Information and Knowledge Exchange in a Multi-Agent System for Stock Trading Yuan Luo School of Computing, Staffordshire University, Stafford, ST18 0DG, UK. Y.Luo@staffs.ac.uk Darryl N. Davis Neural, Emergent and Agent Technologies Group, Department of Computer Science, University of Hull, HU6 7RX, UK. D.N.Davis@dec.hull.ac.uk Kecheng Liu School of Computing, Staffordshire University, Stafford, ST18 0DG, UK. K.Liu@staffs.ac.uk Abstract: A distributed problem solving system can be characterised as a group of individual agents running and co-operating with other agents to solve a problem. As dynamic domains such as stock trading are continuing to grow in complexity, it becomes more difficult to control the behaviour of agents in the domains where unexpected events can occur. This paper presented an information and knowledge exchange framework to support the distributed problem solving in the stock trading domain. It addresses two important issues: (1) How individual agents should be interconnected so that their capacities are efficiently used and their goals are accomplished effectively and efficiently; (2) How the information and knowledge transfer should take place among agents to allow them to respond successfully to user requests and unexpected situations in the outside world. The focus of this paper is dynamic knowledge exchange among MASST agents. The co-ordinator agent together with a decision enabling warehouse acting as a dynamic blackboard plus direct intercommunication among the agents enable facts, commands, and rules to be transferred between MASST agents. Knowledge can be exchanged among the agents by using a combination of facts, rules and commands transfers. Keywords: Distributed Problem Solving, Multi-Agent System, Knowledge Exchange, Stock Trading. 1. Introduction All applications covered by Distributed Problem Solving (DPS) assume that it is possible to carry out a complex task by calling upon an assembly of specialists possessing complementary skills. When the problem is so wide and complex that one person cannot possess all skills to solve the problem, it is necessary to call upon several specialists, who must work together in pursuing a common objective. These specialists co-operate with one another to solve a common problem such as a medical diagnosis [Tu et al., (1995)], the design of an industrial product [Iffenecker and Ferber, (1992)], the acquisition of knowledge, fault finding in nets [Jennings et al., (1995)], the recognition of shapes [Demazeau et al., (1994)], or the understanding of natural language [Sabah, (1990)]. A DPS system can be characterised as a group of individual agents running and co-operating with other agents to solve a problem. As dynamic domains such as stock trading are continuing to grow in complexity, it becomes more difficult to control the behaviour of agents in domains where unexpected events occur. In recent years, there has been considerable growth of interest in the design of intelligent agent architectures for dynamic and unpredictable domains. Most of today's intelligent agent architectures are limited to performing pre-programmed or human assisted tasks. In a multi-agent system that consists of several agents the agents should be able to interact with each other and with their environment in an adaptable manner. Each agent has a local view of the environment, generally has specific goals and alone is unable to solve the system devoted global task. The global characteristics of such a system thus emerges from the co-operation of its component parts. This co- operation, in turn, impinges on the interactions between agents and subtly modifies the properties of the system [Gleizes et al., (2000)]. In order to be more useful in complex real world domains, agents need to be more flexible. They need to learn how to respond promptly to unexpected events while simultaneously carrying out their pre-programmed tasks in response to subtly modified triggers. It is highly possible that an agent with a responsibility in a dynamic environment faces unexpected events. In order to be responsive, the agents should have enough knowledge to deal with unexpected events. If an agent is not able to deal with a particular event on its own, it can take the following actions: 1) Learn how to solve the problem by experimenting with different solution strategies. 2) Let some other knowledgeable agent solve the problem and then use the results. 3) Learn how to solve the particular problem by acquiring the necessary knowledge from other agents capable of solving the problem. 4) Ignore the unexpected event [Cengeloglu et al., (1994)]. For a real time application domain such as the stock trading, action 1 may not be suitable because it may take a long time to obtain the