Self-organizing economic activity with costly information

• Authors: James A. Wilson;Liying Yan
• Venue: Proceedings of the 11th Annual Conference Companion on Genetic and Evolutionary Computation Conference: Late Breaking Papers
• Year: 2009

We describe a multi-agent simulation in which individual boundedly rational agents learn and adapt while competing to capture a complex resource. Our purpose is to explore the fine scale dynamics that emerge as aggregate social structure and dynamics. We use a biophysical model of the Maine lobster fishery to create a complex, dynamic environment. Agents compete by learning how to search for and harvest lobsters. We simulate individual learning with a modified version of John Holland's learning classifier system. At each iteration an agent must decide whether to continue fishing using already acquired knowledge about the resource or whether to acquire new knowledge by exploring on its own or by learning from another agent. Each agent's information about its environment has an opportunity cost that is created by limited time and by restricted (local) observation capabilities. Agents can communicate with and learn by imitating other agents but the cost of communicating with each other agent is an inverse function of the frequency with which the agents encounter one an-other. This 'familiarity' effect generates positive feedback and communication efficiencies that lead to the formation of persistent groups. The sharing of information within these groups gives agents the ability to avoid being trapped in local optima and increases both individual and collective efficiency. Agents develop search strategies that continuously switch between cooperative and autonomous search according to changing conditions of the resource and the costs of communication. We compare the aggregate outputs of the model with those observed in a large data set that tracks the time, location and catch of nearly a million lobster traps.