<%BANNER%>

A Hybrid Heuristcs and Simulation-Based Approach to Decision Support for Robot-Human Team Configuration

Permanent Link: http://ufdc.ufl.edu/UFE0042856/00001

Material Information

Title: A Hybrid Heuristcs and Simulation-Based Approach to Decision Support for Robot-Human Team Configuration
Physical Description: 1 online resource (105 p.)
Language: english
Creator: NIETEN,TERESA LILLIAN
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: HEURISTICS -- PLANNING -- PROGRESSIVE -- REFINEMENT -- ROBOTICS -- SIMULATION -- TEAMWORK
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Missions that involve tasks such as search-and-rescue or reconnaissance have traditionally involved humans, perhaps with the assistance of one or more robots. The robots, or unmanned systems, are typically teleoperated?operated by remote control to inspect a suspicious object, for example. With the advent of newer and less expensive forms of autonomy and improved human-robot communication, the robots are becoming more capable of acting as peers to their human counterparts, rather than just tools. As the diversity of mixed human-robot teams is increased, so is the complexity of trying to answer questions regarding configuration: what robots should be used, how many, and how many humans should be employed in the teaming process? This paper presents the research in search of that hybrid approach. Our solution is a decision support approach that employs a hybrid of simulation and artificial intelligence techniques, using a progressive-refinement queuing model to quickly bypass the least desirable configurations. This approach, realized in a software tool, considers the mission requirements and a priori data to determine the optimal team to perform that mission, given a pool of human and robotic resources.
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 TERESA LILLIAN NIETEN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Fishwick, Paul A.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042856:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042856/00001

Material Information

Title: A Hybrid Heuristcs and Simulation-Based Approach to Decision Support for Robot-Human Team Configuration
Physical Description: 1 online resource (105 p.)
Language: english
Creator: NIETEN,TERESA LILLIAN
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: HEURISTICS -- PLANNING -- PROGRESSIVE -- REFINEMENT -- ROBOTICS -- SIMULATION -- TEAMWORK
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Missions that involve tasks such as search-and-rescue or reconnaissance have traditionally involved humans, perhaps with the assistance of one or more robots. The robots, or unmanned systems, are typically teleoperated?operated by remote control to inspect a suspicious object, for example. With the advent of newer and less expensive forms of autonomy and improved human-robot communication, the robots are becoming more capable of acting as peers to their human counterparts, rather than just tools. As the diversity of mixed human-robot teams is increased, so is the complexity of trying to answer questions regarding configuration: what robots should be used, how many, and how many humans should be employed in the teaming process? This paper presents the research in search of that hybrid approach. Our solution is a decision support approach that employs a hybrid of simulation and artificial intelligence techniques, using a progressive-refinement queuing model to quickly bypass the least desirable configurations. This approach, realized in a software tool, considers the mission requirements and a priori data to determine the optimal team to perform that mission, given a pool of human and robotic resources.
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 TERESA LILLIAN NIETEN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Fishwick, Paul A.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042856:00001


This item has the following downloads:


Full Text

PAGE 1

1 A HYBRID HEURISTICS AND SIMULATION BASED APPROACH TO DECISION SUPPORT FOR ROBOT HUMAN TEAM CONFIGURATION By TERESA NIETEN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMEN T OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2011

PAGE 2

2 2011 Teresa Nieten

PAGE 3

3 To Dan and Brandon

PAGE 4

4 ACKNOWLEDGMENTS I thank my husband, Dan, for his love, support and encouragement; my son Brandon for giving m e lots of encouraging hugs and for not driving me completely batty; and my parents, Tom and Diana Griffith, my sister Vickie, and mother in law Lois Lastinger for their support, encouragement, and occasional pep talks. My brother in law Joe Nieten and my dear friends Dave and Christine Langhorne graciously volunteered to proofread for me. I would also like to thank Dr. Paul Fishwick for not giving up on me. This work was sponsored, in part, by the United States ( US ) Army Research Laboratory under Cooper ative Agreement W911NF 062 0041. Parts of t his work were conducted using the Protg resource, which is supported by grant LM007885 from the United States National Library of Medicine.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLED GMENTS .................................................................................................. 4 LIST OF TABLES ............................................................................................................ 7 LIST OF FIGURES .......................................................................................................... 8 LIST OF ABBREVIATION S ........................................................................................... 10 ABSTRACT ................................................................................................................... 11 C HAPTER 1 INTRODUCTION .................................................................................................... 12 Background ............................................................................................................. 13 Problem Statement ................................................................................................. 19 Solution ................................................................................................................... 19 Contributions to Knowledge .................................................................................... 20 2 RELATED RESEARCH .......................................................................................... 21 Overview ................................................................................................................. 21 Teamwork between Humans and Unmanned S ystems .................................... 21 Team Planning and Scheduling ........................................................................ 25 Mission and Path Planning ............................................................................... 27 Simulation Techniques ..................................................................................... 29 What are the Missing Pieces? ................................................................................ 32 Innovation/Value Added ................................................................................... 33 Contributions .................................................................................................... 35 3 PROTOTYPE .......................................................................................................... 41 Ontology ................................................................................................................. 41 Framework .............................................................................................................. 45 Resource Parameters ...................................................................................... 45 Area of Interest ................................................................................................. 50 In itial Path Selection ......................................................................................... 54 Progressive Refinement ................................................................................... 63 Simulation ......................................................................................................... 66 Scoring Algorithms ........................................................................................... 72 Scenario .................................................................................................................. 73

PAGE 6

6 4 ANALYSIS .............................................................................................................. 75 Comparison ............................................................................................................ 75 Testing Environment ............................................................................................... 75 Scenarios ................................................................................................................ 76 Simulation Based Progressive R efinement, the Early Test .............................. 76 Mission ....................................................................................................... 76 Results ....................................................................................................... 78 Winning team and path .............................................................................. 78 Second place team and path ..................................................................... 78 Final Prototype, Scenario One ......................................................................... 79 Mission ....................................................................................................... 81 Winning scenario, time ............................................................................... 81 Second place scenario, time ...................................................................... 8 1 Winning scenario, cost ............................................................................... 82 Second place team, cost ............................................................................ 82 Results, low ceiling .................................................................................... 83 Results, higher ceiling ................................................................................ 84 Results, all paths ........................................................................................ 84 Final Prototype, Scenario Two ......................................................................... 84 Mission ....................................................................................................... 85 Winning scenario, time ............................................................................... 85 Second place scenario, time ...................................................................... 85 Winning scenario, cost ............................................................................... 87 Second place team, cost ............................................................................ 87 Results, lower ceiling ................................................................................. 87 Results, higher ceiling ................................................................................ 88 Results, All Paths ....................................................................................... 89 Final Prototype, Scenario Three ....................................................................... 89 Mission ....................................................................................................... 90 Winning scenario, time and cost ................................................................ 90 Second place scenario, time, and cost second place team ....................... 90 Results, lower ceiling ................................................................................. 91 Results, higher ceiling ................................................................................ 92 Results, All Paths ....................................................................................... 92 5 CONCLUSIONS ..................................................................................................... 95 Effectiveness of Hybrid Approach ........................................................................... 95 Contributions ........................................................................................................... 96 Future Work ............................................................................................................ 96 LIST OF REFERENCES ............................................................................................... 99 BIOGRAPHICAL SKETCH .......................................................................................... 105

PAGE 7

7 LIST OF TABLES Table page 3 1 Properties ........................................................................................................... 44 3 2 Resource Parameters ......................................................................................... 48 3 3 Sample Formulas ............................................................................................... 73 4 1 Resources available in initial protot ype .............................................................. 77 4 2 Resources available in scenario 12 ................................................................... 80 4 3 Evaluation of approaches, scenario 1 ................................................................. 81 4 4 Evaluation of approaches, scenario 2 ................................................................. 85 4 5 Evaluation of approaches, scenario 3 ................................................................. 90 4 6 Resour ces available in scenario 3 ...................................................................... 94

PAGE 8

8 LIST OF FIGURES Figure page 1 1 Rescue robots used in aftermath of September 11 terrorist attacks ................... 14 1 2 Robot in search and rescue operation ................................................................ 14 1 3 Robots can now take on the role of canary in a coal mine ................................. 15 1 4 Rescue robot carries an injured soldier .............................................................. 15 1 5 Another possible rescue robot that can carry the injured party to safety ............ 16 1 6 Robot assists victim in rescue exercise .............................................................. 16 1 7 Pairing humans and robots allows the robots to do the dangerous jobs. ............ 18 1 8 A team of soldiers and robot on a mission .......................................................... 18 2 1 Teamwork selection to find the optimal team and path combination. ................. 35 2 2 Computing Classification System (CCS) Contributions ...................................... 39 2 3 CCS Subjects of Robotics .................................................................................. 40 3 1 High level arch itecture of prototype .................................................................... 41 3 2 OWL class visualization ...................................................................................... 42 3 3 XML Sample ....................................................................................................... 43 3 4 Entry panel for resource candidate ..................................................................... 46 3 5 Class diagram for Resources ............................................................................. 47 3 6 SPARQL query for team selection ...................................................................... 49 3 7 Diagram of a city area ........................................................................................ 50 3 8 Search team in an urban area ............................................................................ 51 3 9 RDF relations of a named location ..................................................................... 52 3 10 SPARQL results for subject NW_Area1 ............................................................. 52 3 11 Class diagram for map related classes .............................................................. 53 3 12 RDF graph of NW Block, leaf level.. ................................................................... 56

PAGE 9

9 3 13 RDF graph of NW Block, level 2. ........................................................................ 57 3 14 Traversal graph for path planning ....................................................................... 58 3 15 Class diagram for path selection ........................................................................ 58 3 16 Sequence Diagram for path search. ................................................................... 61 3 17 Flow diagram for capped depth first search ........................................................ 62 3 18 Parti al scene graph layout of area ...................................................................... 64 3 19 Class diagram for TeamPlanner and Team classes. .......................................... 65 3 20 Sequence diagram of the progressive simulation ............................................... 67 3 21 Sequence diagram for TeamEvaluation ............................................................. 69 3 22 Simulation of individual path for a single team configuration .............................. 70 3 23 A path through the area at different levels.. ........................................................ 71 3 24 Scene graph representation of queuing model. .................................................. 72 3 25 Diagram of a mission scenario with suggested paths. ........................................ 74 4 1 Mission Area with goals ...................................................................................... 77 4 2 First place team/path .......................................................................................... 78 4 3 Second place team/path ..................................................................................... 79 4 4 Scenario 1 layout. ............................................................................................... 82 4 5 Scenario 1 winning path. .................................................................................... 83 4 6 Layout for scenario 2. ......................................................................................... 86 4 7 Winning path for scenario 2. ............................................................................... 86 4 8 Second place path for scenario 2. ...................................................................... 87 4 9 Winning path, scenario 3 .................................................................................... 91

PAGE 10

10 LIST OF ABBREVIATION S AI Artificial Intelligence CCS Computing Classi fication System OWL Web Ontology Language RDF Resource Definition Framework RDFS RDF Schema SPARQL SPARQL Protocol And RDF Query Language UAV Unmanned Air Vehicle UGV Unmanned Ground Vehicle UMS Unmanned System USAR Urban Search And Rescue WRP Weighted Reg ion Problem XML eXtensible Markup Language

PAGE 11

11 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 A HYBRID HEURISTIC AND SIMULATION BASED APPROACH TO DECISION SUPPORT FOR ROBOT HUMAN TEAM CONFIGURATION By Teresa Nieten April 2011 Chair: PAUL FISHWICK Major: Computer Engineering Missions that involve tasks such as searchand rescue or reconnaissance have traditionally inv olved humans, perhaps with the assistance of one or more robots. The robots, or unmanned systems, are typically teleoperated operated by remote control to inspect a suspicious object, for example. With the advent of newer and less expensive forms of autonomy and improved humanrobot communication, the robots are becoming more capable of acting as peers to their human counterparts rather than just tools. As the diversity of mixed humanrobot teams is increased, so is the complexity of trying to answer quest ions regarding configuration: what robots should be used, how many, and how many humans should be employed in the teaming process? This paper presents the research in search of that hybrid approach. Our solution is a decision support approach that employ s a hybrid of simulation and artificial intelligence techniques, using a progressiverefinement queuing model to quickly bypass the least desirable configurations. This approach, realized in a software tool, considers the mission requirements and a priori data to determine the optimal team to perform that mission, given a pool of human and robotic resources.

PAGE 12

12 CHAPTER 1 INTRODUCTION Teamwork is the ability to work together toward a common vision. The ability to direct individual accomplishments toward org anizational objectives. It is the fuel that allows common people to attain uncommon results -Andrew Carnegie [1] Complex tasks often require more effort than can be put forth by a single person, and the diversity in the sk ills and abilities of the members allows for tasks to be divided into specializations, rather than being replicated. This division of labor works to make an effective, synergistic team provided the team collectively has the abilities and tools required to do the job. In a football game, if an offensive lineman does not perform his job, the quarterback could get sacked, and the team could lose out on the opportunity to score a touchdown. If a scrub nurse does not know how to identify surgical instrument s in an operating room, the patient could die. Tchaikovskys 1812 Overture would end with a whimper instead of a bang if the bass drums were substituted with triangles. Whether it is a symphony, an assembly line, an operating room, a search and rescue operation, a sports team, or a military patrol, the success of the task depends on each member performing his or her assigned duties. Just as crucial to the performance of a team is the preparation that goes into creating the team. No matter how well a gr oup works together, if they collectively do not have the basic skills required to do the job, their effectiveness is limited and the probability of the successful completion of the job is significantly reduced. A college football team could recruit the best quarterback and receivers in the country, but if they neglect to field a center during an offensive play, the team will not be successful. Likewise, an operating room full of surgeons with no anesthesiologist will not yield a

PAGE 13

13 successful operation. F or a quartet to harmonize correctly, it must include a lead, tenor, bass, and baritone. In a search and rescue operation, members of the team must be able to locate victims, assess the safety of the rescue area, and assist victims in leaving the rescue a rea. As man has evolved, he has learned to use tools of increasing complexity. From sharpening sticks to chipping flint to forging steel to developing complex machinery, people have become proficient at crafting tools to help get the job done and as som e of these tools themselves have become increasingly complex, we are arriving at a point that we can now consider some of these tools to be more than just tools. A subset of the tools has now evolved and matured enough that we can now elevate them to the status of individual contributors to the team effort unique, self sustaining team members, working alongside their human teammates to accomplish the task. We know them as either robots or unmanned systems (UMS). Background On a clear Fall day in 2001, a disaster of unthinkable proportions occurred. The twin 110story towers of the World Trade Center (WTC) were destroyed as a pair of hijacked airplanes struck, seventeen minutes apart. At the time of the attacks, there were roughly 17,000 people inside the towers [2]. While many of those people were able to evacuate, thousands more were left trapped in the rubble, including nearly 3000 dead. Among those killed were 411 emergency response workers [3] who were trying to come to the aid of those trapped and injured inside the rubble of the buildings. S ubsequently, s tudies have shown that the dust and pollutants that filled the air in the wake of the attacks have caused a significant increase in respiratory problems in first responders [4] causing problems to surface well after the initial danger had passed.

PAGE 14

14 Some of the rescuers, however, were unaffected by the dust, danger of collapse, or tight spaces. In what is touted to be the first time, a team of rescue robots (Figure 1 1 and 12) was deployed to help search for victims, identify evacuation paths, inspect for structural issues, and detect hazardous materials [5]. Figure 11 Rescue robots used in aftermath of September 11 terrorist attacks (credit: CRASAR) Figure 12. Robot in search and rescue operation

PAGE 15

15 Rescue robots were also used in the aftermath of Hurricane Katrina [6], and have been use d in several mine disaster rescue attempts, including the Pike River mine in New Zealand and Crandall Canyon, Utah [7] By adding robots to the team in a rescue operation, the robots can be sent in ahead of the human members t o sense the environment and detect hazardous conditions that may exist (Figure 13). Robots can also be useful in helping victims get to safety (Figure 14, Figure 15) or even performing first aid (Figure 16). Figure 13. Robots can now take on the r ole of canary in a coal mine during a mine disaster recovery operation. Figure 14. Rescue robot carries an injured soldier (credit army.mil)

PAGE 16

16 Figure 15. Another possible rescue robot that can carry the injured party to safety Figure 16. Robot assists victim in rescue exercise (credit: CRASAR, University of South Florida) Also in 2001, Congress passed a law setting goals for the US Military to have onethird of its operational deep strike force aircraft fleet be unmanned by 2010 and onethird of its operational combat ground vehicles be unmanned by 2015[9]. The ultimate goal is to save the lives of troops by using as many UMS in hazardous situations as possible, allowing the human soldiers to stay in protected areas while the more expendable robotic counterparts neutralize or at least identify the risks much like the goal of the Mine Safety and Health Administration (MSHA) and CRASAR with their mine rescue and recovery programs [7] An a dditional goal set forth by the Joint Robotics

PAGE 17

17 Program (JRP) is to reduce the ratio of human operators to robotic platforms by increasing autonomy levels of unmanned systems [10] thus encouraging the creation o f more autonomous or semi autonomous robots that can operate alongside of human members of the team. The U.S. Armys Army Model and Simulation Office (AMSO) developed a program to quickly and efficiently request and evaluate proposals for modeling and sim ulation research in an urban environment [8] One way to meet the goals set by Congress and the JRP, and to help safeguard humans, whether they are military personnel, rescuers, or people trying to accomplish potentially dang erous tasks such as mining or farming, is to create teams comprised of humans and robots (Figure 1 7 and Figure 18). A mixed team would allow the robots to perform the more dangerous tasks while the human counterparts provide in situ decisions that humans are uncomfortable allowing machines to make, such as authorizing weapons fire or identifying a target. As the autonomy level of unmanned systems increases, the ratio of unmanned vehicles to humans can increase, including having some teams fully composed of unmanned systems. The WTC search and rescue operation showed that this is possible, but was just a first step in having cooperative teams of humans and robots. In a military unit, as well as with most nonmilitary teams, there is an established chain o f command, with one team member dividing tasks among the entire team. As teams become more heavily unmanned, the tasking becomes more complex especially for units composed entirely of unmanned members and the availability of a wider variety of unmanned r esources available for selection. Additionally, as humans and unmanned vehicles are integrated, there is a need for consistent training of human team leads to

PAGE 18

18 learn how to effectively use all members of the team on a mission and, just as important, how to choose the correct members Determining the makeup of a team in an objective manner, based on the overall goals and success criteria of the task, is crucial in obtaining a successful outcome. Figure 17. Pairing humans and robots allows the robots to do the dangerous jobs. Figure 18. A team of soldiers and robot on a search and rescue or reconnaissance mission

PAGE 19

19 Problem Statement There are many factors to consider when choosing members of a team. Having the skills needed to perform one or more of t he tasks within the mission is crucial; each member must physically be able to travel to and from the area of interest, either independently or be transported by another member of the team; in addition, other less obvious factors come into play, such as cost to deploy or replace. Some robots can operate autonomously, semi autonomously, or be remotely operated; still others require operators or technicians to be deployed as part of the team due to maintenance, operability, transportability, and other colloc ation requirements. A given task may require an expertise that can only be developed with extensive and expensive training. These factors all impact the cost, including both financial and availability for other tasks, of assigning that resource to a team. With a large field of resources, both human and robot, and many variables to consider, the task of choosing the optimal configuration, whether that is time, cost, or some other combination, becomes more daunting. Continuous resource planning and replanning becomes a tedious task, one that is well suited to be automated, or at least semi automated. Solution The purpose of this research wa s to create a methodology, realized in a software tool, which considers the mission requirements and a priori data to determine the optimal team to perform that mission, given a pool of human and robotic resources. This method employ s a hybridized approach of simulation and artificial intelligence techniques, using a dynamic progressiverefinement queuing model t o quickly bypass the least desirable configurations. Th e resulting tool can be used during mission planning to allocate the appropriate resources, or in training to help team leads be

PAGE 20

20 consistent in choosing team members; it can be used prior to the operati on or to dynamically reallocate members of the team as situations arise requiring retasking or refocusing. Contributions to Knowledge The contributions from this research are in two areas 1) the hybridization of a heuristics based semi optimal path pl anning and multiple simulations in the presence of two input sets, and 2) workgroup selection automation. The goal of contribution # 1 is to demonstrate that a hybrid progressive refinement simulation and heuristic based planning method finds the desired solution, in this case a near optimal team configuration, faster and using fewer computational resources than either a standalone allpaths progressiverefinement simulation or a heuristic based planning paired with brute force nonrefined simulation. The goal of contribution # 2 is to delineate a clear planning approach to near optimal team based composition where teams have human and robot members with diverse skill sets, operational characteristics, and operating costs.

PAGE 21

21 CHAPTER 2 RELATED RESEARCH Ove rview The research related to this work spans several areas teamwork between humans and UMS, team planning and scheduling, path and mission planning, and simulation. The following sections discuss the relevant related work done in those areas. First, w e will look at research involving the pairing of humans and unmanned systems. Next we will investigate work done in the area of team planning and scheduling, followed by research into mission and path planning. Finally, we will focus on simulation techni ques, specifically where they apply to decision making. Teamwork between H umans and U nmanned Systems There are several groups performing research into humans and UMS working together to solve a problem or complete a task. The Idaho National Engineering and Environmental Laboratory has performed research on developing a control architecture that allows different members of a mixed team of humans and UMSs to hand off the leadership role depending on the task, along with varying the level of autonomy of the UMS members [11] [12] Their research is focused on the interaction between the team members, the shifting roles and responsibilities between team members during an operation, and providing individual autonomy to the UMS members of the team, autonomy that is crucial during periods of poor communication capabilities. The Human Robot Interaction Operating System focuses on how human and robot team members communicate t o solve a problem; it currently handles tasking at an individual resource request level and deals with operational tasks tasks that are highly detailed and well defined for a human/robot pair [13] It does allow for dynamic

PAGE 22

22 team assignments, but only on the level of deciding which UMS is best equipped to handle a single task requested by a human, or which human is best equipped to handle a single task requested by a UMS basically assigning a temporary assi stant to a worker. Levesque studied teamwork using joint commitment, where members of a team work towards a common goal and mathematically negotiate the perceived possibility or more importantly, impossibility of achieving that goal [14] In this work, the importance of team members working towards that common goal takes precedence over an individuals optimal results. In other words, what appears to be the best course of action for a single team member g ets ignored in the hopes that the partner member will assist that member in achieving the goal, even if it appears to be futile, or the partner member will ultimately encourage that member to stop if they agree it is futile. Communication, interpretation coordination, and cooperation between human and robotic team members, where robotic team members primarily function as assistants to the human members, were the primary focus of a research project and resulting multi agent architecture from the Institute for Real Time Computer Systems and Robotics (IPR) [15] In that study, one of the primary goals was to adapt a fully automated situation to allow the cooperation of humans, rather than isolating the robots in a fully automated factory setting. Their assertion was that UMSs are less flexible but more reliable, whereas humans are more flexible but less reliable, in terms of memory, consistency, and fatigue, so by pairing humans and robots together, the strengths of each can be exploited instead of isolating the robots from humans. In addition, the IPR

PAGE 23

23 study included the domain of service robots, where robots operate alongside humans in a cooperative service mode. The Center for Robot Assisted Search & Rescue (CR ASAR) started at the University of South Florida and subsequently moved to Texas A&M, focuses on using robots for search and rescue efforts, and the humanrobot interaction (HRI) required for such efforts [16] The CRASAR wor k is primarily focused on how the humans involved not only the human members of the search and rescue team but also the victims interact with the robots to provide and receive situational awareness. In addition their work brings up the subject of com munication between team members in an environment with limited communication. Subsequent work by CRASAR and MSHA defined and ranked skills, system components, and physical characteristics that are valuable in a rescue robot in different scenarios [7] The Robotics Institute at Carnegie Mellon developed a multi agent system (MAS) for coordinating teamwork between multiple robots and humans, specifically targeted towards an urban search and rescue (USAR) application. In their work developing RETSINA MAS, they explored the difficulty in getting humans and robots to work together seamlessly, for several reasons: it is difficult for robots to readily communicate their intentions to the human members, humans and robots are typical ly trained in different manners, and team level collaboration among multi robot teams is still an emergent technology, so most autonomous or semi autonomous robots are still task driven rather than team driven [17] The Applied Physics Laboratory (APL) at John Hopkins University (JHU/APL) modified swarm based behaviors to study strong autonomy in multi robot teams, relying

PAGE 24

24 on wireless communication and coobservation for situational awareness in an Intelligence, Surveillance, and Reconnaissance (ISR) scenario. In strong autonomy, commanders intent and highlevel operations are provided to the UMS, and the specific course of action is devised onboard the UMS using Dynamic CoFields [18] In swarm behavior, which is based on swarming insect behavior, control is decentralized such that the behavior of the group is influenced by the cooperative actions of the individual members [19] Subsequent work at JUH/APL continued the study of cooperative robotic behaviors using stigmergic potential fields, in which an individual vehicle operates in a locally held model of that vehicle and its environment, and the vehicles and their behaviors are modeled in a hierarchical organizati on. The behaviors are developed as Effects Based Operations (EBO) [20] Both of the JHU/APL efforts studied teams comprised entirely of UMS. The TEAMCARE lab at University of Southern California (USC) presented a chronology of multi agent teamwork research over the last twenty years. They discussed the Belief Desire Intention (BDI) model, which placed the emphasis on executiontime decision making, and the Distributed Constraint Optimization Problem (DCOP) and the Decentral ized Partially Observable Markov Decision Problem (DEC POMDP), which put more effort into planning time decision making. They suggest that a better approach might be to incorporate more executiontime decision making into the planning time approaches, all owing for a more robust system, thus reducing the model uncertainty that produces much of the error in DCOP and DEC POMDP [21] The

PAGE 25

25 TEAMCARE group later devised a hybrid BDI POMDP framework to demonstrate that concept. [22] A group from Wright State Universitys Department of Biomedical, Industrial, and Human Factors Engineering studied human effectiveness in working with unmanned combat aerial vehicles (UCAVs) [23] T heir focus was on reducing operator complexity rather than having humans and UMSs working as peers on a team. Reducing operator complexity allows for more autonomy in the UMS members, thus providing more opportunity for UMS members to participate in complex, team oriented tasks. In all of these cases, the focus of work has been on how to effectively manage existing humanrobot teams or existing multi robot teams. The effort of planning the structure of the team is largely ignored. Team Planning and Scheduling Murphy took the effort of team planning a step further than the others by defining workflow roles and responsibilities in a search and rescue operation and defining possible insertion points of robots into the team [16] Defining the roles and responsibilities can help in two ways: ensuring that each member of the team has a specific role ensures that every member knows what to do and provides assurance that full coverage of needs are assigned; and if used in a team selec tion manner, the defined roles based on abilities can help ensure that the newly created team has the skills needed for the job covered by the members. In the area of team planning and scheduling, there have been several research projects that have used si mulation to evaluate assignment of teams, or crews. For the railroad industry, Canadian National Rail and Circadian Technologies, Inc., used discrete event simulation to assign crews to trains. Their problem involved optimizing

PAGE 26

26 crew schedules with respect to minimizing cost, subject to rules and regulations for worker safety [26] At the same conference, a team from Embry Riddle Aeronautical University discussed using simulation for manpower planning for aircraft maintenance workers [25] Their goal was to develop an optimal assignment of maintenance crews based on aircraft flight and maintenance schedules; however, they assumed all workers had the same skills. Software project planning, fro m resources to project scheduling, was the subject of research by Joslin and Poole. They modeled individual resource skill sets as probability distributions for contributing to the project rather than specific skills needed for a project or task [32] With this approach, there is much work that has to be done at probability assignment time to effectively model the ability a member might have in contributing to the project. The Multiagent Adjustable Autonomy Framework (MAAF) for mixed human/robot teams touched the areas of human robot teamwork and team planning by designing a framework for mixed teams that allows individual autonomy and teamwork, and uses software agents to assign team members to specific tasks at execution t ime, based on the members abilities and availability [24] The actual team selection is performed prior to beginning their process and their assignment is to individual tasks, rather than the composition of the team. Miller [31] discussed the role of simulation in planning with a twostep approach in which partial plans were devised, then revised through simulation, in a constraint optimization problem. The domain he chose to model was planning equipment usage and staging on a manufacturing floor. He suggests that for a planner to be effective, it

PAGE 27

27 cannot be completely domainindependent, but instead requires some knowledge pertaining to the domain in order to choose the more appropriate plan. In each of these scenarios, individual tasks were assigned but there was little effort involved in assigning a subset of members to the team. Instead most dealt more with arrangement and scheduling of existing team members. This indicates a gap in the area of team scheduling. Mission and P ath P lanning One of the inspirations for this research was work done by Lee and Fishwick in the area of Simulation Based Planning to assist in mission planning, using multimodeling to simulate the mission plans at var ying levels of abstraction, depending on time available. The simulations were demonstrated with unit level planning [30] Their later work included simulationbased route planning [34] This work carries their concepts into optimal team planning, with a heuristics based twist. In the area of path planning, Ganapathy and Hill combined heuristics based path planning algorithms with simulation to study the World War II Bay of Biscay U boat hunting mission [27] Hu, Li, Guo, Sun, and Zeng applied a heuristics based transformation to a vehicle routing problem as part of the simulation [36] Another area of research in simulationbased planning includes the Navys Mine Warfare Commands Naval Mine Warfare Simulation, which uses faster than realtime simulations to aid in planning mine countermeasures missions, running numerous simulations using a Monte Carlo method. They also use it to perform force level studies to determine the resource level and types required for a given scenario [37] Another project that combined heuristics based pathplanning algorithms with simulation had simil ar goals to this proposed research, though at a single UAV level

PAGE 28

28 [28] In that project, they used directed acyclic graphs to model roadways in an effort for the UAV to find a groundbased vehicle traveling on those roads. In their case, they traverse all of the nodes and edges of the graph until they find the target. In their survey on graph management and mining algorithms, Aggarwal and Wang discussed the relationship between reachability techniques and optimal routing probl ems in computer networks [42] Sensor management was the initial problem a Swedish team sought to solve using simulationbased planning. In their initial demonstration, they used simulationbased planning to determine correct deployment locations of UAVs and sensors in an attempt to follow a person of interest, but they laid the groundwork for creating a simulationbased planner for allocating resources [38] The EyeRobot project used the heuristics based path planning technique of Weighted Region Problem (WRP) [39] In that project, they determined the globally optimum path for a single robot to travel from one point to another, using navigability factors as weights on polygonal map regions, and then the robot performed local path planning during the traversal to avoid obstacles. Global path planning uses static map data such as roads, houses, fields, and bodies of water to determine that higher level path, and local pa th planning is runtime sensor based planning. EyeRobot used a single optimal pathplanning algorithm, and a polygonal region and test the viability of both vector and grid based planning. Uncertain terrain and risks, along with uneven natural boundaries, makes planning for an autonomous vehicle difficult. One group used a fuzzy approach to dealing with these uncertainties in WRP [40] Their goal was to plan a path through different

PAGE 29

29 weighted terrains and areas that had other risks in the context of the WRP where polygonal regions are assigned weights. Planning a path for a single vehicle has its own challenges, which have been subject of much research to date. Finding an optimal path for multiple, disparate vehicles becom es a significantly more complex problem when the factors of collision avoidance and rendezvous timing are considered. In one attempt, multiple singlevehicle optimal paths were generated using a bounded curvature path planning algorithm, using both Lagrangian duality theory and a penalty function method for convex regions, and the path optimization was extended to multiple vehicles using an elastic multiparticle (vehicle) system to chain the vehicles together [41] As severa l of these projects show, t here has been some work to date on combining heuristic algorithms with simulation. Much of it has centered on single entity movement, when it comes to the path planning domain, or using heuristic algorithms within a simulation t o plan a multi step function, such as stacking then moving palates for shipping [56] Simulation Techniques Simulation is an invaluable tool in decision support. Done with care, it can make the job of the user easier and decrease costs. In their 2003 paper, Hill and Malone discuss some of the common problems encountered in simulation for decision support, and how to avoid them The first issue, conflicting results from different models, can be mitigated by an open dialog between different modelers to ensure free exchange of ideas and opportunity for brainstorming. The second issue, different interpretations of the simulation results, is not a problem exclusive to simulation; the most effective way to

PAGE 30

30 mitigate that risk is to model at the level of the audience, and to communicate regularly [29] A team from Auburn University studied using multisimulation for decision support in an uncertain environment [54] Their work focused on adaptive responses to changes and/or new information in the course of the simulation. T heir incorporation of learning algorithms for an adaptive simulation model is of interest in a dynamic environment, such as for team replanning for a s ubtask that arises during an existing task. Progressive refinement, one of the techniques we use in this project, is most often associated with visualization [43] [44] [45] [46] and dealing with limited resources and large data sets, which is why it is such a good fit for visualization. Attempts at creating an architecture for parallel progressive refinement have demonstrated the scalability of the approach, with linear growth [45] Indeed, scalability is a desired outcome of most projects, which is a reason that technique was chosen for our project. Progressive refinement starts out with a simple model of course granularity, increasing the complexity and granularity of the critical elements of the model at each iteration; the initial definition of the models dimensions and the granularity is key to the success of the problem. Using a granularity that is too fine would be equivalent to flattening the approach out to be bruteforce; too course and the results at the courser levels are inaccurate and provide little to no useful feedback for subsequent levels [52] In progressive refi nement radiosity, the unshot residual energy of a scene is processed for each element, by way of determining the amount of energy reflected back. The accumulated and residual energy values are tallied, and when the

PAGE 31

31 calculations for a single element are ex hausted, its residual energy is reset and the algorithm iterates to the next element, until the residual energy is determined to be sufficiently low [43] This allows the calculations to stop when sufficient detail has been uncovered without doing exhaustive and expensive calculations. Blanford discusses approaches to estimate image portions during a slow download of a large image, using spatial and grayscale variance techniques [47] This progres sive refinement approach allows the recipient of the image to get a good idea of what it looks like before it is fully downloaded. Going beyond visualization, 20 years after writing The Mythical Man Month, Brooks revisited his theories and drew a correlat ion between iterative, or incremental, development of a software project to progressive refinement. He compares early testing with stubs and incrementally adding new features to the software to the progressive refinement approach both give an early, though vague, view of the results [52] Rosenbaum and Shumann suggest that the tour through the data ability to incrementally preview data is beneficial enough to use a progressive refinement approach, regardless of the system resources [48] An even more compelling use for progressive refinement is demonstrated in Jet Propulsion Laboratory (JPL) research on using progressive refinement in support vector machines (SVM). In that project, they used a reduced set SVM as initial classification and used a slower, more detailed SVM with more support vectors on the results from the reduced SVM with courser granularity to correct the errors. The intended application was image classification with large dat a sets (pixels) [49]

PAGE 32

32 The Center for Risk and Reliability, in an effort to perform simulationbased risk assessment, created a simulator that performed hierarchical planning and multi level scheduling. The multi level scheduli ng is a variation on progressive refinement, in that the simulation models are generated with multiple levels of detail. Instead of iterating through the levels, however, the multi level planner chooses a single level to use for each model during the simul ation, depending on the importance of that component for that particular scenario [35] Game engines, such as the one used at the National Institute of Standards (NIST) USAR Test Facility, are an increasingly popular way to develop simulations. Game Engines are used by both game and simulation developers to reuse the common functionality required by a simulation after all, a computer game is a form of simulation to reduce the development cost [ 50] The development of the model used by the simulation can be separated from the development of the engine itself, and requires subject matter expertise in the area being simulated. The OneSAF Objective System (OOS) is the U.S. Armys targeted Computer Generated Force simulation system used for training and testing across a large variety of operations. At I/ITSEC, Karr described the challenges associated with translating real world into computer models for simulation [51] What are the Missing Pieces? The main focus of most of the research into human/robot teaming has been how to get an existing team to achieve its common goal and how to communicate effectively between human and UMS members of the team. The focus of o ur research is to determine how to create an efficient, effective team, not necessarily using all available resources, and give guidance on the best path to take. Most of the research to date in

PAGE 33

33 humanrobot interaction has been on how to develop effective communication between the members, or how to divide tasks among the members. Very little research has been published on methodologies to choose an optimal set of members to populate that team in a consistent manner. The distinguishing f actor here is that our projects goal was to optimize based on a set of disparate team members, including not only the proposed path to be taken but also the makeup of the team. In addition, our project is at the level of global path planning, since the major goal of the project is to select an optimal team, not perform the actual task. Common sense suggests that a team should be comprised of members that have at least a subset of the skills required to do the job. For a team to complete a task successfully, the entire s et of required skills must be covered among the members of the team. The question remains of how to create a team that can perform the assigned task in an optimal manner. Our research is focused on staffing a single team for a task. For future work, it c ould be enhanced to allow optimization across multiple tasks and multiple teams, but the goal of our research is not to assign every member to a task. Innovation/ V alue A dded Heuristics based path planning can be highly complex and highly specific to a predefined heuristic. When evaluating multiple combinations of team resources using purely heuristics based algorithms, the complex heuristics algorithm must be run once for every combination of resources, leading to a very timeconsuming process that is di fficult to re fit to new criteria. A simulationbased approach would estimate a result for every team going down ever y path and choose the best one. With high speed computers, memory, and time, this can be done for a small sample; however, it

PAGE 34

34 becomes unwi eldy for a large pool of resources and a mission that can take many paths. Somewhere between AI path planning and massive multiplesimulation using progressive refinement there is a well balanced approach to determining an optimal team for a given task. For this hybridization to be effective, it must reduce the complexity of the heuristics portion and reduce the number of iterations required for the simulation portion. This research search es for that hybrid approach. The results of this research includes a flexible, datadriven method to determine the optimal configuration of a human robot team for a given task and a tool that can be used in team planning, member re tasking, and leadership training. This method employ s a hybridized approach of simulatio n and artificial intelligence techniques, using a progressiverefinement queuing model to quickly bypass the least desirable configurations. For a task involving movement of the team, this combination of techniques allow s the user to specify paths at any f idelity, from start/end points to a series of waypoints, and skills required to perform the mission. From the points provided, the program uses modified pathplanning techniques that employ some, but not all, of the variables used in the cost algorithm, t o determine a set of viable paths, rather than a single, optimal path, since the exact distance in this case, score is different based on the characteristics of the individual team and the scoring method may be changed depending on the priorities of the mission The variables, or heuristics, chosen for the pathplanning algorithm were chosen because they have a similar impact on all team configurations; those that are heavily influenced by the team characteristics were left to the simulation and scoring algorithm. Th e resulting set of paths is then

PAGE 35

35 converted to a series of queuing models that have progressively more levels of detail, and simulations are performed for all possible team configurations that fulfill the basic mission requirements. Fig ure 21 illustrates the highlevel concept of team and path selection using multiple simulations. Figure 21. Teamwork selection starts with a pool of candidate members, a set of required skills, and a goal. Multiple simulation approach tests each viable team against each path to find the optimal team and path combination. Contributions Beyond a prototype that will likely sit on a shelf, for the research effort to be worthwhile, something good must come of the research. This research crosses multiple disciplines within the realm of c omputer science. Major classifications as designated by the Association for Computing Machinerys (ACM) Computing Classification System (CCS) covered include I.2.8 Artificial Intelligence/Heuristics, I.2.9 Artificial

PAGE 36

36 Intel ligence/Robotics/Workcell organization and planning, and I.6.8 Discrete Event Simulation [55] These classifications are illustrated in Figure 22. The domain of the problem space is I.2.9 Workcell organization and planning in this case, planning the organization/structure of work groups, or teams, that contain both humans and robots, with each member having a specialized skill or set of skills that are needed to contribute to the tasks goal. The first contribution from th is research is in the automation of that team planning, which is aided by the designation and breakdown of skills as mission and member parameters and operational and environmental characteristics that define how the members interact with the surroundings such as member and passage widths, or member weights and passage weight limits Figure 2 3 shows the specific subjects within the I.2.9 classification. The domain of the solution space of this work encompasses both I.2.8 (AI/Heuristics) and I.6.8 (Disc rete Event Simulation). These are the areas that we use to address the problem solution how to best perform the team planning A pure simulation approach would involve simulations of every reasonable path that every valid team configuration could take Assuming N teams and an average of M possible paths per team, this would involve N*M simulation runs on top of the exhaustive path search, which has a complexity of O( n !), where n is the number of intersections [56] If n N and M are relatively small, this may not be an issue; however, this solution would not scale very well. Scalability becomes an issue when there are many options from both a number of teams and number of possible paths perspective. In a very large or v ery complex location, the number of possible paths could become quite large even when the paths are constrained to simple, noncyclic paths. Traditional consumer global

PAGE 37

37 positioning system (GPS) units and online mapping products attempt to stay on a main road for as long as possible, then perform more finetuned path planning when near the destination. This is sufficient for most recreational road travel; however, in a task such as the hostage rescue scenario, it may be undesirable to stay on a main thor oughfare for a long period of time, as this leaves the team more exposed to hostile activity. In addition, in some areas of the world, there are no larger roads to use for the initial part of the journey. With the prospect of having to send the team thr ough a myriad of back roads and alleys through a large area, scalability becomes a factor. In addition, if the tool is used to dynamically replan, fast processing time becomes crucial. A purely heuristic approach would require a significant number of weights to be calculated based on skills and terrain features much of the same calculations that are involved in the simulation. Each team configuration would require its own set of weights for the path nodes, and the heuristic algorithm would still have to run N times. With a best first search that runs in linear time, that might sound appealing; however, the calculations of the heuristics, which must be done on each edge e must also enter into the equation. A more compelling reason to avoid the solut ion to run the best first search for each team configuration is that the hybrid heuristic/simulation approach gives alternative solutions that may be almost as good, or even equally good, as we will show in chapter 4. This hybrid heuristic and progressive refinement simulation approach is the second contribution from this work, providing a faster way to reach the answer. To reduce the computing time and complexity required to find the optimal team, this research combines the strengths of multi simulation faster than realtime estimate of results, ability to safely run multiple scenarios, single pass calculations with the

PAGE 38

38 strengths of a heuristic approach quickly eliminating less desirable paths, the ability to use different weighting factors. Our go al was simple reduce the computing time to find the best team and the most favorable path, without compromising the result. By using a heuristic approach at a higher level to reduce the number of paths followed in simulation, we were able to achieve our goal, contributing to the areas of team planning in a hybrid human/robot team, and a hybrid of heuristics based path planning and multiple discrete event simulation.

PAGE 39

39 Figure 22. Computing Classification System Contributions

PAGE 40

40 Figure 23. CCS Subjects of Robotics

PAGE 41

41 CHAPTER 3 PROTOTYPE The research into using simulation via progressive refinement to determine an optimal humanrobot team resulted in a prototype that allows the user to define required skills, user skills, and a layout of the area in quest ion. The prototype is designed in a modular fashion, allowing for different scoring functions, goals, and path selections to be tested. The prototype, as shown in Figure 31, contains an editor to create and view Resources, a mission editor to select mission properties, including skills and start/end nodes, and a team planner that performs the team and path selection and simulation. Figure 31. High level architecture of prototype Ontology The data model shown in Figure 32 was created as a Resourc e Description Framework (RDF) [58] Schema (RDFS), using Protg, an opensource ontology editor from Stanford [59] It displays the highlevel classes that are used to represent the data. An ontology is a representation of the concepts and relationships that can be used to describe and model data specific to a particular domain. It allows the data modeler to specify object and class

PAGE 42

42 properties to define the relationships, including domain and range. Having the formal definition facilitates reasoning across the data, using automated reasoners or first order logic in the form of queries such as SPARQL or Prolog [60] RDFS is a World Wide Web Consortium (W3C) recommended general purpose language for representing webbased data. It can be represented in eXtensible Markup Language (XML), which is a standard web language for representing data in a structured format [62] XML provides a standard format for exchanging data. Namespaces within XML allow multiple RDF Schemas to be linked and referenced, providing common definitions to be shared. An XML primitive would be specified by value or the shorthand, . XML insta nces can be grouped together into complex objects to provide a complete definition of an entity or idea. Figure 31 shows a portion of the XML version of the RDFS generated from the data model. Table 32 lists the data properties needed by Resources, Ter rain, and Mission classes and is the basis of the RDFS definition. Figure 32. OWL class visualization

PAGE 43

43 Figure 33. XML Sample 1 1 1

PAGE 44

44 Table 31. Properties Resource Terrain Mission Name Level Skills required Size Name Sta rting location Weight Location coordinates Goal location Speed Parent Scoring function Cost to deploy Children Cost to replace Neighbors Hourly rate Danger probability Skills Smallest width Traversal length Weight limit Since RDF and XM L are standardized data formats, using RDF for defining the data makes conversion to and from external programs less complex, resulting in potential integration savings. In addition, there are several open source tools that allow for data visualization and manipulation in RDF, making it far easier to build a reasonable data set. The instance data is asserted as subject predicateobject facts, or triples, in an AllegroGraph triple store [61] Both Protg and AllegroGraph fol low the RDF standard. Storing the data as RDF triples allows the use of SPARQL to easily query the data, and to perform the initial reasoning across the facts; for instance, the set of candidate teams is generated directly from the triple store using a SP ARQL query. SPARQL queries can be used to discover information about a subject, such as all information related to a Resource or set of Resources; link subjects together by common objects, such as retrieving all Resources who have a specific skill or fit into a given weight range; find information by predicate, such as retrieving a list of all Resources and their associated skills; or by any combination of linked or nested subjects, predicates, or objects.

PAGE 45

45 Framework Resource Parameters For a team to meet its goals all members have at least one skill that is required to complete the specified task ; in addition, all members have certain operational characteristics that affect their ability to perform the mission or the overall cost of the mission. For i nstance, speed is used as a factor in computing the estimated time required for the mission; weight and width are used in conjunction with weight limits and widths of passages in the terrain to determine if a given team can go down the path. For the purposes of this discussion, cost could be monetary or temporal, or even projected loss. In the prototype, the skills are used to match resources to the current mission, and operational characteristics are used to calculate navigability and overall score of the mission. Figure 35 shows the classes created for building the team. The CreateResourceBean and ResourceBean classes are used by the Resource Editor to define and store resources. A Team consists of one or more Resources, and a Resource contains one or more Skills. For the initial prototype, we used a small set of skills for each potential m ember of the team, and subsequently added more skills and operational characteristics. The prototype is designed to allow skills to be added both to the resourc e pool and to the mission goals as needed, rather than having a static, predefined set of skills. Each available resource is also assigned physical and operational characteristics, including size, maximum speed, and weight using the Resource Editor shown in Figure 34 For some scoring functions cost to deploy and cost to replace are also considered. Since operational characteristics

PAGE 46

46 are an integral part of the scoring function, they are not as dynamic as skills. Table 32 lists a set of sample parameters that are used in the prototype. Figure 34. Entry panel for resource candidate

PAGE 47

47 Figure 35. Class diagram for Resources

PAGE 48

48 Table 32 Resource Parameters Classification Resource Parameters Operational Characteristics Width Weight Speed Len gth Cost to deploy Cost to replace Hourly cost to deploy Skills Dynamic list The set of all possible teams is generated at the beginning of the simulation using a dynamically built SPARQL query, based on mission requirements. Figure 36 shows a sam ple query for a mission with 4 defined skills, displayed in AllegroGraphs AGWebview tool. Only a subset of the results is displayed since the query returns 120 results. In the cases where a team member has multiple skills, the current implementation as sumes the member can perform all of the duties; thus a result of Barney, Cylon, M1, and Barney will yield a 3member team. Future enhancements to the tool could force all members to be unique, i.e., use only one of their defined skills per mission, to allow the definition of exclusive skills to allow some overlap for skills such as carry radio and identify target, but exclude some skills such as carry weapon and carry wounded, or to allow a minimum number of members with a given skill. To enforce unique members, a FILTER could be applied to the SPARQL query in the form of FILTER (?idx != ?idy) for each pair (x,y) of resources. To enforce multiple unique members with the same skill, a FILTER could be applied as above on ids with the same skill defined. If uniqueness were already being enforced, the SPARQL query would simply have multiple entries for the same skill.

PAGE 49

49 Figure 36. SPARQL query for team selection

PAGE 50

50 Area of Interest Figure 37 shows a humanreadable mapbased display of the area, much like the area depicted in Figure 38. While this is easy for a human to identify paths and features, it is virtually useless to a computer in that form. The pertinent facts about each feature are translated into RDF triples. At runtime, those facts are t ranslated into a traversable graph for path planning and simulation. Figure 37. Diagram of a city area. The main City area is divided into 4 large Blocks, NW, NE, SW, SE, with major roads to the North, South, and crossing the center. Each block is divided into Sections with roads dividing the sections. Each section is further divided into Areas with smaller roads. Each Area contains houses, apartment buildings, or other buildings.

PAGE 51

51 Figure 38. Search team in an urban area (courtesy U.S. Marines ) The terrain layout is defined as an RDF graph, using a set of RDF triples to model the individual map features, such as buildings, roads, and rooms, as well as transition points between the features. Each node in the RDF graph is represented by a subje ct in the RDF graph; features of a node are represented by predicates and the individual characteristics of that node are specified in the object portion of the RDF triple. Figure 39 shows the node NW_Area1 as it appears an RDF graph. The same node i s displayed as SPARQL output as the Subject and Object of individual RDF triples in Figure 310. The description of the area itself occurs when NW_Area1 is in the subject. Associations with other nodes identified by NW_Area1 being in the object. The layout can then be displayed as an RDF graph with paths between the neighboring nodes, as displayed in Figures 312 and 313.

PAGE 52

52 Figure 39. RDF relations of a named location Figure 310. SPARQL results for subject NW_Area1

PAGE 53

53 Figure 311. Class di agram for maprelated classes

PAGE 54

54 Once the scenario has started, the map data from the triplestore is loaded into a set of classes to be used by the simulation and path selection. The classes related to the terrain data are displayed in Figure 311. The Map singleton contains a set of MapNodes which represent each map feature, including dimensions, coordinates, and neighbors. Once the initial path selection is complete, a Path contains a series of MapNodes and corresponding Waypoints used to generate the pat h at each level. Initial Path Selection Once the set of candidate teams is created using skill matching, a set of paths must be generated. The mission is defined with starting and ending locations. The initial set of paths is determined based on the settings of the heuristic pathplanning module. At initialization time, a SPARQL query retrieves all leaf node Entry/Path item (i.e., road, block, area) pairs. The query results are used to create a traversable graph, with the Entry being a graph node and the path items being edges on the bidirectional graph. Figure 312 gives an RDF graph representation of this data; the resulting traversable graph is shown in Figure 314. The heuristics based search function discovers a reduced set of candidate paths us ing the traversable graph and makes it available to the simulation portion of the program. The classes involved in the search are shown in Figure 315. Several well known pathplanning algorithms were evaluated prior to implementation of the path selectio n portion of the prototype. One of the most common algorithms is Dijkstras algorithm. In the variation of Dijkstras that looks for the shortest path from source to target as opposed to finding the shortest

PAGE 55

55 path from source to all other nodes, the algorithm finds the single shortest path [63] A suggested variation that finds a set of near optimal paths removes a single node from the solution and recalculates without that node; the logistics of performing the removal and re plan were deemed too cumbersome for the prototype. In addition to Dijkstras algorithm, A* also finds the shortest path from source to target using an admissible heuristic that is based on a distance that is likely not obtainable. A traditional A* shor test path algorithm was also eliminated from consideration for the same reason as Dijkstras the goal of this research was to stop at a near optimal subset of paths and simulate the rest. However, the consideration of the admissible heuristic was adapt ed into our search algorithm [64] Breadth first search was another consideration for the pathplanning algorithm, since it could be expanded reasonably to continue searching beyond the optimal path to explore near optimal paths as well. The depthlimited depth first search was chosen above this algorithm for a number of reasons. First, the distance of the path is not directly dependent on the number of nodes traversed in the graph; so a path with fewer nodes, due to length or danger ratings, could in fact be more expensive than a path that travels more nodes. Second, the depthfirst search has a higher space complexity than depthfirst [65] Other approaches including BellmanFord and FloydWar shall algorithms were eliminated because they are of most use in negativeweight graphs, which was not necessary for this problem.

PAGE 56

56 Figure 312. RDF graph of NW Block, leaf level. Larger squares are transitions, represented by the Entry class, between the roads or areas.

PAGE 57

57 Figure 313. RDF graph of NW Block, level 2.

PAGE 58

5 8 Figure 314. Traversal graph for path planning, derived from Figure 311. In this graph, the transition between two map nodes is a circular graph node; the edges are the streets in the terrain, with the length of the segment as the basis for the cost of traversing that edge. The path search function is a modified depthlimited depthfirst search. Several configurations of this search were tested, from no heuristics, which finds all possible paths between the two graph nodes, and a pair of cost limited searches with varying limits that find a set of the shortest paths capped by the weighted distance between start and goal. The classes involved in the path search are shown in Figure 315. The search interface allows for replacing search algorithms. Figure 315. Class diagram for path selection

PAGE 59

59 For the heuristics enhanced searches, the center waypoints of the start and goal nodes are used to calculate the floor length, in di stance, between the two nodes. This as the crow flies distance is guaranteed to be less than or equal to the shortest path. In this case, the shortest path refers to the travel distance between the two points, without regard to the number of graph node s involved in the path. The crow flies distance is then multiplied by a limiting factor that places a ceiling length on the search once a path reaches that threshold without reaching the goal, that path is abandoned. Factors tested were 3 and 4 tim es the absolute distance, but could vary depending on the complexity of the terrain and whether there are localized hot spots of dangerous activity. A more even terrain with fewer navigational limitations would be better served by having a lower threshold, closer to 2.5 3 times the absolute distance. One of the initial goals of this project was to find an optimal threshold between the heuristic planning and the progressive refinement; through the course of the project, however, we determined that there is no single best answer to that question but instead a guideline on how to set those heuristics and what factors to consider. During the recursive depthfirst search, the length of each edge is added to the length of the path traversed to that point; i f the new length is greater than the ceiling length, that edge is skipped. This heuristic is sufficient in an area where dangerous conditions, such as possible debris in a search and rescue operation after a natural disaster or a hostage extraction opera tion where there is terrorist or gang activity in certain areas and the danger probability is high; or in an area with varied terrain that affects navigation, as used in the EyeRobot project [39]

PAGE 60

60 Since the danger probabilit y is a factor in the scenario we used, the length of the edge is multiplied by the danger probability. The more dangerous a passage, the higher the calculated cost of that segment of the path is. In a scenario that involves varied terrain as opposed to avoiding hostile activity, a navigability factor could be used instead. The algorithm is displayed in sequence diagrams in Figure 316 and 317. The algorithm used for heuristic path planning is: heuristicSearch(start, goal) determine floor and ceiling costs search ( queue, start, goal, 0) Generate paths from results search (queue, node, goal, cost) mark node visited node queue for each edge on node if edge cost + path cost < ceiling apply edge cost to path cost retrieve connected_node if connected_node is goal reverse path, place in paths list clear visited return search ( queue, connected_node, goal, cost) node queue remove node cost from path cost clear node visited flag

PAGE 61

61 Figure 316. Sequence Diagram for path search initializes the heuristic cap, initiates the search, and processes the paths at the end.

PAGE 62

62 Figure 317. Flow diagram for capped depth first search. Cost is calculated at each node and the path is rejected if it goes beyond that cost.

PAGE 63

63 Progressive Refinement The layout of the city is st ored in a set of scene graphs. Scene graphs are often used in graphics to drill down into a scene from generic/highlevel overview to more and more detailed lower level. For instance, drilling down into a city block yields exterior buildings and streets, then interior buildings with room layout, to room detail. Each node of the graph, which is arranged in tree form, contains a specific feature. A nodes children are parts of that node in higher detail. For instance, a building nodes children would be rooms within that building [63] This allows all operations on a node to propagate to its children and thus reduce computation [64] A partial scene graph, as view ed through an RDF graph, of the urban layout from Figure 37 is depicted in Figure 318. The scene graph nodes are created in the MapNode class, which is detailed in Figure 3 11. The overall progressive refinement and simulation functionality is initiali zed and contained within the TeamPlanner and Team classes, which are displayed in Figure 319. We use d scene graph technology to allow the simulation to start with higher level estimations and iterate into more and more detailed estimations. The lower the level, the more accurate the estimation should be. The depth of the simulation depends on the amount of time available to run simulations; a team built well in advance of a mission will have more time to simulate to the lowest level, but a mission needing immediate tasking may not be able to go as deep. In addition, some details may not be known for some features of the city, so the simulation must stop at the higher level for those areas thus using less accurate scoring As each level completes, the estimates for team performanc e are sorted

PAGE 64

64 Figure 318. Partial scene graph layout of area. City is level 0, blocks are level 1, sections are level 2, and smaller sections of roads are level 3. The RDF graphs relating levels 2 and 3 are displayed in Figures 310 and 311.

PAGE 65

65 Figure 319. Class diagram for TeamPlanner and Team classes. TeamPlanner initializes the simulation and runs the progressive refinement. Team contains the team members and si mulates each path for that team. Results are stored in SimResults.

PAGE 66

66 and a configurable percentage of the teams deemed to be the most optimal are carried into the next level. This allows the number of team/path combinations to be reduced at lower levels. In a fairly shallow area layout with few levels, the pr ogressive refinement can add overhead; however, a more complex layout can benefit significantly from the progressive refinement approach. The sequence of events from a scene graph perspective is depicted in Figure 320. Simulation Each level of the simulation starts with a required path or set of path choices for the mission at that level. This could be a detailed set of waypoints, a set of high level waypoints, or even a start and end waypoint, depending on the mission definition and the results of the previous levels simulation. Before simulating any team configurations, the program uses those waypoints to determine a set of feasible paths to the target position starting with the paths selected by the heuristic approach described above at the first level and a subset of those paths as the lower levels are simulated, based on the scores from previous, higher levels ; this reduces the overhead of finding all of the potential paths for each team configuration. In addition, i nformation is stored with each path that allows a quick check of whether a team member can go down that path. If a path has a narrow alley that team members over a certain dimension cannot pass through, this automatical ly eliminates those paths from any team configuration containing members over that size this constraint applies primarily to UMS members, though it is feasible for a path to exist that is too small for a human member Weight limits and member weights ma y also be used to constrain paths, since some unmanned vehicles can be much heavier

PAGE 67

67 Figure 320. Sequence diagram of the progressive simulation

PAGE 68

68 than their human counterparts. Unknowns, such as enemy presence dangerous terrain, or other potential dangers, are represented by a probability matrix and are based on a priori data. The sequence diagram in Figure 321 shows the activities performed in evaluating a single team. For each team configuration tested, the mission is simulated for each of the previously sel ected paths. Projected failures are discarded, and projected successes are ranked according to the assigned criteria. At the end of a level, a percentage of the teams with the best scores move on to the next level. The mission simulation is performed i n Team::runMission(), which is shown as a sequence diagram in Figure 322. The simulation retrieves the length, width, and danger probability then calculates the traversal time for that node and the transition to the next node, and then applies those times to the running score. The throughput is determined using a generalized deterministic queuing network. The constraint, or service time, is based on the minimum width and lengt h of the passage and the width and speed of the team members [65] Unlike traditionally stochastic queuing models that must account for entries arriving at different times, ours assumes the team members arrive at once and is used to determine how quickly they can pass through. Team members are allowed to travel abreast, provided their combined width is smaller than the passageway. Figure 323 shows a portion of the urban layout with two levels of detail the higher level block view and the lower level street view, with a sample path. Figure 324 converts those maps into queuing models for the paths at both levels of abstraction. The more abstract queuing model in ( a ) i s obtained in one

PAGE 69

69 Figure 321. Sequence diagram for TeamEvaluation

PAGE 70

70 Figure 322. Simulation of individual path for a single team configuration of two ways: by partit ioning the lower level queuing model in ( b ), or by using a higher level within the scene graph defining the geometry of this area. The queuing model identifies a partial path from the entry into the NW Block on NW E Street, continuing as it turns into NW W Street, and turning in to NW Section 4 with the final goal of NW Area 1. Part (a) shows the queuing model at the block level, and part (b) shows it at the section level. The block level combines queuing nodes that represent major roads that pass through that block, the section level includes smaller local roads and sections off the larger roads at

PAGE 71

71 intersections; a more detailed level would break out building queuing nodes, with as much detail as is known, such as closet and even furniture. A B Figure 323. A path through the area at different levels. A) shows the level 2 structure of the map, and B) shows level 3.

PAGE 72

72 Figure 324. Scene graph representation of queuing model A) shows a more abstract queuing model of the cutout area di splayed in Figure 323 A, and B) shows the lower level queuing model of the same area, corresponding to Figure 323 B. Scoring Algorithms How do we determine an optimal team? There are many factors to consider in deciding the winning configuration. An obvious factor is the potential for succeeding in the mission goals. Of the team configurations that have the ability to complete the mission, the optimal team should have qualities that make it stand out from the others. Depending on mission parameters, it could be the team that can complete the mission fastest; the team with the fewest members overall and the fewest human members; the team that is least expensive to deploy; or the team that is most expendable or has the best ability to protect its human team members from injury or loss of life. Deployment cost might involve the cost associated with transporting a distant specialist, having to pay a highly skilled contractor a high bill rate, maintenance and operation costs of the unmanned systems, or re placement costs for a member that is lost in a dangerous mission, either through manufacturing cost of a UMS or training cost of a human member. The modular approach to the design allows us to assign an importance to each factor at runtime for a simple calculation, or to plug in a different, more complex formula to calculate the score for the team. Several of

PAGE 73

73 these formulas are listed in Table 33. The results of the different scenarios and scoring models are discussed in the next chapter. Table 33. S ample Formulas Formula Type Score Weighted Prototype T*time + N*size Weighted expanded T*time + N*size + X*survival + C*cost Cost sum(Resource:costToDeploy) + sum(time*Resource:hourlyRate) T = importance of time to complete. N = importance of size of team X = importance of survival rate. C = importance of cost to deploy Scenario The initial target demonstration of the team selection tool wa s an urban search and rescue scenario. In this scenario, t he goal of the mission was to travel undetec ted through an urban area to a building known to house hostages, enter the building, locate the hostages, free them, and get them to safety. Figure 325 shows a diagram with details of the mission, including the area layout, location of hostages and host iles, and several suggested paths. The p osition of hostages is marked with a star. Portions of the Northeast and Southeast sections of the area have known or suspected snipers. Possible paths are marked by different color lines. Streets are numbered, w ith East/West being Street and North/South being Road. For simplicity only the NW block is displayed, and buildings have been hidden from view. If building floor plans were available for a scenario, those details would be considered in the next level down. An arrow indicates the ingress route. For this mission the assumed priority wa s the safety of the hostages, then the human team members, then the robotic team members. The initial prototype used a simple approach to success or failure. Loss or projected loss of any

PAGE 74

74 hostages i s deemed mission failure. Loss or projected loss of any human team member is undesirable but may be counted as a success in the absence of simulations that project no loss of human members. Loss or projected loss of unmanned human members is c onsidered more costly/lower score than no loss, but counts as mission success. The success criteria are localized and could easily be changed based on mission or task constraints. Figure 325. Diagram of a mission scenario with suggested paths Starting point is the Arrow in the upper right, goal is NW Area1 which is marked with a star. Areas of higher danger probability are m arked with a jagged circle.

PAGE 75

75 CHAPTER 4 ANALYSIS Comparison To test the feasibility of the hybrid heuristic/progressive refinement approach, a bruteforce simulation approach, a pure progressive refinement simulation approach, a hybrid heuristic/bruteforce simulation approach, and the hybrid heuristic/progressive refinement approach were used to test mul tiple scenarios and cost functions. Their results were evaluated for performance, scal ability, and accuracy. Memory needed for each approach was also tallied and evaluated for suitability in laptops and even smaller mobile devices. On the simulation side, the bruteforce simulation approach tests every team against every available path, with no intermediate levels that reduce the data set. The progressive refinement approach is tested in two configurations reducing the bottom 10% of the teams after each level, and reducing the bottom 20% of the teams after each level. Each remaining team is tested against every available path. On the path planning side, there are three configurations used all paths, and two heuristic based path searches using a heuristic based on length and danger of each pathway, subject to a cap of either 3 or 4 times the straight line distance from the start to the end. Each scenario is tested against 9 combinations each of the three path searches paired with each of the three simulation configurations. Units used in the final prototype are generic scaled units based meters and meters per second. Testing Environment The prototype was written in Java using JSF 2.0 and PrimeFaces components. Since the prototype ran in a Java v irtual machine, the application had to be redeployed

PAGE 76

76 between runs so the garbage collection delays in a running program did not skew the memory footprint results. The test runs were performed on a MacBook Pro with a 3.06 GHz Intel Core 2 Duo processor, 4 GB 1067 MHz DDR3 Memory, using Mac OS X 10.6.6. The web application was run from within Netbeans in a Glassfish 3.0.1 application server. Scenarios Multiple scenarios were run to test performance at different sizes of the mission area, complexities of available paths, number of skills required, and number of available team combinations. Also included for comparison, but not included in the final results, are the initial test results from the early prototype. The early prototype contained a small set of members and skills, a small mission area, and an initial subset of paths instead of a generated path set. This prototype followed the pure progressive refinement approach. The current prototype contains more members, a larger set of physical characteristics, and more skills to choose from. Skills, for the purposes of the prototype, are defined with a simple label, such as Carry Wounded, and are used to choose which team members are relevant to each mission. Simulation Based Progressive Refinemen t, the Early Test As part of the prototype phase, an initial set of test data was created. The results are provided below. The initial prototype was written in C++ with file based data and was subsequently converted to Java for easier display creation. Mission The team must travel from the intersection of First & Main to the hostage location in Blue House, room 1 (BH Room 1) using one of 4 predefined paths. The area layout

PAGE 77

77 is displayed in Figure 4 1, and the initial set of 12 potential team members is in Table 41. Units for length and width are in meters. Figure 41. Mission Area with goals Skills required: Weapon TargetId WeaponAuth HeatSensor Gurney Table 41. Resources available in initial prototype Name Type Length Width Weight Speed Skill Fred Human 0.5 0.5 180 5.0 TargetId Barney Human 0.5 0.5 180 5.0 WeaponAuth Mitchell Human 0.5 0.5 180 5.0 Weapon Huey UMS 2.0 2.0 25.0 5.0 HeatSensor Talon UMS 2.0 3.0 30.0 6.0 HeatSensor, TargetId Cylon UMS 4.0 4.9 1 20.0 8.0 Weapon Viper UMS 3.0 4.0 85.5 10.0 Weapon, TargetId B9 UMS 5.0 5.0 35.5 8.0 HeatSensor Robbie UMS 5.0 4.5 35.5 15.0 HeatSensor Mule UMS 5.0 5.0 35.5 10.0 Gurney M1 UMS 3.0 2.5 35.5 5.0 Gurney T2 UMS 1.0 1.0 35.5 15.0 WeaponAuth

PAGE 78

78 Results Thi s simulation generated 144 possible team combinations. The scores for level 2 ranged 148.587 to 313.218. In this scenario, the same team won the top two spots, with different recommended paths. Scores were computed based on navigational concerns only. Total simulation time was 0.96 seconds Winning team and path Score 148.587 Team members: Mitchell, Talon, M1, T2 Path is shown in Figure 42. Second place team and path Score 152.58 Team members: Mitchell, Talon, M1, T2 Path is shown in Figure 43. Figure 42. First place team/path

PAGE 79

79 Figure 43. Second place team/path Final Prototype, Scenario One The first scenario included 15 potential members, displayed in Table 42, four skills, and a simple mission area with 4 levels Nine variations were run against this data set progressive refinement and brute force simulation each run with a distance ceiling of 3 and 4 times the direct distance, and with all paths. All of the progressive refinement variations were run with the bottom 10% and 20% of the teams removed after each level, starting at level 1 (the second level). The runtime results are displayed in Table 43. All nine variations had the same winners and second place team/path combinations. The run with a ceiling of three times the direc t distance and a 20% removal rate had the best time. The only surprise was in the all paths scenario, the 20% removal performed slightly worse than the 10% removal.

PAGE 80

80 Table 42. Resources available in scenario 12 Name Type Length Width Weight Speed De ploy Cost Hourly Rate Skill B9 Robot 0.751 0.751 35 2.00 600 60 Heat Sensor Barney Human 0.152 0.152 0.80 2200 220 Authorize Fire Control Drone Cylon Robot 0.747 0.610 120 1.60 100 80 Carry Weapon Fred Human 0.152 0.152 150 0.78 1500 150 Contr ol Drone ID Target H1 Human 0.152 0.152 150 0.80 1900 190 Authorize Fire ID Target H2 Human 0.152 0.152 180 0.90 1500 275 Control Drone Radio Huey Robot 0.305 0.305 23 1.10 500 75 Heat Sensor M1 Robot 0.457 0.381 35 1.10 1000 75 Carry Wounded Mitchell Human 0.152 0.152 160 0.90 1100 200 Carry Weapon Authorize Fire Mule Robot 1.000 1.000 35 2.20 1200 110 Carry Wounded R1 Robot 0.305 0.305 30 1.90 400 60 Authorize Fire Heat Sensor Robbie Robot 0.751 0.686 35 1.00 700 70 Heat Sensor Talon Robo t 0.457 0.305 30 1.20 600 66 Heat Sensor ID Target T2 Robot 0.152 0.152 35 1.50 750 250 Authorize Fire Control Drone Viper Robot 0.610 0.457 85 2.20 1500 150 ID Target

PAGE 81

81 Table 43. Evaluation of approaches, scenario 1 Approach Runtime, ms Hybrid, low c eiling, 20% cut 7235 Hybrid, low ceiling, 10% cut 9214 Brute force, low ceiling 9828 Hybrid, high ceiling, 20% cut 12147 Hybrid, high ceiling, 10% cut 15870 Brute force, high ceiling 20952 Progressive refinement, all paths, 20% cut 50371 Progressive refinement, all paths, 10% cut 63092 Brute force, all paths 69829 Mission The team must travel from NW E Street 3 to the hostage location in NW Area 1. The area layout is displayed in Figure 44. There are trouble spots near the intersection of the m ain roads that pass through the NW Block, as indicated on the map. Skills required: Authorize Fire Carry Weapon Carry Wounded Heat Sensor Winning scenario time Score time = 107.55, cost = $3105.97 Team: R1, Mule, Viper Path: As shown in Figure 45. Second place scenario time Score time = 127.71, cost = $1707.10 Members: Cylon, R1, Mule Path: As shown in Figure 45.

PAGE 82

82 Figure 44. Scenario 1 layout. Starting point is at the arrow on NW E Street 3, goal is NW Area 1, marked with a star. Danger areas and their values are marked. Winning scenario cost Score cost = $1510.32, time = 185.77 Team: R1, M1, Cylon Path: As shown in Figure 45, followed by several other paths up to 642.62 time, $1535.70 cost Second place team, cost Score time = 127.71, cost = $1707.10 Members: Cylon, R1, Mule Path: As shown in Figure 45.

PAGE 83

83 Figure 45. Scenario 1 winning path. Path displayed in red is the winning path. Both first and second place teams in timing and cost chose the sa me path. Results, low ceiling This simulation generated 150 possible team combinations and 27 leaf level paths, using a ceiling of 3 times the direct distance. The scores for level 4 ranged a time of 107.55 to 1249.50 and cost of $1510.32 to $5636.70. Overall scores were computed based on speed and cost independently. In this scenario, there was a single time winner and a single different cost winner with multiple paths, so the priority, either speed or cost, would determine which team was chosen. How ever, the same team/path came in second place for both cost and time, so with a cost function that provided equal balance between cost and time, the second place team would likely be chosen.

PAGE 84

84 With a bruteforce simulation, all 150 teams are simulated agai nst all 27 paths. In the case where 10% of the lowest performing teams are cut each level, the first level runs 150 teams with 1 path, the second level has 135 teams and 23 paths, and the third has 121 teams and 27 paths. By cutting the bottom 20% teams with each level, the same winning teams are generated with 150, 120, and 96 teams for each level. This configuration performed the best, with no loss of accuracy. Results, higher ceiling This simulation generated 150 possible team combinations and 68 lea f level paths, using a ceiling of 4 times the direct distance. The scores for level 4 ranged a time of 107.55 to 1477.14 and cost of $1507.94 to $5663.57. Overall scores were computed based on speed and cost independently. The winners of this scenario, which was less restrictive on the paths included in the simulation, were identical to the more restrictive ceiling. This outcome shows that, for this data set, the bounding of 3 times the ceiling was sufficient. The number of teams run for each level was consistent with the low ceiling results. Results, all paths This simulation generated 150 possible team combinations and 287 leaf level paths, using no cost ceiling. The scores for level 4 ranged a time of 107.55 to 1881.06 and cost of $1510.32 to $5687.10. Overall scores were computed based on speed and cost independently. The winning results, as expected, were unchanged from the heuristic based runs; however the run time was significantly higher. Final Prototype, Scenario Two The second scenario included the same 15 potential members and the same area, with the exception of a new obstruction. The same set of 9 configurations was run

PAGE 85

85 against that scenario. The results shown in Table 44 show that once again the hybrid approach with a 20% reduction in teams per level performed the best. Table 44. Evaluation of approaches, scenario 2 Approach Runtime, ms Hybrid, low ceiling, 20% cut 5455 Hybrid, low ceiling, 10% cut 7479 Brute force, low ceiling 7840 Hybrid, high ceiling, 20% cut 10007 Hybri d, high ceiling, 10% cut 11255 Brute force, high ceiling 13742 Progressive refinement, all paths, 20% cut 32857 Progressive refinement, all paths, 10% cut 49261 Brute force, all paths 42343 Mission The team must travel from NW E Street3 to the hostag e location in NW Area 1. The area layout is displayed in Figure 4 6 There are trouble spots near the intersection of the main roads that pass through the NW Block and a vehicle is partially blocking the intersection between NW_3ERoad1, NW_3ERoad2, and N W_5Street2. This blockage allows only the humans and a few of the UMS members to travel that route. Skills required: Authorize Fire Carry Weapon Carry Wounded Heat Sensor Winning scenario time Score time = 132.79, cost = $ 3107.37 Team: R1, Viper, Mule Path: As shown in Figure 47. Second place scenario time Score time = 142.55, cost = $ 3107.92 : Team: R1, Viper, Mule Path: As shown in Figure 48.

PAGE 86

86 Figure 46. Layout for scenario 2. There is an obstruction in the int ersection between NW 3 E Road 1, NW 5 Street 2, and NW 3 E Road 2 that reduces the width of the passage. Figure 47. Winning path for scenario 2. Because of the blockage and the higher percentage of danger, the most effective path is to go up a bloc k to NW 4 Street 2 and back down.

PAGE 87

87 Winning scenario cost Score time = 229.37, cost = $ 1512.74 Team: M1, Cylon, R1 Path: As shown in Figure 47, plus several other paths for same team at increasing costs Second place team, cost Score time = 157.69, cost = $ 1708.76 Team: Cylon, R1, Mule Path: As shown in Figure 47. Figure 48. Second place path for scenario 2. Results, lower ceiling This simulation generated 150 possible team combinations and 27 leaf level paths, using a ceiling of 3 times the direc t distance. The scores for level 4 ranged a time of 142.55 to 1231.91 and cost of $ 1512. 92 to $ 5651 07. Overall scores were

PAGE 88

88 computed based on speed and cost independently. In this scenario, there was a single time winner and a single different cost winner with several different paths so the priority, speed or cost, would determine which team was chosen. However, the same team/path came in second place for both cost and time, so given a cost function that adequately combined cost and time, the second place team would likely be chosen. The lower ceiling, however, caused the winning path/team combination for both time and cost to be eliminated. Results, higher ceiling This simulation generated 150 possible team combinations and 68 leaf level paths, usi ng a ceiling of 4 times the direct distance. The scores for level 4 ranged a time of 132 79 to 1455.43 and cost of $ 1512. 7 4 to $5653 2 1. Overall scores were computed based on speed and cost independently. In thi s scenario, one team took first and secon d places with 2 different paths, with scores ranging from 132. 79 to 1 49.44. A different team led the cost with 40 different paths and costs ranging from $1512.74 to 1535.46 This scenario, with the higher ceiling on the heuristic, came in with a differen t winner than the lower ceiling, with a lower time score and a lower cost score. The early, heuristics based path planning is designed to choose the paths most likely to be favorable, without knowledge of individual team constraints. In this particular c ost function, the danger level was used to choose the type of movement efficient, slow and cautious, or a slower leapfrogging move called bounding overwatch. These movements are dependent on individual resource speeds and widths, and overall team values as opposed to the calculation done in the path selection, which is based purely on length and danger. The higher ceiling and the all paths results were the same, so

PAGE 89

89 the heuristics based planning was still an improvement; it just needed a higher ceiling to get the correct answer. Results, All Paths This simulation generated 150 possible team combinations and 287 leaf level paths, using no cost ceiling. The scores for level 4 ranged a time of 132.79 to 1858.08 and cost of $1512.74 to $5678.30. Overal l scores were computed based on speed and cost independently. All three versions progressive refinement with 80% retention, progressive refinement with 90% retention, and brute force got the same values for the winners, with more results on the slow s ide for the progressive refinement steps. Final Prototype, Scenario Three The third scenario included the same 15 potential members and the same area as scenario two, with a change in the skills. The list of resources is displayed in Table 46. In the first two scenarios, skills were assigned to members somewhat randomly, including authorization to fire being held by both human and UMS resources. Due to the estimated costs and speeds of the UMS resources, this caused all of the winning members to be unmanned a scenario that technology is not ready for. By the same token, technology and society is not ready, nor may it ever be ready, to give nonhumans the power to authorize weapons fire. This scenario was created to remove authorization to fi re from the UMS members both to accurately reflect the technology, and to provide a mechanism to force at least one human onto the team. The same set of 9 configurations was run against that scenario. The results shown in Table 45 show that once again the hybrid approach with a 20% reduction in teams per level performed the best.

PAGE 90

90 Table 45. Evaluation of approaches, scenario 3 Approach Runtime, ms Hybrid, low ceiling, 20% cut 3474 Hybrid, low ceiling, 10% cut 4210 Brute force, low ceiling 5656 H ybrid, high ceiling, 20% cut 6689 Hybrid, high ceiling, 10% cut 7980 Brute force, high ceiling 7680 Progressive refinement, all paths, 20% cut 20888 Progressive refinement, all paths, 10% cut 23169 Brute force, all paths 22805 Mission The team must travel from NW E Street3 to the hostage location in NW Area 1. The area layout unchanged from the previous scenario, is displayed in Figure 4 6 There are trouble spots near the intersection of the main roads that pass through the NW Block and a vehicle is partially blocking the intersection between NW_3ERoad1, NW_3ERoad2, and NW_5Street2. This blockage allows only the humans and a few of the smaller UMS members to travel that route. Skills required: Authorize Fire Carry Weapon Carry Wounded Heat Sensor Winning scenario time and cost Score time = 224.22, cost = $ 2512.46 Team: R1, Mitchell, M1 Path: As shown in Figure 49. Second place scenario time, and cost second place team Score time = 224.22, cost = $ 2612.46 Team: Huey, Mitchell, M1 Path: As shown in Figure 49.

PAGE 91

91 Figure 49. Winning path, scenario 3 Results, lower ceiling This simulation generated 60 possible team combinations and 27 leaf level paths, using a ceiling of 3 times the direct distance. The scores for level 4 ranged a time of 142.55 to 1231.91 and cost of $1512. 92 to $ 5651. 07. Overall scores were computed based on speed and cost independently. In this scenario, three teams tied for first place, with slightly different costs one of which was also the cost winner. As was the case with the other two scenarios, the team winning in cost had several more paths that came in before the next team. All nine versions progressive refinement with 80% retention, progressive refinement with 90% retention, a nd brute force; all paths, high

PAGE 92

92 ceiling, low ceiling got the same values for the winners, with more results on the slow side for the progressive refinement steps. The progressive refinement with an 80% retention and the lower ceiling had the best runti me with no loss in accuracy for the winning team/path combinations. The 80% retention progressive refinement approach processed 60 teams and 1 path for the first pass, 48 teams with 23 paths for the second pass, and 38 teams with 27 paths for the third pass. The 90% retention progressive refinement approach processed 60 teams and 1 path for the first pass, 54 teams with 23 paths for the second pass, and 48 teams with 27 paths for the third pass. Results, higher ceiling This simulation generated 60 possi ble team combinations and 68 leaf level paths, using a ceiling of 4 times the direct distance. The scores for level 4 ranged a time of 224.22 to 1858.08 and cost of $2512. 46 to $ 5653. 2 1. Winning results are identical to the lower ceiling results across all three versions. The 80% retention progressive refinement approach processed 60 teams and 1 path for the first pass, 48 teams with 36 paths for the second pass, and 38 teams with 68 paths for the third pass. The 90% retention progressive refinement a pproach processed 60 teams and 1 path for the first pass, 54 teams with 36 paths for the second pass, and 48 teams with 68 paths for the third pass. Results, All Paths This simulation generated 60 possible team combinations and 287 leaf level paths, using no cost ceiling. The scores for level 4 ranged a time of 224.22 to 1858.08 and cost of $2512.46 to $5678.30. Overall scores were computed based on speed and cost independently. The runtime was slightly faster for the brute force than the 90% retention progressive refinement, but the 80% retention progressive refinement had the

PAGE 93

93 lower runtime. This is due to the smaller number of teams with the given configuration, so fewer teams are cut from the 90% progressive refinement, thus more iterations are run than the brute force, which runs through all of the teams, all of the leaf nodes. The 80% retention progressive refinement approach processed 60 teams and 1 path for the first pass, 48 teams with 78 paths for the second pass, and 38 teams with 287 paths f or the third pass. The 90% retention progressive refinement approach processed 60 teams and 1 path for the first pass, 54 teams with 78 paths for the second pass, and 48 teams with 287 paths for the third pass.

PAGE 94

94 Table 46. Resources available in scenario 3 Name Type Length Width Weight Speed Deploy Cost Hourly Rate Skill B9 Robot 0.751 0.751 35 2.00 600 60 Heat Sensor Barney Human 0.152 0.152 0.80 2200 220 Authorize Fire Control Drone Cylon Robot 0.747 0.610 120 1.60 100 80 Carry Weapon Fre d Human 0.152 0.152 150 0.78 1500 150 Control Drone ID Target H1 Human 0.152 0.152 150 0.80 1900 190 Authorize Fire ID Target H2 Human 0.152 0.152 180 0.90 1500 275 Control Drone Radio Huey Robot 0.305 0.305 23 1.10 500 75 Heat Sensor M1 Robot 0.457 0.381 35 1.10 1000 75 Carry Wounded Mitchell Human 0.152 0.152 160 0.90 1100 200 Carry Weapon Authorize Fire Mule Robot 1.000 1.000 35 2.20 1200 110 Carry Wounded R1 Robot 0.305 0.305 30 1.90 400 60 Heat Sensor Robbie Robot 0.751 0.686 35 1.00 70 0 70 Heat Sensor Talon Robot 0.457 0.305 30 1.20 600 66 Heat Sensor ID Target T2 Robot 0.152 0.152 35 1.50 750 250 Control Drone Viper Robot 0.610 0.457 85 2.20 1500 150 ID Target

PAGE 95

95 CHAPTER 5 CONCLUSIONS Effectiveness of Hybrid Approach The results of t he testing shows that adding a heuristic semi optimal path selection to the progressive refinement approach does significantly improve the efficiency of the algorithm in large data sets. In each testing scenario, the hybrid approach using the most restrictive cap on the path during path selection and the most aggressive culling of poorly performing teams was the fastest overall. In one of the scenarios the more restrictive, aggressive hybrid eliminated the winning path, but the most optimal team was still a front runner using a different path. That missing path was found when the cap on the path length was raised slightly. Our testing shows that with properly chosen heuristics, the hybrid approach can achieve the same accuracy, with far lower cost, as t he progressive refinement alone or the brute force simulation. With a complex scoring function, the path selection heuristics must be chosen with care. This also holds true for paths with an indirect route. For instance, if the starting and goal points are fairly close together but separated by an impenetrable boundary, the cap on path length must be increased to get any paths. In addition, a heuristic in that scenario might be the change in distance between current node and goal node while some movement away from the goal is necessary, a constant trending in the wrong direction is likely to yield a costly path. The heuristic multiple path search, as opposed to all paths or single best path search, showed an improvement in processing time when combined with each of the simulation methods the tighter the cap, the faster the ensuing simulation. Conversely, the manner of the simulation also makes a difference. In a scenario where

PAGE 96

96 there are few teams, a leaf layer bruteforce simulation can sometimes be faster than a conservative progressive refinement simulation that is, a progressive refinement where fewer of the lower performing teams are eliminated before the next level. A more aggressive elimination was faster than brute force in all instances however. On their own, the heuristic multiplepath search and progressive refinement each displayed performance gains with no loss to accuracy; combined, they made even greater performance improvements. Our testing indicates that the performance impr ovements are greater with a bigger data set, especially with the aggressive progressive refinement. Contributions Our goal in this research was to contribute to the knowledge in two broad areas, as designated by the ACMs Computing Classification System Artificial Intelligence and Simulation. The contribution to Artificial Intelligence was twofold the first contribution was the definition of a methodology to determine a near optimal team from a larger pool of resources, both human and robot, each in dividual possessing a set of useful skills and having unique operational and physical characteristics. The second contribution was a bridge between Artificial Intelligence and Simulation in the form of a hybrid heuristic path planning/multiple discrete event simulation approach, using progressive refinement. We have demonstrated in the prototype that there is an efficiency advantage to creating that hybrid; in a dynamic immediateneed situation, that efficiency advantage could be key. Future Work The scoring scenarios used in our study used simplistic team movement. A future enhancement to this simulation would be to incorporate more realistic team movements

PAGE 97

97 that include having members split up to go different routes, including the frequently used sce nario where the UMS travels ahead of the human members to test environmental factors, such as air quality and ground stability. Subject matter experts in the field of the task would need to be interviewed to accurately model the team movements. Also, a s different team operational scenarios are devised, more complex scoring functions should be developed to include factors such as a members proficiency at a given skill. Other intangibles that should be considered are minimum or maximum requirements for h umans or robots. Based on the cost functions that relied exclusively on speed and cost, the winning teams all ended up comprised of all robots. This prompted the third scenario, in which one of the skills was limited to human members. Current technology while improving, is not mature to the point of having all robots on an autonomous team. Identifying skills that only a human, or only a robot, can perform would ensure that at least one human and at least one robot are members of the final team. Vi sualization plays a large part in conveying information and understanding. While graphical user interfaces were built for the prototype for entering resources and selecting mission parameters, the task of building a visual, interactive display of the area fell beyond the scope of our work. This would be a most welcome addition to the project, not only for aesthetics and quicker conveyance of results, but also to aid in the task of entering the map data, which was a tedious, timeconsuming task of hand cre ating ntriples. A good visual editor with the capability to add other metadata like weight limits and probability of danger and to generate paths and entryways automatically would have eliminated mistakes and a great deal of eye strain. Automatic generation of map data is crucial for larger areas of interest.

PAGE 98

98 Another improvement would be to add the third dimension to the prototype, and to include unmanned air vehicles (UAV). UAVs are increasingly crucial resources in search and rescue operations and in military operations, and the technology is maturing to the point where UAVs can act more autonomously. The addition of UAVs would require not only 3D modeling and planning, but also more complex team behaviors, since limitations on UAVs are different t han limitations on groundbased team members. The heuristic graphbased search fit the pathplanning domain, but could be used for other applications as well. Social network analysis, or friendof a friend discovery, is a common graphwalking application. Used to discover relationships between people, entities, or events, heuristics can be applied based on strength of relationships, allowing the user to investigate several possible linkages between two entities. For instance if John knows Mary, and Mary is Anns sister, and Fred shops at the same store as Mary, the strength of the relationship between John and Mary is stronger than Fred and Mary, so John may be more directly correlated to knowing Ann as well. The work done in team planning could go bey ond the urban search and rescue scenario used in the prototype. It could be used in an agricultural or factory setting to determine which combination of equipment would function most efficiently for a given task. The costing function might be adapted to penalize teams that have idle time faster members having to wait for slower members, resulting in a more evenly matched team as opposed to the fastest overall.

PAGE 99

99 LIST OF REFERENCES [1] A. Carnegie. (2009, November 23). Quote s for Teamwork [Online]. Availab le: http://www.motivatingquotes.com/teamwork.htm [2] J. Averill, et al "Occupant Behavior, Egress, and Emergency Communications Final Reports of the Federal Building and Fire Investigation of t he World Trade Center Disaster National Institute of Standards and Technology, Gaithersburg, MD, Rep. NIST NCSTAR 1 7, Sept. 2005. [3] Wikipedia contributors. (2009, November 18). September 11 attacks Wikipedia, The Free Encyclopedia [Online]. Available: http://en.wikipedia.org/wiki/September_11_attacks [4] The Mount Sinai Hospital / Mount Sinai School of Medicine (2009, November 3). World Trade Center Responders Plagued With Asthma; 9/11 Responders Twice As Likely To Have Asthma, ScienceDaily [Online]. Available: http://www.sciencedaily.com /releases/2009/11/091103144818.htm [5] University of South Florida. (2009, November 18). History of Rescue Robots Center for Robot Assisted Search & Rescue [Online]. Available: http://www.crasar.org/rescuerobots/history.htm [6] A. Boyle, (2005, August). How hightech is coming to the rescue: Scientists bring gadgets to post Katrina disaster scene [Online]. Available: http://www.msnbc.msn.com/id/9131498/ [7] R. Murphy, J. Kravitz, S. Stover, and R. Shoureshi, Mobile Robots in Mine Rescue and Recovery, IEEE Robotics & Automation Magazine, vo l 16, no. 2, pp. 91103, 2009. [8] T. Cioppa, J. Willis, N. Goerger, L. Brown, Research Plan Development for Modeling and Simulation of Military Operations in Urban Terrain, Proceedings of the 2003 Winter Simulation Conference, vol. 1, pp. 10461051, 2003. [9] Floyd D. Spence National Defense Authorization Act for Fiscal Year 2001, HR.4205.ENR, Section 220 (2001). [10] OUSD (AT&L) Defense Systems/Land Warfare and Munitions (2005). Joint Robotics Program Master Plan [Online]. Available: http://www.jointrobotics.com/library01.php [11] D. J. Bruemmer, D. D. Dudenhoeffer, J. Marble, Dynamic Autonomy for Urban Search and Rescue, Proceedings of the 2002 AAAI Mobile Robot Workshop, Edmonton, Canada, August 2002, p 3337.

PAGE 100

100 [12] D.J. Bruemmer, and M. Walton. Collaborative Tools for M ixed Teams of Humans and Robots, Proceedings of the Workshop on Multi Robot Systems Washington D.C., March 2003, pp. 219229. [13] T. Fong, C. Kunz, L Hiatt, M Bugajska, The HumanRobot Interact ion Operating System, Proc. 1st ACM SIGCHI/SIGART conference on Humanrobot interaction, Salt Lake City, Utah, 2001. [14] H. J. Levesque, P. R. Cohen, and J. H. T. Nunes. On Acting Together, Proceedings of AAAI 90 Boston, MA, 1990 pp. 94 99. [15] T. Laengle, T. Hoeniger, L. Zhu, Co operation in HumanRobot Teams, Proceedings of the IEEE International Symposium on Industrial Electronics, ISIE '97 Guimaraes, Portugal vol.3, pp. 12971301, 1997. [16] R. Murphy HumanRobot I nteraction in Rescue Robotics, IEEE Tran sactions on Systems, Man, and Cybernetics Part C: Applications and Reviews, vol. 34, n o. 2 pp. 138 153, May 2004. [17] I. Nourbakhsh, et al, HumanRobot Team ing for Search and Rescue, IEEE Pervasive Computing: Mobile and Ubiquitous Systems pp. 7278, 20 05. [18] R. Chalmers, D. Scheidt, T. Neighoff, S. Witwicki, R. Bamberger, C ooperating Unmanned Vehicles, Proceedings of the AIAA 1st Intelligent Systems Technical Conference, 2004. [19] F. Lacombe, Modeling Swarm Behavior, P HYS org.com 21 February 2006. [Online] Available: http://www.physorg.com/news11060.html [20] D. Scheidt, J. Stipes, Cooperating Unmanned Vehicles, Proceedings of the 2005 IEEE International Conference on Networking, Sensing and Control pp. 326331, 2005. [21] M. Taylor, M. Jain, C. Kiekintveld, J. Kwak, R. Yang, Z. Yiu, M. Tambe, Two Decades of Multiagent Teamwork Research: Past, Present, and Future Postproceedings of Collaborative Agents R E search and Development (CARE) 2010 workshop 2011 [22] T. Gupta, P. Varakantham, T. Rauenbusch, M. Tambe, Demonstration of Teamwork in Uncertain Domains using Hybrid BDI POMDP Systems, Proc. 6th Int. Joint Conf. on Autonomous Agents and Multiagent Systems (AAMAS), art icle 264, 2007. [23] S. Prabhala, J. Gallimore, S. Narayanan, Human Effectiveness Issues in Simul ated Uninhabited Combat Arial Vehicles Proceedings of the 2003 Winter Simulation Conference, v ol. 1, pp. 1034 1038, 2003. [24] A. Freedy, O. Sert, E. Freedy, J. McDonough, G. Weltman, M. Tambe, T. Gupta, W. Grayson, P. Cabrera, Multiagent Adjustable Autono my Framework (MAAF) for

PAGE 101

101 Multi Robot, Multi Human Teams, Proceedings of the Symposium on Collaborative Technologies & Systems, pp. 498 505, 2008. [25] M. BazargainLari, P. Gupta, S. Young A Simulation Approach to Manpower Planning, Proceedings of the 2003 Winter Simulation Conference, pp. 16771685, 2003. [26] R. Guttkuhn, T. Dawson, U. Trutschel, J. Walker, M. Moroz, A Discrete Event Simulation for the Crew Assignment Process in North American Freight Railroads, Proceedings of the 2003 Winter Simulation Conf erence, pp. 16861692, 2003. [27] S. Ganapathy, R. Hill, Dynamic Path Planning for Search and Destroy Missions The Bay of Biscay Scenario, Proceedings of the 2003 Winter Simulation Conference, pp. 999 1003, 2003. [28] F. Kamrani, A. Rassul, SimulationAided Pat h Planning of UAV Proceedings of the 2007 Winter Simulation Conference, pp. 13061314, 2007. [29] C. Hill, L. Malone, Caveats For Simulation Modeling In Support of Decision Making, Proceedings of the 2003 Winter Simulation Conference, pp. 11021109, 2003. [30] J. Lee, P. Fishwick, Real Time SimulationBased Planning for Computer Generated Force Simu lation, Simulation, vol. 63, pp. 299 315, 1994. [31] D. Miller, The Role of Simulation in Planning, Proceedings of the 1987 Winter Simulation Conference pp. 530 533, 1 987. [32] D. Joslin, W. Poole, Agent Based Simulation for Software Project Planning, Proceedings of the 2005 Winter Simulation Conference, pp. 10591066, 2005. [33] G. Drake, J. Smith, Simulation System for Real Time Planning, Scheduling, and Control, Proceedin gs of the 2006 Winter Simulation Conference pp. 10831090, 2006. [34] J. Lee, P. Fishwick, Simulation Based Real Time Decision Making for Route Planning, Proceedings of the 1995 Winter Simulation Conference, pp. 10871095, 1995. [35] H. Nejad, D. Zhu A. Mosleh, Hierarchical Planning and Multi Level Scheduling for Simulation Based Probabilistic Risk Assessment, Proceedings of the 2007 Winter Simulation Conference pp. 11891197, 2007. [36] X Hu, Y Li, J Guo L Sun and A. Zeng A Simulation Optimization Algorithm With Heuristic Transformation and its Applicati on to Vehicle Routing Problems, International Journal of Innovative Computing, Information and Control, v ol. 4 n o. 5 pp. 11691181, May 2008.

PAGE 102

102 [37] R. Hoeckley, B. K. Williams, and J. A. Kent, The Monte Carlo Ap proach to Mine Warfare Navy League of the United States, February 2001. [Online]. Available: http://www.navyleague.org/sea_power/feb_01_05.php [38] P. Svenson, C. Martenson, SB Plan: Simulat ion Based Support for Resource Allocation and Mission Planning Proceedings of the Conference on Civil and Military Readiness (CIMI 2006) Enko ping, Sweden, 2006. [39] B.J.H. van Basten, Path Planning and Online Obstacle Avoidance in Weighted Regions, Master s Thesis, Utrecht University August 2006. [40] X. Ning, S. Shihuang, F. Xizhou, A Fuzzy Approach to the Weighted Region Problem for Autonomous Vehicles, Proceedings of the 1993 International Symposium on Intelligent Control Chicago, IL, August 1993. [41] N. Mote e, A. Jadbabaie, G. Pappas, A Duality Theory for Path Planning for Multiple Vehilcles, IEEE Transactions on Robotics (under review), 2011. [42] C. Aggarwal, H. Wang, Graph Data Management and Mining: A Survey of Algorithms and Applications, Managing and Mining Graph Data, Advances in Database Systems 40, DOI 10.1007/9781 441960450_2, Springerlink, 2010. [43] G. Coombe, M. Harris, Global Illumination Using Progressive Refinement Radiosity, GPU Gems 2 Taunton, MA: AddisonWesley, 2005, ch. 39. [44] M. Cohen, S. C hen, J. Wallace, D. Greenberg, A Progressive Refinement Approach to Fast Radiosity Image Generation, Proceedings of the 15th Annual Conference on Computer Graphics and Interactive Techniques New York, NY, pp. 7584, 1988. [45] S. Nandy, R. Narayan, V. Visvanathan, P. Sadayappan, P. Chauhan, A Parallel Progressive Refinement Image Rendering Algorithm on a Scalable Multithreaded VLSI Processor Array, International Conference on Parallel Processing, 1993. ICCP 1993, vol. 3, pp. 9497, Aug 1993. [46] N. Holzschuch, F. Sillion, G. Drettakis, An Efficient Progressive Refinement Strategy for Hierarchical Radiosity, 5th Eurographics Workshop on Rendering, 1994. [47] R. Blanford, Progressive Refinement Using Local Variance Estimators, IEEE Transactions on Communications vol. 41, no. 5, pp 749759, 1993. [48] R. Rosenbaum, H. Schumann, Progressive Refinement: More than a Means to Overcome Limited Bandwidth, Proceedings of SPIE 7243, San Jose, CA, 2009. [49] K. Wagstaff, M. Kocurek, D. Mazzoni, B. Tang, Progressive Refinement for Support Vector Machines, Data Mining and Knowledge Discovery 2009.

PAGE 103

103 [50] J. Wang, M. Lewis, J. Gennari, A Game Engine Based Simulation of the NIST Urban Search and Rescue Arenas, Proceedings of the 2003 Winter Simulation Conference, vol. 1, pp. 10391045, 20 03. [51] C. Karr, Conceptual Modeling in OneSAF Objective System (OOS), Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2005, vol. 2005, no. 1, 2005. [52] M. Morgan, M. Henrion, M. Small, Uncertainty: A Guide to Dealing with Uncertai nty in Quantitative Risk and Policy Analysis 1990, Cambridge University Press, 1990, pp. 277278. [53] F. Brooks, Jr., Mythical ManMonth, The: Essays on Software Engineering, Anniversary Edition 1995, p. 118. [54] B. Mitchell and L. Yilmaz Symbiotic A daptive M ultisimulation: An A utonomic S imulation F ramework for R eal T ime D ecision S upport U nder Uncertainty, ACM Trans. Model. Comput. Simul. vol. 19, no. 1, pp. 1 31, 2008. [55] Association for Computing Machinery, The 1998 ACM C omputing Classification System, [O nline]. Available: http://www.acm.org/about/class/1998/ [56] D. Shin, E. Chong, H. Siegel, A Multiconstraint QoS Routing Scheme using the DepthFirst Search Method with Limited Crankbacks, 2001 IEEE Workshop on High Performance Switching and Routing, pp. 385389, 2001. [57] S. Lim, S. Yu, M. Kang, C. Ham, Robot Palletizing Simulation Using Heuristic Pattern Generati on and Trajectory Optimization, SICE ICASE 2006, International Joint Conference, Bexco, Busan, Korea pp. 22272232, 2006. [58] ( February 2004), G. Kkyne, J. Carroll, Resource Description Framework (RDF): Concepts and Abstract Syntax W3C Recommendation [Online]. Available: http://www.w3.org/TR/rdf co ncepts/ [59] (2011, Feb. 12), Stanford Center for Biomedical Ontology, Protg [Online]. Available: http://protege.stanford.edu/ [60] B. Smith, C. Wety, Ontology: Towards a New Synthesis, Proceedings of the Inter national Conference on Formal Ontology in Information Systems (FOIS2001). ACM Press, 2001 [61] (2011, Feb. 12), Franz, Inc. [Online]. Available: http://www.franz.com/ [62] (2011, Feb. 12), N. Walsh, A Technical Introduction t o XML OReilly XML.com October 1998. Available: http://www.xml.com/pub/a/98/10/guide0.html [63] G. Brassard, P. Bratley, Fundamentals of Algorithms Englewood Cliffs, NJ: Prentice Hall, 1996, pp. 198 202.

PAGE 104

104 [64] S. Russell, P. Norvig, Artificial Intelligence A Modern Approach. Upper Saddle River, NJ: Prentice Hall, 2003, pp. 94 105. [65] S. Russell, P. Norvig, Artificial Intelligence A Modern Approach. Upper Saddle River, NJ: Prentice Hall, 2003, pp. 75 76. [66] A. Bar Zeev, Scengraphs: Past, Present and Future, Reality Prime 2003 [Online] Available: http://www.realityprime.com/scenegraph.php [67] Wikipe dia contributors, Scene Graph, Wikipedia, Th e Free Encyclopedia, 2007 Feb 1, 11:16 UTC [cited 2007 Feb 2]. Available: http://en.wikipedia.org/w/index.php?title=Scene_graph&oldid=104818809. [68] P. Fishwick, Simulation Model Design and Execution. Englewood Cliffs, NJ: Prentice Hall, 1995, pp. 165169.

PAGE 105

105 BIOGRAPHICAL SKETCH Teresa Nieten received her Bachelor of Science degree in computer and information science and engineering at the University of Florida in 1992, and her Master of Science in Computer Engineering at the University of Florida in 2004. She has worked on command and control systems in energy management, spaceflight, and robotics areas; simulation systems for mission planning and robotics; and telecommunications. She is currently a Senior Systems Engineer with Modus Operandi, Inc., in Melbour ne, FL working with natural language processing. Teresa is married to fellow PhD student Daniel Nieten and has a son, Brandon.