Convergence of reinforcement learning algorithms and acceleration of learning - art. no. 026706

摘要

The techniques of reinforcement learning have been gaining increasing popularity recently. However, the question of their convergence rate is still open. We consider the problem of choosing the learning steps alpha(n), and their relation with discount gamma and exploration degree epsilon. Appropriate choices of these parameters may drastically influence the convergence rate of the techniques. From analytical examples, we conjecture optimal values of alpha(n) and then use numerical examples to verify our conjectures. [References: 20]