Coevolutionary Learning and the Design of Complex Systems

Complex systems composed of a large number of loosely coupled entities, with no central coordination offer a number of attractive properties like scalability, robustness or massively distributed computation. However, designing such complex systems presents some challenging issues that are difficult to tackle with traditional top-down engineering methodologies. Coevolutionary learning, which involves the embedding of adaptive learning agents in a fitness environment that dynamically responds to their progress, is proposed as a paradigm to explore a space of complex system designs. It is argued that coevolution offers a flexible framework for the implementation of search heuristics that can efficiently exploit some of the structural properties exhibited by such state spaces. However, several drawbacks have to be overcome in order for coevolutionary learning to achieve continuous progress in the long term. This paper presents some of those problems and introduces a new strategy based on the concept of an "ideal" trainer to address them. This presentation is illustrated with a case study: the discovery of cellular automata rules to implement a classification task. The application of the "ideal" trainer paradigm to that problem resulted in a significant improvement over previously known best rules for this task.