Learning With A Competitive Population

In the typical machine learning system, a teacher, in the form of a collected training environment or an expert human or program, is provided to supply directed feedback. Unfortunately, in many learning situations no such teacher is available, or the expense of constructing one far outweighs the benefits of the learning system. Recent work by researchers have explored several teacherless, or unsupervised, training methods. In the domain of games, this can involve self-play. Learning through such a reflexive competition is of both theoretical and practical interest. But self-play by itself is flawed and can lead learning systems toward nonoptimal concepts that minimize their own feedback. In this paper, we argue that a population of hierarchically competitive learners provides the advantages of selfplay while removing this flaw. We demonstrate our model with several experiments using a population of learners that induce lisp programs to play Tic Tac Toe from scratch. Submitted to IJCAI ..