Temporal Difference-Based Policy Iteration for Optimal Control of Stochastic Systems

In this paper, a unified policy iteration approach is presented for the optimal control problem of stochastic system with discounted average cost and continuous state space. The approach consists of temporal difference learning-based potential function approximation algorithms and performance difference formula-based policy improvement. The approximation algorithms are derived by solving the Poisson equation-based fixed-point equation, which can be viewed as continuous versions of least squares policy evaluation algorithm and least squares temporal difference algorithm. The simulations are provided to illustrate the effectiveness of the approach.