Bo Pang

This research, whose principal author is Ph.D. student Bo Pang, was directed by Zhong-Ping Jiang, professor in the Department of Electrical and Computer Engineering. 

As an important and popular method in reinforcement learning (RL), policy iteration has been widely studied by researchers and utilized in different kinds of real-life applications by practitioners. 

Policy iteration involves two steps: policy evaluation and policy improvement. In policy evaluation, a given policy is evaluated based on a scalar performance index. Then this performance index is utilized to generate a new control policy in policy improvement. These two steps are iterated in turn, to find the solution of the RL problem at hand. When all the information involved in this process is exactly known, the convergence to the optimal solution can be provably guaranteed, by exploiting the monotonicity property of the policy improvement step. That is, the performance of the newly generated policy is no worse than that of the given policy in each iteration. 

However, in practice policy evaluation or policy improvement can hardly be implemented precisely, because of the existence of various errors, which may be induced by function approximation, state estimation, sensor noise, external disturbance and so on. Therefore, a natural question to ask is: when is a policy iteration algorithm robust to the errors in the learning process? In other words, under what conditions on the errors does the policy iteration still converge to (a neighborhood of) the optimal solution? And how to quantify the size of this neighbourhood?

This paper studies the robustness of reinforcement learning algorithms to errors in the learning process. Specifically, they revisit the benchmark problem of discrete-time linear quadratic regulation (LQR) and study the long-standing open question: Under what conditions is the policy iteration method robustly stable from a dynamical systems perspective? 

Using advanced stability results in control theory, they show that policy iteration for LQR is inherently robust to small errors in the learning process and enjoys small-disturbance input-to-state stability: whenever the error in each iteration is bounded and small, the solutions of the policy iteration algorithm are also bounded, and, moreover, enter and stay in a small neighbourhood of the optimal LQR solution. As an application, a novel off-policy optimistic least-squares policy iteration for the LQR problem is proposed, when the system dynamics are subjected to additive stochastic disturbances. The proposed new results in robust reinforcement learning are validated by a numerical example.

This work was supported in part by the U.S. National Science Foundation.