Neuroevolution of Controllers for Self-Organizing Mobile Ad Hoc Networks

This paper describes a study in the use of neuroevolution to discover controllers for a simulated mobile ad hoc network. Neuroevolution is a technique whereby an evolutionary algorithm is used to produce artificial neural networks that solve a user-defined task. Here, we use neuroevolution to study a generic coverage-based problem, where agents in the network are to maximize the area covered by the largest connected component of the network. An example application for this work is the discovery of control algorithms for an ocean-monitoring mobile network. While this is a challenging problem domain for neuroevolution, results of our experiments reveal three important characteristics to be considered when using such an approach. Specifically, we found that approaches that implicitly reduce entropy, while explicitly addressing self-organization and scalability, are capable of discovering behaviors that remain stable even when they control networks of different sizes than were evaluated during evolution. This result suggests that neuroevolution may be a viable strategy for discovering controllers for self-organizing multi-agent systems.