ADAPTIVE AGENTS FOR SUPPLY NETWORKS

Gavin Finnie, Jeff Barker

School of Information Technology, Bond University, Gold Coast, Australia

Keywords: Agents, Case-Based Reasoning

, Supply Chain Management

Abstract: Dynamic information flow in esupply networks require

s that buyers and suppliers have the ability to react

rapidly when needed. Using intelligent agents to automate the process of buyer/seller interaction has been

proposed by a number of researchers. One problem in providing intelligent automated collaboration is

incorporating learning capability i.e. an agent should be capable of adapting it’s behaviour as conditions

change. This paper proposes a scalable multi-agent system which uses case-based reasoning as a framework

for at least part of its intelligence. Tests with a simulated system show that such an agent is capable of

learning the best supplier and also capable of adapting if supply conditions change.

1 INTRODUCTION

Supply chain and supply networks can be of

arbitrary size and complexity. In an electronic

business environment, information flows at high

speed and organisations must be capable of rapid

reaction and reorganisation in response to dynamic

information relating to any changes in constraints or

conditions (McClellan 2003).

This paper will describe an agent-based approach

for in

telligent automation of inter-organisational

interaction in the supply chain. Any organisation

will have some history of dealing with problems

relating to orders and perturbations in the supply

chain and the solutions applied, as well as some

formal processes for dealing with these. In order to

automate the response to any stochastic event,

software must be capable of reacting as one would

expect a human agent to do. In many cases, a human

agent responds by working from and possibly

adapting solutions to previously encountered

situations similar to the present problem i.e. a

process of reasoning from prior cases or Case-Based

Reasoning (CBR). A model is proposed in which the

interface between an organisation and the outside

world is controlled by a number of agents, each of

which acquires at least part of its intelligence by

applying CBR.

2 OVERVIEW OF CBR

Case based reasoning (CBR) solves new problems

by adapting previously successful solutions to

similar problems. The appeal of CBR as a problem

solving approach lies in its familiarity - in many

problem solving situations a solution will be based

on a similar problem solved by us in the past. As an

example, doctors would not usually start all

diagnoses from first principles. They would in most

cases recall similar cases of patients with the same

symptoms and also recall what treatments have

worked in the past. Treatments may be modified for

the specific circumstances of this patient eg

difference in ages, sex, weight, medical history, etc.

might all suggest some need for adaptation of a past

solution.

A new problem (the target case) is matched

against cases

in the case-base. The importance

attached by the user to various features (indexes) of

the case may be used to guide the matching process.

One or more similar cases are retrieved from the

case base. A solution suggested by these cases is

reused and tested for success. If necessary, the

retrieved case(s) will probably be revised to produce

a new case which can then be retained in the case

base.

480

Finnie G. and Barker J. (2004).

ADAPTIVE AGENTS FOR SUPPLY NETWORKS.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 480-483

DOI: 10.5220/0002611104800483

 SciTePress

3 THE CASE AT THE INTERFACE

CBR has been primarily used in scheduling as an aid

to creating and adapting specific schedules, usually

within the organisation. This paper proposes the use

of CBR for intelligence at each stage of a schedule

within a specific supply chain. The interface

between an organisation and its suppliers will be

controlled by a number of buyer agents, each of

which will have access to CBR to provide intelligent

processing of supply needs on the basis of prior

experience. Coordinating and controlling the

activation and operation of the buyer agents is a

buyer interface control agent which again utilises

CBR to select a suitable strategy for finding all

components required for a particular product i.e. it

will review the bill of materials, decide on suitable

suppliers and set up agents to control the interaction

with each supplier. There is one buyer interface

agent for each organisation. The buyer (and the

supplier) interface agents are “middle agents” which

act as brokers between buyers and seller (Wong and

Sycara, 2000). It will also have responsibility for

ensuring that all components are suitably sourced i.e.

a failure procedure must be in place to backtrack if a

specific supplier fails to ensure supply.

At the supplier interface, there will be one seller

agent per transaction. These are relatively short-

lived agents responsible for monitoring the progress

of a specific request for materials. A request to

purchase from an organisation may itself trigger

adaptations in the internal schedule for that

organisation and in turn cause its buyer agents to

negotiate with its suppliers. To coordinate the

actions of supplier agents there is a supplier

interface control agent for each supplier. This has

responsibility for checking, also using CBR, whether

the product can be supplied. The supplier interface

agents will check on the impact of an order i.e. can it

be realistically scheduled and processed. This may

in turn generate a procurement need, causing a

spreading activation of agents.

The supplier agent will also retain a base of prior

cases i.e. what did we do last time. Agents will also

need to have fall back positions i.e. if there is no

suitable information in the case base, there must still

be a response – either by appealing for human

intervention or going to other forms of reasoning

e.g. rule-based.

The Buyer Agent Cycle

Cases relate to specific products and suppliers and

the basic cases will be indexed by product (or

product class). There may need to be some form of

generic or template cases which provide basic

reasoning.

The Buyer interface Agent cycle will be:

1. An order is received

2. The case base is checked for previous

suppliers of the product

3. An agent is initiated to control the buyer

cycle.

4. A message is broadcast to the “web”

looking for prospective suppliers. This

assumes a standardised structure to define

suppliers.

5. Prospective suppliers are ordered in terms

of some priority scheme and either:

(a) the order is sent to the supplier

(b) there is a call for quotes

3.2 The Supplier Interface Agent Cycle

A supplier agent will receive a request for an order

or a quote and will need to initiate a process to

determine if and when the order could be filled. This

may require rescheduling of production and ordering

of new inputs. Each supplier interface agent will also

maintain a case history of prior dealings with buyers.

On the basis of history (if it exists) and any other

intelligence provided, the agent will decide to:

(a) Decline the order or quote

(b) Agree to fill the order/quote without

adjusting existing schedules

meet an order or estimate impact if a

quote is required. In this case a

supplier watch agent is initiated to

monitor the progress.

If (c) is selected, there may be a need to initiate a

purchase cycle for input materials. This will require

the company buyer agents to initiate PO’s or RFQ’s

and any response by the supplier agent will be

delayed until the necessary information is available.

Once an order is shipped and payment received

and processed, the case base for the supplier agent

will be updated.

ADAPTIVE AGENTS FOR SUPPLY NETWORKS

481

Figure 1: Average Unit Cost over time

4 TEST IMPLEMENTATION

In order to test the case based approach, a simple

scenario was set up and a buyer agent modelled. The

buyer agent has its own case base which was

implemented using a relational database and a

simple nearest-neighbour search strategy. Cases

simply record information on order size, order time

(in days), previous delays and a price for the order

type.

Three suppliers of a particular product exist i.e.

S1, S2 and S3. S1 is a low cost supplier ($10) but

has a number of problems with delivery delays and

inability to meet order requests. S2 has a better

record but charges a higher price ($15) while S3 can

meet all deadlines but has a high price ($22). The

delays were modelled as follows:

• S1 and S2 can meet a new request for

an order 80% of the time while S3 can

always meet an order.

• If there is a delay in meeting an order,

S1 has a 60% chance of a one day delay,

20% chance of two days and 20% chance of

three. S2 has a 50% chance of a one day

delay and a 50% chance of a two day delay.

• On order delivery, S1 has a 40%

chance of delivery on schedule, 40%

chance of a one day delay and 20% chance

of a two day delay. S2 has an 80% chance

of no delay and a 20% chance of a one day

delay.

• Delays are assumed to cost a fixed

amount per day (modelled as $4 per unit)

Calculating the expected values for these

distributions give an expected cost per unit for S1 of

$14.10, for S2 of $17.00 and $22.00 for S3.

The simulation was run for 500 random cases. If

a case was judged to be sufficiently different from a

case in the case base, it was added to the set of

cases. If it was reasonably similar to an existing

case, the existing case had its price for the order

adjusted as the average of the new price and two

times the existing price (this has the effect of giving

significant weight to the most recent case). 59 new

cases were added overall.

Figure 1 shows the average cost per unit as each

case is dealt with (line Scenario 1) for the first 400

cases. Figure 2 shows the frequency of use of the

supplier data for each case i.e. S1 showS the number

of times a case for supplier one is used as the basis

for the new order. S2 shows the number of times a

case for S2 is used as the basis. Under this scenario,

the average price per unit rapidly converges to close

to the expected price of $14.10. From Figure 2, it is

apparent that S1 is by far the preferred supplier. S3

does not enter the reasoning as its price is too high.

To determine whether the CBR system is capable

of learning to change, the scenario was altered by

assuming that after 100 cases had been processed,

S1 (probably due to its low prices and consequent

high demand) has shown a decline in being able to

meet orders from 80% to 50%. S1’s unit price has

also increased from $10.00 to $12.00. S2 on the

other hand has been able to improve its ability to

meet orders to 90% and has managed to cut its price

to $14.00. This gives a new expected price per unit

for S1 of $17.36 and for S2 of $15.40. S3 is

unaffected by the change.

ICEIS 2004 - SOFTWARE AGENTS AND INTERNET COMPUTING

482

Figure 2: Use of Suppliers over time

Figure 1 shows the effect on average cost per

unit with line Scenario 2 showing the impact after

case 100 has been processed. The average cost per

unit is recalculated from case 101 and rapidly

stabilises around the new minimum expected price

of $15.40. Figure 2 shows the relative use of S1 and

S2 as the basis for new ordering decisions (S1 New

and S2 New). It is apparent that S2 replaces S1 as

the preferred supplier to accommodate the new

pricing realities.

As a further comparison, the average price per

unit of a random supplier selection policy (including

S3) was simulated over 500 cases and stabilised at

$23.50 i.e. the CBR approach can find the most cost

efficient alternative.

5 CONCLUSIONS

To use dynamic information effectively in inter-

enterprise supply chain management, decisions will

need to be made automatically and effectively. The

multi-agent system approach proposed in this paper

provides a suitable architecture for rapid and agile

response to any event. An agent is this environment

must be capable of intelligent reasoning and

learning. The CBR approach provides a suitable

framework for at least part of the intelligence, and is

capable of learning dynamically i.e. as a new case is

encountered it will be added to the case base for that

specific product in a specific company. As shown

above, the CBR system adapts if the conditions

change.

The framework proposed here requires further

development and testing both in simulated and real

environments. The initial results are encouraging

and suggest that an MAS approach with CBR could

be a powerful tool to further automating supply

chain management.

REFERENCES

Cunningham, P. and B. Smyth (1996). "Case-Based

Reasoning in Scheduling: Reusing Solution

Components."

International Journal for Production

Research 35(11): 2947-2961.

Dorn, J. (1995). "Iterative Improvement Methods for

Knowledge-based Scheduling."

AI Communications

8(8): 20-34.

Finnie, G. and Z. Sun (2003).

A knowledge-based model

of multiagent CBR systems. Proc. Int Conf on

Intelligent Agents, Web Technologies, and Internet

Commerce (IAWTIC'2003), Vienna, Austria.

Fox, M. S., M. Barbuceanu, et al. (2000). "Agent-Oriented

Supply-Chain Management."

The International Journal

of Flexible Manufacturing Systems 12: 165-188.

McClellan, M. (2003).

Collaborative Manufacturing:

Using Real-Time Information to Support the Supply

Chain

. Boca Raton, St Lucie Press.

Wong, H.C and K Sycara. (2000) A Taxonomy of Middle-

Agents for the Internet, Proceedings of the Fourth

International Conference on Multiagent Systems, pp

465-466.

ADAPTIVE AGENTS FOR SUPPLY NETWORKS

483