A GPSS SIMULATION MODEL OF INTERACTIONS IN A MARKET-BASED MULTI-AGENT SYSTEM Eugenia Kalisz and Adina Magda Florea Department of Computer Science “Politehnica” University of Bucharest Spl. Independentei, 313, Bucharest, 77206 Romania E-mail: {ekalisz,adina}@cs.pub.ro KEYWORDS Multi-agent system, GPSS model, agent abilities. ABSTRACT Academic and industrial system designers who consider using agent technology to solve application problems are faced with a wide variety of agent system paradigms. Choosing a paradigm or another is not an easy endeavour, as little knowledge or not too many guidelines exist on how to make the good choice. The paper presents a multi-agent system simulation framework for testing different time and load performances of agents depending on some essential characteristics of the agent model, e.g., abilities, negotiation time, and response time to requests. Agents are self- interested and cost-based, as in a market-like model. Their behaviour is driven by the necessity of achieving their own goals while minimising the payment for actions executed by other agents on their behalf. A system facilitator is supporting agent interactions. Starting from a core GPSS model, we have developed several simulation scenarios to simulate and evaluate interaction performances in a multi- agent system, under different assumption of agent behaviour and the tasks assigned to the facilitator. Some comparative results of several simulation experiments are presented and discussed. INTRODUCTION Academic and industrial system designers who consider using agent technology to solve application problems are faced with a wide variety of agent paradigms (Muller 1998). Once a given paradigm is chosen, the designer must develop the application, either experimentally or following an existing methodology (Iglesias et. al. 1998). In both cases, agent and system modelling is an essential part of this process. The data and world conceptual model and the agent identification for a particular application are some of the difficult tasks that are faced by a multi-agent systems (MAS) application designer, for which little knowledge or not too many guidelines exist. The work undertaken aims to contribute in solving part of these difficulties. The paper presents a MAS simulation framework for testing different time and load performances of agents depending on some essential characteristics of the agent model, e.g., abilities, negotiation time, and response time to requests, and on the system architecture. Starting from a core GPSS model, we have developed several simulation scenarios to simulate and evaluate interaction performances in a multi-agent system, under different assumption of agent behaviour and facilitator assigned tasks. The simulation framework includes a set of parameters that may be set by the user to describe the system under evaluation, while the obtained results and statistics help the user to select the appropriate paradigm. CONCEPTUAL MODEL The multi-agent system we consider is formed of a society of self-interested agents that are acting in a distributed environment. Each agent is pursuing its own goals but, in order to achieve them, it has to cooperate with other agents in the system. An agent is endowed with a set of characteristics (abilities) that are relevant for the simulation model: the actions the agent is able to perform; a cost associated to each action, that may be either how much it is willing to pay for this action to be executed by another agent or how much it would like to receive if it is requested to perform the action; a time delay in which one action is to be executed. All the other characteristics of the agents are hidden in the simulation model (as being irrelevant for our focus of interest) and they may range from simple internal representations and negotiation strategies to more elaborate ones. The only impact these internal characteristics have on the model is at the level of the time necessary to issue a request and to take a decision following an answer to a request. The characteristics to be considered in the simulation framework are selected from the agent model presented in (Florea 1999) but they are quite common to many market-based models of MAS. The agent behaviour is driven by the necessity of achieving their own goals while minimising the payment for actions executed by other agents on their behalf, as in a market-like environment. The agents are interacting by sending request messages in which they ask for the execution of a given action and by receiving answer messages of acceptance or rejection from the other agents in the system. The system comprises a facilitator that supports agent interactions, as described bellow. The facilitator has knowledge about the agents in the system (their names) and about their associated abilities: actions, with associated costs and time delays, specific to each agent.