Co-evolutionary Classifier Systems for Multi-Agent Simulation

In this paper, MAZCS – a MultiAgent system that learns using ZCS – is used for social modelling on the “El Farol” Bar problem. Experiments with ten agents and different goal settings for the problem show that MAZCS is always able to solve it and emergent behaviour derives from the autonomous control of each agent. The results are divided into two different analysis scopes: macro and micro-level. The former providing the overall performance evaluation, the latter the detailed ZCS rule evolution and cause for the agent’s answers. Analysis of the values of the rules’ performance show that it is the amount of reward received by each ZCS combined with its reinforcement mechanism which cause the emergent behaviour. MAZCS has proved to be a good modelling tool for social simulation, both because of its performance and providing the explanation for the actions. Furthermore the system solved the problem when using a hundred agents, assuring the scalability for bigger simulation needs.