EVOLVING ARTIFICIAL NEURAL NETWORKS TO CONTROL CHAOTIC SYSTEMS

We develop a genetic algorithm that produces neural network feedback c ontrollers for chaotic systems. The algorithm was tested on the logist ic and Henon maps, for which it stabilizes an unstable fixed point usi ng small perturbations, even in the presence of significant noise. The network training method [D. E. Moriarty and R. Miikkulainen, Mach. Le arn. 22, 11 (1996)] requires no previous knowledge about the system to be controlled, including the dimensionality of the system and the loc ation of unstable fixed points. This is the first dimension-independen t algorithm that produces neural network controllers using time-series data. A software implementation of this algorithm is available via th e World Wide Web.