and

Abstract

1 Introduction

Stuart Watt Zdenek Zdrahal

Mike Brayshaw

Multiple Agent Systems for Configuration Design

Knowledge Media Institute
Open University, Milton Keynes, UK.

School of Computer Science
University of Birmingham, Birmingham, UK.

This paper compares and contrasts traditional methods and multiple agent architectures as techniques for performing the task of configuration design. Configuration design is a common engineering problem, involving component selection and arrangement, and thus presents an interesting area to apply AI technology. One particular example, the VT elevator design system (Marcus et al., 1988), has been proposed as the Sisyphus-2 benchmark for testing and comparing various AI techniques and methodologies. Taking this as a starting point we will demonstrate how a multiple agent architecture gives extra flexibility in use and transparency of design choice over a conventional system. Our argument will be structured as follows. First we shall describe the engineering task and domain under discussion. Then we shall discuss how traditional approaches would go about solving the problem.

1

This paper investigates how the task of configuration design can be carried out using concepts of multiple agency. Configuration design is the task of selecting components from a predefined set to complete a system which meets a given functional specification and other design constraints. It is a class of task which is conventionally solved using a single agent reflecting an arbitrary balance of the design criteria chosen by the system designer. To study the efficacy of the multiple agent approach, we show how more of the original domain knowledge can be applied in an alternative form where each agent in a multiple agent design system has individual design goals and acts in negotiation with others in order to achieve those goals. Each agent's goal represents different domain axes upon which design decisions are based. At any point where different design decisions can be made, negotiation between these agents enables a balancing between the different agent's goals, and therefore these axes, to be achieved. Using the Sisyphus-2 benchmark configuration design problem, we will compare and contrast these methods to identify their relative merits.