Learning Sequential Decision Tasks

This paper presents a new approach called SANE for learning and performing sequential decision tasks. Compared to problem-general heuristics, SANE forms more e ective decision strategies because it learns to utilize domain-speci c information. SANE evolves neural networks through genetic algorithms and can learn in a wide range of domains with minimal reinforcement. SANE's evolution algorithm, called symbiotic evolution, is more powerful than standard genetic algorithms because diversity pressures are inherent in the evolution. SANE is shown to be e ective in two sequential decision tasks. As a value-ordering method in constraint satisfaction search, SANE required only 1/3 of the backtracks of a problemgeneral heuristic. As a lter for minimax search, SANE formed a network capable of focusing the search away from misinformation, creating stronger play.