Environment Driven Dynamic Decomposition for Cooperative Co-evolution of Multi--Agent Systems

Cooperative co-evolutionary approaches have shown the ability to evolve specialised agent behaviours in multi-agent systems. A key aspect of applying cooperative co-evolution as an optimisation technique is the decomposition of the overall problem into smaller, interacting sub-components. Typically, problem decomposition occurs a priori, with a common approach being to decompose by agent, where each agent's behaviour corresponds to a separate evolving individual. Such approaches have inherent scalability limitations as the number of agents increases. In this work, we demonstrate a novel dynamic decomposition scheme based on the hierarchal subdivision of the environment, rather than by agent. In our approach, decomposition is performed simultaneously with the evolution of behaviour, starting with behaviour for the whole environment, followed by environment sub-division and behaviour specialisation. The approach was evaluated by evolving the behaviour of a set of cooperating combat unit agents in a real- time strategy game. The game was set in an environment composed of areas, each requiring a different behaviour for optimal gameplay. Using the behaviour evolved by our approach, the agents consistently defeated a numerically superior opponent and demonstrated location-appropriate behaviour. This approach may benefit other multi-agent systems where distinct environment sections call for different specialisation.