Cooperative Versus Competitive System Elements in Coevolutionary Systems

This paper examines systems that have coevolved elements. For any given environmental state, the overall action of such systems can be formulated in one of two ways: the system's elements can cooperate to jointly formulate an action, or the elements can compete, with a winner formulating an action. In both cases, underlying coevolution may be necessary to create a group of elements that can implement a complete action repertoire. However, one would expect that the emergent relationships between elements would differ between competitive and cooperative systems. These differences have important implications for both natural and artificial adaptive systems. This paper examines this issue by using neural networks that evolve under the action of a genetic algorithm. These neural networks are functionally similar to learning classifier systems. Cooperative and competitive versions of these networks are evaluated in a simulated mobile robot test problem. Experimental results lead us to the conjecture that certain symbiotic relationships between elements (e.g., default hierarchies and chains) develop more easily in cooperative systems than in competitive systems. Implications of this conjecture, and future directions of inquiry are discussed.