Reinforcement Learning with Partial Parametric Model Knowledge
This work addresses the challenge of improving reinforcement learning efficiency for control tasks by leveraging partial model information, though it appears incremental as it builds on existing RL and control methods.
The paper tackles the problem of bridging the gap between complete ignorance and perfect knowledge in reinforcement learning for continuous control by proposing PLSPI, which uses partial model knowledge and data-driven adaptation, showing effectiveness in linear quadratic regulator experiments.
We adapt reinforcement learning (RL) methods for continuous control to bridge the gap between complete ignorance and perfect knowledge of the environment. Our method, Partial Knowledge Least Squares Policy Iteration (PLSPI), takes inspiration from both model-free RL and model-based control. It uses incomplete information from a partial model and retains RL's data-driven adaption towards optimal performance. The linear quadratic regulator provides a case study; numerical experiments demonstrate the effectiveness and resulting benefits of the proposed method.