Bottleneck Analysis as a Heuristic for Self-Adaption in Multi-Agent Societies -Extended Abstract - Christian Gerber German Research Center for Artificial Intelligence (DFKI) Stuhlsatzenhausweg 3, 66123 Saarbruecken, Germany Tel. +49 681 302 4578, Fax +49 681 2235 gerber@dfki.de During recent years, due to advances in the theory of multi-agent systems, but also due to the explosion of hardware power, more and more complex multi-agent systems can be realized. This movement induces the need to introduce structures into such large systems which do not violate the paradigm of agent automony since this paradigm has been proven, tried and tested. Introducing a structure to a group of agents that pursue a common goal, leads to the definition of quantities such as the size of the group, distribution of specialized agents, command structures, types of communication protocols, etc. However, for the sake of agent autonomy, the agents' reasoning processes should be left untouched. As multi-agent systems are usably intended to operate for a rather long period of time (sometimes even ad infinitum), a system which was originally adjusted to work at a high performance level may lose its performance as the environment changes. Therefore, anautomated self-adaption mechanism which reorganizes the system is of great value. An approach uniting macro-level (i.e., society) aspects and micro-level. To achieve higher performance, heuristics can be integrated to such a mechanism. In this paper, we demonstrate how to incorporate bottleneck analysis as a heuristic. Traditional bottleneck analysis approaches try to derive a rather abstract mathematical model of a complex technical system and the work flow within that system. Usually, a directed graph is used to represent such a system: atomic system components are expressed by nodes; work flow is represented by arcs. Operations Research methods can then be applied on that model in order to detect local overloads. Such an approach can easily be used to analyze a multi-agent society by interpreting agents as nodes, agent cooperation for achieving a common goal as arcs in the graph and agent actions as work load transitions. However, this kind of approach has some drawbacks: Only agent overloads can be detected; underloads leading to waste of agent power remain uncovered. Furthermore, this approach bases on estimation of the system's behavior: faulty estimations may lead to inaccurate system modifications. Therefore, we propose a new approach to incorporate bottleneck analysis to a agent society self-adaption mechanism: for multi-agent systems we do not need to built up a mathematical model of the system, we can collect bottleneck data directly from the agents representing components or component groups. -- Christian Gerber German Research Center for Artificial Intelligence(DFKI) Office:Room 221, Building 36 (CS) Project MAS Phone: +49-681-302-4578 Stuhlsatzenhausweg 3 Fax: +49-681-302-2235 D-66123 Saarbruecken email: gerber@dfki.de Germany http://www.dfki.de/~gerber/ Private: Foersterstr. 38 Dresdner Str.1 D-66111 Saarbruecken D-66955 Pirmasens Germany Germany Phone: +49-681-376426 Phone: +49-6331-42901