Learning to achieve co-operation by temporal-spatial fitness sharing

We propose a co-operative GA-based learning system that would make real-world heterogeneous agents feasible with the minimum amount of communication hardware. The problem is identical to a distributed GA implemented on processors connected by local and very slow communication lines. We have developed an extension of the fitness sharing method that incorporates sharing over temporally-spatially distributed populations. Restricting an agent's task to the inter-agent conflict avoidance, this sharing is realised by exchanging estimated fitness values over all agents. The mechanism of finding conflict avoidance actions is similar to that of a self-organisation mechanism of a Kohonen-type network. Our results from simulations of a bump-avoidance task for multiple mobile robots show that it elicits a notable performance improvement compared to normal classifier systems.