Abstract: A new algorithm is proposed, which immolates the optimality of control policies potentially to obtain the robusticity of solutions. The robusticity of solutions maybe becomes a very important property for a learning system when there exists non-matching between theory models and practical physical system, or the practical system is not static, or the availability of a control action changes along with the variety of time. The main contribution is that a set of approximation algorithms and their convergence results are given. A generalized average operator instead of the general optimal operator max (or min) is applied to study a class of important learning algorithms, dynamic programming algorithms, and discuss their convergences from theoretic point of view. The purpose for this research is to improve the robusticity of reinforcement learning algorithms theoretically.

Key words: reinforcement learning, risk-sensitive, generalized average, algorithm, convergence