003 )extendtheBendersdecompositionframeworksothatconstraintlogicprogramscanbeusedassubproblemstogeneratecutsthatareaddedtoamixed-integerlinearmasterproblem.Arecentworkby Hooker ( 2007 )useslogic-basedBendersdecompositiontosolveseveralplanningandschedulingproblems. 22

PAGE 23

2.2 ,wedevelopamixed-integerprogrammingformulationforthestochasticedge-partitionproblem,andprovidecuttingplanesthatcanbeusedwithinatwo-stagedecompositionalgorithm.InSection 2.3 ,weprescribeanalternativethree-stagealgorithmtoovercomethecomputationaldicultiesassociatedwiththeweaknessoftheproposedcuttingplanes.Finally,wecomparetheecacyofthesealgorithmsinSection 2.4 onasetofrandomlygeneratedtestinstances. 2.4 ).Wedenotethevectorofnode-to-subgraphassignmentsbyx.Let~wdenotetherandomvectorofedgeweightswithknowndistribution,andwdenotearealizationwithcomponentswij.Wedenebinarydecisionvariablesyijk=1ifedge(i;j)isassignedtosubgraphk.Givenanallowedviolationprobability2(0;1)theprobabilisticedge-partitionproblemcanbeformulatedasfollows:MinimizeXi2NXk2Kxik whereG(x;w)=Minimizez 23

PAGE 24

Theobjective( 2{1 )minimizesthetotalnumberofnodesassignedtosubgraphs.Constraints( 2{2 )limitthenumberofnodesassignedtoeachsubgraph.Constraints( 2{6 )requirethattheedgesbepartitionedamongthesubgraphs.Constraints( 2{7 )computethemaximumassignedweightoverallsubgraphs.Constraints( 2{8 )requirethatnoedgecanbeassignedtoasubgraphunlessbothofitsincidentnodesareassignedtothatsubgraph,and( 2{3 )and( 2{9 )statelogicalrestrictionsonthevariables.Byconvention,theoptimalvalueG(x;w)oftheintegerprogram( 2{5 ){( 2{9 )is+1iftheproblemisinfeasible.Givenanode-to-subgraphassignmentvectorxandedgeweightvectorwthereexistsafeasibleedge-to-subgraphassignmentifandonlyifG(x;w)b,i.e.,theweightassignedtoanysubgraphdoesnotexceedb.Thustheprobabilisticedge-partitionproblem( 2{1 ){( 2{4 )seeksaminimumcostnode-to-subgraphassignmentsuchthattheprobabilitythattherewillbeafeasibleedge-to-subgraphassignmentwhentheedgeweightsarerealizedissucientlyhigh. Tobuildascenarioapproximationoftheprobabilisticedge-partitionproblem( 2{1 ){( 2{4 ),wegenerateani.i.d.sampleof~wdenotedbyfwqgq2Q(wecalleachrealizationascenario).Thescenarioapproximationisthen:MinimizeXi2NXk2Kxik 24

PAGE 25

LuedtkeandAhmed ( 2008 ),providesjusticationforconsideringthescenarioapproximation. 2{10 ){( 2{13 )isfeasibletotheprobabilisticedge-partitionproblem( 2{1 ){( 2{4 )withprobabilityatleast1. Proof. 2{2 )and( 2{3 ),letXdenotethesetoffeasiblesolutionstotheprobabilisticedge-partitionproblem(satisfying( 2{2 ){( 2{4 )),andletXQdenotethesetoffeasiblesolutionstothestochasticedge-partitionproblemcorrespondingtoasampleQ(satisfying( 2{11 ){( 2{13 )).WewanttoboundjQjsuchthatPrfXQXg1. Considerasolutionx2XnX,i.e.,PrfG(x;~w)bg<1.Thenx2XQifandonlyifG(x;wq)bforallq2Q.Sincethewqforq2Qarei.i.d.itfollowsthatPrfx2XQg(1)jQj.NowPrfXQ6Xg=Prf9x2XQs.t.PrfG(x;~w)bg<1gPx2XnXPrfx2XQgjXnXj(1)jQjjXj(1)jQj: 1: 25

PAGE 26

Nextwedescribeanextensiveformmodelofthestochasticedge-partitionproblem.LetEqbethesetofedgeswithnon-zeroweightsunderscenarioq.Wedenebinarydecisionvariablesyqijk=1ifedge(i;j)isassignedtosubgraphkinscenarioqand0otherwise,8q2Q;(i;j)2Eq,andk2K.Thestochasticedge-partitionproblemcanthenbeformulatedasfollows:MinimizeXi2NXk2Kxik Observethatifoneweretosolvetheaboveextensiveformproblemgivenby( 2{15 ){( 2{21 ),integralityrestrictionsneedonlybeimposedonthey-variables,whichwouldinturnenforcetheintegralityofthex-variablesatoptimality.Notealsothatgivenaxedsetofx-values,thisproblemdecomposesintojQjseparableintegerprograms,wherethesubproblemcorrespondingtoscenarioq2Qisgivenby:Sq(x)=Maximize0 (2{22) 26

PAGE 27

2{16 ),( 2{18 ),( 2{19 ),and( 2{21 ): Undertheforegoingmodel,itisusefultodenevijk=minfxik;xjkgasapartoftherst-stagedecisionvariables,8(i;j)2E;k2K.Thepresenceofthesevariablesallowustoformulatestrongercuttingplanesthanwouldbepossiblewithjustx-variables(seealso Smithetal. ( 2004 )).Assumingthat[q2QEq=E,theextensiveformproblemisnowequivalentto: MinimizeXi2NXk2Kxik subjecttoXi2Nxikr8k2K where subjecttoXk2Kyqijk=18(i;j)2Eq Thevalidinequalities( 2{27 )requirethatforeachedge(i;j)2E,bothiandjmustbeassignedtosomecommonsubgraph,andareusefulinimprovingthecomputationalecacyofthedecompositionalgorithmthatwepropose.Notethatanoptimalsolutionexistsinwhichvijk=minfxik;xjkg8(i;j)2E;k2K,withoutenforcingintegralityrestrictionsorlowerboundsonthev-variables. 27

PAGE 28

SheraliandSmith 2001 ; Sheralietal. 2000 ).Toreducemodelsymmetrywecanrewritethecardinalityconstraints( 2{25 )(or( 2{17 )fortheextensiveformproblem)byusingthefollowinginequalities: Forascenarioqandagivenvector^v,theproblem( 2{30 ){( 2{33 )isessentiallyanidenticalparallelmachineschedulingproblemtominimizemakespan(P==Cmax)(withsomeassignmentrestrictions).Inparticular,therewouldbejKjmachinesandjEqjjobs,whoseprocessingtimesaregivenbywqij;8(i;j)2Eq.Eachjobmustbeassignedtoexactlyonemachine,andthev-variablesimposesomerestrictionsontheassignments.Theintegerprogrammingschemedevelopedin Smith ( 2004 )istailoredforasimilarprobleminwhichthe(weighted)numberofdemandsthatcannotbeplacedononeofthesesubgraphsisminimized(i.e.,minimumweightednumberoftardyjobs).Thisisnotequivalenttosolvingaminimummakespanproblem;however,theoptimalsolutionofFq(^v)isnomorethanbifandonlyiftheminimumnumberoftardyjobsisequalto0.Ifapositivelowerboundtotheproblemofminimizingthenumberoftardyjobsisestablished,onecanterminatethesubproblemalgorithmandconcludeinfeasibility. Wenowpresentacuttingplanealgorithmforsolving( 2{24 ){( 2{29 ).Theschemerelaxesconstraints( 2{29 )andaddscuttingplanesasnecessarytoenforcefeasibilitytothesubproblems.Letuscalltheproblem( 2{24 ){( 2{28 )themasterproblem(MP). 1. SolveMP.IfMPisinfeasiblethenSTOP;theproblemisinfeasible.Otherwiselet^vbeanoptimalsolutionofMP. 2. Forq2Q,computeFq(^v).IfFq(^v)bforallq,thenSTOP;thecurrentsolutionisoptimal.Otherwise,continuetoStep3. 3. UpdateMPbyaddingacuttingplaneoftheform( 2{37 )aspresentedinRemark1,andreturntoStep1. 28

PAGE 29

LaporteandLouveaux ( 1993 )forthisclassofproblemsisgivenby 2{35 )bydividingbothsidesby(F^q(^v)L^q)androundingdowntoobtain 2{36 ): 2{37 )isvalid,considerasolutionv0thatdoesnotsatisfytheaboveinequality,i.e.,v0ijk=0forall(ijk)2O(^v).Therefore,v0ijk^vijkforall(ijk).ThenFq(v0)Fq(^v)>b,andv0isnotfeasible.Inequality( 2{37 )dominates( 2{36 )sincetheleft-hand-sideof( 2{37 )isnotmorethanthatof( 2{36 ),andtheright-hand-sidesarebothequalto1.Thus,( 2{37 )servesasacuttingplanethatcanbeusedinStep3oftheabovealgorithm. 2{37 ).Intheparlanceofmachinescheduling,insteadoftryingtominimizethemaximummakespan,wemaywishtominimizethetotalsumoftardiness.Letck;8k2Kbeanonnegativevariablethatdenotestheamountofcapacitydecitinsubgraphk.Then,theproblemofminimizingthetotalcapacitydecitcanbeformulated 29

PAGE 30

Clearly,Fq(v)bifandonlyifTq(v)=0,andsowecanreplacemasterproblemconstraints( 2{29 )withtherestrictionsthatTq(v)=0forallscenariosq2Q.IfsubproblemsTq(v)areusedinlieuofFq(v),wewouldobtain( 2{36 )(directly,thistime)fromLaporteandLouveaux'sintegerfeasibilitycut.However,wecanstateastrongercuttingplaneforasolutionvector^vhavingT^q(^v)>0forsomescenario^q,byrequiringthatthetotalamountofadditionalcapacitythatmustbeallocatedtothecollectionofsubgraphsisatleastT^q(^v).Thisinequalityisformallystatedinthefollowingproposition. 2{37 ): 30

PAGE 31

For(i;j)2E^q,ifyijk=1and^vijk=1,thenset^yijk=1aswell. 2. For(i;j)2E^q,ifyijk=1and^vijk=0,thenset^yij^k=1forany^k2Kforwhich(ij^k)2I(^v).(Notethat(ijk)2O(^v)since^vijk=0.) 3. Setallother^yijk=0. Inotherwords,^yisconstructedintwophases.Intherstphase,weensurethatifedge(i;j)wasassignedtosubgraphkinsolutiony,then(i;j)isassignedtokin^yaswell,unless^vijk=0(prohibitingthisassignment).Inthesecondphase,ifyijk=1but^vijk=0,thenweassign(i;j)toany^ksuchthat^vij^k=1.Notethatthisassignmentresultsinasolutionfeasibleto( 2{39 ),( 2{40 ),and( 2{43 ).Next,letusconstruct^c.Observethatintherstphaseofassigningedgestosubgraphsbasedon(ijk)2I(^v)forwhichyijk=1,nosubgraphcapacitiesareviolatedsinceck=0,8k2K,andsoweinitialize^ck=0,8k2K.Inthesecondphase,weguaranteefeasibilityto( 2{41 )(andmaintainfeasibilityto( 2{42 ))byincreasing^c^kbyw^qij.Thus(^y;^c)isafeasiblesolutiontoMTq(^v). Attheendofthesecondphaseofassignments,wehavePk2K^ck=P(ijk)2O(^v)w^qijvijk,sincePk2K^ckisincreasedbyw^qijonlywhenbothvijk=1and(ijk)2O(^v).However,byassumption,wehavethatP(ijk)2O(^v)w^qijvijk
PAGE 32

2{44 )mightbenetfromderivingmultiplecutsforeachinfeasiblescenario,sincethesecutscouldbedistinct. ( 2004 )exploretheinclusionof\warmingconstraints"inthemasterproblem,whichenforcesimplenecessaryconditionsforfeasibilitytoSONETproblems.Denotethedegreeofnodei2Nbydeg(i),andthesetofnodesadjacenttoibyA(i). Leeetal. ( 2000b )showthatnodeimustbeassignedtoatleastldeg(i) 2{34 )needtobeadjustedsothattheyareenforcedseparatelyforsubgraphs1;:::;`^{,and`^{+1;:::;jKj. Sheralietal. ( 2000 )showcomputationallythatsuchavariable-xingschemeimprovessolvabilityofprobleminstances. Smithetal. ( 2004 )notethatanodeicannotbeassignedtoasubgraphkinanoptimalsolutionunlessanadjacentnodeisalsoassignedtothesamesubgraph.Therefore,wealsoincludethefollowingconstraintsinMP: 32

PAGE 33

( 2005 )describesvalidinequalitiesthatcanbederivedbyanalyzingthetopologyofthegraph.First,consideranedge(i;j)2Esuchthat`i=`j=1.LetA(i;j)=A(i)[A(j)fi;jgdenotethesetofdistinctnodesthatareadjacenttoiorj.IfjA(i;j)jr1,theniorjmustbeassignedtoatleasttwosubgraphs.Similarly,wedeneWq(i;j)=Pk2A(i;j)(wqik+wqjk)+wqij,andnotethatifWq(i;j)>bforsomeq2Q,thenwecannotfeasiblyassignnodesiandjtoasinglesubgraph.IfA(i;j)r1orWq(i;j)>b,thenwestatethefollowingvalidinequality: Smith ( 2005 )showsthatnodesiandjcollectivelyneedtobeassignedtoatleastfoursubgraphs,whichwestateas: 2.2 arepreferabletosolvingstochasticedge-partitioninstancesbytheextensiveformproblemgivenby( 2{15 ){( 2{21 ),asweshowinSection 2.4 .However,thetwo-stagecuttingplanealgorithmsstillsuerfromseveralcomputationaldiculties.First,themasterproblem,MP,containsjNjjKjbinaryvariables,jEjjKjcontinuousvariables,andO(jEjjKj)constraints,whichresultsinlargeintegerprograms.Second,thelinearprogrammingrelaxationofMPisquiteweakformanyprobleminstances.Furthermore,thelowerboundimprovesslowlyascutsofthetype( 2{37 )or( 2{44 )areaddedtoMPineachiteration.ThemainreasonforthisslowconvergenceistheexistenceofsymmetryinMP.Inequalities( 2{34 )reduce,butdonotcompletelyeliminate,symmetricsolutionsinMP.Therefore,whenasolutionofMPisfoundtobeinfeasibletoasubproblem,MPoftensimplyswitchestoasymmetricsolution 33

PAGE 34

Inthissectionwedevelopanewdecompositionframeworktoremedythesediculties.Wecombatsymmetryduetoreshuingofsubgraphsbyrepresentingsubgraphsascongurations.AcongurationcisidentiedbyasubgraphnodesetNc(weallowNc=;)andapositiveintegerc,whichgivesthenumberofsubgraphshavingnodesetNc.AsolutionisrepresentedbyacongurationmultisetCwhoseelementsarepairs(Nc;c).Weeliminatesymmetrybyensuringthatnoisomorphiccongurationmultisets(i.e.,thosethatareidenticalafterreindexingcongurationindices)areencounteredinoursearch. AcongurationmultisetCsatisesthefollowingnecessaryfeasibilityconditions. AmultisetCthatsatisesF1,F2,andF3representsafeasiblesolutionifalledgescanbepartitionedonthesetofsubgraphscorrespondingtoCwithoutviolatingtheweightrestrictionsforanyscenario.NotethatthenumberofdistinctcongurationsinC,whichwedenotebyjCj,isdynamicallydeterminedinouralgorithm. Wenowprovideanoverviewofourthree-stagehybridalgorithm. 1. Therst-stageproblemdetermines(viaoptimalsolutionofamixed-integerprogram)thenumberoftimesweassigneachnodetothecongurationsinC.Forinstance,intheexamplegiveninFigure 2-1 a,wecouldspecifythatwemustusetwocopiesofnodes4and5,andonecopyoftheothernodes. 2. Inthesecondstage,weseekamultisetCthatusesexactlythenumberofnodeassignmentsspeciedintherstphaseandsatisesF1,F2,andF3.Intheexamplementionedabove,amultisetChavingcongurationsf1;2;4g,f3;4;5g,andf5;6g(eachwithmultiplicityone)couldbegeneratedbasedontherst-stagesolution. 34

PAGE 35

Finally,inthethirdstage,wedeterminewhetherCisfeasible.IfCisfeasiblethenwestopwithanoptimalsolution.Else,wereturntothesecondstage,andgenerateadierentmultisetmeetingthestatedcriteria.Ifnosuchmultisetexists,acutisaddedtotherst-stageproblem,whichisthenre-solved.Fortheexamplegivenabove,themultisetyieldsafeasiblesolution(seeFigure 2-1 b). (2{51)`izijKj8i2N (2{53) where`iisalowerboundonthenumberofcopiesrequiredfornodei,asgivenin( 2{45 ).Toformulatetherst-stageproblemasanintegerprogram,werewrite( 2{51 )asanexponentialsetoflinearinequalitiesbyconsideringthez-vectorsthatviolateit.Werstneedtointroduceauxiliarybinaryvariablestik;8i2N;k=`i;:::;jKj,sothattik=1ifzi=k.Then,givenavector^zthatdoesnotinduceafeasiblemultiset,wenotethatnozsuchthatzi^zi;8i2N,inducesafeasiblemultiset.Hence,atleastonecomponentof^zmustbeincreased,andso isavalidinequality.Ourrst-stageproblemcannowbeexpressedasthefollowingintegerprogram:MinimizeXi2Nzi 35

PAGE 36

whereZisthesetofallz-vectorsthatdonotinduceafeasiblemultiset.(Thez-variablesareinfactunnecessaryinthisformulation,butwekeepthemforeaseofexposition.)Inouralgorithmwerelaxconstraints( 2{58 )intherst-stageproblem,andaddtheminacuttingplanefashion.Ineveryiterationwesolvetherst-stageproblemtond^z,andsolvethesecond-andthird-stageproblemstoseekafeasiblemultisetinducedby^z.Ifafeasiblemultisetisfound,then^zinducesanoptimalsolutionandwestop.Otherwise,weaddacutoftype( 2{58 )andre-solvetherst-stageproblem. Smith ( 1995 ), LustigandPuget ( 2001 ),and Rossietal. ( 2006 )forathoroughdiscussionofconstraintprogrammingtechniques. 36

PAGE 37

Forinstance,considerave-nodegraph,andletthez-vectorobtainedbytherst-stageproblembe^z=(2;3;1;4;3).Supposethatnodes1,2,and3havebeenprocessed,andthefollowingpartialmultisetwithjCj=3hasbeenobtained: Supposethatweprocessnode4bychoosingitsdistributionvectoras^4=(2;1;1).Addingnode4totwoofthevecopiesofN1createsanewcongurationN01whosenodesetconsistsonlyofnode4(withmultiplicitytwo)andreducesthemultiplicityofN1bytwo.Aftersimilarlyaddingonecopyofnode4toN2andonecopyofnode4toN3,weobtainthefollowingpartialmultisetwithjC0j=5: Ingeneral,whenweprocessnodeibychoosingadistributionvectori,weupdatethepartialmultisetCasfollows.Foreachcongurationc2Cific=0,thennochanges 37

PAGE 38

InthebeginningofthesecondstageweinitializeourmultisetCwithasinglecongurationhavingN1=;and1=jKj.Eachnodecanonlybeaddedtotheloneconguration,andsothedomainfornodeiisinitiallythesingleone-dimensionalvectori=(^zi).Ouralgorithmnextprocessessomenodei2Nandupdatestheexistingsetofcongurations:N1=;;1=jKj^ziandN2=fig;2=^zi.Next,thedomainsofallunprocessednodesareupdatedtoreectthechangesinC.Foreachunprocessednodej,weenumerateallpossiblewaysofpartitioning^zjcopiesintonodesetsN1andN2.Thislogicisrepeatedatallfuturestepsaswell.Forinstance,intheexamplegivenabove,supposethat^5=(2;0;1)wastheonlyvectorinthedomainofnode5beforeprocessingnode4.Sinceprocessingnode4modiestherstcongurationbyreducing1andgeneratesanewconguration(N01;01),weexpandthedomainofnode5byenumerating 38

PAGE 39

ToenforceF2,thepropagationalgorithmidentiesallcongurationstowhichrnodeshavebeenassigned.Foreachsuchcongurationc,weremovealldistributionvectorsjhavingjc>0fromthedomainsofallunprocessednodesj2N.ToenforceF3,thepropagationalgorithmiteratesoverthedomainsoftheunprocessednodesjadjacenttoi,andremovesalldistributionvectorsthatdonotaddatleastonecopyofjtoanycongurationinCi.Otherwise,thecongurationscontainingnodeiwouldbedisjointfromthosecontainingnodej,whichviolatesF3. 39

PAGE 40

vanBeek 2006 ). Wecallonesuchtestimpliednodeassignmentanalysis.Supposethatweidentifyaprocessednodeisuchthat^zi=1,andthecongurationctowhichihasbeenassigned.ByconditionF3itfollowsthatallunprocessednodesjadjacenttoimustalsobeassignedtocongurationc.Weusethisanalysistoaugmentpartialcongurationswithimpliednodeassignments,andthencheckwhetheranyaugmentedcongurationcontainsmorethanrnodes,andhenceviolatesF2. Wealsoperformanimpliededgeassignmentanalysisbyndingalledgesthatcanonlybeassignedtoasingleconguration.Foreach(i;j)2E,ifbothnodesiandjhavebeenprocessed,thenwecheckwhetherbothiandjareinasinglecongurationcforwhichc=1.Inthiscaseedge(i;j)canonlybeassignedtocongurationc.Ontheotherhandif(withoutlossofgenerality)nodeihasbeenprocessedbutnodejhasnotyetbeenprocessed,and^zi=1,thenedge(i;j)canonlybeassignedtothecongurationtowhichihasbeenassigned.Afterndingallimpliededgeassignments,wecheckwhetherF3isviolatedforanyscenario. Finally,weconsiderasingletonnodeanalysis,inwhichweensurethateachnodeisadjacenttoatleastoneothernodeineachconguration.Foreachprocessednodei,andforallcongurationsc2Ci,weseekanodejadjacenttoisothateitherj2Nc(ifjalsohasbeenprocessed),orjc>0forsomedistributionvectorinthedomainofj(ifjhasnotbeenprocessed).Ifnosuchjcanbefoundforacongurationc2Ci,thenthecurrentpartialsolutioncannotleadtoanoptimalsolution;nodeicanultimatelyberemoved 40

PAGE 41

LustigandPuget 2001 ; Smith 1995 ).Especiallyforinfeasiblesecond-stageprobleminstances,processingthe\problematic"nodesrstcanquicklyleadtothedetectionofinfeasibilityandcanresultinsignicantsavingsincomputationaltime.Weemployadynamicnodeselectionruleinwhichtheorderofnodesconsideredcanvaryindierentsectionsofthesearchtree.Inaccordancewiththe\fail-rst"principlewidelyusedinconstraintprogrammingalgorithms( HaralickandElliott 1980 ; vanBeek 2006 ),ournodeselectionrulerstpicksanunprocessednodethat 1. hasthefewestnumberofdistributionvectorsinitsdomain, 2. hasthefewestnumberofcopiestobepartitioned,and 3. hasthelargestnumberofunprocessedadjacentnodes, breakingtiesinthegivenorder.Inthismanner,wecanquicklyenumerateallpossibledistributionvectorsofafewkeynodes,allowingconstraintpropagationtoquicklyreducethesizeoftheremainingsearchspace. 41

PAGE 42

2{38 ){( 2{43 )tosolvethethird-stageproblem. Notethatthistransformationre-introducessymmetryintothethird-stageproblem.However,thesolutionofthethird-stageproblemsdoesnotconstituteabottleneckinthealgorithm,andsymmetry-breakingconstraintsappendedtothetransformedsubproblemwillnotimpactthecomputationalecacyoftheoverallalgorithm. 2{58 )statethatthenumberofcopiesofsomenodemustbeincreased,buttheydonotcontainanyinformationaboutwhichnodesneedtobeadded.Weobservethattheprogressofoursecond-stagealgorithmcanbeanalyzedtoidentifya\problematic"subsetofnodeswhosecorrespondingz-valuescauseinfeasibilityregardless 42

PAGE 43

Constraints( 2{61 )clearlydominate( 2{58 )foranyPN,andgetstrongerasjPjdecreases.Basedonthisobservation,weupdateournodeselectionrulebygivingpreferencetoselectingnodesthathavealreadybeenaddedtoP.Ourrevisednodeselectionrulerstpicksanodethat 0. hasbeenaddedtoPinapreviousbacktrackingstep, 1. hasthefewestnumberofdistributionvectorsinitsdomain, 2. hasthefewestnumberofcopiestobepartitioned,and 3. hasthelargestnumberofunprocessedadjacentnodes, againbreakingtiesinthestatedorder. 3 canbeadaptedtotherst-stageproblemtoeliminatethez-vectorsthatviolatethecorrespondingnecessaryfeasibilityconditions.Inparticular,constraints( 2{46 )translatetosimplelowerbounds( 2{52 )onthez-variables.Constraints( 2{48 ),whicharewrittenfornodepairsthatsatisfythe 43

PAGE 44

3 ,canbewrittenas: Similarly,eachconstraintoftype( 2{49 )canbeequivalentlyrepresentedasfollowing: Smith ( 2005 )discussesanadditionalvalidinequality,whichcannotberepresentedusingthex-variablesinourtwo-stagealgorithm,butcanbewrittenintermsofthez-andt-variablesintherst-stageproblemofourhybridalgorithm.Fornodesi2Nandj2N,if(i;j)=2E;deg(i)r1;deg(j)r1;jA(i;j)jr1,andthereexistsacommonneighbork2Nsothatk2A(i);k2A(j);deg(k)r,andifi;j;khavemorethan2r4distinctneighborsintotal,thenzi=1;zj=1implieszk3.Thisconditioncanbewrittenas: whichreducestozk3forzi=zj=1,andisredundantotherwise. 2{62 ){( 2{64 )toobtainaninitialsolution^z,andexecutethesecond-andthird-stagealgorithmstoseekafeasiblemultiset.Ifoneisfound,weterminatewithanoptimalsolution.Otherwise,weinvestigatethesetofprocessednodes^PN,andpickanode^{2^Phavingthefewestnumberofcopies(breakingtiesbypickinganodehavingthelargestdegree).Wethenset^z^{=^z^{+1andre-invokethesecond-andthird-stagealgorithms.Thisalgorithmeventuallyndsafeasiblemultisetorconcludesthattheentireproblemisinfeasibleaftergeneratingthesolution^zi=jKj;8i2

PAGE 45

2{61 )foreach^zgeneratedbeforeafeasiblemultisetisfound,whichweaddtotherst-stageproblemtoimprovethelowerbound. 2{61 ),andagainresumethesolutionprocess.Thesameideaisalsoapplicabletothemasterproblem(MP)ofthetwo-stagealgorithmdiscussedinSection 2.2 Inourtests,thisapproachturnedouttobemoreeectivethansolvingtherst-stageproblemtooptimalityineachiteration,addingacut,andre-solvingit.Thereasonisthattheproblemissolvedusingasinglebranch-and-boundtree,whichwetightenbyaddingcutsasnecessaryonintegralnodes,insteadofrepeatedlygeneratingabranch-and-boundtreeineachiteration.Italsoallowsustoobtaingoodfeasiblesolutionsforprobleminstancesthataretoodiculttosolvetooptimality. Wenotethatthisapproachrequiresaminormodicationtothesecond-stagealgorithm.Allconstraintpropagation(Section 2.3.2.3 )andforwardcheckingrules(Section 2.3.2.4 )exceptforsingletonnodeanalysisarebasedonnecessaryconditionsforfeasibilityofcongurations,andthereforetheyarevalidforanyintegral^z.However,singletonnodeanalysisisbasedonanoptimalityconditionandhencecanonlybeusedif^zisacandidateoptimalsolutiontotherst-stageproblem. jNj(jNj1))takesvalues0.2,0.3,and 45

PAGE 46

Goldschmidtetal. ( 2003 ); Sheralietal. ( 2000 ); Smith ( 2005 )).ChoosingjKjtoosmallmaymaketheprobleminfeasible,andlargevaluesofjKjincreasedicultyoftheproblem.Inourtests,wechosejKjsucientlylargetoyieldafeasibleedgepartitionineachprobleminstance.Ingeneratinginstanceswerstpickedarandomsubsetofedgestohaveapositiveweight,andthenweassignedaweightuniformlydistributedbetween1and10toeachedgeineachscenario.Wegeneratedveprobleminstancesforeachproblemsize,whichisdeterminedbytheexpectededgedensity,thenumberofnodes,andthenumberofscenarios.ThedatasetnamesanddetailsusedinourexperimentsaregiveninTable 2-1 Table2-1. Descriptionsoftheprobleminstancesusedforcomparingalgorithms 2{27 ),( 2{46 ){( 2{49 ),andthesymmetry-breakingconstraints( 2{34 ),andusesthemodelgivenby( 2{38 ){( 2{43 )forthesubproblem,whichistheformulationthatminimizesthetotaltardiness.Inourbasesettingforthethree-stagealgorithm,weusedourheuristictondaninitialfeasiblesolution,generatedvalidinequalities( 2{62 ){( 2{64 ),and(similartothetwo-stagealgorithm)weusedformulation( 2{38 ){( 2{43 )forthethird-stageproblem.WeusedcallbackfunctionsofCPLEXtogenerateasinglebranch-and-boundtreeforbothtwo-stageandthree-stagealgorithmsasdiscussedinSection 2.3.5 .Weimposedahalf-hour 46

PAGE 47

Ourrstexperimentcomparestheperformanceoftheextensiveform,two-stage,andthree-stagealgorithms.Table 2-2 summarizestheresultsofthesethreealgorithmsonlowdensitygraphshavingexpectededgedensity0:2.Foreachproblemsize,wereportthefollowingstatisticscalculatedoververandominstances:(i)thenumberofproblemssolvedtooptimality(\Solved"),(ii)theaverageoptimalitygapobtainedattherootnode(\RootGap"),(iii)theaveragenaloptimalitygapforinstancesthatcouldnotbesolvedwithintheallowedtimelimit(\FinalGap"),(iv)theaverageamountoftimespentbyeachalgorithmontheinstancesthatweresolvedtooptimality(\Time").Outofthe75instancesinthisdataset,CPLEXcouldsolvetheextensiveformtooptimalityfor61instances,whilebothtwo-stageandthree-stagealgorithmssolvedall75instancestooptimalitywithinafewseconds.Theresultsrevealthattheperformanceoftheextensiveformformulationdeterioratesrapidlyasthenumberofscenariosincreases,buttheeectofthenumberofscenariosismitigatedforthetwo-stageandthree-stagealgorithms.Weobservethattheaverageoptimalitygapobtainedbythethree-stagealgorithmattherootnodeis1:46%,whichissignicantlylessthantheinitialgapsobtainedusingotherapproaches. Table2-2. Comparisonofthealgorithmsongraphshavingedgedensity=0:2 Two-Stage Three-Stage RootFinal RootFinal RootFinalName SolvedGapGapTime SolvedGapGapTime SolvedGapGapTime 5-1 50.00%-0.1 55.00%-0.1 50.00%-0.15-30 518.33%-6.6 54.00%-0.2 50.00%-0.15-100 512.38%-5.4 511.00%-0.6 50.00%-0.38-1 525.90%-0.4 56.67%-0.1 50.00%-0.18-30 512.89%-4.0 53.64%-0.2 50.00%-0.18-100 537.61%-223.1 514.84%-1.2 50.00%-0.310-1 519.58%-0.5 517.80%-0.4 50.00%-0.110-30 557.01%-147.3 510.71%-0.8 50.00%-0.210-100 430.35%7.14%684.1 513.94%-2.0 50.00%-0.412-1 547.25%-8.1 524.66%-2.2 50.00%-0.112-30 455.09%25.00%507.3 517.99%-4.3 53.08%-0.312-100 262.21%24.88%713.1 536.28%-4.7 52.11%-0.815-1 531.85%-33.0 564.38%-16.4 54.56%-0.215-30 165.29%21.65%864.6 539.13%-27.1 57.29%-0.615-100 057.33%28.47%524.49%-20.4 54.86%-1.2

PAGE 48

2-3 and 2-4 comparethethreeapproachesondensergraphshavingedgedensity0:3(mediumdensity)and0:4(highdensity),respectively.Weobservethatperformancesofallthreealgorithmsdeteriorateastheedgedensityincreases,whichisnotsurprisingduetothenatureoftheedge-partitionproblem.Thenumberofinstances Table2-3. Comparisonofthealgorithmsongraphshavingedgedensity=0:3 Two-Stage Three-Stage RootFinal RootFinal RootFinalName SolvedGapGapTime SolvedGapGapTime SolvedGapGapTime 5-1 50.00%-0.1 52.86%-0.1 50.00%-0.15-30 525.76%-10.4 56.15%-0.5 50.00%-0.15-100 510.00%-3.1 510.77%-0.4 50.00%-0.28-1 531.30%-0.5 511.20%-0.1 50.00%-0.18-30 542.57%-18.0 57.48%-0.4 50.00%-0.28-100 439.37%7.14%110.0 516.19%-1.3 51.43%-0.310-1 532.42%-3.7 516.27%-0.6 51.18%-0.110-30 451.33%21.05%953.0 540.82%-8.2 55.83%-0.310-100 261.24%29.05%382.6 535.43%-302.7 58.89%-0.512-1 553.85%-312.0 539.49%-16.8 54.65%-0.212-30 063.41%27.06%546.98%-120.5 59.31%-0.812-100 084.24%65.50%442.78%4.35%89.0 511.99%-1.415-1 446.88%11.54%460.4 572.86%-250.5 512.93%-0.915-30 066.01%42.41%272.20%16.02%30.4 513.51%-3.515-100 080.76%74.48%053.05%13.21%516.31%-4.1 2-4 clearlyshowsthatthethree-stagealgorithmdominatestheotherapproaches,andthetwo-stagealgorithmprovidesbetterresultsthandirectlysolvingtheextensiveformulation.OuranalysisofoptimalsolutionsobtainedfortheprobleminstancesshowninTables 2-2 { 2-4 showedthattheaverageobjectivefunctionvalueforthedeterministic(single-scenario)probleminstancesis14.8.Thisvalueissmallerthantheaverageobjectivefunctionvaluefor30-scenarioand100-scenarioinstances(15.52and15.6,respectively).Wealsoobservethatseveralsubgraphscanbeemptyinanoptimalsolution. 48

PAGE 49

Comparisonofthealgorithmsongraphshavingedgedensity=0:4 Two-Stage Three-Stage RootFinal RootFinal RootFinalName SolvedGapGapTime SolvedGapGapTime SolvedGapGapTime 5-1 55.00%-0.1 50.00%-0.1 50.00%-0.15-30 524.67%-2.6 519.79%-0.3 50.00%-0.15-100 512.38%-5.6 58.31%-0.6 50.00%-0.28-1 541.32%-2.0 53.33%-0.1 50.00%-0.18-30 548.89%-140.9 517.68%-1.1 51.43%-0.18-100 347.23%22.50%113.0 521.08%-8.7 52.50%-0.410-1 545.08%-48.3 532.36%-3.5 52.16%-0.110-30 061.52%20.64%556.55%-39.5 58.45%-0.410-100 064.47%50.91%354.82%7.50%151.7 512.73%-1.512-1 167.13%14.30%33.2 540.60%-327.3 57.86%-0.512-30 088.61%46.74%542.93%-160.9 53.16%-0.812-100 084.37%68.24%551.54%-583.7 513.91%-1.715-1 260.11%11.21%369.6 353.01%5.56%410.0 511.57%-0.915-30 085.29%65.29%066.72%22.66%318.03%4.74%120.215-100 096.00%86.92%062.62%24.58%319.98%6.45%173.8 Descriptionsoftheprobleminstancesusedforanalyzingthree-stagealgorithm 2-5 .Similartoourpreviousexperiments,we Table2-6. Three-Stagealgorithmongraphshavingedgedensity=0:2 RootFinalHeuristic RootFinalHeuristicName 5 40750.00%-0.82.86% 219450.00%-3.50.00%8 83750.00%-2.21.54% 434350.00%-12.73.33%10 1169510.88%-5.65.09% 600651.33%-19.61.33%12 172451.11%-13.24.19% 877953.00%-58.53.33%15 214057.24%-22.12.74% 10858512.41%-170.94.37%17 2417510.19%-41.04.78% 1224559.82%-211.18.01%20 2833516.55%-79.87.01% 14324512.40%-403.74.51%22 3110514.77%-128.26.49% 15710515.78%-699.96.46% 1 .Hence,inequality( 2{14 )ensuresthatwecanbe95%(99%,respectively)certainthatalldemandscanbesatised95%(99%,respectively)ofthetime.Wegeneratedverandominstancesforeachdataset, 49

PAGE 50

Three-Stagealgorithmongraphshavingedgedensity=0:3 RootFinalHeuristic RootFinalHeuristicName 5 40750.00%-0.82.86% 219450.00%-3.50.00%8 83750.00%-2.62.86% 434353.33%-13.02.86%10 116956.58%-7.58.99% 6006511.86%-27.44.80%12 172458.61%-16.14.51% 877958.22%-71.74.31%15 2140515.45%-45.13.05% 10858418.53%3.45%176.44.25%17 2417513.63%-42.43.43% 1224559.93%-189.32.68%20 2833517.47%-362.53.24% 14324518.13%-639.13.32%22 3110416.18%4.76%159.35.82% 15710515.86%-738.53.45% 2-2 ,Tables 2-6 2-7 ,and 2-8 showtherelativegapbetweenthequalityofthesolutionfoundbyourinitialheuristic(Section 2.3.5.2 )andthebestlowerboundobtained(\HeuristicGap").Ouralgorithmcansolve206instancesoutof240tooptimality,andprovidesanaverage Table2-8. Three-Stagealgorithmongraphshavingedgedensity=0:4 RootFinalHeuristic RootFinalHeuristicName 5 40750.00%-0.82.22% 219450.00%-3.20.00%8 83757.71%-2.72.43% 434357.25%-17.35.33%10 1169516.38%-9.45.71% 6006515.84%-52.19.73%12 1724516.71%-63.25.41% 8779514.13%-118.45.45%15 2417416.24%2.86%338.84.07% 12245216.55%4.71%549.86.86%17 2140123.46%8.46%993.79.23% 10858124.77%11.44%1515.513.07%20 2833018.83%9.46%-9.46% 14324020.01%11.30%-11.81%22 3110018.36%11.05%-11.47% 15710017.83%9.90%-10.29% 50

PAGE 51

3-1 (a) (b) Figure3-1. (a)Amultileafcollimatorsystem(b)Theprojectionofanapertureontoapatient 51

PAGE 52

Sincethemid1990's,large-scaleoptimizationoftheuenceappliedfromanumberofbeamorientationsaroundapatienthasbeenusedtodesigntreatmentsfromMLC-equippedlinearaccelerators.AtypicalapproachtoIMRTtreatmentplanningistorstselectthenumberandorientationsofthebeamstouseaswellasanintensityproleoruencemapforeachofthesebeams,wheretheuencemaptakestheformofamatrixofintensities.Thisproblemhasbeenstudiedextensivelyandcanbesolvedsatisfactorily,inparticularwhen(asiscommoninclinicalpractice)thebeamorientationsareselectedmanuallybythephysicianorclinicianbasedontheirinsightandexpertiseregardingtreatmentplanning.Foroptimizationapproachestotheuencemapoptimizationproblemwithxedbeamorientationswerefertothereviewpaperby Shepardetal. ( 1999 ).Morerecently, Romeijnetal. ( 2006 )proposednewconvexprogrammingmodels,and HamacherandKufer ( 2002 )and Kuferetal. ( 2003 )consideredamulti-criteriaapproachtotheproblem. Leeetal. ( 2000a 2003 )studiedmixed-integerprogrammingapproachestotheextensionoftheuencemapoptimizationproblemthatalsooptimizesthenumberandorientationsofthebeamstobeused.However,toenabledeliveryoftheoptimaluencemapsbytheMLCsystem,theyneedtobedecomposedintoacollectionofdeliverableapertures.(Forexamplesofintegratedapproachestouencemapoptimization,alsoreferredtoasaperturemodulation,wereferto Shepardetal. ( 2002 ), Preciado-Waltersetal. ( 2004 ),and Romeijnetal. ( 2005 ).) ThevastmajorityofMLCsystemscontainacollectionofleavesthatcanbemovedinparallel,therebyblockingpartoftheradiationbeam.Thisarchitectureimpliesthatwecanvieweachbeamasamatrixofbeamletsorbixels(thesmallestdeliverablesquarebeamthatcanbecreatedbytheMLC),sothateachaperturecanberepresentedbyacollectionofrows(or,byrotatingtheMLChead,columns)ofbixels,eachofwhich 52

PAGE 53

Baataretal. ( 2005 )):264364215375: 53

PAGE 54

Bortfeldetal. ( 1994 )proposedthesweepmethod,which AhujaandHamacher ( 2005 )(whoderivedanequivalentmethod)showedtoindeedyieldanoptimalsolution;otherexactalgorithmswereproposedby Kamathetal. ( 2003 ),and Siochi ( 1999 ).Inaddition, Baataretal. ( 2005 ), Bolandetal. ( 2004 ), Kalinowski ( 2005a ), Kamathetal. ( 2004a b c d ), Lenzen ( 2000 ),and Siochi ( 1999 )studiedtheproblemofminimizingbeam-on-timeunderadditionalhardwareconstraints,while Kalinowski ( 2005b )studiedthebenetsofallowingrotationoftheMLChead. AlthoughthetimerequiredbytheMLCsystemtotransitionbetweenaperturesformallydependsontheaperturesthemselves,thefactthatthesetimesaresimilarandthepresenceofsignicant(aperture-independent)vericationandrecordingoverheadtimesjustiestheuseofthetotalnumberofsetups(or,equivalently,thetotalnumberofapertures)tomeasurethetotalsetuptime.Inaddition,deliveringIMRTwithasmallnumberofaperturesprovidestheadditionalbenetsoflesswear-and-tearonthecollimators(lessstoppingandstarting)andalesserror-pronedeliveryasIMRTdeliveryerrorsareknowntobeproportionaltothenumberofapertures(see Stelletal. ( 2004 )).Theproblemofdecomposingauencemapintotheminimumnumberofrow-convexapertureshasbeenshowntobestronglyNP-hard(see Baataretal. ( 2005 )),leadingtothedevelopmentofalargenumberofheuristicsforsolvingthisproblem.Notableexamplesaretheheuristicsproposedby Baataretal. ( 2005 )(whoalsoidentifysomepolynomially 54

PAGE 55

AgazaryanandSolberg ( 2003 ), DaiandZhu ( 2001 ), Que ( 1999 ), Queetal. ( 2004 ), Siochi ( 1999 2004 2007 ), VanSantvoortandHeijmen ( 1996 ), XiaandVerhey ( 1998 ).Inaddition, Engel ( 2005 ), Kalinowski ( 2005a ),and LimandChoi ( 2007 )developedheuristicstominimizethenumberofapertureswhileconstrainingthetotalbeam-on-timetobeminimal.Finally, Langeretal. ( 2001 )developedamixed-integerprogrammingformulationoftheproblem,while Kalinowski ( 2004 )proposedanexactdynamicprogrammingapproachfortherelatedproblemofminimizingthenumberofaperturesthatyieldstheminimumbeam-on-time. Baataretal. ( 2007 )describedintegerprogrammingandconstraintprogrammingmodelsforthesameproblem,and Ernstetal. ( 2009 )proposedaconstraintprogrammingapproachforminimizingthenumberofapertures.However,computationalstudiesreportedin Baataretal. ( 2007 ); Ernstetal. ( 2009 ); Kalinowski ( 2004 ); Langeretal. ( 2001 )revealthattheseapproachescanonlybeusedtoecientlysolvesmallprobleminstancestooptimality.Ourprimarycontributionisthatwedeveloptherstalgorithmcapableofsolvingclinicalprobleminstancestooptimality(ortoprovablynear-optimality)withinclinicallyacceptablecomputationaltimelimits. Inthischapter,ourfocusisontheproblemofndingadecompositionofauencemapintorow-convexaperturesthatminimizestotaltreatmenttime,asmeasuredbythesumofthetotalsetuptimeandbeam-on-time.InSection 3.2 wedevelopourdecomposition-basedsolutionapproach.InSection 3.3 wediscusstheapplicationofouralgorithmonacollectionofclinicalandrandomlygeneratedtestdata,andcompareitsperformancewithalternativeexactandheuristictechniques. 55

PAGE 56

WestartthissectionbydescribingadecompositionframeworkfortheoptimalleafsequencingprobleminSection 3.2.1 andusethistoformulateourmasterprobleminSection 3.2.2 .WeintroduceoursubprobleminSection 3.2.3 ,proveitscomplexity,andprovideacombinatorialsearchalgorithmforitssolution.WethenenhancetheempiricalperformanceofourdecompositionalgorithmbyintroducingclassesofvalidinequalitiestothemasterprobleminSection 3.2.4 ,andnallydescribeanalgorithmforconstructingafeasiblesolutionwithmedicallydesiredpropertiesinSection 3.2.5 Ehrgottetal. ( 2008 )).However,thisapproachisnotdirectlyapplicablewhentheobjectiveistominimizethetotaltreatmenttime. Eventhoughtheoptimalleafsequencingproblemisnotdirectlydecomposablebyrows,thefactthatleavescorrespondingtodierentrowscanbepositionedindependentlycanstillbeexploited.Denoteaparticularpositioningofleftandrightleavesforarowasaleafposition;anapertureiscomposedofaleafpositionforeachrowofB.Ourmainobservationisthatgivenacollectionofintensities,whichcanbeusedinaperturesthatcollectivelycovertheuencemap,therowsareindependentofoneanother.Thatis,wecandeterminetheleafpositionstobeusedforcoveringeachrowindependently,andthenformaperturesforcoveringtheentireuencemapbycombiningindividualleafpositionsforeachrowthatareassignedtothesameintensity. 56

PAGE 57

Considertheallowableintensitymultisetf1,3,5g.Assigningeachofthesevaluestoatmostoneleafposition,therstrowcanbedecomposedas [148]=1[110]+3[011]+5[001]; sothattheallowableintensitymultisetiscompatiblewiththerstrow.Similarly,thesecondrowcanbedecomposedas [385]=3[110]+5[011]: However,therstbixelinthethirdrowmustbecoveredbytwoleafpositionsassignedtointensities1and3,andthesecondbixelmustbecoveredbyasingleleafpositionassignedtointensity5.Therefore,allallowableintensitiesmustbeusedtocoverthersttwobixels,andthethirdbixelwithrequiredintensity3cannotbecovered.Hence,theallowableintensitymultisetisnotcompatiblewiththethirdrow.Alternatively,consideranallowableintensitymultisetthatcontainsthevalues1,3,and4forthesameuencemap.Therowscanbedecomposedas [148]=1[111]+3[011]+4[001];

PAGE 58

(3{4) [453]=1[010]+3[001]+4[110]: Themasterproblemcanthereforesuccinctlybewrittenas minimizew1LX`=1x`+w2LX`=1`x`(3{6)subjectto 58

PAGE 59

whereofcourse~Tcannotbelessthantheminimumachievablebeam-on-time~z(whichcanbefoundinpolynomialtimeusingthealgorithmsmentionedinSection 3.1 ). Toformulateourmasterproblemasanintegerprogrammingproblem,weintroducebinaryvariablesy`r,8`=1;:::;L,r=1;:::;R`,wherey`r=1ifandonlyifx`=r,andR`isanupperboundonthenumberofapertureshavingintensity`usedinanoptimalsolution.(WecancomputeR`bycomputinganinitialupperboundontheoptimalobjectivefunctionvalueviaanyoftheheuristicsmentionedinSection 3.1 ,andthensettingR`tothelargestvaluesuchthatw1R`+w2`R`isnomorethanthisbound.)Usingthesedecisionvariables,wecanreformulatethemasterproblem(MP)asfollows: minimizew1LX`=1x`+w2LX`=1`x`(3{10)subjectto Wenextformulate( 3{13 )asasetoflinearinequalitiesbyderivingvalidinequalitiesthatcutopreciselythosevectorsxthatviolate( 3{13 ).Tothisend,consideraparticular 59

PAGE 60

Sinceallintegersolutionsexceptfor^xsatisfy( 3{16 ),itisindeedavalidinequality.Constraint( 3{16 )canbetightenedbyobservingthatifthesolution^xisincompatiblewithrowi,thenanysolutionxsuchthatx`^x`,8`=1;:::;L,isalsoincompatiblewithrowi.Therefore,werequirethatxcontainatleastonecomponentthatislargerthanitscorrespondingcomponentin^x,whichyieldsthestrongervalidinequality Constraint( 3{17 )can,inturn,betightenedfurtherbyexplicitlyconsideringtherowsforwhichxisincompatible.LetLi=maxj=1;:::;nbijbethemaximumintensityintheuencemapforrowi.Bythesameargumentasabove,ifthecurrentsolution^xisincompatiblewithrowi,thenanysolutionxsuchthatx`^x`,8`=1;:::;Li,isalsoincompatiblewithrowi,sincenoleafpositionswithintensitygreaterthanLicanbeusedindecomposingrowi.Therefore,werequirethatxislargerthan^xinatleastonecomponent1;:::;Li: Since( 3{18 )isstrongerthan( 3{16 )or( 3{17 ),weusethelatterinequalitiesinourmodel.Notealsothat( 3{18 )statedforrowi1dominatesacutgeneratedforrowi2ifLi1
PAGE 61

3.2.3 Sincethecollection( 3{18 )containsanexponentialnumberofvalidinequalities,weaddthemonlyasneededinacuttingplanefashion.Inparticular,thismeansthatwerelax( 3{18 ),solvetherelaxationof(MP)andgenerateanx-solutionrepresentingacandidateallowableintensitymultiset.Wethensolveasubproblemforeachbixelrowtodetermineiftheallowableintensitymultisetisincompatiblewiththatrow.Ifnot,wehavefoundanoptimalsolutionto(MP).Otherwise,weaddaconstraintoftheform( 3{18 )to(MP)thatcutsothatsolution. Werepresentafeasibledecompositionasacollectionofn-dimensionalbinaryvectorsv`r(`=1;:::;L;r=1;:::;x`).Thevaluesofv`rthatequal1correspondtothe(consecutive)exposedbixelsintherthaperturehavingintensity`.Forexample,thedecompositioninequation( 3{2 )correspondstov11=(1;1;0),v31=(0;1;1),v51=(0;0;1),andv`r=0forother`;r.(Notethatthisdecompositionwouldbefeasibleaslongasx1;x3;x51.)Thesubproblemcanthenformallybepresentedasfollows: QUESTION:Dothereexistn-dimensionalbinaryvectorsv`r(`=1;:::;L;r=1;:::;x`)thatsatisfytheconsecutive-onespropertysuchthatPL`=1Px`r=1`v`r=b?

PAGE 62

Inprinciple,theC1-Partitionproblemcanbeformulatedandsolvedasanintegerprogram.However,wehavedevelopedacomputationallymoreeectivebacktrackingalgorithmthatfocusesonpartitioningintensityrequirementsindividuallyforeachbixel.Anintegervectorpj=(pj1;:::;pjL)providesabixeldecompositionofbixelj2f1;:::;nginrowbifandonlyifbj=PL`=1`pj`.Wethenattempttoformacollectionofleafpositionsthatrealizestheindividualbixelpartitions.Wecallsuchacollectionofleafpositionsaleafdecompositionofb. Tomoreeectivelyconductoursubproblemsearches,wedescribeapropertythatholdsinsomeleafdecomposition(ifoneexists)thatsatisesthegivencollectionofbixeldecompositions. Proof. Wenextderiveanecessaryconditionthatanyfeasiblebixeldecompositionmustsatisfysothatthecorrespondingsetofleafpositionsiscompatiblewithagivenallowableintensitymultisetx.SimilartotheideabehindLemma 1 ,ifpj`>pj+1`,thenpj`pj+1`leafpositionshavingintensity`mustexposebixeljbutnotj+1.Lemma 2 formalizesthisidea.

PAGE 63

Wenextdescribeourbacktrackingalgorithm.Inthisalgorithm,werstenumerateallpossiblewaysofdecomposingthebixelintensitiesinbusingasubsetoftheallowableintensitymultisetgivenbyx.WedenotethesetofallcandidatebixeldecompositionsforbixeljbyPj,whereforeachp2[nj=1Pj,wemusthavep`x`;8`=1;:::;L. ThebacktrackingalgorithmforsolvingthesubproblemisstatedformallyinAlgorithm 1 .WebeginbyenumeratingeachpossibleelementofPj,8j=1;:::;n.WedenotethesetofprocessedbixelsbyF(forwhichacandidate\active"bixeldecompositionhasbeenestablished),andthesetofunprocessedbixelsbyR.Ineachiteration,wechecktoseeifthesetofcandidatebixeldecompositionsPjforanyj2Risempty.Ifso,thecurrentactivebixeldecompositionsdonotyieldafeasiblesolution,andthealgorithmbacktracks.Otherwise,weconsideranunprocessedbixel^|2R,andchooseanuntriedbixeldecompositionp^|2P^|tobeactiveforbixel^|.Next,wemove^|fromRtoF,creatingupdatedsetsR0andF0,andinvokeLemma 2 toupdatethesetofbixeldecompositionsforthebixelsinR0.Specically,foreachj2R0andpj2Pj,wecalculatethenumberofleafpositionsthatwouldberequiredduetoselectingpjastheactivebixeldecompositionforbixelj,inadditiontothosealreadyselectedforbixelsinF0.Weeliminatepjifaconditionoftype( 3{19 )isviolated.Wethenrecursivelycalltheproceduretocontinuewithanewbixelj02R0. Westopeitherwhenwendafeasiblebixeldecompositionforallbixels,orwhenweexhaustallbixeldecompositionswithoutndingafeasiblesolution.Intheformercase,aleafdecompositionthatrealizesthebixeldecompositionsforbixelsj2f1;:::;ngcanbe 63

PAGE 64

3 ,whichisbasedontherepeatedapplicationofLemma 1 .ToseethatAlgorithm 3 recoversafeasibleleafdecomposition,notethatAlgorithms 1 and 2 providebixeldecompositionsthatsatisfyLemma 2 ,andinparticular,thecondition Algorithm 3 recoversafeasibleleafdecompositionif,intheouterwhile-loopcorrespondingtoeach`=1;:::;L,thecounterrisneverincrementedmorethanx`times.Notethatrisincrementedeachtimetheinnerwhile-loopterminates,whichoccurseitherwhen~|>n(atotalofpn`times),orwhenp~|`=0(p~|1`p~|`times)for~|=2;:::;n.Thetotalnumberoftimesthatrisincrementedintheouterwhile-loopfor`=1;:::;Listhustheleft-hand-sideof( 3{20 ),whichisnomorethanx`,asrequired. Ifweexhaustallbixeldecompositionswithoutndingafeasiblesolution,weconcludethatthecurrentallowableintensitymultisetisincompatiblewiththecurrentrow. fThisalgorithmndswhetherthereexistsaC1-Partitionofbcompatiblewithxg F;fFisthesetofprocessedbixelsg Rf1;:::;ngfRisthesetofunprocessedbixelsg SinceAlgorithm 1 isabacktrackingalgorithm,andthereforeintheworstcaseinvestigatesallpossiblebixeldecompositions,itisofexponentialtimecomplexity(asexpected,duetoProposition 3 ).However,theempiricalrunningtimeofthealgorithmcanbereducedusingthefollowingobservations: (i) Iftwoadjacentbixelsinarowhavethesamerequiredintensityvalue,theremustexistanoptimalsolutioninwhichtheyareexposedbythesameleafpositions.ThisresultcanbeproveninasimilarwayasLemma 1 ,andisthereforeomittedforbrevity.Thisobservationimpliesthatwecanpreprocessthedatabymerging 64

PAGE 65

returntruefallbixelshavebeenprocessed,Prepresentsafeasiblesolutiong if9j2R:Pj=;then returnfalsefthereisnoremainingwayofdecomposingbixeljg F0F[f^|g,R0Rnf^|g 2 ,giventheactivedecompositionsp~|for~|2F0g returntruefafeasiblesolutionthatusesp^|todecomposebixel^|isfoundg v`r0 forall`2f1;:::;Lgdo ifpj`>0then

PAGE 66

(ii) Inchoosingthenextbixeltobeprocessed,wepickabixelj2Rhavingthesmallestnumberofremainingcandidatebixeldecompositions.Inthismanner,wecanquicklyenumerateallpossiblebixeldecompositionsforafewkeybixelsandeliminateasignicantportionofbixeldecompositionsfortheremainingbixelswithoutwastingeortbyunnecessarybacktrackingsteps. (iii) Inchoosingthenextcandidatebixeldecompositionpj2Pjforachosenbixelj2R,weselectanuntriedbixeldecompositionhavingthefewestnumberofintensityvalues.Sinceeachintensityvalueusedindecomposingabixelneedstobeassignedtoadierentaperture,thisrulefavorsabixeldecompositionusingthefewestnumberofaperturestodecomposethechosenbixel.Therefore,ittendstoretaintheavailabilityofmoreelementsoftheallowableintensitymultiset(andhenceapertures)fortheremainingbixels,makingiteasiertondafeasiblesolution(ifoneexists). 3{18 )haveyetbeenaddedtothemodelwillsetallvariablesequaltozero,whichisclearlyincompatiblewithallrows.Inthissection,wederivesomecharacteristicsofallfeasiblesolutionsandusethesetodenevalidinequalitiesfor(MP).Inthisway,weattempttoimprovetheconvergencerateofthedecompositionalgorithmbyeliminatingsomeclearlyinfeasiblesolutionsbeforetheinitialexecutionofthemasterproblem. 66

PAGE 67

WeformulateanintegerprogrammingmodeltodetermineTi(w01;w02)foragivenrowi.First,denotethesetofpossibleleafpositionsforthatrowbyK,anddenen-dimensionalbinaryvectorsvkfork2K(wherejKj=O(n2)),suchthatvkj=1ifandonlyifbixeljisexposedbyleafpositionk.Inadditiontodecisionvariablesx`asin(MP),denebinarydecisionvariableszk`,8k2K,`=1;:::;Lksuchthatzk`=1ifandonlyifleafpositionkisusedwithintensity`(whereLk=minj:vkj=1bjisanupperboundontheintensityofleafpositionk.)ThenTi(w01;w02)istheoptimalobjectivefunctionvalueofthefollowingoptimizationproblem,(SR): minimizew01LX`=1x`+w02LX`=1`x`(3{22)subjectto Constraints( 3{23 )ensurethateachbixelreceivesexactlyitsrequiredamountofdosewhileconstraints( 3{24 )guaranteethateachleafpositioniseithernotusedorisassignedtoasingleintensityvalue.Finally,constraints( 3{25 )relatethex-andz-variables. Apracticaldicultyinimplementingthevalidinequalitiesoftheform( 3{21 )isthatwemustdetermineappropriatevaluesfortheweightsw01andw02.However, Baatar ( 2005 )showsthat,whendecomposingasinglebixelrow,thereexistsasetofleafpositions 67

PAGE 68

3{21 )by Itiseasytoseethatwecancaptureallofthesevalidinequalitiesbyrestrictingourselvestothecoecientpairs(w01;w02)=(1;0)and(0;1)only: Wecangeneralizethisideaasfollows.LetR(L)denotethesetofrowsforwhichthemaximumintensityrequirementisboundedbyLforsomeL2f1;:::;Lg,i.e.,R(L)=fi2f1;:::;mg:LiLg.SinceintensityvaluesgreaterthanLcannotbeusedindecomposingtherowsinR(L),asimilarapproachtotheoneabovecanbeusedtoderivethefollowingfamilyofvalidinequalities Finally,notethatthevaluesofNiand~zicanbefoundbysolving(SR)withw01=1;w02=1orbyusingthemethodof Kalinowski ( 2004 ),sincethereexistsasolutionthatminimizesbothbeam-on-timeandthenumberofapertures( Baatar 2005 ). 3{16 ){( 3{18 )representnecessaryconditionsforfeasibilityofanallowableintensitymultisetwithrespecttoaparticularrow.Itispossibletodevelopstronger 68

PAGE 69

3 showsthat,ifagivenallowableintensitymultisetisincompatiblewithasubsequencesofrowi,thenitalsomustbeincompatiblewithrowi. Proof. NotethatwecaninvokeLemma 3 toassociateasubproblemwitheachoftheO(2n)subsequencesofabixelrowb.Eachofthesesubproblemscanthenbeusedtogeneratecuttingplanesoftheform( 3{18 ),aswellasvalidinequalitiesoftheform( 3{31 )and( 3{32 ).However,sincethestrengthof( 3{18 ),( 3{31 )and( 3{32 )dependonthelargestintensityvalueinabixelrow,weformsubsequencesofeachbixelrowby,forL=1;:::;L,consideringonlythosebixelshavingrequiredintensitylessthanorequaltoL.ThevalidinequalitiesgeneratedbytheO(min(n;L))subsequencesgeneratedinthisfashionimplyallO(2n)validinequalitiesassociatedwithallpossiblesubsequences. 3 tondaleafdecompositionforeachrow.Wecanthengenerateaperturematricesbyarbitrarilycombiningleafpositionsusingthesameintensityvaluesindierent 69

PAGE 70

Sinceanypairofleafpositionsassignedtothesameintensityvalueindierentrowscanbecombined,thereareuptoQL`=1(x`!)maperturematricesthatcanbeconstructedfromagivenfeasibleleafdecompositionforeachrow.Eventhougheachsuchchoicerepresentsanalternativeoptimalsolutiontotheoptimalleafsequencingproblem,somematrixdecompositionsmayclinicallybepreferabletoothersbasedontheirstructuralproperties.Perhapsthemostchallengingstructuralconsiderationpertainstotheso-called\tongue-and-groove"eectobservedinMLCs.Wereferthereadertotheworksof Dengetal. ( 2001 )and Queetal. ( 2004 )fortechnicaldetailsofthetongue-and-grooveeectindynamicMLCdosedelivery.Forthepurposesofthisstudy,itissucienttounderstandthatleavesinadjacentrowsofteninterlockwithatongueonthebottomofonerowslidingalongagrooveinthetopofanotherrow.Tongue-and-grooveunderdosageoccurssincealeaf'stongueblocksdosageintendedforcellsbeneathit.Therefore,itisdesirabletolimitsuchunderdosages. Tomeasuretheamountoftongue-and-grooveeectinatreatmentplan, Queetal. ( 2004 )notethatitisgenerallynotdesirabletodeliveroneapertureinwhichsomebixel(i;j)isblockedbyaleafwhilebixel(i+1;j)isnotblocked,ifanotherapertureisbeingdeliveredwhere(i;j)isnotblockedbyaleafwhile(i+1;j)isblocked.Basedonthisobservation, Queetal. ( 2004 )derivethefollowingtongue-and-grooveindex(TGI).SupposeatreatmentplanconsistsofKaperturesdescribedbybinaryvaluesvikj,wherevikj=0ifcell(i;j)isblockedbyaleafinaperturekandvikj=1otherwise,foreachi=1;:::;m,j=1;:::;n,k=1;:::;K.LetIkbetheintensitydeliveredinaperturek=1;:::;K.ThentheTGIofamatrixdecompositionisdenedas:m1Xi=1nXj=1K1Xk=1KX`=k+1minfIk;I`ghvikj(1vi+1;kj)(1vi`j)vi+1;`j+(1vikj)vi+1;kjvi`j(1vi+1;`j)i: 70

PAGE 71

Givenapairofadjacentrows,weattempttomatchindividualleafpositionsinthetworowstominimizetheTGIinducedbytheadjacentrowpair.Tolimitcomputationaloverheadinthisphaseofouralgorithm,wereduceTGIindirectlybythefollowingscheme.Letusdenotealeafpositionforrowibyabinaryn-vectorvi,wherevij=1iftheleafpositionexposesbixeljinrowi.Wemeasuretheoverlapbetweentwoleafpositionshavingthesameintensityvalueinconsecutiverowsbycountingthenumberofcolumnsthatbothleafpositionsexposesimultaneously.Formally,wedenetheoverlapbetweenleafpositionsviandvi+1as(vi;vi+1)=Pnj=1vijvi+1j.OurapproachistoheuristicallyminimizeTGIbymaximizingthetotaloverlapbetweenallleafpositionpairs,whichcanecientlybesolvedasanassignmentproblem.Theeciencyoftheassignmentproblemscanbefurtherimprovedbynotingthattheproblemdecomposesovertheintensityvalues`2f1;:::;Lg,sinceonlyleafpositionshavingthesameintensityvaluecanbecombined.Therefore,wecangenerateamatrixdecompositionbyndingaleafdecompositionforeachrow,andthenmatchingleafpositionsinadjacentrowshavingthesameintensityvaluebysolvinganassignmentproblemsothatthetotaloverlapismaximized. TheTGIminimizationstepdescribedinthepreviousparagraphcanbeimprovedasfollows.Typically,multiplebixeldecompositionsexistforeachrowthatarecompatiblewithagivenfeasibleintensitymultiset.Algorithm 2 canbemodiedinastraightforwardmannersothatitndsallleafdecompositionsofarow,insteadofstoppingoncetherstfeasiblebixeldecompositionforallbixelsisfound.Sincedierentbixeldecompositionsforabixelrowcorrespondtodierentleafdecompositions,consideringalternativebixeldecompositionscanleadtoamatrixdecompositionhavingasmallerTGI. 71

PAGE 72

2 byterminatingonce250feasiblebixeldecompositionshavebeenidentied.Next,notethatastraightforwardacyclicshortestpathimplementationprocesseslayersoneatatime,anddoesnotgenerateafeasibleS{Fpathbeforeprocessingthelastlayer.Sincebeingabletospecifyatimelimitisadesiredfeatureinapracticalsetting,weuseahybridalgorithmforsolvingtheshortestpathproblem.Ouralgorithmstartsbyprocessinglayersone-by-one,updatingnodelabelsasusual.Ifashortestpathisnotfoundwhenagiveninitialtimelimitexpires,ouralgorithmswitchestoadepth-rst-search(DFS)procedure,whichweterminateafteragivennaltimelimit.WestartDFSfromanunprocessednodeNidhavingasmallestlabel,selectaminimum-costarc(Nid;N(i+1)d0)exitingthatnode,andupdatethelabelofN(i+1)d0ifwehavefoundanewshortestS{N(i+1)d0path.Else,the

PAGE 73

3.3.1ProblemInstances 3-1 reportstheproblemcharacteristicsfortheseprobleminstancesintermsofthematrixdimensionsmandn.ThemaximumintensityvalueisL=20foralltheseinstances.Inaddition,toallowcomparisonofourresultswithpublishedresultsonotherapproachestotheproblem,wegenerated100randomprobleminstancesofdimensions2020havingmaximumintensityvalueL=10. However,sincetheseprobleminstancesaregenerallytoolargetobesolvablebytheintegerprogrammingmodelfrom Langeretal. ( 2001 )anditsmodicationdescribedinAppendixA,wealsorandomlygeneratedeightinstances(\test5x5a",:::,\test6x7b")todemonstratethecomputationallimitationsofthelatterapproaches.Unlessotherwisespecied,weusedw1=7andw2=1astheobjectiveweightsforthenumberofaperturesandbeam-on-time,respectively. Table3-1. Dimensionsofclinicalprobleminstances c2b11820c3b12217c4b11922c5b11516c1b21115 c2b21719c3b21519c4b21324c5b21317c1b31515 c2b31818c3b32017c4b31823c5b31416c1b41515 c2b41818c3b41917c4b41723c5b41416c1b51115 c2b51718c3b51519c4b51224c5b51217 73

PAGE 74

3{31 )and( 3{32 )describedinSection 3.2.4.2 .WealsouseEngel'sheuristic( Engel 2005 ),whichexecutesinwellunderoneCPUsecondforeachinstanceandgeneratesasolutionhavingminimumbeam-on-time,to(i)obtainaninitialupperboundand(ii)computetheupperboundsR`(`=1;:::;L). Langeretal. ( 2001 )andtothemodicationoftheirmodelasdescribedinAppendixA.Wechooserandomlygeneratedtestinstancesofvariousdimensionstoidentifytheproblemsizesthatcanbesolvedbyeachalgorithm,aswellasfourofthesmallestclinicalinstancestocomparetheeectivenessofthealgorithmsonclinicalinstances.Weimposedaone-hourtimelimitpastwhichwehaltedtheexecutionofanalgorithm.FortheseexperimentswedisabledtheuseofEngel'sheuristicasaninitialheuristictotesttheabilityofthesemodelstoecientlyndgood-qualityupperbounds. Table 3-2 summarizestheresultsofthesethreealgorithmsintermsoftheexecutiontime,thebestupperandlowerboundsfoundwithinthetimelimit,andtheoptimalitygap(calculatedasthedierencebetweentheupperandlowerboundasapercentageoftheupperbound).Ourdecompositionalgorithmcansolveall15instancesinthisdatasetwithinafewseconds,whereasonlysixinstancescanbesolvedtooptimalitywithinanhourbyeitherintegerprogrammingformulation.Weconcludethat,eventhoughtheintegerprogrammingformulationgivenin( Langeretal. 2001 )cansolvesmallinstancestooptimality,itcannotbeusedtosolveclinicalprobleminstancestooptimalitywithinpracticalcomputationtimelimits. 74

PAGE 75

Comparisonofourbasealgorithmwith Langeretal. ( 2001 )model Langer ModiedLangerNamemnL CPUUBLBGap CPUUBLBGap test3x3338 0.129 1.62929.000.0% 0.92929.000.0%test3x4348 0.137 5.23737.000.0% 1.63737.000.0%test4x4448 0.136 30.43636.000.0% 10.73636.000.0%test5x5a5510 0.245 2069.64545.000.0% 86.44545.000.0%test5x5b5515 0.250 198.25050.000.0% 92.65050.000.0%test5x6a5610 0.255 36006133.5345.0% 36005540.9525.5%test5x6b5618 0.471 36008451.5838.6% 36007758.6723.8%test6x6a6613 0.355 36005545.6317.0% 36005548.0012.7%test6x6b6613 0.352 36005743.8223.1% 36005750.0012.3%test6x7a7610 0.245 690.04545.000.0% 435.14545.000.0%test6x7b6715 0.474 36009435.6962.0% 36008047.8840.1%c1b1151420 1.3111 360033648.5885.5% 360027342.0084.6%c1b2111520 0.8104 360028038.2686.3% 360013239.5570.0%c1b5111520 3.1104 360028046.2083.5% 360014049.2964.8%c5b4141620 2.5124 360036034.0090.6% 360036039.1189.1% Siochi ( 2007 ), Engel ( 2005 ),and XiaandVerhey ( 1998 ),whichweexecutedonthesamedataset.(Theresultswepresentfrom Siochi ( 2007 )refertotheVariableDepthRecursion(VDR)algorithmwithouttongue-and-grooveconstraints,usingtheparametersrecommendedinthepaper.Wediscusstheeectofincludingtongue-and-grooveconsiderationsinthealgorithmbelow.) Figure 3-2 summarizesthetotaltreatmenttimesassociatedwiththesolutionsgeneratedbythesixalgorithmswetested.Eachalgorithmisrepresentedbyacurvethatdepictsqualityofthesolutionsobtainedbythecorrespondingalgorithm.ForeachvalueToftotaltreatmenttimeonthehorizontalaxis,eachcurveplotsthenumberofprobleminstancesforwhichthecorrespondingalgorithmwasabletondasolutionhavingtotaltreatmenttimenomorethanT.Forinstance,Figure 3-2 showsthatSiochi'sheuristicfoundasolutionwithatotaltreatmenttimeofatmost175timeunitsin5%oftheprobleminstances,whileanoptimalsolution(representedby\TotalTime")has 75

PAGE 76

Comparisonoftotaltreatmenttimesonrandomdata thesamequalitylevelin97%oftheprobleminstances.Weobservethatallthreeexactalgorithmsndsolutionshavingsimilartreatmenttimes.SolutionqualitiesgeneratedbytheEngelandSiochiheuristicsaresimilar,withtheSiochiheuristicbeingslightlybetter.AcomparisonoftheheuristicsolutionswithoptimalsolutionsrevealsthataverageoptimalitygapsforSiochi,EngelandXia-Verheyheuristicsare10.1%,12.0%,and51.5%,respectively. Figure 3-3 comparesthealgorithmswithrespecttothenumberofaperturesusedintheirrespectivesolutions.Wenotethatouralgorithmthatminimizestotaltreatmenttime(\TotalTime")ndsasolutionthatalsominimizesthenumberofaperturesformostprobleminstances.Asexpected,lexicographicminimizationofthetwoobjectivefunctions 76

PAGE 77

Comparisonofthenumberofaperturesonrandomdata resultsinanincreasednumberofapertures.Forthisobjectivethe\#Apertures"algorithmndsoptimalsolutions.AverageoptimalitygapsfortheheuristicsofSiochi,Engel,andXia-Verheyare15.6%,18.9%,and62.3%,respectively. Weanalyzethebeam-on-timevaluesofthesolutionsgeneratedbyeachalgorithminFigure 3-4 .SincebothEngel'sheuristicandour\Lexicographic"algorithmndoptimalsolutionshavingminimumbeam-on-time,theircurvesoverlap.WeobservethattheSiochiheuristicandour\TotalTime"algorithmtendtogeneratesolutionshavingsmallbeam-on-timevalues,butthesolutionsgeneratedbyour\#Apertures"algorithm,andbytheXia-Verheyheuristichavehigherbeam-on-timevalues.Wecalculatedtheaverageoptimalitygapsforthelattertwoalgorithmsas12.6%and32.1%,respectively. 77

PAGE 78

Comparisonofbeam-on-timevaluesonrandomdata ThenalmeasureofsolutionqualitythatweconsiderisTGI,whichisameasureofthetongue-and-grooveeectgivenby( 3{33 ).Figure 3-5 revealsthatthesolutionsobtainedbyallthreevariantsofourdecompositionalgorithmhavesignicantlylowerTGIvaluesthantheheuristicprocedures.Thisresultimpliesthat,eventhoughourTGI-reductionalgorithmdescribedinSection 3.2.5 doesnotguaranteeaminimumTGI,itishighlyeectiveinndingsolutionswithTGIvaluessuperiortotheotherheuristicapproaches.Toestimateoptimalitygapsfortheheuristicswecompareheuristicsolutionswiththesolutionsgeneratedbyour\Lexicographic"algorithm,whichprovidesthebestTGIamongallmethodsmentionedabove.WenotethataveragegapsforSiochi,EngelandXia-Verheyheuristicsare162.1%,164.4%,and205.4%,respectively.Wealsonote 78

PAGE 79

ComparisonofTGIvaluesonrandomdata thattheseheuristicsdonotattempttominimizeTGI,anditmightbepossibletomodifythemtoobtainsolutionswithlowerTGIvalues.ItisinterestingtonotethatavariantofSiochi'salgorithm( Siochi 2007 )iscapableofcompletelyeliminatingTGIattheexpenseofcreatingadditionalapertures.Thisvariantisreportedtoincreasethenumberofaperturesby10%to30%relativetothevariantthatdoesnotremoveTGI( Siochi 2007 ). Finally,theEngelandXia-VerheyheuristicstooklessthanonesecondofCPUtimeinallinstanceswetested.TheaverageCPUtimeforSiochi'sheuristic,\TotalTime"algorithm,\#Apertures"algorithm,and\Lexicographic"algorithmwere31.5,963.1,414.8,and421.4seconds,respectively.Wenotethatallvariantsofourtwo-stagealgorithmshoweda\heavy-tail"behavior,whereabout80%oftheprobleminstancesweresolvedto 79

PAGE 80

3.1 thatinclinicalpractice,wecandelivereachuencemapusingadecompositionintoeitherrow-convexorcolumn-convexapertures,wherethelatterrequiresrotationoftheMLChead.Ournalsetofexperimentscomparesthealgorithmsonclinicalprobleminstancesinourdataset,allowingforMLCheadrotation. Werstshowtheresultsofapplyingourdecompositionalgorithmtodecomposeeachofthe25clinicaluencemapsintorow-convexapertures,andcolumn-convexapertures,wherethelatterisachievedbyapplyingouralgorithmtothetransposeofeachuencemap.Table 3-3 reportstheperformanceofouralgorithmwhentheobjectivefunctionissettominimizetotaltreatmenttime,anddisplaysthenumberofapertures(\nAper"),beam-on-time(\BOT"),totaltreatmenttime(\Time"),tongue-and-grooveindex(\TGI"),andCPUtimeused(\CPU")forthealgorithm. Ouralgorithmndsanoptimalsolutiontoseveralinstanceswithinafewsecondswhilefourinstancestakemorethan10minutesofCPUtimetobesolvedtooptimality.Comparingthesolutionsobtainedforrow-convexandcolumn-convexdecompositions,weobservethatrotatingtheMLCheadismostbenecial(intermsoftreatmenttime)forinstancesinwhichthenumberofrowsismuchsmallerthanthenumberofcolumns.Thesebenetsaremostapparentoninstancesc4b2andc4b5,whererotatingtheMLCheadcanresultinmorethan50%reductionintotaltreatmenttime.Wealsonotethatseveralprobleminstancesrequiremuchlesscomputationaltimetosolveforacolumn-convexdecompositioncomparedtoarow-convexdecomposition. 80

PAGE 81

EectofrotatingtheMLChead Column-ConvexName nAperBOTTimeTGICPU nAperBOTTimeTGICPU c1b1 10411111021.1 1138115505.5c1b2 1034104800.8 82379140.7c1b3 11311089711.4 92891591.0c1b4 11331107437.0 11371141467.0c1b5 10341041334.3 83288491.2c2b1 143413213426.5 123011418711.5c2b2 134113215920.1 11331101928.0c2b3 134914024514.7 11281051513.1c2b4 145114931687.3 12341181488.3c2b5 1341132217395.6 1027971202.0c3b1 1341132323310.0 144013825423.0c3b2 14461443204759.8 82379861.1c3b3 134914053310373.9 124012436018.6c3b4 1244128481524.9 1240124327428.2c3b5 13341251333.3 92790752.6c4b1 164015221634.9 124613024410.6c4b2 166918145020901.0 9279014915.8c4b3 144113913044.7 10321021293.3c4b4 1444142246164.3 1027971638.0c4b5 177619547014511.4 92487484.0c5b1 102696680.5 1035105410.5c5b2 12411255914.3 82581270.6c5b3 10341041553.1 92386421.0c5b4 12401241052.2 1032102874.3c5b5 124613015151.9 83187170.8 Computationalresultsforourbasealgorithm #Apertures LexicographicName nAperBOTTGICPU nAperBOTTGICPU nAperBOTTGICPU c1b1 10411024.7 10411022.3 1138504.8c1b2 823141.1 823141.1 823141.1c1b3 928593.0 928594.5 928594.5c1b4 11337441.2 113712827.1 11337412.2c1b5 832492.1 834561.3 92691.9c2b1 123018715.6 123018714.9 123018714.4c2b2 113319210.8 11381616.9 11331467.8c2b3 11281498.9 11281139.9 112819710.8c2b4 123414816.8 123414816.8 123414817.1c2b5 10271206.1 10311556.2 10271206.2c3b1 1341323315.0 125152162.1 134132531.4c3b2 823864.4 826874.5 823625.6c3b3 124036027.4 1240360894.7 124036520.1c3b4 1240327442.2 1246284548.8 133892855.1c3b5 927755.6 927755.4 927755.7c4b1 124624416.8 124622710.6 124622711.3c4b2 92714945.5 93215056.2 92713535.0c4b3 103212915.7 103410814.9 103212915.6c4b4 102716332.0 102811232.6 11267229.9c4b5 9244827.8 9244827.7 9244827.0c5b1 1026681.2 1026681.2 1026681.2c5b2 825271.1 825271.0 92381.1c5b3 923423.6 924453.2 923833.1c5b4 1032875.8 10411012.7 1032872.8c5b5 831171.4 833161.2 831711.1

PAGE 82

3-4 showsthenAper,BOT,TGI,andCPUmetricsobtainedfromouralgorithmenhancedwiththeaboveboundingscheme,correspondingtothe\TotalTime,"\#Apertures,"and\Lexicographic"objectives.Observethatall25instances,underanymetric,terminateinunder15minutesofCPUtimewithasolutionthatisoptimalwithrespecttothecorrespondingobjective,andallinstancesaresolvedtooptimalitywithinaminuteusingthe\Lexicographic"algorithm. Recallthatthe\BOT"columnin\Lexicographic"reportstheminimumachievablebeam-on-time,andthe\nAper"columnundertheobjective\#Apertures"reportstheminimumnumberofaperturesneededtodecomposeeachinstance.Perhapssurprisingly,incomparingthesevalueswiththeresultsof\TotalTime,"weobservethatthereexistsasolutionthatminimizesboththenumberofshapesandthebeam-on-timesimultaneouslyin19ofthe25instances. Finally,weanalyzeperformanceofthethreeheuristicsonclinicaldata,whereweexecuteeachheuristiconeachprobleminstanceanditstranspose(correspondingtorow-convexandcolumn-convexdecompositions),andpickthesolutionyieldingthesmallesttreatmenttime.Table 3-5 showsthenumberofapertures,beam-on-time,TGI 82

PAGE 83

Comparisonofheuristicalgorithmsonclinicaldata Engel Xia-VerheyName nAperBOTTGICPU nAperBOTTGICPU nAperBOTTGICPU c1b1 113824514.0 1238261<1 1340219<1c1b2 8231093.0 823127<1 1032133<1c1b3 9282134.0 1028192<1 1234198<1c1b4 12343069.5 1137398<1 1442355<1c1b5 9261033.6 926175<1 1235124<1c2b1 123065211.2 1230738<1 1545635<1c2b2 123339517.8 1233464<1 1545460<1c2b3 122862534.8 1228429<1 1543459<1c2b4 123462843.2 1234723<1 1856417<1c2b5 112746315.3 1127465<1 1441375<1c3b1 144382836.3 15401054<1 1755765<1c3b2 92314311.1 923127<1 1236289<1c3b3 1440131640.8 1440869<1 19601038<1c3b4 134867833.3 1438765<1 1755553<1c3b5 9282637.0 927325<1 1345261<1c4b1 134661729.4 1446625<1 1862531<1c4b2 102929573.8 1027466<1 1444350<1c4b3 113233919.4 1132365<1 1448428<1c4b4 112648913.5 1126540<1 1546424<1c4b5 92423689.8 924328<1 1544328<1c5b1 11261884.6 1226176<1 1238185<1c5b2 9231296.9 923100<1 1033145<1c5b3 9262015.1 1023293<1 1232189<1c5b4 113221811.2 1132322<1 1346243<1c5b5 8322177.2 931211<1 1135138<1 3-4 revealsthateventhoughtheheuristicsconsistentlygeneratedhigh-qualitysolutions,theSiochiandEngelheuristicswereabletondanoptimalsolutioninonlyveprobleminstances,andXia-Verheyheuristiccouldnotndanoptimalsolutiontoanyinstance. 83

PAGE 84

External-beamradiationtherapyisdeliveredfrommultipleanglesbyadevicethatcanrotatearoundapatient.Theuseofmultiple(typically3{9)anglesisoneofthetoolsthatallowforthetreatmentofdeep-seatedtumorswhilelimitingtheradiationdosetosurroundingfunctioningorgans.Conventionalconformalradiationtherapythenfurtherusesblocksandwedgestoshapethebeams(see,e.g., Lim ( 2002 )and Limetal. ( 2004 2007 )).IMRTisamorepowerfultherapythatinsteadmodulatesbeamintensity.Themostcommontechniqueforachievingthismodulationistodynamicallyshapebeamswiththehelpofamultileafcollimator(MLC)system.Suchsystemscandynamicallyformmanycomplexaperturesbyindependentlymovingleafpairsthatblockpartoftheradiationbeam.Unfortunately,MLCsystemsareverycostlyandtechnologicallyadvanced,andarethereforedicultandexpensivetooperateandmaintain.Moreover,MLCsystemsarecurrentlyonlyavailableforusewithaso-calledlinearacceleratorthatgenerateshigh-energyphotonbeamsfortreatment.However,theuseofradioactive60Co(Cobalt)sourcesforradiationtherapyisstillubiquitousinmanypartsoftheworldandis 84

PAGE 85

Earletal. ( 2007 )).Recently,researchershavebeguntoexploretheclinicalfeasibilityofdeliveringIMRTusingconventionaljawsthatarealreadyintegratedintoradiationdeliverydevicesandcancreateaperturesthatarerectangularinshape(see,e.g., Earletal. ( 2007 ), Kimetal. ( 2007 ),and Menetal. ( 2007 )).Successfulapplicationofthismuchsimplerdeliverytechniquedependscriticallyontheabilitytoecientlydeliverhigh-qualitytreatmentplans.Wethereforedevelopandtestnewoptimizationapproachestominimizethetreatmenttimerequiredforaparticulartreatmentplanusingrectangularaperturesonly. Solvingaso-calleduencemapoptimizationproblemyieldsanoptimalIMRTtreatmentplanthatresolvesdierent,andconicting,clinicalmeasuresoftreatmentplanqualityrelatedtotumorcontrolandsideeects(see,e.g., Shepardetal. ( 1999 )forareview; Leeetal. ( 2000a 2003 )formixed-integerprogrammingapproaches; Romeijnetal. ( 2006 )forconvexprogrammingmodels;and HamacherandKufer ( 2002 )and Kuferetal. ( 2003 )foramulticriteriaapproach).Atreatmentplanthenconsistsofacollectionofnonnegativeintensitymatrices,oftenreferredtoasuencemaps,onecorrespondingtoeachbeamangle.Tolimittreatmenttime,eachofthesematricesisthenexpressedasamultipleofanintegraluencemapinwhichthemaximumelementisontheorderof10{20.Toallowdeliveryofthetreatmentplan,eachoftheseuencemapsshouldbe 85

PAGE 86

AhujaandHamacher ( 2005 ), Bolandetal. ( 2004 ), Kamathetal. ( 2003 ), Engel ( 2005 ), Kalinowski ( 2005a ),and Tasknetal. ( 2009b ).(Notethatintegratedapproachestouencemapoptimization,alsoreferredtoasaperturemodulation,havebeenproposedaswell;wereferto,e.g., Preciado-Waltersetal. ( 2004 ), Romeijnetal. ( 2005 ),and Menetal. ( 2007 ).) Theproblemthatwestudyisthedecompositionofanintegraluencemapintorectangularaperturesandcorrespondingintensities.While DaiandHu ( 1999 )proposedastraightforwardheuristicforavariantofthisdecompositionproblem,wedeveloptherstcomputationallyviableoptimizationapproachtothisproblem.InSection 4.2 weconsiderthecoreproblemofdecomposingan(integral)uencemapwhileminimizingthenumberofrectangularapertures.InSection 4.3 wethenextendourmodelstotheproblemsof(i)minimizingtotaltreatmenttime(asmeasuredbythesumoftherequiredaperturesetuptimesandthebeam-on-time,i.e.,theactualtimethatradiationisbeingdelivered);and(ii)minimizingthenumberofaperturessubjecttobeam-on-timebeingminimal.Finally,Section 4.4 discussesourcomputationalresultsonacollectionofclinicaluencemaps. 4.2.1 byformallydescribingtheoptimizationmodelunderinvestigationandmodelingitwithamixed-integerprogrammingformulation.WenextdescribeseveralclassesofvalidinequalitiesinSection 4.2.2 .Finally,wediscussmethodsforpartitioningtheinputmatrixinSection 4.2.3 ,whichleadstoeectivelowerandupperboundingtechniques. 86

PAGE 87

4-1 showsanexampleuencemap,whichweusethroughoutthischapter. Figure4-1. Exampleuencemap LetRbethesetofallO(n2m2)possiblerectangularapertures(i.e.,submatricesofBhavingcontiguousrowsandcolumns)thatcanbeusedtodecomposeB,excludingthosethatcontainazero-bixel.Foreachrectangler2Rwedeneacontinuousvariablexrthatrepresentstheintensityassignedtorectangler,andabinaryvariableyrthatequals1ifrectanglerisusedindecomposingB(i.e.,ifxr>0),andequals0otherwise.LetCrbethesetofbixelsthatisexposedbyrectangler.WedeneMr=min(i;j)2Crfbijgtobetheminimumintensityrequirementamongthebixelscoveredbyrectangler.Furthermore,wedenotethesetofrectanglesthatcoverbixel(i;j)byR(i;j).Giventhesedenitions,wecanformulatetheproblemasfollows:IPR:MinimizeXr2Ryr 87

PAGE 88

4{1 )minimizesthenumberofrectanglesusedinthedecomposition.Constraints( 4{2 )guaranteethateachbixelreceivesexactlytherequireddose.Constraints( 4{3 )enforcetheconditionthatxrcannotbepositiveunlessyr=1.Finally,( 4{4 )statesboundsandlogicalrestrictionsonthevariables.Notethattheobjective( 4{1 )guaranteesthatyr=0whenxr=0inanyoptimalsolutionofIPR. FormulationIPRcontainstwovariablesandaconstraintforeachrectangle,resultinginalarge-scalemixed-integerprogramforprobleminstancesofclinicallyrelevantsizes.Furthermore,theMr-termsinconstraints( 4{3 )leadtoaweaklinearprogrammingrelaxation;withnovalidinequalitiesorbranchingyetperformedontheproblem,wehavethatyr=xr=MratoptimalitytothelinearprogrammingrelaxationofIPR.AnalternativeformulationthatdoesnotrequireMr-termsemploysadecompositionmethod.Recallthatweinvestigatedtheproblemofdecomposinganintegermatrixinto\consecutive-ones"matricesinChapter 3 ,whereineachdecomposedmatrixallnonzerovaluestakethesamevalueandappearconsecutivelyoneachrow.Ourcomputationalresultsshowedthatsolvabilityoftheproblemissignicantlyimprovedbyapplyingabi-leveloptimizationalgorithm.Asimilarapproachfortheproblemweconsiderinthischapterwouldformulateamasterproblemas:MP:MinimizeXr2Ryr (4{6)yrbinary8r2R; whereweaddresstheformof( 4{6 )inthesequel.Givenavector^y,wecancheckwhetherconstraint( 4{6 )issatisedbysolvingthefollowinglinearprogram:SP(^y):Minimize0 (4{8)subjectto:Xr2R(i;j)xr=bij8i=1;:::;m;j=1;:::;n 88

PAGE 89

Associatingvariablesijwith( 4{9 ),andrwith( 4{10 ),weobtainthedualformulation:DSP(^y):MaximizemXi=1nXj=1bijij+Xr2RMr^yrr OurBendersdecompositionstrategyrstsolvesMP,whichyields^y.IfSP(^y)isfeasible,then^ycorrespondstoafeasibledecompositionandisoptimal.Else,DSP(^y)isunbounded(sincethetrivialall-zerosolutionguaranteesitsfeasibility).Let(^,^)beanextremedualrayofDSP(^y)suchthatPmi=1Pnj=1bij^ij+Pr2RMr^yr^r>0.Then,ally-vectorsthatarefeasiblewithrespectto( 4{6 )mustsatisfymXi=1nXj=1bij^ij+Xr2R(Mr^r)yr0: Weadd( 4{16 )inacuttingplanefashionasnecessary.

PAGE 90

4.2.1 (a) (b) Weobservethatthereexistsanoptimalsolutioninwhichnotwoadjacentrectanglesareusedinthedecomposition.Toseethis,assumethatadjacentrectanglesr1andr2haveintensitiesxr1andxr2,respectively,wherexr1xr2withoutlossofgenerality.Inthiscase,analternativeoptimalsolutioncanbeconstructedbyextendingr1intor2.Specically,letr0betherectangleforwhichCr0=Cr1[Cr2.Analternativeoptimalsolutionthatdoesnotcontainanyadjacentrectanglesusesr2havingintensityxr2xr1,andr0havingintensityxr1.Thisdominancecriterioncanbewrittenas:yr1+yr218adjacentrectanglesr1;r2; whichstatesthatnopairofadjacentrectanglescanbeselectedinanoptimalsolution. 90

PAGE 91

4-2 .Sinceb43=2,thetotalintensitydeliveredto(5;3)byallrectanglesthatstartinrowsi=1;:::;4cannotexceed2.However,b53=14>2,andhenceatleastonerectanglethatstartsinrow5isrequiredtocoverbixel(5;3).Similarly,b53>b52impliesthatatleastonerectanglethatstartsincolumn3isrequiredtocoverthesamebixel.Theseresultscanbestrengthenedbyconsideringboth(4;3)and(5;2)simultaneously.Sinceb53>b43+b52,weconcludethatatleastonerectanglethatstartsatbixel(5;3)isrequiredinanyfeasibledecompositionoftheuencemap.Ingeneral,arectanglemuststartat(i;j)ifbij>b(i1)j+bi(j1)issatised.Figure 4-3 illustratesasimilaridea,wherewe Figure4-2. Examplestartindex comparetheintensityrequirementofbixel(6;4)withthebixelbelowit,andtheoneonitsright.Usingargumentssimilartotheonesregardingstartingindices,weconcludethatarectanglemustend(i.e.,havealower-rightcorner)at(6;4)sinceb64>b74+b65. Figure4-3. Exampleendindex Startingandendingindexconditionscanbegeneralizedfurtherasfollows.Assumethatthereexistintegersu2[0;i1],d2[i+1;m+1],l2[0;j1],andr2[j+1;n+1] 91

PAGE 92

4-4 illustratesaboundingboxforbixel(6;3)(markedindarkgray),whichcorrespondsto(l;u;r;d)=(2;4;5;7).Thefourbixelsthatrepresentthebordersofaboundingboxaremarkedinlightgray.Wenotethatanyrectanglethat Figure4-4. Exampleboundingbox containsbixel(6;3),anddoesnotstartinsidetheboundingbox(at(5,3)or(6,3))orendinsidetheboundingbox(at(6,3)or(6,4)),hastocontainatleastoneofthefourbixelsontheborder.Therefore,thesumofintensitiesofthoserectanglesisboundedbythetotalrequiredintensityofthebixelsinlightgray.Sincetheintensityofthedarkgraybixelcannotbesatisedbythoserectanglesalone,itfollowsthatatleastonerectanglecontainedwithintheboundingboxmustbeusedtocoverbixel(6;3).LetBBijrepresenttheinteriorofaboundingboxforbixel(i;j),i.e.,given(l;u;r;d)allbixelsattheintersectionofrowsu+1;:::;d1andcolumnsl+1;:::;r1.WedenotethesetofrectanglesinR(i;j)thatarecontainedwithinBBijbyR(BBij).Inthiscase,thefollowinginequalityisvalid:Xr2R(BBij)yr1: Notethat(0;0;n+1;m+1),whichcorrespondstotheinputmatrix,isaboundingboxforanybixel.Thereforetherecanbemultipleboundingboxesassociatedwitheachbixel.LetBBijandBB0ijbetwoboundingboxesforbixel(i;j).WesaythatBBijdominatesBB0ijifR(BBij)R(BB0ij).Sincetheinequality( 4{18 )thatcorrespondstoadominated 92

PAGE 93

4-5 displaysanothernondominatedboundingboxforthebixelconsideredinFigure 4-4 Figure4-5. Anothernondominatedboundingboxseededat(6,3) Togeneratenondominatedboundingboxes,werstmakethefollowingobservation.Anondominatedboundingboxforbixel(i;j)isminimalinthesensethatnoneofitsedgescanbeshiftedcloserto(i;j)withoutviolatingtheboundingboxintensityproperty.Weusethisobservationtodesignanalgorithmthatndsseveralnondominatedboundingboxesassociatedwithagivenbixel.Inouralgorithm,westartatabixel(i;j),andrstmoveinaverticalorhorizontaldirectionuntilweencounterabixel(i0;j0)havingbi0j0
PAGE 94

Wenotethatinequalities( 4{19 )areimpliedby( 4{2 )and( 4{3 )inIPR.However,( 4{19 )canbeusedtotightenthemasterproblemoftheBendersdecompositionapproachdiscussedinSection 4.2.1 .Furthermore,varioustighteningprocedurescanbeappliedto( 4{19 )foruseineitherthedirectsolutionofIPRorintheBendersmasterproblem.Inourimplementation,weapplyaChvatal-Gomoryroundingprocedure(see,e.g., NemhauserandWolsey ( 1988 ))inwhichwedividebothsidesoftheinequalitybythesmallestMrcoecientontheleft-hand-side(unlessbijisdivisiblebythatnumber),androundupcoecientsonbothsidesoftheinequality.IfbijisdivisiblebythesmallestMr-coecientontheleft-hand-sideof( 4{19 ),thentheroundingprocedureyieldsaninequalityimpliedby( 4{19 ),andhencewedonotgenerateit. 4.2.2.3 dividesbothsidesof( 4{19 )bybij1,providedthatbij2,andthenroundsupallcoecientsandtheright-hand-side.Notingthatallcoecientsontheleft-hand-sideareboundedfromabovebybij,thisprocessyields:Xr2R(i;j):Mr
PAGE 95

4{21 ),andthusallvcorrespondingtoafeasiblesolutionmustalsosatisfy( 4{21 ). Similarly,considersubmatricesoftheformqLqR; 4 revealsthatthefollowinginequalityisvalid:2Xr2A=Lyr+2Xr2A=Ryr+2Xr2A=2yr+Xr2A
PAGE 96

4.2.2.4 forothersubmatricesaswell.However,thisprocessisverytedious,andthereisalargenumberofpossiblesubmatrixcombinations.Inthissectionwedescribeasimilarsetofinequalities,whichareweakerthanthosedescribedintheprevioussection,butareeasiertogenerate.WerstobservethattheformulationIPRcanbesolvedquicklyforsmallinputmatrices.LetSdenoteasubmatrixoftheinputmatrix,andR(S)representthesetofrectanglesthatcoveratleastonebixelinS.LetLB(S)bealowerboundonthenumberofrectanglesrequiredtodecomposeS.SinceLB(S)constitutesalowerboundonthetotalnumberofrectanglesrequired,thefollowinginequalityisvalidforanysubmatrixS:Xr2R(S)yrdLB(S)e: WecanobtainLB(S)byformulatinganauxiliaryintegerprogrammingproblemoftypeIPRforS,andsettingalimitonthemaximumsolutiontime. 96

PAGE 97

4.2.2 Weobserveonclinicaldatathatseveralregionsoftheinputmatrixarecompletelysurroundedbyzero-bixels.Sincenorectanglecancoverazero-bixel,eachoftheseregionscanbesolvedindependently.Aconnectedsubsetoftheinputmatrixobeysthepropertythatarectilinearpathexistsbetweenanytwononzero-bixelsofthesubset,suchthateachbixelinthepathisalsoanonzero-bixelthatbelongstothesubset.Wecallaconnectedsetofnonzero-bixelsacomponentoftheinputmatrixifitisadjacenttozero-bixelsacrossallofitsboundaries(i.e.,ifthesubsetisnotcontainedwithinalargerconnectedsubset). Toidentifythecomponentsoftheinputmatrix,wegenerateagraphGinwhicheachnonzero-bixelhasacorrespondingnode.Weaddanarcbetweenapairofnodesifandonlyifthecorrespondingbixelsareadjacentintheinputmatrix.WethenidentifyconnectedcomponentsonGbyrunningastandarddepth-rst-searchalgorithm.EachconnectedcomponentonGcorrespondstoacomponentoftheinputmatrix,whichcanbesolvedindependentlyofothercomponents.Figure 4-6 depictsthecomponentsoftheuencemapgiveninFigure 4-1 97

PAGE 98

Twocomponentsofauencemap 4-7 ,theregionswithlightanddarkgraybackgroundareindependent.IfwesolveIPRseparatelyoverallindependentregions,thesumofrectanglesrequiredtodecomposeeachindependentregionyieldsalowerboundontheobjectivefunctionforthecorrespondingcomponent. Figure4-7. Regionsofaconnectedcomponent Ingeneral,therearemultiplewaysofpartitioningacomponentintoindependentregions,witheachyieldingpossiblydierentlowerbounds.Theproblemofndingapartitionthatyieldsthebestlowerboundcanbethoughtofasa\dual"ofndingtheminimumnumberofrectanglestodecomposeacomponent.Tosolvethisdualproblem,weneedtobalancetwoconictingcriteria: 98

PAGE 99

Weuseaheuristicproceduretopartitionacomponentintoindependentregions,whichemploysanauxiliaryobjectiveofmaximizingthenumberofcomponentbixelscoveredbyanindependentregion.Eachbixel(i;j)iscalled\committed"ifiteitherbelongstoanindependentregion,orif(i;j)iscontainedwithinsomerectangleinRthatalsocoversbixelsinanindependentregion(andhence,(i;j)cannotbelongtoanotherindependentregion).Allotherbixelsarecalled\uncommitted."Weselectourindependentregionsoneatatime,untilnomoreuncommittedbixelsremain.Theprocedure'sdetailsaredescribedasfollows. 99

PAGE 100

Inouralgorithmforsolvingacomponent,weexecutetheforegoingheuristictondasetofindependentregions.WeformulateIPRforeachregion,withalimitonthemaximumsolutiontime.Wethenusethelowerboundobtainedforeachregiontogenerateaninequalityoftype( 4{24 ).(Itisoftenprudenttoskipthisstepifonlyoneregioniscomputedforacomponent.) 4-7 isadependentregion.LetDrepresentthesetofbixelsinadependentregion,andletR(D)representthesetofrectanglesthatcoveronlyasubsetofthebixelsinD. Toimproveourlowerbound,wewishtocomputetheminimumnumberofrectanglesrequiredtocoverD;however,wewishtoavoiddouble-countingthoserectanglesusedtocoverbixelsinindependentregions.Accordingly,weseektheminimumnumberofrectanglesinR(D),perhapsinconcertwithrectanglesoutsideR(D),requiredtocoverthebixelsinD.Usingthex-andy-variablesasbefore,weformulatethefollowingvariationofIPRtondtheminimumnumberofrectanglesinR(D)requiredtopartition 100

PAGE 101

(4{28)xr08r2R;yrbinary8r2R(D) (4{29) Objective( 4{25 )minimizesthenumberofrectanglesinR(D)usedinthesolution.Constraints( 4{26 )ensurethatthebixelsinDgetpartitionedexactly,where( 4{27 )limittheintensitydeliveredtotheremainingbixels.Constraints( 4{28 )relatethex-andy-variablesasdoneinIPR,andnally( 4{29 )denevariabletypes.Asbefore,wesetatimelimitforthesolutionofDPR,andobtainalowerboundontheobjectivefunctionvalue,whichwedenotebyLB(D).Giventhisvalue,thefollowinginequalityisvalid:Xr2R(D)yrdLB(D)e: Inourexample,theoptimalvalueofDPRfortheblack(dependent)regionis1sincetheintensityrequirementofbixel(1;4)cannotbesatisedcompletelybyrectanglesthatcoverbixelsinthegray(independent)regions(infact,thisresultcanalsobeseenduetotheboundingboxconstraintimplyingthatonerectanglerepresentingthesingletonbixel(1,4)mustappearinanyfeasiblesolution).WenotethattherectanglesinR(D),bydenition,donotintersectanyother(dependentorindependent)regions.Therefore,thelowerboundsobtainedforallregionscanbesummedtoobtainalowerboundontheminimumnumberofrectanglesrequiredtodecomposeacomponent. 101

PAGE 102

4.2.3.2 .FeasiblesolutionsfordependentregionscanbeextractedfromsolutionsoftheformulationgivenbyDPR.However,sinceDPRminimizesthenumberofrectanglesthatarecontainedwithinadependentregion,andnotnecessarilythetotalnumberofrectanglesrequiredtodecomposeadependentregion,thesolutionsobtainedfromDPRpotentiallyuseanunnecessarilylargenumberofrectanglesnotcontainedinR(D). AbetterwayofobtainingfeasiblesolutionsfordependentregionsistoformulatetheproblemIPRforeachdependentregion.SinceIPRexplicitlyminimizesthetotalnumberofrectanglesrequired,weexpectthisapproachtoresultinfeasiblesolutionsofhigherquality.However,thisapproachdoesnotconsiderthefactthatsomeoftherectanglesthatarealreadyusedfordecomposingindependentregionscanbeextendedintodependentregionswithoutincreasingthetotalnumberofrectangles.Topermittheuseofrectanglesthatintersectindependentanddependentregions,werequirearevisedintegerprogrammingformulation. Inourapproach,wesolvetheintegerprogrammingformulationsfordecomposingtheindependentregionsrst,andstorethebestfeasiblesolutionsfoundwithintheallowedtimelimit.Letxrrepresenttheintensityassignedtorectanglerfordecomposingindependentregions.Next,wegenerateafeasiblesolutionforeachdependentregion,oneatatime,asfollows.Werstndthesetofrectanglesthatcanbeextendedintothecurrentdependentregion,anddeterminehowthoserectanglescanbeextended.LetE(D;r)representthesetofrectanglesinRthatextendrectanglerintodependentregion 102

PAGE 103

(4{34)xr0;yrbinary8r2R(D) (4{35)zrebinary8r2R;e2E(D;r): Wegenerateafeasiblesolutionbycombiningthreetypesofrectangles:(i)rectanglesusedtodecomposeindependentregionsthatarenotextendedbyEPR;(ii)rectanglesobtainedbyextendingrectanglesfromindependentregionsintodependentregionsbyEPR;and(iii)rectanglesinR(D)usedbyEPR. NotethattheoptimalvalueofEPRforthedependentregiongiveninFigure 4-7 is1.Thiscanbeseenbyobservingthattherectangle(s)thatcoverbixel(3;4)canbeextendeduptofullysatisfytheintensityrequirementofbixel(2;4)withoutanypenaltyontheobjectivefunctionofEPRformulatedforthedependentregion.Therefore,asinglerectanglecontainedinthedependentregionsolvesEPRoptimally.SincetheoptimalvalueofDPRforthedependentregionisalso1,ourpartitionsolvestheproblemofndingtheminimumnumberofrectanglestooptimality. 103

PAGE 104

wherewisaparameterthatrepresentstheaveragesetuptimeperaperturerelativetothetimerequiredtodeliveraunitofintensity. TheBendersdecompositionprocedurediscussedinSection 4.2.1 alsoneedstobeadjustedaccordingly.WerstaddacontinuousvariablettoMP,which\predicts"theminimumbeam-on-timethatcanbeobtainedbythesetofrectangleschosenbyMP.Theupdatedmasterproblemcanbewrittenasfollows.MPTT:MinimizewXr2Ryr+t (4{39)tminimumbeam-on-timecorrespondingtoy 104

PAGE 105

NotethatSPTTisobtainedbysimplychangingtheobjectivefunctionofSP.IfSPTT(^y)isinfeasible,thenweaddaBendersfeasibilitycutoftype( 4{16 )asbefore,andre-solveMPTT.Otherwise,letthevalueoftinMPTTbe^t,andtheoptimalobjectivefunctionvalueofSPTTbet?.If^t=t?,then(^y;^t)isanoptimalsolutionofMPTTthatminimizesthetotaltreatmenttime.However,if^t>t?,thenweaddthefollowingBendersoptimalitycuttmXi=1nXj=1bij^ij+Xr2R(Mr^r)yr; where^ijand^rareoptimaldualmultipliersassociatedwithconstraints( 4{43 )and( 4{44 ),respectively. 105

PAGE 106

ThemodicationsrequiredfortheBendersdecompositionalgorithmarealsostraightforward.Toenforcetheminimumbeam-on-timerestriction,weadd( 4{47 )toSP,whichcheckswhetheragivensetofrectanglescandecomposetheuencemap.TheupdatedfeasibilitycutisgivenbymXi=1nXj=1bij^ij+Xr2R(Mr^r)yr+T?^0; whereisthedualvariableassociatedwith( 4{47 )inSP.Finally,weneedtocheckwhetherthesolutiongeneratedbyourheuristicdiscussedinSection 4.2.3.4 satisesconstraint( 4{47 );ifso,thenitcanbeusedasaninitialupperbound. OurpreliminarycomputationaltestsshowedthatthenaiveimplementationofourBendersdecompositionapproach,inwhichweaddacutandre-solvethemasterproblemineachiteration,wasnotcomputationallycompetitivewithsolvingtheexplicitintegerprogrammingformulation.Thisisduetothefactthatrepetitivelysolvingthemasterproblem,whichisanintegerprogrammingproblem,iscomputationallyveryexpensive.WeinsteadusedcallbackfunctionsofCPLEXtogenerateasinglebranch-and-bound 106

PAGE 107

OurrstexperimentquantiestheeectsofthevalidinequalitiesdiscussedinSection 4.2.2 ,andthepartitioningapproachdiscussedinSection 4.2.3 onsolutionqualityandexecutiontime.InTable 4-1 ,thesetofcolumnslabeled\DefaultCPLEX"showstheresultsweobtainedbysolvingtheformulationIPRoneachprobleminstanceusingdefaultCPLEXoptions.The\+Validinequalities"columnsrepresenttheIPRformulationenhancedwiththeadjacentrectangleinequalities( 4{17 ),boundingboxinequalities( 4{18 ),strengthenedaggregateintensityinequalities( 4{19 )and( 4{20 ),and12submatrixinequalities( 4{21 )and( 4{22 ).(Additionalcomputationalresultsshowedthatthe22submatrixinequalities( 4{23 )andthearbitrarysubmatrixinequalities( 4{24 )didnotimprovethesolvabilityofthemodel.)Thesetofcolumnslabeled\+Partitions"showstheresultsweobtainedbypartitioningtheproblemintoseparablecomponents(Section 4.2.3.1 ),furtherpartitioningeachcomponentintoindependentanddependentregions(Sections 4.2.3.2 and 4.2.3.3 ),andusingourupperboundingheuristic(Section 4.2.3.4 )inadditiontothevalidinequalitiesusedforthetestsintheprevioussetofcolumns.Werefertothelattersettingsasourbasealgorithmintheremainingcomputationaltests. EachsetofcolumnsinTable 4-1 displaysthetimespentforeachprobleminstance(\CPU"),andupperbound(\UB"),lowerbound(\LB"),andoptimalitygap(\GAP")obtained.Wealsoreporttheaverageandmaximumgapsoverallprobleminstances.WeobservethatnoneoftheprobleminstancesweresolvedtooptimalityusingthedefaultCPLEXoptions,whereasc1b2andc5b2weresolvedtooptimalityafteraddingthevalidinequalitiesofSection 4.2.2 .Anadditionalinstance(c5b5)wassolvedusingthepartitioningstrategydescribedinSection 4.2.3 .Wenotethateventhoughourapproach 107

PAGE 108

Eectofvalidinequalitiesandthepartitioningstrategy +Validinequalities +PartitionsNamemnL CPUUBLBGap CPUUBLBGap c1b1151420 180066600.09 180063620.02 180048470.02 138.248480 180057540.05 180057540.05 180061520.15 180061530.13 180059550.07c1b5111520 180047450.04 180046450.02 1800114790.31 1800119850.29 1800103870.16c2b2171920 180095690.27 180096810.16 180094820.13c2b3181820 180098730.26 1800103770.25 180094770.18c2b4181820 1800114800.3 1800115840.27 1800105880.16c2b5171820 180094640.32 180098720.27 180091720.21c3b1221720 1800121690.43 1800134790.41 1800119790.34c3b2151920 180073460.37 180071520.27 180070520.26c3b3201720 1800119690.42 1800119750.37 1800107770.28c3b4191720 1800103690.33 1800106730.31 180099780.21c3b5151920 180073550.25 180071580.18 1800106790.25 1800107890.17 180088540.39 180099580.41 180091580.36c4b3182320 180095710.25 180099750.24 180093770.17c4b4172320 1800103780.24 1800102810.21 180098830.15c4b5182420 180093620.33 180093660.29 180087670.23c5b1151620 180066640.03 180066650.02 180058570.02 102.158580 180063540.14 180068560.18 180065570.12c5b4141620 180063570.1 180064590.08 180062590.05c5b5121720 180053470.11 180051480.06 36.249490 OurnextexperimenttestsourbasealgorithmundertheextensionsdiscussedinSection 4.3 .Thesetofcolumnslabeledas\TotalTime"inTable 4-2 presentstheextensioninwhichtheobjectivefunctionisdenedasalinearcombinationofthebeam-on-timeandthenumberofrectangles.Theactualvalueofwdependsontheparticulartreatmentdeliveryequipmentusedintheclinic,wherevaluesofwintherange1{10aretypical(see,e.g., DaiandHu ( 1999 ),and Tasknetal. ( 2009b )).Inourexperiments,wethereforeusedw=7asarepresentativevalue.Thenextsetofcolumns(\Lexicographic")isdedicatedtotheextensioninwhichwerstminimizebeam-on-time,T?,andthenndtheminimumnumberofrectanglesthatyieldstheminimumbeam-on-time.Thecolumn\BOT"representsthevalueofT?,and\TotalTime"representsthetotaltreatmenttimeassociatedwiththesolutionfound,whereweagainusew=7astheaveragesetuptimeperrectangle.Weobservethatouralgorithmcouldsolvemoreprobleminstancestooptimalityforbothextensionscomparedtothe 108

PAGE 109

Computationalresultsonmodelextensions LexicographicNamemnL CPUUBLBGapBOTTotalTime c1b1151420 255.96216210 36.566660176638c1b2111520 330.34594590 132.650500121471c1b3151520 1800548542.720.01 130.462620147581c1b4151520 1800557542.490.03 186.962620136570c1b5111520 1800451443.630.02 30.953530115486c2b1182020 1800962814.240.15 18001071040.03194943c2b2171920 1800883797.740.1 180096920.04207879c2b3181820 1800918797.60.13 180096880.08237909c2b4181820 18001028889.360.13 18001111060.052581035c2b5171820 1800890721.130.19 180092830.1207851c3b1221720 18001161858.90.26 18001161030.112661078c3b2151920 1800668533.240.2 180070640.09151641c3b3201720 18001066847.090.21 1800111950.142781055c3b4191720 18001023857.910.16 1800103950.082871008c3b5151920 1800722610.010.16 204.476760182714c4b1192220 18001044918.570.12 18001081050.032751031c4b2132420 1800895656.150.27 180095760.2232897c4b3182320 1800858743.620.13 180092890.03189833c4b4172320 1800943834.320.12 1800101960.05235942c4b5182420 1800913740.190.19 180086770.1260862c5b1151620 271.46266260 5.571710158655c5b2131720 33.45975970 19.963630156597c5b3141620 1800623597.960.04 869.268680180656c5b4141620 1800584571.150.02 192.466660145607c5b5121720 90.45035030 37.257570147546 Baataretal. ( 2005 )).Ontheotherhand,minimizingthebeam-on-timeisapolynomiallysolvableproblem(seeSection 4.3 ).Therefore,weexpectthattheproblemshouldbecomeeasierastheweightofthebeam-on-timetermintheobjectivefunctionincreases.Thereasonthelexicographicminimizationproblemiseasiertosolvethantheothertwovariationsisbecausetheadditionalbeam-on-timeconstraintconsiderablyshrinksthefeasiblesolutionspace. Anotherwayoflookingattheproblemofbalancingthenumberofaperturesandthebeam-on-timeistoviewtheproblemasamulticriteriaoptimizationproblem.Inthissetting,weareinterestedinconstructingtheParetoecientfrontierofsolutionswiththepropertythatneitherofthetwocriteriacanbeimprovedwithoutdeterioratingtheother.NotethatthelexicographicapproachthatweconsideredabovedeterminesaparticularParetooptimalsolutiontothemulticriteriaproblem.Togenerateothernon-dominated 109

PAGE 110

4-1 )withacorrespondingbeam-on-timeof160,whiletheminimumbeam-on-timeforthisprobleminstanceis147(seeTable 4-2 )whichrequires57apertures.Figure 4-8 thendepicts(i)thenon-dominatedsolutions;(ii)theParetoecientfrontierforvaluesof2[49;57],and(iii)the(boundaryofthe)convexhulloftheParetoset.Thesolutionsonthelatteraretheoptimalsolutionstotheproblemofminimizingtotaltreatmenttimethatcanbeobtainedwithdierentvaluesofw. Figure4-8. Ecientfrontierfornumberofaperturesandbeam-on-time OurnalexperimentanalyzestheeectofthemaximumintensityvalueL.Usuallyuencemapsareobtainedbysolvinganonlinearoptimizationproblemforeachbeamangletodetermineanintensityproleforeachbeamangle,whichisrepresentedbya 110

PAGE 111

Eectofmaximumintensityvalueonsolvability CPUUBLBGap CPUUBLBGap c1b11514 4.43053050 22.94414410 1800539536.990c1b21115 1.42382380 4.43203200 498.53943940c1b31515 9.62872870 228.63773770 1800495487.670.01c1b41515 5.82692690 1800393377.420.04 1800513493.130.04c1b51115 2.42162160 28.93263260 316.74114110c2b11820 31.34404400 1800648635.10.02 1800826732.680.11c2b21719 51.54484480 1800625599.840.04 1800759690.670.09c2b31818 144.44284280 1800645615.640.05 1800800704.610.12c2b41818 1593.14874870 1800755678.490.1 1800919796.130.13c2b51718 197.94294290 1800606538.620.11 1800728621.90.15c3b12217 1359.44804800 1800747662.470.11 18001080779.160.28c3b21519 1800280274.970.02 1800414376.010.09 1800532461.60.13c3b32017 1800461446.770.03 1800731641.20.12 1800908755.250.17c3b41917 1800463456.420.01 1800713634.260.11 1800900762.020.15c3b51519 1800332325.870.02 1800481466.60.03 1800582532.20.09c4b11922 39.95295290 1800758719.530.05 1800899827.730.08c4b21324 1800422408.140.03 1800595503.920.15 1800764582.690.24c4b31823 126.34094090 1800579564.50.03 1800695666.230.04c4b41723 321.44444440 1800662649.10.02 1800815742.810.09c4b51824 1194.74144140 1800636573.020.1 1800794662.740.17c5b11516 5.93423420 5.64424420 28.95845840c5b21317 4.32892890 5.54394390 49.35165160c5b31416 1574.42942940 1800453441.440.03 1800566538.220.05c5b41416 3.62392390 1800473468.490.01 1800534524.310.02c5b51217 2.12522520 9.33543540 63.94414410 4-3 showstheresultsofourexperiments.WeobservethatouralgorithmproducessmalleroptimalitygapsasLdecreases,whichisnotsurprisingsinceIPRbecomestighterastheMr-coecients(whichareboundedbyL)decrease.Furthermore,deliveryeciencyisalsohigherforsmallvaluesofL.Theaveragetreatmenttime(calculatedoverthelowerbounds)forallprobleminstancesincreasesfrom366:05forL=5to513:79forL=10,609:79forL=15,and684:96forL=20,whichiscalculatedusingthesetofcolumnslabeled\TotalTime"inTable 4-2 .Ourresultsshowthatthechoiceofgranularitychosenforroundinghasasignicanteectonthetreatmenttime.Foreachindividualpatient,therisksassociatedwiththedeteriorationintreatmentplanqualityduetotheroundingofintensitiesneeds 111

PAGE 112

112

PAGE 113

Thegraphsearchproblemwasinitiallydenedby Parsons ( 1978 )inthecontextofseekingapersonlostinacave.Thecaveisrepresentedasagraph,wheretunnelsofthecavecorrespondtoedgesofthegraph.Searchershavetosweepedgesofthegraphtolocatethemissingperson,whoisassumedtobewanderingunpredictablyorispurposefullytryingtoevadesearchers.Thesearchnumbers(G)ofagraphGisdenedtobetheminimumnumberofsearchersneededsothatthemissingpersoncanbefoundevenifhecouldmoveinnitelyfastalonganypathnotoccupiedbysearchers( Parsons 1978 ).Computings(G)isNP-hardforgeneralgraphs( BienstockandSeymour 1991 ; LaPaugh 1993 ; Megiddoetal. 1988 ),butitcanbecomputedinlineartimefortrees( Alspach 2004 ; Megiddoetal. 1988 ; Pengetal. 2000 ).Thesearchnumberofagraphhasbeenshowntoberelatedtootherimportantparameterssuchastree-width,path-width,andvertexseparation( Dendrisetal. 1997 ; Ellisetal. 1994 ; SeymourandThomas 1993 ). 113

PAGE 114

Flocchinietal. 2008 ; LaPaugh 1993 ; PenuelandSmith 2009 ).Inrendezvousproblemsdierentplayers,whoarenotawareofthelocationofothers,trytomeetatacommonnodeasquicklyaspossible( Alpern 1995 ; AlpernandGal 2003 ; KikutaandRuckle 2007 ).Hide-and-seekproblemsconsideranintruderthat\hides"inastationarylocation,whilethesearcherstrytolocatetheintruderinminimumtime( Alpern 2008 ; JotshiandBatta 2008 ).Suchproblemsalsoariseinsearch-and-rescuesettings( Benkoskietal. 1991 ).Pursuitevasion(or\cops-and-robber")gamesmodelanintruderthattriestoavoidbeingcapturedbysearchers( AignerandFromme 1984 ; Alspachetal. 2008 ; Hahn 2007 ; IslerandKarnad 2008 ).Insomeapplicationsnodesofagraphneedtobepatrolledforprotectionorsupervision( Chevaleyreetal. 2004 ; Saketal. 2008 ).Inparticular,oneinterestingapplicationcoordinatesautomatedsoftwaresearcherssothattheypatroltheInternettondwebsitesthatexploitbrowservulnerabilities( Wangetal. 2005 ).Wereferthereaderto AlpernandGal ( 2003 ); Alspach ( 2004 ); FominandThilikos ( 2008 )fordetailedsurveysoftheliteratureonsearchproblemsandapplicationsinvariouspracticalsettings. Mostofthepreviousresearchongraphsearchproblemshasfocusedontheoreticalaspectsoftheproblems(e.g. Chevaleyreetal. ( 2004 ); Dendrisetal. ( 1997 ); Ellisetal. ( 1994 ); GoldsteinandReingold ( 1995 ); SeymourandThomas ( 1993 ))ordesigningalgorithmsforsolvingtheproblemsonspecialgraphstructures(e.g. Alpern ( 2008 ); Flocchinietal. ( 2008 ); KikutaandRuckle ( 2007 ); Pengetal. ( 2000 )).Ourcontributionisanexactoptimizationalgorithmforsolvingseveralvariantsofthesearchproblemongeneralgraphs(seealso PenuelandSmith ( 2009 )foradecontaminationprobleminwhichtheintruderlocationhasbeendetermined).Inparticular,weconsiderthreespecicgraphsearchproblems:(i)ahide-and-seekproblem,(ii)apursuitevasionproblem,and(iii)a 114

PAGE 115

Avariantofthebranch-and-boundalgorithm,whichaddscuttingplanestolinearprogrammingrelaxationstotightendualboundsiscalledbranch-and-cut,andisemployedinmostcommercialsolversforsolvingintegerprograms( Marchandetal. 2002 ; NemhauserandWolsey 1988 ; Wolsey 1998 ).Aneectivemethodforsolvingintegerprogramshavingalargenumberofvariablesisbranch-and-price,whichisbasedondynamiccolumngeneration( Barnhartetal. 1998 ).Branch-cut-priceisessentiallyanalgorithmthatcombinesdynamiccolumngenerationwithdynamicrowgeneration( JungerandThienel 2000 ). Theremainderofthischapterisorganizedasfollows.InSection 5.2 wedescribeahide-and-seekproblemandproposeacolumngenerationalgorithmforsolvingitslinearprogrammingrelaxation.Similarly,Sections 5.3 and 5.4 analyzethepursuitevasionandpatrolproblems,respectively.WedescribesomebranchingrulesthatcanbeusedinallthreealgorithmstoobtainanoptimalsolutiontotheseproblemsinSection 5.5 .Finally,wegivecomputationalresultsinSection 5.6 Ahujaetal. 1993 ).Wedenethelengthofawalkasthenumberoftraversededgesonthewalk.LetP(T)denotethesetofallpossiblewalksof 115

PAGE 116

(5{3) Theobjectivefunction( 5{1 )minimizesthenumberofselectedsearchers,andconstraints( 5{2 )guaranteethateachnodeiscoveredbyatleastonesearcherwithintheallowedtimeframe.Wenotethataspecialcaseofthisproblemforwhichasearcherlocatedatnodei2Ncanobservenodeianditsneighbors,andweneedtoguaranteeimmediatedetectionoftheintruder(i.e.,whenS(i)=A(i);8i2NandT=0),isequivalenttotheminimumdominatingsetproblem,whichisknowntobeNP-hard( GareyandJohnson 1979 ).Therefore,thehide-and-seekproblemthatweconsiderisNP-hard. 116

PAGE 117

whereupperboundsonthe-variablesarenotnecessaryatoptimality.Givenanoptimaldualvector^,thereducedcostofp,whichwedenotebycp,canbecalculatedas1Pi2N^idpi.Since^isanoptimaldualvector,cp0forallp2P0(T).WecanconcludethatthecurrentsolutionofLHSLPisalsooptimalforthelinearprogrammingrelaxationofHSifcp0forallp2P(T).Ontheotherhand,ifc^p<0forsome^p2P(T)nP0(T),thenadding^ptoP0(T)canpotentiallydecreasethevalueoftheobjectivefunction( 5{4 ).Wediscussourpricingproblem,whichseekssucha^p,inthenextsection. 5{5 )correspondingtonodei2N.Also,letyibeadecisionvariablethatequals1ifnodeiis\seen"byasearcherfollowingawalkthatwegenerate,and0otherwise,8i2N.Givenanoptimaldualvector^,wesolvethefollowingpricingproblemtoseeka-variablehavinganegativereducedcost:maxPi2N^iyi,subjecttotherestrictionthat(y1;:::;yjNj)correspondstoasetofnodesobservedbyawalkoflengthnomorethanT. Thepricingproblemcanbeformulatedasamixed-integerprogrammingproblemonatime-expandednetworkconsistingofT+1stages.Inparticular,wecreateanodeNitforeachi2N;t=0;:::;T.WecreateanarcfromnodeNit,8i2N;t=0;:::;T1tonodesNj(t+1)forallj2A(i).Forthisproblemitiseasytoseethatanoptimalsolutionexistsinwhichallsearchersmoveateachtimeperiod.Therefore,weomitarcsbetweennodesNitandNi(t+1)foreachi2N.Figure 5-1 displaysasimpleexamplegraph,andthecorrespondingtime-expandednetworkforT=2. Toformulatethepricingproblemasamixed-integerprogram,weintroducebinaryvariablesxti=1ifthesearcherisatnodeiattimet.Then,anintegerprogramming 117

PAGE 118

(b) Figure5-1. (a)Anexamplegraph(b)Time-expandednetworkforT=2 formulationoftheproblemcanbegivenas:maximizeXi2N^iyi Constraints( 5{8 )representthefactthatthesearchercanvisitonlyonenodeatatime.Constraints( 5{9 )ensurethatnodeicanbevisitedattimetonlyifoneofitsneighborshasbeenvisitedattimet1.Constraints( 5{10 )forcethevalueofyitozerounlessnodeicanbeobservedbythesearcheratsometimeperiod.Notethatthey-variableswilltakeonbinaryvaluesinanoptimalsolution,andthereforewerelaxthemascontinuousvariables.Iftheoptimalobjectivefunctionvalueof( 5{7 ){( 5{12 )isgreaterthan1,then 118

PAGE 119

WeinitializeouralgorithmbygeneratingastationarysearcherthatstaysatnodeiforTperiods,foreachi2N.Eventhoughtheseelementarysearchersarenotlikelytobeselectedinanoptimalsolution,theyguaranteethefeasibilityofLHSLP.WediscussseveralbranchingstrategiesthatcanbeusedforStep1inSection 5.5 119

PAGE 120

(5{14)p2f0;1g8p2P(T); wherepagainequals1ifandonlyifasearcherisassignedtofollowwalkp.Theobjectivefunction( 5{13 )minimizesthenumberofsearchers.Constraints( 5{14 )ensurethatforeachpossibleintruderwalkoflengthT+1,atleastonesearcherisselectedtodetectit. 5{14 )correspondingtointruderwalkr2R0(T).Also,letyrbeadecisionvariablethatequals1ifanintruderfollowingwalkrisdetectedbyasearcherfollowingawalkthatwegenerate,and0otherwise,8r2R0(T).Givenanoptimaldualsolution^ofthelinearprogrammingrelaxationtoLPE,wesolvethefollowingpricingproblemtoseeka-variablehavinganegativereducedcost:maxPr2R0(T)^ryr,subjecttotherestrictionthat(y1;:::;yjR0(T)j)correspondstoasetofintruderwalksdetectedbyasearcherwalkoflengthnomorethanT. 120

PAGE 121

(5{19)0yr18r2R0(T) (5{20)xti2f0;1g8i2N;t=1;:::;T: Constraints( 5{17 )ensurethatthesearchercannotbelocatedatmultiplenodessimultaneously.Constraints( 5{18 )modelthefactthatthesearchercaneitherstayatthesamenode,orcanmovetoanadjacentnodeateachperiod.Constraints( 5{19 )representtheconditionthatthesearcherdetectsintruderr2R0(T)onlyifitmovestoanodewhereitcandetecttheintruderduringthepursuit.Wenotethatthey-variablescanberelaxedascontinuousvariablesinthiscase,too.Thispropertyallowsthenumberofbinaryvariablesinthesearcher'sproblemtostayconstantasnewevasionpathsfortheintruderarediscovered.Asbefore,iftheoptimalobjectivefunctionvalueof( 5{16 ){( 5{21 )isgreaterthan1,thenwehavefoundavariablewhosereducedcostisnegative. 121

PAGE 122

Afterconstructingthetime-expandednetworkasdescribed,weseekafeasibles{qpathonthenetworkbyastandardbreadth-rst-searchalgorithm,whichworksinO(N2T)timeintheworstcaseifGisdense.Ifsuchapathexists,thenitcorrespondstoawalkrthattheintrudercantaketoavoiddetectionforT+1timeunits.Inthiscase,weaddrtoR0(T),andgeneratetheassociatedconstraintoftype( 5{14 ).Ontheotherhand,ifnosuchpathexists,then^isafeasiblesolutionofPA. 5{14 ),andgobacktoStep1.Else,stopprocessingthecurrentsubproblemwithanintegralsolution. 122

PAGE 123

5.5 5.4.1ProblemDescription 123

PAGE 124

(5{23)p2f0;1g8p2Pc(K) (5{24) Weproposeabranch-cut-pricealgorithmsimilartothepursuitevasionproblemforsolvingthisproblem.WestartwithasubsetofpatrolroutesP0c(K)andevasionwalksR0(T),andsolvetheresultinglimitedpatrolproblem(LPP).Wegeneratenewpatrolroutesandevasionwalksasneeded. 5{23 )correspondingtointruderwalkr2R0(T).Wedeneyrtobeadecisionvariablethatequals1ifanintruderfollowingwalkrisdetectedbyasearcherfollowingapatrolcircuitthatwegenerate,and0otherwise,8r2R0(T).Givenanoptimaldualsolution^ofthelinearprogrammingrelaxationtoLPP,wesolvethefollowingpricingproblemtoseeka-variablehavinganegativereducedcost:maxPr2R0(T)^ryr,subjecttotherestrictionthat(y1;:::;yjR0(T)j)correspondstoasetofintruderwalksdetectedbyasearcherfollowingapatrolcircuitofperiodnomorethanK. Wecansolvethesearcher'sproblembysolvingaseriesofintegerprogramsasfollows.Letdenotethelengthofthecurrentcircuitunderconsideration.Byconsideringdierentvaluesof2f1;:::;Kgwecanndacircuitthatoptimizesthesearcher'sproblem.NotethatsomevaluesofmaynotcorrespondtoanycircuitsinG.Foreachvalueof,wegenerateatime-expandednetworkcontaining+1levels,wheretherstlevelcorrespondstotheinitialdeploymentofthesearcher,andthelastlayerisadummylayerthatweusetomodeltherecurringpatrolpatterns.Weconnecteachnodetoitsneighborsinthenextstage.Asbefore,wedeneabinaryvariablextiforalli2N;t=0;:::;,whichequals1ifthesearcherislocatedatnodeiattimet.Wealsodeneaparameterdtir=1ifasearcherlocatedatnodeiattimetcandetectanintruderfollowingwalkr2R0(T).We 124

PAGE 125

(5{29)0yr18r2R0(T) (5{30)xti2f0;1g8i2N;t=0;:::;: Constraints( 5{26 )and( 5{27 )ensurethateachfeasiblesolutioncorrespondstoawalk.Constraints( 5{28 )guaranteethattherstandthelastnodesvisitedbythesearcherarethesame,andhencethesearcher'swalkformsacircuit.Finally,Constraints( 5{29 )relatethex-andy-variables,wherewecanonceagainrelaxintegralityrestrictionsonthey-variables. Integerprogramscorrespondingtodierentvaluesofcanbesolvedinanysequence.Wenotethatagoodsolutionobtainedbysolvingthesearcherproblemforaparticularvalueofcanbeusedtopruneproblemstobesolvedlaterfordierentvaluesofbybound.Therefore,wecanstartbysolvingasearcherproblemforthelargestvalueof,sinceasearcherfollowingalongercircuitismorelikelytodetectmoreintruderwalks.Also,weskipany~ifwedeterminethatnocircuitoflength~existsinG. 125

PAGE 126

Tosolvetheintruder'sproblem,wecangenerateatime-expandednetworkconsistingofLstagessimilartothepursuitevasionproblem.Weconnecteachnodetothecopyofitselfanditsneighborsinthenextstagebyadirectedarchavinglength1.WealsoconnectthenodescorrespondingtostageLtothenodestotheirneighborsintherststagewithdirectedarchavinglength1(modelingthefactthattheoverallsearchpatternrepeatsafterLperiods).Weaddadummystartnodesandadummyendnodeq.Weconnectstoallnodesbyadirectedarchavinglength0(reectingourassumptionthattheintrudercanenterthesystematanytimeandlocation),andconnectallnodestoqbyadirectedarchavinglength0.Finally,wetraceeachselectedsearcher'scircuit,andremovenodesandarcsfromtheintruder'snetworkthatwouldleadtothedetectionoftheintruderbythesearcher. Wecansolvetheintruder'sproblemonthegeneratedgraphbyseekingalongests{qpath.Werstseekatopologicalorderingofthenodesusingastandarddepth-rst-searchalgorithm,whosecomplexityisO(N2L)foradenseG.Sinceadirectedgraphisacyclicifandonlyifitishasatopologicalorder,thisstepidentieswhetherthegraphiscyclic.Ifthereisacycleinthisgraph,thentheintrudercanstayinthesystemforeverwithoutbeingdetectedbythesearchers.Inthiscase,wegenerateacutoftype( 5{23 ),andstop.Otherwise,thegraphisacyclic,andgivenatopologicalorderofthenodes,alongestpathcanbefoundinpolynomialtimebyadynamicprogrammingalgorithmwhosecomplexityisO(N2L)( Ahujaetal. 1993 ).IfthelengthofalongestpathisgreaterthanT,thentheintrudercansuccessfullyevadethesearchers.Wecanusesuchalongestwalktogenerateacutoftype( 5{23 ). 126

PAGE 127

5{23 ),andgobacktoStep1.Else,stopprocessingthecurrentsubproblemwithanintegersolution. Asbefore,weinitializeouralgorithmbygeneratingastationarysearcherthatstaysatnodei,andastationaryevaderthatstaysinnodeiforT+1periods,foreachi2N.WediscussseveralbranchingstrategiesthatcanbeusedinStep1inthenextsection. 127

PAGE 128

Barnhartetal. 1998 ; JungerandThienel 2000 ). However,ourmasterproblemisasetcoverproblem,andtheapproachdescribedaboveisnotdirectlyapplicable.AnyconstraintonagroupofvariablesforoursetcoverproblemwouldnecessitateaddingabranchingconstraintlikePx2Xx~inonebranch,andPx2Xx~+1(forsomesuitable~)intheotherbranch.Sincetheseconstraintscannotbeaddedimplicitly,weneedtohandleanewdualvariableforeachbranchingconstraintinthesubproblem. Asanalternative,weproposeamulti-tieredbranchingrule.Givenafractionalsolution^,werstevaluatethevalueofeachconstraintexpressionvr=Pp2P0(T)dpr^p.Ifthereexistsafractionalvrvalueforsomer2R0(T),thenwebranchontheconstraint( 5{23 )correspondingtorasfollows.Intheup-branch,wesimplychangetheright-hand-sidevalueoftheconstraintasdvre.Onthedown-branch,wesettheupperboundoftheconstraintexpressiontobvrc(andconvertittoanequalityconstraintifbvrc=1).Notethatbranchingonaconstraintinthismannerdoesnotintroduceanynewdualvariablesorconstraintsthatneedtobeconsideredinanyofthepricingproblems.Anotherbenetofthisbranchingschemeisthefollowing.IfwebranchdownonaconstraintandobtainPp2Pdprp=1,wecanthenuseset-partitiontypeofbranchingschemesonthecorrespondingsubproblem.Thisallowsustobranchonagroupof-variablesinthesubsequentbrancheswithoutdestroyingthepricingproblemstructure. Notethatsignofthedualvariablercanchangeafterbranchingdownontheconstraint( 5{23 )correspondingtor,whichmakesyr=0optimalregardlessofthevaluesofx-variables.Therefore,weneedtoaddconstraintsthatforceyrto1ifthesearcher'schosenpathdetectsintruderr.Hence,weaddconstraintsyixtj8i2N;j2S(i);t=0;:::;T 128

PAGE 129

(5{34) tothesearcher'spricingproblemforthehide-and-seek,pursuitevasion,andpatrolproblems,respectively. Ingeneralitispossibletohaveafractionalsolution^forwhichallv-valuesareintegral,andthereforethebranchingruledescribedabovecannotbeapplied.Insuchcaseswecanapplyasimplevariable-basedbranchingrule.Ifthereisavariable^pwhosevalueisfractional,wecansimplycreatetwobrancheswith^p=0and^p=1.Inthedown-branch,wesimplyeliminatethecolumncorrespondingto^pfromthesetcoveringformulation.Intheup-branchweneedtoadjusttheright-hand-sidevectorofoursetcoveringproblembeforeeliminating^p.Ineithercase,weneedtoadjustthepricingproblemssothatthesamevariablecannotberegenerated.Recallthatthesolutionmethodsweproposeforallthreepricingproblemsarebasedonatime-expandednetworkformulation.LetusdenotebyW(p)thesetoftime-expandednodeindicescorrespondingtoasearcherwalk(orcircuit)p.Wecanenforcetheconditionthatawalkpisnotgeneratedagainbyaddingthefollowingconstrainttothecorrespondingsearcher'spricingproblem:X(i;t)2W(p)xtijW(p)j1: Branchingonasinglevariableislikelytobequiteweakonthedown-branchsinceonlyoneparticularsearcherwalk(orcircuit)isavoided.Hence,weonlyapplythisruleifourconstraint-basedbranchingrulefails.Alsonotethatthedicultyofsolvingthesetcoveringproblemdoesnotincreaseundereitherbranchingrule,sincenoconstraintsareaddedexplicitlywhilebranching. 129

PAGE 130

Sieketal. 2001 ).Ourbasesetoftestprobleminstancesconsistsof150randomlygeneratedprobleminstancesforwhichtheexpectededgedensityofthegraph(measuredasjEjjNj jNj(jNj1),wherewedonotconsiderself-loopedgesincalculatingedgedensity)is10%,thenumberofnodesNrangesfrom5to25,andthemaximumallowedtimetodetectionTrangesfrom0to5.Ingeneratinginstanceswerstpickedarandomsubsetofedgessothattheedgedensityis10%,andifnecessaryaddedtheminimumnumberofedgesneededtomakethegraphconnected(see Sieketal. ( 2001 )).Wethenaddedself-loopedges,andwegeneratedveprobleminstancesforeachproblemsize,whichisdeterminedbythenumberofnodes.Finally,wesolvedeachprobleminstancewithdierentvaluesofT2f0;:::;5gforthehide-and-seek,pursuitevasionandpatrolproblems.Ineachcase,weassumethatasearcherlocatedatnodei2Ncanobservenodeiandallnodesadjacenttoit,andhencewesetS(i)=A(i),foralli2N.Weimposeda1200-secondtimelimitpastwhichwestoppedtheexecutionofanalgorithminallourexperiments. RecallthatallproblemsthatweconsiderinthischapterreducetotheminimumdominatingsetproblemforT=0.WeusethispropertytocalculateaninitialupperboundbysolvingourLHSformulation,whereweinitializeR0(0)byaddingasearcherlocatedateachnodei2N. Ourrstexperimentfocusesonthehide-and-seekproblem.Forthisexperiment,weexecutedourbranch-and-pricealgorithmdescribedinSection 5.2.2.2 onourbasedataset.All150probleminstancesinourdatasetweresolvedtooptimalitywithin12.8seconds.Table 5-1 displaystheaveragenumberofbranch-and-boundnodesevaluatedinourbranch-and-pricealgorithmfordierentvaluesofNandT.Eachvaluerepresentstheaverageoftheresultsobtainedonverandomlygeneratedgraphs.Weobservethatthenumberofbranch-and-boundnodesthatareexploredincreasesasNincreases,which 130

PAGE 131

Averagenumberofbranch-and-boundnodesexploredforhide-and-seekproblem 51111111011.41.81.41.41151111112012.63.41.81.812517.42.22.61.41 5-2 Table5-2. Averagenumberofsearchersneededforthehide-and-seekproblem 521.81111103.42.821.81.61154.63.22.2221.6204.63.62.6221.8255.83.82.8222 5-2 revealsthatthenumberofsearchersneededincreasesasthegraphgetslarger,anddecreasesasthemaximumallowedtimetodetecttheintruderincreases. Weanalyzetheperformanceofourbranch-cut-pricealgorithmdescribedinSection 5.3.2.3 forthepursuitevasionproblem.Table 5-3 showsthatouralgorithmwasabletosolve128outofthe150instanceswithintheprescribedtimelimit.Asexpected,thedicultyoftheproblemincreasesasNandTincrease,sincethissettingallowsformoreevasionroutesfortheintruder,andhencerequiresthesearcherstodevelopmoresophisticatedroutes. Table5-3. Numberofinstancesthataresolvedwithintimelimitforthepursuitevasionproblem 555555510555555155555552055531225554300 5-4 displaystheaveragenumberofbranch-and-boundnodesevaluatedinourbranch-cut-pricealgorithmforthepursuitevasionproblem.Wenotethatforthis 131

PAGE 132

Averagenumberofbranch-and-boundnodesexploredforthepursuitevasionproblem 51111111011.831.81.8115153.46.24.2320128.2109334.610114.225123.4869324.641.411.4 5-5 displaystheaveragenumberofsearchersneededforthisproblem,whereweusethebestknownsolutionsforinstancesthatwerenotsolvedtooptimalitywithinthetimelimit.A Table5-5. Averagenumberofsearchersneededforthepursuitevasionproblem 521.81111103.432.2222154.63.42.6222204.6432.82.82.6255.843.43.23.23 5-2 and 5-5 revealsthatthenumberofsearchersneededforthehide-and-seekproblemislessthanthatforthepursuitevasionproblem.Thisresultisnotsurprisingsincetheintruderisstationaryintheformerproblem,whileitcanmovetoavoidthesearchersinthelatterproblem. Table5-6. Numberofinstancesthataresolvedwithintimelimitforthepatrolproblem 555555510555555155555552051111025500000 5.4.2.3 forthepatrolproblem.Table 5-6 revealsthatouralgorithmforthepatrolproblemcansolvefewerinstancesinourdatasetwithinthetimelimitcomparedtoouralgorithmsfortheotherproblems.Thiscanbeexplainedby(i)ourassumptionthattheintrudercanpickatimetoenterthesystem,and(ii)oursolutionalgorithmforthesearcher'sproblem,whichrequiressolvingmultiplemixed-integerprograms.Therstfactormakesiteasierfor 132

PAGE 133

5-7 Table5-7. Averagenumberofbranch-and-boundnodesexploredforthepatrolproblem 511111110111.41.41.41.415144.620.26.211.4320112.25.44.23.4125153.22.11.21 5-8 .Asbefore,ourcalculationsarebasedonthebestknownsolutionsanddonotnecessarilycorrespondtooptimalsolutionsfortheinstancesthatwerenotsolvedwithinthetimelimit.However,weobservethatthenumberofsearchersneededforthepatrolproblemislargerthantheothertwoproblemsasexpected. Table5-8. Averagenumberofsearchersneededforthepatrolproblem 521.81.61.211103.432.82.62.62.6154.63.432.82.82.6204.64.23.83.132.8255.85.34.73.83.63.2

PAGE 134

Inthischapter,wefocusonfutureresearchareasregardingtheproblemswehavedescribedinthepreviouschapters.Weanalyzethetechniquesweemployed,identifyassociatedweaknesses,andsuggestimprovements.Wealsodiscussaresearchtopicthatisbasedonatechniqueforreformulatingthemasterproblemforaclassofbi-leveloptimizationproblems. 2 ,seealso Tasknetal. ( 2009a ))werstdevelopedanintegerprogrammingformulationoftheproblem,andprescribedabi-levelreformulationoftheproblemthathasintegervariablesinbothstages.Ourcomputationaltestsshowedthatboththedirectsolutionoftheintegerprogrammingformulationandourintegerprogramming-basedcuttingplanealgorithmforthebi-levelformulationarecapableofsolvingonlysmallprobleminstancestooptimality.Wethendesignedahybridintegerprogramming/constraintprogrammingalgorithmtoovercomethecomputationaldicultiesencounteredbythersttwoapproaches.Ourhybridapproachrstallocatesnodecopiesthataretobedistributedacrosscongurationsusinganintegerprogrammingformulation,andthenassignsnodestosubgraphsusingaconstraintprogrammingalgorithm.Afterassigningnodestosubgraphs,itpartitionsedgestosubgraphsforeachscenarioinathirdstage,usinganotherintegerprogrammingformulation.Ourcomputationalexperimentsshowthatthehybridapproachsignicantlyoutperformstheotherapproaches. Inourstudy,wehaveassumedthatthenumberofsubgraphs,jKj,isapartoftheprobleminput.InSONETnetworkdesignapplicationthereisnopracticallimitonthenumberofsubgraphs,butalimitisspeciedtomodeltheproblem( Goldschmidtetal. 2003 ; Sheralietal. 2000 ; Smith 2005 ).ChoosingjKjtoosmallmaymaketheprobleminfeasibleorsuboptimal,andchoosingjKjtoolargeincreasesthedicultyoftheproblem 134

PAGE 135

2.4 .Inourexperiments,wemanuallychosejKjsucientlylargetoyieldafeasiblesolutionineachprobleminstance.AnareaforfutureresearchistotreatjKjnotasaparameterbutasavariable,andtondthesmallestpossiblevalueofjKjthatguaranteestheexistenceofanoptimaledgepartitionthatminimizesthenumberofnodecopies.Thisproblemappearstobeverydicultingeneral,butsomeupperboundscanbederivedforourproblem.WerstnotethatchoosingjKj=jEjisguaranteedtoyieldafeasibleedgepartition(andnoteliminateanyfeasibleedgepartitionsfromconsideration),sinceeachedgecanbepartitionedintoauniquesubgraph.Furthermore,ifafeasibleedgepartitionhavinganobjectivefunctionvalue^zisknown,thenb^z=2cisanupperboundonjKj.Thisboundfollowsfromthefactthatthereexistsanoptimalsolutionthatcontainsatleasttwonodesineachnon-emptysubgraph.Sucha^zcanbecalculatedbyasimpleimprovementheuristic.WestartwithjKj=jEj,andinitiallyassigneachedgetoauniquesubgraph.Wethenseektwosubgraphsthatcanbemergedwithoutviolatinganyconstraintsinanyscenario,whilealsoimprovingtheobjectivefunctionvalue.Ifwendsuchsubgraphs,wemergethemandrepeatthisprocessuntilnomoresubgraphscanbemerged.Sincethisalgorithmstartswithafeasiblesolution,andretainsfeasibilityineachiteration,ityieldsafeasiblesolution.However,ourpreliminaryanalysissuggeststhattheboundonjKjthatwegetusingthisapproachisquiteweak.ImprovingboundsonjKjandndingthesmallestpossiblenumberofsubgraphsthatyieldsanoptimaledgepartitionisafutureresearcharea. 3 wehavedescribedexactdecompositionalgorithmsforsolvingleafsequencingproblemsarisinginIMRTtreatmentplanning(alsosee Tasknetal. ( 2009b )).Oursolutionalgorithmforthematrixdecompositionintoaperturessatisfyingtheconsecutive-onespropertyisbasedonanintegerprogrammingmodelforndingasetofintensityvaluestobeassignedtoapertures,andabacktrackingalgorithmthatformsaperturesbyndingcompatibleleafpositionsforeachrow.Ourcomputationalresults 135

PAGE 136

Oursolutiontechniquefortheconsecutive-onesmatrixdecompositionproblemandourthree-stageapproachtothestochasticedge-partitionproblemarebasedonasimilaridea.Inbothproblems,weadd\aggregatevariables"toourformulations,whichdescribeimportantstructuralpropertiesofsolutions,butarenotenoughbythemselvestoencodecompletesolutions.Ineachcasewerepresenttheoptimizationproblemintermsofouraggregatevariablesinamasterproblem,andprovideasubproblemthatseeksacompletefeasiblesolutioncorrespondingtothevaluesoftheaggregatevariableschosenbythemasterproblem.Inbothapplications,ourmasterproblemsarediscreteoptimizationproblems,whichwesolveusingintegerprogrammingmethods,andoursubproblemsarediscretefeasibilityproblems,whichwesolveusingconstraintprogrammingmethods.Separatingcriticaloptimizationdecisionsfromfeasibilitydecisions,andutilizingstrengthsofintegerandconstraintprogrammingtechniquesinahybridalgorithmhasallowedustoobtainsignicantlybetterresultsthanothermethods.Amajorthemeinourfutureresearchisgoingtobeongeneralizingthishybridapproachtohandleabroaderclassofproblems. InChapter 4 westudiedadierentvariantofthematrixdecompositionproblem,inwhichtheaperturematricesneedtoberectangularinshape(seealso Tasknetal. ( 2008 )).RectangularaperturescanbeformedbyusingconventionaljawsalreadyintegratedintoIMRTtreatmentdevices,anddonotneedanadvancedMLCsystem,whichiscostlytomanufactureandoperate.InChapter 4 ,weproposedanexactoptimizationalgorithmthatcanbeusedtoanalyzewhetherajaws-onlytreatmentsystemcandeliveruencemapseciently.Ouralgorithmisbasedonanintegerprogramming 136

PAGE 137

Wealsoderivedabi-levelBendersdecompositionalgorithmforthisproblem.ThemasterproblemofourBendersdecompositionapproachchoosesasubsetoftherectangularshapesthatcanbeusedindecomposingtheinputmatrix.Later,asubproblemcheckswhethertheselectedsubsetofrectanglescancompletelydecomposetheinputmatrix.Unfortunately,ourcomputationaltestsshowedthatourBendersdecompositionalgorithmiscomputationallyinferiortotheintegerprogrammingapproach.Themainreasonofslowconvergenceistheweaknessofcutsgeneratedineachiteration.Specically,givenaninfeasiblesubsetofrectangleschosenbythemasterproblem,oursubproblem(whichisalinearprogrammingproblem)detectsinfeasibility,andreturnsacut,whichisgeneratedbasedonadualextremeray.However,thisextremerayisamathematicalproofofinfeasibility,butdoesnotnecessarilyidentifytheunderlyingrea-sonofinfeasibility.Inotherwords,itdoesnotidentifywhichbixelsintheinputmatrixcannotbepartitionedwiththeselectedsubsetofrectangles.Furthermore,therearetypicallymanyextremedualraysforasingleinfeasiblemasterproblemsolution,fromwhichseveralnon-dominatedcutscanbederived.Onewayofimprovingtheconvergencecanbeapplyingatwo-dimensionalbinarysearchalgorithmontheinputmatrixtondoutwhichregionoftheinputmatrixcannotbepartitionedwiththeselectedrectangles.Thatis,iftheentirematrixcannotbepartitioned,wetrytopartitiontheleft-hand-sideandtheright-hand-sidehalvesofthematrixindependently.Ifoneofthesesubmatricescannotbepartitioned,weimmediatelyhaveamorespecicreasonforinfeasibility(andhenceastrongercut),becausethisresultimpliesthattheinfeasibilityinasubmatrixneedstobexedusingasubsetoftherectangles,whichcoverthatpartofthematrix.Thisideacanbeappliedrecursivelytondpossiblymultipleinfeasibleregionsofthematrix.Thecutsassociatedwiththeseinfeasiblesubmatricesaremuchstrongerthanasinglecutderivedbasedontheentirematrix.Furthermore,theinformationregardingthe 137

PAGE 138

Inourstudy,wehavedevelopedsolutiontechniquesfortwoversionsofthematrixdecompositionproblem,whichapplytomostavailableIMRTtreatmentmachinery.However,thereareothertypesofmachinerythathavedierentapertureshaperestrictions,suchasinterdigitationorconnectednessconstraints(seee.g., Lim ( 2002 )).Inarelatedlineofresearch,weareplanningtodesignexactoptimizationalgorithmstosolveothervariantsofthematrixdecompositionproblemtooptimality.Quantifyingtheeectofseveralshapeconstraintsenforcedbydierenttypesofmachineryonradiationdeliveryeciencywouldbeavaluablecontributiontothemedicalphysicseld. 5 weconsideredthreevariantsofagraphsearchproblem:(i)ahide-and-seekproblem,whereasetofsearchersseekastationaryintruder,(ii)apursuitevasionproblem,wheretheintrudermovestoavoiddetection,and(iii)apatrolproblem,wheresearchersfollowrecurringpatrolroutes.Theaimineachproblemistondtheminimumnumberofsearchersneededsothattheintrudercannotstayinthesystemwithoutbeingdetectedforlongerthanaprespecieddurationoftime.Weproposedabranch-cut-pricealgorithm,whichcanbeadaptedtoallthreeproblemswithcertainmodications.Ourmaincontributionisthatwedonotmakeanyassumptionsonthetopologyoftheinputgraph,andouralgorithmsworkongeneralgraphs. EventhoughallthreeproblemsthatweconsiderareNP-hard,ourcomputationaltestsshowthatthehide-and-seekandpursuitevasionproblemscanbesolvedtooptimalityformodestproblemssizeswithinreasonablecomputationaltime.However, 138

PAGE 139

Oursetcoveringformulation,whichisbasedontheideathatatleastonesearchermustbechosenforeachroutetheintrudercantake,canbegeneralizedtoothervariantsofthegraphsearchproblem.Inparticular,aproblemthathasbeenwidelystudiedintheliteratureassumesthattheintrudercanresideattheedgesofthegraph.Thisproblemhasbeeninvestigatedfromatheoreticalperspective(see,e.g., Dendrisetal. ( 1997 ); Ellisetal. ( 1994 ); SeymourandThomas ( 1993 ));however,tothebestofourknowledgenoexactoptimizationmethodthatworksongeneralgraphshasbeenproposed.Weareplanningtoextendouralgorithmsothatitcanalsosolvethisproblem. 139

PAGE 140

Smithetal. 2004 ),andinaproductintroductionandinterdictiongame, Smithetal. ( 2008 )considertheadditionofaquadratic-ordersetofvariablesinthemasterproblem.Thesenewvariablesarepassedtothesubproblem,andaBenderscutisgeneratedintermsofthenewvariablesthatimpliesallofthe(exponentially-many)Benderscutsthatcouldhavebeengeneratedintheoriginalvariablespace.ThismasterproblemreformulationtechniquehasthepotentialtodramaticallyreducethenumberofiterationsrequiredbyBendersdecompositiontoconvergetoanoptimalsolution,withonlyamodestincreaseinthesizeoftheformulation.Inbothcasesmentionedabove,thetrade-oofincreasingmodelsizetoimprovethestrengthofBenderscutswascomputationallybenecial. LetusdescribetheideainmoredetailinthecontextofaSONETnetworkdesignproblem( Smithetal. 2004 ),whichissimilartotheedge-partitionproblemdiscussedinChapter 2 .Thereexistasetofdemandpairs(i;j)2Ethatmaybesatisedonasinglecommunicationsnetwork(alldemandpairshavetobesatisedinouredge-partitionproblem).Thecommunicationsnetworkiscomposedof\rings"k=1;:::;K.Ifbothclientsiandjhavebeenlinkedtoringk,thenwemaychoosetosatisfythedemandrequestbetweeniandjonringk.Deneyqijktobeacontinuousvariablethatrepresentsthefractionofthedemandbetweeniandjthatissatisedonringkinscenarioq(thesevariablesaredenedtobebinaryinouredge-partitionproblemsinceweassumethateachdemandpairneedstobesatisedonasinglering).LetEqbethesetofdemandpairs 140

PAGE 141

wheretheexistenceofscenariosq2Qisduetouncertaindemandsbetweenclientsiandj.Usingastraightforwarddecompositionapproach,theproblemcanbedecomposedsothatthex-variablesarepartsofthemasterproblemformulation,whiletheyq-variablesaredeterminedinsubproblemscorrespondingtoscenariosq2Q.Constraints( 6{1 )and( 6{2 )essentiallystatethatinorderforthecommunicationdemandbetweencustomersiandjtobeassignedtoringk,bothcustomersiandjhavetobeassignedtoringk.Unfortunately,cutsenforcingthisrelationshipcannotberepresentedintheoriginalspaceofx-variables. Smithetal. ( 2004 )showthattherecanexistanexponentialnumberofalternativedualsolutionsassociatedwithamasterproblemsolutionrepresentedby^x,eachleadingtoanon-dominatedBenderscut.Thentheyreformulatethemasterproblembyaddinguijkvariables,whichrepresenttheminimumofxikandxjk.Inotherwords,uijk=1ifbothcustomersiandjareassignedtoringk.Giventheu-variables,theconstraints( 6{1 )and( 6{2 )canbereplacedbyyqijkuijk8q2Q;(i;j)2Eq;k2K: Smithetal. ( 2004 )showthatasingleBenderscutbasedontheu-variablesdominatesanexponentialnumberofcutsbasedontheoriginalx-variables.Inotherwords,addingaquadraticnumberofvariablestothemasterproblemcansaveanexponentialnumberofiterationsoftheBendersdecompositionalgorithm.Inthisparticularproblem,theauthorsrecognizedthatthey-variablesinthesubproblemaredependentonmin(xik;xjk),andusedthisnonlinearrelationshipbetweenthex-variablestoreformulatethemaster 141

PAGE 142

wherex1andx2arevariablesofthemasterproblem.Sinceboth( 6{4 )and( 6{5 )areoftype,wecantakethecomponent-wiseminimumofthetwoconstraintstoobtainmin(a11;a21)y1+min(a11;a21)y2++min(a1n;a2n)ynx1 Sincetheleft-hand-sidesof( 6{6 )and( 6{7 )arethesame,wecancombinethetwoconstraintsintomin(a11;a21)y1+min(a11;a21)y2++min(a1n;a2n)ynmin(x1;x2): Constraint( 6{8 )describesanonlinearrelationshipbetweenthey-andx-variables.Atthispoint,themasterproblemcanbereformulatedbyaddingavariablev12=min(x1;x2),andthesubproblemcanbereformulatedbyusingthisvariableasmin(a11;a21)y1+min(a11;a21)y2++min(a1n;a2n)ynv12: Eventhough( 6{9 )isweakerthan( 6{4 )and( 6{5 ),thisreformulationmightimprovetheconvergenceofthealgorithmduetothe\exponential-cut"behavior,especiallyinthebeginningiterationsoftheBendersprocess.Inourfutureresearchweareplanning 142

PAGE 143

143

PAGE 144

Inthisappendixwediscussanintegerprogrammingapproachtodecomposingauencemapintoanumberofaperturesandcorrespondingintensitiesthatisbasedonamodelproposedby Langeretal. ( 2001 ).Givenamaximumnumberofunit-intensityapertures,sayT,thisformulationdeterminesthepositionsoftheleftandrightleavesineachrowofeachoftheseapertures.Wedevelopthemodelbyseparatelystudyingfourcomponents: Inparticular,constraints( A{2 )statethateachbixeliseithercoveredbyaright-handleaf,coveredbyaleft-handleaf,oruncovered(wherethed-variablesareincludedonlyforconvenienceandcanbesubstitutedoutoftheformulation).Constraints( A{3 )and( A{4 )statethatifanybixel(i;j)iscoveredbyaright-handleaf(resp.left-handleaf),thenbixel(i;j+1)(resp.(i;j1))shouldbecoveredbyaright-handleaf(resp.left-handleaf)aswell. 144

PAGE 145

While Langeretal. ( 2001 )imposethefollowingconstraintstoensurethatthesevariableshave(atleast)theirdesiredvalue: wenotethatthefollowingstrongerformulation,whichwouldactuallynotrequireenforcingthez-variablestobebinary,canbeobtainedbydisaggregating( A{6 ).dtijzt8i=1;:::;m;j=1;:::;n;t=1;:::;T:( A{6 Notethatthismodelallowszttobeequaltooneevenifinaperturetnobixelsareexposed,sothatformallyspeaking( A{5 )isanupperboundonthebeam-on-time.Theobjectivefunctionensuresthatthez-variablestakeontheirminimumpossiblevalue. (Ifanyapertureisusedmorethanoncebutseparatedbyanotherone,weconsiderthesecondoccurrenceoftheaperturetobeanewsetup.However,whenminimizingtotaltreatmenttimetherealwaysexistsanoptimalsolutioninwhichidenticalaperturesaredeliveredsequentially.)Nowletctijandutijbeauxiliarybinaryvariablessuchthattheformerisequaltooneifbixel(i;j)isexposedinaperturetbutnotinaperturet+1andzerootherwise,andthelatterisequaltooneifbixel(i;j)iscoveredinaperturetbutnotinaperturet+1.Thisrelationshipisstatedby Langeretal. ( 2001 )thenusethefollowingconstraintstoensurethatthevariablesgthave(atleast)theirdesiredvalue: 145

PAGE 146

A{9 Similartothecaseofthebeam-on-time,thismodelallowsgttobeequaltooneevenifaperturestandt+1areidentical,althoughourobjectivefunctionensuresthattheg-variablesarechosensucientlysmall. Langeretal. ( 2001 )thenstudytheproblemofminimizingthenumberofsetups( A{7 )subjecttotheconstraints( A{2 ){( A{4 ),( A{6 ),( A{9 ),theconstraintthatthebeam-on-timeisminimal: andbinaryconstraintsonthevariables,wherewerecommenddetermining~zviaoneofthepolynomial-timeproceduresmentionedinSection 3.1 .WenotethatanequivalentmodelisobtainedbysimplysettingT=~z,whichreducestheproblemdimension,andhenceshouldbemoreecientthanaddingabeam-on-timeconstraint. Wewishtominimizethetotaltreatmenttimeasmeasuredby subjecttoconstraints( A{2 ){( A{4 ),( A{6 A{9 A{10 )). 146

PAGE 147

3 .C1-PartitionisstronglyNP-complete. Proof. LetKdenotethesetofallO(n2)n-dimensionalbinaryvectorswhoseonesappearconsecutively,wherevkisthebinaryvectorcorrespondingtok2K.Consideraguessedsolutionthatconsistsofjj-dimensionalnonnegativeintegervectorsdk,8k2K,wheredk`denotesthenumberoftimesleafpositionvk,k2K,isassignedtointensity`2.Sincealldk`Linsomefeasiblesolution,werestricttheguessedd-vectorsassuch.ThesizeoftheguessedvectorsisthusO(n2jjlog2(L)).WecanverifywhetherornotPk2KP`2dk`vk=binO(n2jj)additions.Therefore,C1-PartitionisinNP. ToshowthatC1-PartitionisNP-complete,wereduce3-Partitiontoit.3-PartitionisastronglyNP-completeproblemandseekswhetheragivenmultisetofintegerscanbepartitionedintotripletshavingthesamesum.Formally,itcanbedenedasfollows(see GareyandJohnson ( 1979 )): QUESTION:CanAbepartitionedintodisjointmultisetsA1;:::;AsuchthatPj2Aiaj=Bfori=1;:::;? Givenanarbitraryinstanceof3-Partition,weconstructaninstanceforC1-Partitionasfollows.First,wedene^xtobeanintegervectorwhose`thcomponent,^x`,isequaltothenumberofindicesiforwhichai=`.Furthermore,weletbbea(21)-dimensional 147

PAGE 148

Assumethatthe3-Partitioninstanceisayes-instance,andhencethereexistmultisetsA1;:::;AsuchthatPj2Aiaj=B.Inthiscase,afeasiblesolutionoftheC1-Partitioninstanceletsd2j1;`bethenumberofelementsofintensity`inAj,foreachj=1;:::;,andassignsdk`=0forallotherk.Similarly,supposethattheC1-Partitioninstanceisayes-instance.Sinceallpositivevaluesinbareadjacentto0,inanyfeasiblesolutiontotheinstanceofC1-Partition,wemayonlyuseleafpositionsthatexposeasingleodd-indexbixel.Also,sinceB=4
PAGE 149

Agazaryan,N.,T.D.Solberg.2003.Segmentalanddynamicintensity-modulatedradiotherapydeliverytechniquesformicro-multileafcollimator.MedicalPhysics30(7)1758{1765. Ahuja,R.K.,H.W.Hamacher.2005.Anetworkowalgorithmtominimizebeam-on-timeforunconstrainedmultileafcollimatorproblemsincancerradiationtherapy.Networks45(1)36{41. Ahuja,R.K.,T.L.Magnanti,J.B.Orlin.1993.NetworkFlows:Theory,Algorithms,andApplications.PrenticeHall,UpperSaddleRiver,NJ. Aigner,M.,M.Fromme.1984.Agameofcopsandrobbers.DiscreteAppliedMathematics8(1)1{12. Alpern,S.1995.Therendezvoussearchproblem.SIAMJournalonControlandOptimiza-tion33(3)673{683. Alpern,S.2008.Hide-and-seekgamesonatreetowhichEuleriannetworksareattached.Networks52(3)162{166. Alpern,S.,S.Gal.2003.TheTheoryofSearchGamesandRendezvous,InternationalSeriesinOperationsResearch&ManagementScience,vol.55.KluwerAcademicPublishers,Dordrecht,TheNetherlands. Alspach,B.2004.Searchingandsweepinggraphs:abriefsurvey.LeMatematiche595{37. Alspach,B.,D.Dyer,D.Hanson,B.Yang.2008.Timeconstrainedgraphsearching.TheoreticalComputerScience399(3)158{168. Andreas,A.K.,J.C.Smith.2009.Decompositionalgorithmsforthedesignofanonsimultaneouscapacitatedevacuationtreenetwork.Networks53(2)91{103. Baatar,D.2005.Matrixdecompositionwithtimeandcardinalityobjectives:Theory,algorithms,andapplicationtomultileafcollimatorsequencing.Ph.D.thesis,DepartmentofMathematics,TechnischeUniversitatKaiserslautern,Kaiserslautern,Germany. Baatar,D.,N.Boland,S.Brand,P.J.Stuckey.2007.Minimumcardinalitymatrixdecompositionintoconsecutive-onesmatrices:CPandIPapproaches.IntegrationofAIandORTechniquesinConstraintProgrammingforCombinatorialOptimizationProblems.Berlin,1{15. Baatar,D.,H.W.Hamacher,M.Ehrgott,G.J.Woeginger.2005.Decompositionofintegermatricesandmultileafcollimatorsequencing.DiscreteAppliedMathematics152(1)6{34. 149

PAGE 150

Benders,J.F.1962.Partitioningproceduresforsolvingmixedvariablesprogrammingproblems.NumerischeMathematik4238{252. Benkoski,S.J.,M.G.Monticino,J.R.Weisinger.1991.Asurveyofthesearchtheoryliterature.NavalResearchLogistics38469{494. Bienstock,D.,P.Seymour.1991.Monotonicityingraphsearching.JournalofAlgorithms12(2)239{245. Bockmayr,A.,N.Pisaruk.2006.Detectinginfeasibilityandgeneratingcutsformixedintegerprogrammingusingconstraintprogramming.Computers&OperationsResearch33(10)2777{2786. Boland,N.,H.W.Hamacher,F.Lenzen.2004.Minimizingbeam-ontimeincancerradiationtreatmentusingmultileafcollimators.Networks43(4)226{240. Bortfeld,T.R.,D.L.Kahler,T.J.Waldron,A.L.Boyer.1994.X-rayeldcompensationwithmultileafcollimators.InternationalJournalofRadiationOncologyBiologyPhysics28(3)723{730. Calaore,G.C.,M.C.Campi.2005.Uncertainconvexprograms:Randomizedsolutionsandcondencelevels.MathematicalProgramming10225{46. Care,C.C.,R.Schultz.1999.Dualdecompositioninstochasticintegerprogramming.OperationsResearchLetters2437{45. Chevaleyre,Y.,F.Sempe,G.Ramalho.2004.Atheoreticalanalysisofmulti-agentpatrollingstrategies.AAMAS'04:ProceedingsoftheThirdInternationalJointConferenceonAutonomousAgentsandMultiagentSystems.IEEEComputerSociety,Washington,DC,USA,1524{1525. Codato,G.,M.Fischetti.2006.CombinatorialBenders'cutsformixed-integerlinearprogramming.OperationsResearch54(4)756{766. Dai,J.,Y.Hu.1999.Intensity-modulationradiotherapyusingindependentcollimators:analgorithmstudy.MedicalPhysics26(12)2562{2570. Dai,J.,Y.Zhu.2001.Minimizingthenumberofsegmentsinadeliverysequenceforintensity-modulatedradiationtherapywithamultileafcollimator.MedicalPhysics28(10)2113{2120. Dendris,N.D.,L.M.Kirousis,D.M.Thilikos.1997.Fugitive-searchgamesongraphsandrelatedparameters.TheoreticalComputerScience172233{254. 150

PAGE 151

Earl,M.A.,M.K.N.Afghan,C.X.Yu,Z.Jiang,D.M.Shepard.2007.Jaws-onlyIMRTusingdirectapertureoptimization.MedicalPhysics34(1)307{314. Ehrgott,M.,H.W.Hamacher,M.Nubaum.2008.Decompositionofmatricesandstaticmultileafcollimators:Asurvey.C.J.S.Alves,P.M.Pardalos,L.N.Vicente,eds.,OptimizationinMedicine.Springer,NewYork,NY,25{46. Ellis,J.A.,I.H.Sudborough,J.S.Turner.1994.Thevertexseparationandsearchnumberofagraph.InformationandComputation113(1)50{79. Engel,K.2005.Anewalgorithmforoptimalmultileafcollimatoreldsegmentation.DiscreteAppliedMathematics152(1)35{51. Ernst,A.T.,V.H.Mak,L.A.Mason.2009.AnexactmethodfortheminimumcardinalityproblemintheplanningofIMRT.INFORMSJournalonComputingForthcoming. Flocchini,P.,M.J.Huang,F.L.Luccio.2008.Decontaminationofhypercubesbymobileagents.Networks52(3)167{178. Fomin,F.V.,D.M.Thilikos.2008.Anannotatedbibliographyonguaranteedgraphsearching.TheoreticalComputerScience399(3)236{245. Garey,M.R.,D.S.Johnson.1979.ComputersandIntractability;AGuidetotheTheoryofNP-Completeness.W.H.Freeman&Co.,SanFrancisco,CA. Goldschmidt,O.,D.S.Hochbaum,A.Levin,E.V.Olinick.2003.TheSONETedge-partitionproblem.Networks4113{23. Goldstein,A.S.,E.M.Reingold.1995.Thecomplexityofpursuitonagraph.TheoreticalComputerScience143(1)93{112. Hahn,G.2007.Cops,robbersandgraphs.TatraMountainsMathematicalPublications361{14. Hamacher,H.W.,K.-H.Kufer.2002.Inverseradiationtherapyplanning{amultipleobjectiveoptimizationapproach.DiscreteAppliedMathematics118(1)145{161. Haralick,R.M.,G.L.Elliott.1980.Increasingtreesearcheciencyforconstraintsatisfactionproblems.ArticialIntelligence14(3)263{313. Hooker,J.N.2007.Planningandschedulingbylogic-basedBendersdecomposition.OperationsResearch55588{602. Hooker,J.N.,G.Ottosson.2003.Logic-basedBendersdecomposition.MathematicalProgramming9633{60. 151

PAGE 152

Isler,V.,N.Karnad.2008.Theroleofinformationinthecop-robbergame.TheoreticalComputerScience399(3)179{190. Jain,V.,I.E.Grossmann.2001.AlgorithmsforhybridMILP/CPmodelsforaclassofoptimizationproblems.INFORMSJournalonComputing13258{276. Jotshi,A.,R.Batta.2008.Searchforanimmobileentityonanetwork.EuropeanJournalofOperationalResearch191(2)347{359. Junger,M.,S.Thienel.2000.TheABACUSsystemforbranch-and-cut-and-pricealgorithmsinintegerprogrammingandcombinatorialoptimization.SoftwarePrac-ticeandExperience30(11)1325{1352. Kalinowski,T.2004.Thealgorithmiccomplexityoftheminimizationofthenumberofsegmentsinmultileafcollimatoreldsegmentation.Tech.rep.,FachbereichMathematik,UniversitatRostock,Rostock,Germany. Kalinowski,T.2005a.Adualitybasedalgorithmformultileafcollimatoreldsegmentationwithinterleafcollisionconstraint.DiscreteAppliedMathematics152(1)52{88. Kalinowski,T.2005b.Reducingthenumberofmonitorunitsinmultileafcollimatoreldsegmentation.PhysicsinMedicineandBiology50(6)1147{1161. Kamath,S.,S.Sahni,J.Li,J.Palta,S.Ranka.2003.Leafsequencingalgorithmsforsegmentedmultileafcollimation.PhysicsinMedicineandBiology48(3)307{324. Kamath,S.,S.Sahni,J.Palta,S.Ranka.2004a.Algorithmsforoptimalsequencingofdynamicmultileafcollimators.PhysicsinMedicineandBiology49(1)33{54. Kamath,S.,S.Sahni,J.Palta,S.Ranka,J.Li.2004b.Optimalleafsequencingwitheliminationoftongue-and-grooveunderdosage.PhysicsinMedicineandBiology49(3)N7{N19. Kamath,S.,S.Sahni,S.Ranka,J.Li,J.Palta.2004c.Acomparisonofstep-and-shootleafsequencingalgorithmsthateliminatetongue-and-grooveeects.PhysicsinMedicineandBiology49(14)3137{3143. Kamath,S.,S.Sahni,S.Ranka,J.Li,J.Palta.2004d.Optimaleldsplittingforlargeintensity-modulatedelds.MedicalPhysics31(12)3314{3323. Kikuta,K.,W.H.Ruckle.2007.Rendezvoussearchonastargraphwithexaminationcosts.EuropeanJournalofOperationalResearch181(1)298{304. 152

PAGE 153

Kufer,K.-H.,M.Monz,A.Scherrer,H.L.Trinkaus,T.Bortfeld,C.Thieke.2003.Intensitymodulatedradiotherapy{alargescalemulti-criteriaprogrammingproblem.ORSpektrum25223{249. Langer,M.,V.Thai,L.Papiez.2001.ImprovedleafsequencingreducessegmentsormonitorunitsneededtodeliverIMRTusingmultileafcollimators.MedicalPhysics28(12)2450{2458. LaPaugh,A.S.1993.Recontaminationdoesnothelptosearchagraph.JournaloftheACM40(2)224{245. Laporte,G.,F.V.Louveaux.1993.TheintegerL-shapedmethodforstochasticintegerprogramswithcompleterecourse.OperationsResearchLetters13133{142. Lee,E.K.,T.Fox,I.Crocker.2000a.Optimizationofradiosurgerytreatmentplanningviamixedintegerprogramming.MedicalPhysics27(5)995{1004. Lee,E.K.,T.Fox,I.Crocker.2003.Integerprogrammingappliedtointensity-modulatedradiationtreatmentplanning.AnnalsofOperationsResearch119165{181. Lee,Y.,H.D.Sherali,J.Han,S.Kim.2000b.Abranch-and-cutalgorithmforsolvinganintra-ringsynchronousopticalnetworkdesignproblem.Networks35(3)223{232. Lenzen,F.2000.Anintegerprogrammingapproachtothemultileafcollimatorproblem.Master'sthesis,DepartmentofMathematics,UniversityofKaiserslautern,Kaiserslautern,Germany. Lim,G.J.2002.Optimizationinradiationtreatmentplanning.Ph.D.thesis,ComputerSciencesDepartment,UniversityofWisconsin-Madison,Madison,WI. Lim,G.J.,J.Choi.2007.Atwo-stageintegerprogrammingapproachforoptimizingleafsequenceinIMRT.Tech.rep.,DepartmentofIndustrialEngineering,UniversityofHouston,Houston,TX. Lim,G.J.,M.C.Ferris,D.M.Shepard.2004.OptimizationtoolsforradiationtreatmentplanninginMatlab.M.L.Brandeau,F.Sainfort,W.P.Pierskalla,eds.,OperationsResearchandHealthCare:AHandbookofMethodsandApplications.KluwerAcademicPublishers,Boston,MA,775{806. Lim,G.J.,M.C.Ferris,S.J.Wright,D.M.Shepard,M.A.Earl.2007.Anoptimizationframeworkforconformalradiationtreatmentplanning.INFORMSJournalonComput-ing19(3)366{380. Luedtke,J.,S.Ahmed.2008.Asampleapproximationapproachforoptimizationwithprobabilisticconstraints.SIAMJournalonOptimization19674{699. 153

PAGE 154

Marchand,H.,A.Martin,R.Weismantel,L.Wolsey.2002.Cuttingplanesinintegerandmixedintegerprogramming.DiscreteAppliedMathematics123397{446. Megiddo,N.,S.L.Hakimi,M.R.Garey,D.S.Johnson,C.H.Papadimitriou.1988.Thecomplexityofsearchingagraph.JournaloftheACM35(1)18{44. Men,C.,H.E.Romeijn,Z.C.Taskn,J.F.Dempsey.2007.AnexactapproachtodirectapertureoptimizationinIMRTtreatmentplanning.PhysicsinMedicineandBiology52(24)7333{7352. Migdalas,A.,P.M.Pardalos.1996.Hierarchicalandbilevelprogramming.JournalofGlobalOptimization8(3). Migdalas,A.,P.M.Pardalos,P.Varbrand,eds.1997.MultilevelOptimization:AlgorithmsandApplications.KluwerAcademicPublishers. Nemhauser,G.L.,L.A.Wolsey.1988.IntegerandCombinatorialOptimization.JohnWiley&Sons,NewYork,NY. Nemirovski,A.,A.Shapiro.2005.Scenarioapproximationofchanceconstraints.G.Calaore,F.Dabbene,eds.,ProbabilisticandRandomizedMethodsforDesignUnderUncertainty.Springer,London,3{48. Ntaimo,L.,S.Sen.2005.Themillion-variable\march"forstochasticcombinatorialoptimization.JournalofGlobalOptimization32(3)385{400. Parsons,T.D.1978.Pursuit-evasioninagraph.TheoryandApplicationsofGraphs.Springer,Berlin,426{441. Peng,S.,C.Ho,T.Hsu,M.Ko,C.Y.Tang.2000.Edgeandnodesearchingproblemsontrees.TheoreticalComputerScience240(2)429{446. Penuel,J.,J.C.Smith.2009.Modelsandcomplexityanalysisforthegraphdecontaminationproblemwithmobileagents.Tech.rep.,DepartmentofIndustrialandSystemsEngineering,UniversityofFlorida,Gainesville,FL. Preciado-Walters,F.,R.Rardin,M.Langer,V.Thai.2004.Acoupledcolumngeneration,mixedintegerapproachtooptimalplanningofintensitymodulatedradiationtherapyforcancer.MathematicalProgramming101(2)319{338. Que,W.1999.Comparisonofalgorithmsformultileafcollimatoreldsegmentation.MedicalPhysics26(11)2390{2396. Que,W.,J.Kung,J.Dai.2004.\Tongue-and-groove"eectinintensitymodulatedradiotherapywithstaticmultileafcollimatorelds.PhysicsinMedicineandBiology49(3)399{405. 154

PAGE 155

Romeijn,H.E.,R.K.Ahuja,J.F.Dempsey,A.Kumar.2006.Anewlinearprogrammingapproachtoradiationtherapytreatmentplanningproblems.OperationsResearch54(2)201{216. Rossi,F.,P.vanBeek,T.Walsh,eds.2006.HandbookofConstraintProgramming.Elsevier. Sak,T.,J.W.,S.K.Goldenstein.2008.Probabilisticmultiagentpatrolling.AdvancesinArticialIntelligence{SBIA2008.Springer,Berlin,124{133. Sen,S.,J.L.Higle.2005.TheC3theoremandaD2algorithmforlargescalestochasticintegerprogramming:Setconvexication.MathematicalProgramming1041{20. Seymour,P.,R.Thomas.1993.Graphsearchingandamin-maxtheoremfortree-width.JounalofCombinatorialTheory,SeriesB58(1)22{33. Shepard,D.M.,M.A.Earl,X.A.Li,S.Naqvi,C.Yu.2002.Directapertureoptimization:aturnkeysolutionforstep-and-shootIMRT.MedicalPhysics29(6)1007{1018. Shepard,D.M.,M.C.Ferris,G.H.Olivera,T.R.Mackie.1999.Optimizingthedeliveryofradiationtherapytocancerpatients.SIAMReview41(4)721{744. Sherali,H.D.,B.M.P.Fraticelli.2002.AmodicationofBenders'decompositionalgorithmfordiscretesubproblems:anapproachforstochasticprogramswithintegerrecourse.JournalofGlobalOptimization22319{342. Sherali,H.D.,J.C.Smith.2001.Improvingzero-onemodelrepresentationsviasymmetryconsiderations.ManagementScience47(10)1396{1407. Sherali,H.D.,J.C.Smith,Y.Lee.2000.Enhancedmodelrepresentationsforanintra-ringsynchronousopticalnetworkdesignproblemallowingdemandsplitting.INFORMSJournalonComputing12(4)284{298. Siek,J.G.,L.Lee,A.Lumsdaine.2001.TheBoostGraphLibrary:UserGuideandReferenceManual.Addison-Wesley. Siochi,R.A.C.1999.Minimizingstaticintensitymodulationdeliverytimeusinganintensitysolidparadigm.InternationalJournalofRadiationOncologyBiologyPhysics43(3)671{680. Siochi,R.A.C.2004.ModicationstotheIMFASTleafsequencingoptimizationalgorithm.MedicalPhysics31(12)3267{3278. Siochi,R.A.C.2007.Variabledepthrecursionalgorithmforleafsequencing.MedicalPhysics34(2)664{672. 155

PAGE 156

Smith,B.M.2005.Symmetryandsearchinanetworkdesignproblem.RomanBartak,MichelaMilano,eds.,IntegrationofAIandORTechniquesinConstraintProgrammingforCombinatorialOptimizationProblems.Springer,Berlin,336{350. Smith,J.C.2004.Algorithmsfordistributingtelecommunicationtraconamultiple-ringSONET-basednetwork.EuropeanJournalofOperationalResearch154(3)659{672. Smith,J.C.,C.Lim,A.Alptekinoglu.2008.Optimalmixed-integerprogrammingandheuristicmethodsforabilevelStackelbergproductintroductiongame.Tech.rep.,DepartmentofIndustrialandSystemsEngineering,UniversityofFlorida,Gainesville,FL. Smith,J.C.,A.J.Schaefer,J.W.Yen.2004.Astochasticintegerprogrammingapproachtosolvingasynchronousopticalnetworkringdesignproblem.Networks44(1)12{26. Stell,A.M.,O.A.Zeidan,J.G.Li,J.F.Dempsey.2004.Anextensivelog-leanalysisofstep-and-shootIMRTsubelddeliveryerrors.InternationalJournalofRadiationOncologyBiologyPhysics31(6)1593{1602. Sutter,A.,F.Vanderbeck,L.A.Wolsey.1998.Optimalplacementofadd/dropmultiplexers:Heuristicandexactalgorithms.OperationsResearch46(5)719{728. Taskn,Z.C.,J.C.Smith,S.Ahmed,A.J.Schaefer.2009a.Cuttingplanealgorithmsforsolvingastochasticedge-partitionproblem.DiscreteOptimizationForthcoming. Taskn,Z.C.,J.C.Smith,H.E.Romeijn.2008.Mixed-integerprogrammingtechniquesfordecomposingIMRTuencemapsusingrectangularapertures.Tech.rep.,DepartmentofIndustrialandSystemsEngineering,UniversityofFlorida,Gainesville,FL. Taskn,Z.C.,J.C.Smith,H.E.Romeijn,J.F.Dempsey.2009b.OptimalmultileafcollimatorleafsequencinginIMRTtreatmentplanning.OperationsResearchForthcoming. Thorsteinsson,E.S.2001.Branch-and-check:Ahybridframeworkintegratingmixedintegerprogrammingandconstraintlogicprogramming.TobyWalsh,ed.,CP'01:Proceedingsofthe7thInternationalConferenceonPrinciplesandPracticeofConstraintProgramming.Springer,16{30. vanBeek,P.2006.Backtrackingsearchalgorithms.F.Rossi,P.vanBeek,T.Walsh,eds.,HandbookofConstraintProgramming.Elsevier,85{134. VanSantvoort,J.P.C.,B.J.M.Heijmen.1996.Dynamicmultileafcollimationwithout\tongue-and-groove"underdosageeects.PhysicsinMedicineandBiology41(10)2091{2105. 156

PAGE 157

Wolsey,LA.1998.IntegerProgramming.Wiley-Interscience,NewYork,NY. Xia,P.,L.J.Verhey.1998.Multileafcollimatorleafsequencingalgorithmforintensitymodulatedbeamswithmultiplestaticsegments.MedicalPhysics25(8)1424{1434. 157

PAGE 158

Z.CanerTasknwasborninBalkesir,TurkeyonSeptember8,1981.HeearnedhisB.S.andM.S.degreesinIndustrialEngineeringfromBogaziciUniversity,_Istanbulin2003and2005,respectively.Beforestartinghisdoctoratestudy,heworkedforICRONTechnologiesasaproductconsultant,wherehetookroleinadvancedplanningandschedulingprojectsforcustomersinseveralindustriesincludingsteel,automotive,electronicsandglassmanufacturingindustries.HewillnishhisPh.D.degreeinIndustrialandSystemsEngineeringattheUniversityofFloridainAugust2009.Followinggraduation,hewilljoinDepartmentofIndustrialEngineeringatBogaziciUniversityasafacultymember. 158