Cultural evolution for sequential decision tasks: Evolving tic-tac-toe players in multi-agent systems

Sequential decision tasks represent a difficult class of problem where perfect solutions are often not available in advance. This paper presents a set of experiments involving populations of agents that evolve to play games of tic-tac-toe. The focus of the paper is to propose that cultural learning, i.e. the passing of information from one generation to the next by non-genetic means, is a better approach than population learning alone, i.e. the purely genetic evolution of agents. Population learning is implemented using genetic algorithms that evolve agents containing a neural network capable of playing games of tic-tac-toe. Cultural learning is introduced by allowing highly fit agents to teach the population, thus improving performance. We show via experimentation that agents employing cultural learning are better suited to solving a sequential decision task (in this case tic-tac-toe) than systems using population learning alone.