Robust Linear Quadratic Reinforcement Learning by Filtering

This paper investigates the robustness of linear quadratic reinforcement learning when unmodeled dynamics are included, and how the performance can be improved by applying filtering. We examine both model-free and model-based approaches, and it is shown that the model-free approach greatly suffers from the inclusion of unmodeled dynamics when no filtering is used. With the inclusion of filtering, similar performance is, however, achieved by both approaches. It is also concluded that model-based reinforcement learning has some notable other advantages over the model-free approach, thus making model-based the preferable approach when unmodeled dynamics are present.