An Information-Theoretic Class of Stochastic Decision Processes

Abstract

Stochastic decision processes in reinforcement learning are usually formulated as Markov decision processes which are stationary and ergodic. However, in fact, some of the stochastic decision processes are not necessarily Markov, stationary, and/or ergodic. In this paper, using an information-theoretic property, we show a class of stochastic decision processes in reinforcement learning in which return maximization occurs with a positive probability. The class would be useful in considering reinforcement learning applications.