JOurn.31 ofr iLn.err.3dua.3- Re--search
,,Oluiine 5, S iue I - 1.3, 2i:1114
Agent-Based Framework for Dynamic Supply Chain Configuration
Supply Chain Management has gained renewed interest among researchers in recent years. This is primarily due
to the availability of timely information across the various stages of the supply chain, and therefore the need
to effectively utilize the information for improved performance. In this paper we develop a framework, with
machine learning, for automated dynamic supply chain configuration. Recent developments in
eCommerce applications and faster communication over the Internet in general necessitate dynamic (re)
configuration of supply chains over time to take advantage of better configurations. The supply chain models
each actor as an agent who makes independent decisions based on information gathered from the next
level upstream. Examples show performance improvements of the proposed adaptive supply chain
configuration framework over static configurations.
Mismatches in demand and supply arise primarily due to market volatility. And, there are opportunity costs that
are associated with these mismatches (e.g., Radjou, 2002). Examples include decrease in quarterly earnings in
1996 by $900 million for General Motors due to an 18-day labor strike at a brake supplier factory that idled
workers at 26 assembly plants and Boeing's $2.6 billion loss in 1997 due to failure of two key suppliers to
deliver critical parts on time. Information and inventory have been identified as factors that work synergistically
to enable better performance of supply chains (e.g., Alles et al., 2000; Cachon and Fisher, 2000; Hariharan
and Zipkin, 1995; Mukhopadhyay et al., 1997; Whang, 1993; Woolley, 1997).
Automated supply chain configuration is beneficial when there are changes in cost of products/services,
resource availability, and customer demands. This assumes that for a given order there are several feasible
supply chain configurations that can deliver the product. The number of such feasible configurations increases
with the number of stages, products, suppliers, etc.
This paper aims to address the following specific question using examples: Does dynamically switching
among appropriate nodes in a stage result in improvements in (a) revenue, (b) effectively serving the
customer based on the percentage of orders fulfilled as desired by the customer?
The proposed framework learns to associate the best node(s) at each stage of the network for each combination
of order attributes (price, lead-time, quantity, etc.) in the system. It assumes that the products or parts that
pass through each node in every stage are of the same quality.
Automated Supply Chain Configurer (ASCC)
Additinal VERlOLMASCE Kaowledge
ï¿½a>? | ,E"E..E r
SAMPLER N EAI -BASE
w order 3farmnIIMn
afrom Dtt (dwnsuf*m)
Bfs choe of uode(i)
from prwoui (uipitu)
Figure 1. Automated Supply Chain Configurer (ASCC) Framework
AUTOMATED SUPPLY CHAIN CONFIGURER FRAMEWORK
The proposed Automated Supply Chain Configurer (ASCC) framework is given in Figure 1. ASCC itself can
be considered an agent that resides at every node (except for the final node upstream, without loss of generality
in this study, since those nodes are assumed to not make choice decisions) in the supply chain. Each of these
agents makes myopic decisions based on the information they have about the nodes in the next stage upstream
to them and the order information that comes from a stage downstream from them. Although the local
decisions made are myopic in character, it can be shown that these series of myopic decisions do indeed lead to
the best overall performance from start to finish. This is achieved by selecting the best available option at each
stage. The following section illustrates the ASCC framework.
EXAMPLES USING THE PROPOSED FRAMEWORK
A two-stage supply chain illustrates the proposed framework. Assume a transaction begins when customers send
in their orders through a web interface. Based on the order specifications (e.g., product and quantity of each
product ordered, length of time the customer is willing to wait till the order is shipped, price the customer is willing
to pay, etc.), the order is routed to the most appropriate supplier.
pt% aim (up&inga)
Figure 2.Two-stage supply chain with two suppliers
A Two-Stage Supply Chain Example
Consider the case where there is only one product (P) that is supplied by two types of suppliers SA and SB
(Figure 2). Assume that the suppliers are capable of supplying different quantities of the product with different
lead-times and prices. When overlaps in price/quantity/lead-time combinations occur, the system chooses the
most appropriate supplier for that particular combination. Table I provides the information for choosing the
most appropriate supplier.
The data in Table I are read as follows: The first line indicates that supplier SA can supply up to 30 units of product
P with a lead-time of 1-5 days at a unit cost of 97.
Product Information for Each Supplier
Lead Time (Days)
Clearly neither SA nor SB is the best choice under all circumstances. For example, if an order requires 100 Ps to
be delivered in 1 day, it is not feasible in this problem context. In such cases, since the suppliers in this system
(SA and SB) cannot fulfill the order, the customer balks unless the order can be filled from inventory.
The data in Table I to generates training examples for the learning module in ASCC. An example of the set of
rules used to select supplier SA or SB using just quantity and lead-time information is as follows:
1. IF (Quantity >= 30) & (LeadTime >= 3) THEN SA.
2. IF (Quantity >= 30) & (LeadTime > 10) THEN SA.
3. IF (Quantity >= 30) & (3 < LeadTime >= 10) THEN SB.
4. IF (Quantity > 30) & (8< LeadTime >= 10) THEN SB.
5. IF (Quantity > 30) & (LeadTime > 10) THEN SA*
6. IF (30 < Quantity >= 70) & (LeadTime >= 8) THEN Sg.
7. IF (Quantity > 100) & (LeadTime >= 8) THEN SA*
8. IF (70 < Quantity >= 100) & (LeadTime < 6) THEN SA*
9. IF (70 < Quantity >= 100) & (5 < LeadTime < 9) THEN SA*
An over-simplified example here illustrates the proposed framework.
On arrival of the order, the pattern (e.g., want 110 units of product in 12 days) in the order is
matched with the most appropriate rule in the rule base and the most appropriate action is taken as
per this rule (e.g., supplier SA based on the rule 4).
Each supplier begins with a set amount of inventory, e.g. 75 units, purchased at a cost of 91 per unit.
The supplier first fills orders from inventory. When the inventory drops below a certain threshold, e.g.
53 units, the supplier replenishes inventory according to the framework in Table I. The inventory
costing applies the LIFO principle. When using two different batches of inventory to fill one order,
the supplier assesses the cost of the inventory by using a weighted average of the batch costs. Thus
the cost of filling an order of 130 units where the first 30 units cost 95 and the last 100 units cost
90 would be calculated as follows:
cost per unit = [(30*95) + (100*90)]/130
The framework uses Table I and assumes the following: orders arrive several times per day, as
generated by a lognormal(2,0.5) function where 2 represents the mean and 0.5 represents the
standard deviation; the amount of product P requested in each order varies uniformly from 1
through 150; the lead-time requested in each order varies uniformly from 1 through 15. As a
benchmark to compare the proposed system, the study uses two cases: one where every order is sent
to supplier SA, and another where every order is sent to supplier SB. Of course, the proposed
system sends the order to either SA or SB as per the specifications in the order.
The framework models the cost of inventory to include the cost of warehousing excess inventory.
This cost is added to the general unit cost to get the effective unit cost. Let a = a penalty constant
of 0.01. The effective cost of each order is modeled as:
[unit cost * quantity] + [a * unit cost * inventory]
The unit selling price is 110. The study simulates the process for 372 days with a warm-up period of
7 days, and collects necessary statistics for 365 days. Results are provided in Table 2.
Results for Two-Stage Supply Chain
Inventory Level Supplier With No
Inventory Level n to
Supplier (Max, Min) Profit % Balked
SAB (75, 53) $3,382,568 19.56
Suppe Wh I r Supplier With Inventory
Supplier With Inventory Penalty
Profit % Balked Profit % Balkedp
$3,554,909 7.50 $3,393,808 7.50
SA (75, 53)
SB (75, 53)
SAB (97, 75)
SAB (75, 53)
Results for profits in the inventory case are listed under Profiti for profits without any penalty
for inventory and Profitip for profits with a penalty for inventory. Here, the inventory volume varied
as given in the second column, with the first number representing the maximum inventory level and
the second number representing the minimum threshold level inventory reaches where
reordering begins. Supplier SA refers to the case where all the orders were directed to supplier
SA. Supplier SAB refers to the case when the proposed framework was used to direct the
orders appropriately to suppliers SA or SB. Profit is the overall profit (110-cost) without
inventory summed over all Ps that were dispatched through the system over 365 days. %Balked is
the percentage of orders that were not fulfilled because of lead-time/quantity constraints.
For the case where no inventory is modeled, results are as expected. Here, based on the way
the capabilities of suppliers SA and SB are modeled, always sending the orders to SA resulted in the
least profit and those through the proposed system framework resulted in the most profit. Similarly,
the number of orders that were not fulfilled is the most for supplier SA and least using the
proposed system framework. On the other hand, once inventory is introduced, the trends
continue. However, the inventory greatly reduces the bulk rate. The profits with the inventory
penalty remain greater than those when the same suppliers do not carry inventory. The profit for SB
in the case of the maximum inventory level of 97 remains slightly higher for the inventory cases than
in the SAB case.
The presented framework dynamically forms and reconfigures a supply chain as per the dictates of
order specifications. With the involvement of inventory in the process, no definitive way to place
orders exists such that freely choosing from any supplier or any assembler always results in the
optimal profits and the minimum balks. The inventory reduces the balk rate but does not
completely remove the presence of balks. When directing an order that would otherwise have balked
to an assembler or supplier in the chain, the possibility that the order will still balk remains up to
chance. If the order can be filled, then the supplier inventory level may be at such a low level
afterwards that it cannot be replenished fast enough to keep the next order from balking.
Therefore, always ordering from the same assembler or supplier could perform better than
dynamically choosing between assemblers and supplier for the same set of orders.
When carrying inventory, dynamically choosing a supplier or assembler does not necessarily
increase profits over using only one supplier or assembler combination. The orders placed in order
to replenish inventory in the case of an order that would otherwise balk can be much smaller than
the actual order. In this case, the inventory cost remains higher than the cost of a large order.
Therefore, reduced profits result. The results show that dynamically choosing, while not always
resulting in optimal profits and maximum ability to fill orders, can be much more profitable than
choosing the wrong assembler and supplier combination.
The results do show the general trends that carrying inventory reduces the percentage of balks,
thus satisfying the customer to a greater degree because a greater percentage of orders can be
filled. Even with the cost of storing the inventory taken into account, the profits still remain higher
than in the absence of inventory. The effects of discounting factors such as goodwill also play a role
in the profits of a dynamic supply chain as discussed in the extended version of this paper.
Future research may focus on extending the framework to handle more stages, several nodes in
each stage, order variability, and orders with a mix of various types of products.
The complete references can be found in the extended version of this paper, Agent-Based Framework
for Dynamic Supply Chain Configuration, presented at the Thirty-Seventh Annual Hawaii
International Conference on System Sciences, January 2004.
Back to the Journal of Undergraduate Research
College of Liberal Arts and Sciences I University Scholars Program I University of
W UNIVERSITY of
Th ' .,dsiho K"' N- Galr ' c 4?iou r
ï¿½ University of Florida, Gainesville, FL 32611; (352) 846-2032.