TOWARDS MINING FOR INFLUENCE IN A MULTI AGENT ENVIRONMENT Robert Logie Osaka Gakuin University, 2-36-1 Kishibe-Minami,Suita-shi,Osaka 564-8511,Japan Jon G. Hall and Kevin G. Waugh The Open University, Walton Hall, Milton Keynes MK7 6AA England ABSTRACT Multi agent learning systems pose an interesting set of problems: in large environments agents may develop localised behaviour patterns that are not necessarily optimal; in a pure agent system there is no globally aware element which can identify and eliminate retrograde behaviour; and as systems scale they may produce large amounts of data, a system may have in the order of 10 6 cells with 10 5 agents, each generating data. This position paper introduces research that combines data mining with a logical framework to allow agents in large systems to learn about their environment and develop behaviours appropriate to satisfying system norms. We build from traditional multi agent systems, adding a novel process algebraic approach to co-operation using data mining techniques to identify co-operative behaviours worth learning. The result is predicted to be a learning system in which agents form collectives increasing their ‘mutual influence’ on the environment. KEYWORDS Agent-systems, stit, multi-agent learning, coaching. 1. INTRODUCTION Multi-agent learning theory is arcane and even in practice multi agent learning systems pose serious computational problems: each agent is an autonomous computational entity governed by what it knows and how it has been told to respond to its environment. Learning in multi-agent systems is, typically, achieved by individual agents having some ability to learn. If goals are known and simple, such learning is often adequate. However, for some problems, goals may be specified at a high level which is agent agnostic and agents with limited abilities may need to co-operate to achieve these goals. Without a globally aware planning entity to guide them, such co-operative behaviours are difficult to identify in the masses of data that can accrue, and even when co-operative behaviours are identified, determining which will lead to system goals remains difficult. Our response to these problems is a system containing ant-like agents, viz., perceptive (Page 190, Ferber 1999), modally reactive (Page 284, Ferber 1999) agents with simple communication abilities. In nature modified ant behaviour is determined by slowly acting evolution, we provide coaches that observe populations of ants, looking for evidence of influence enhancing behaviour (ieb), co-operative behaviour that strictly increases the influence that agents exert on their environment. Influence is measurable, and corresponds to certain observable patterns in agent behaviour. Detecting an increase in influence is simply a matter of coaching agents mining for those patterns. On identifying an ieb, coaching agents will then seed appropriate co-operative behaviours in the environment so that workers may pick them up and adapt. There are three complications in this approach:  the density of the influence-revealing patterns in raw data is very low: random behaviour in a simple test typically yielded one ieb sequence per thousand or lower (depending on many factors); IADIS European Conference Data Mining 2008 97