An Investigation into Island Model Rule Migration for a Number of Mobile Autonomous Learning Classifier System Agents

This work presents an investigation into combining the Island Model Genetic Algorithm and Learning Classifier System paradigms into a Multi-Agent approach to reinforcement learning. A multi-agent system is defined in which a number of interacting Zeroth Level Learning Classifier System agents inhabit a two-dimensional virtual test environment and are expected to perform a food-finding task. On occasion, a communication of behaviours between the agents is made, that is, rule migration. It is hoped that this communication between agents will enhance their performance as a group over non-communicating agents. There are many different policies available for communication characterised by aspects of timing, quantity and spatial reference. A direct implementation of the Island Model paradigm using multiple ZCS-agents is presented in which the agents can detect or affect other agents. Results indicate that rule migration between ZCS-agents is an effective way to improve learning speed and solution quality.