Physics-informed Dyna-style model-based deep reinforcement learning for dynamic control

Model-based reinforcement learning (MBRL) is believed to have much higher sample efficiency compared with model-free algorithms by a predictive model of the environment. However, performance MBRL highly relies on quality learned model, which usually built in black-box manner and may poor accuracy outside data distribution. The deficiencies prevent policy from being fully optimized. Although some uncertainty analysis-based remedies been proposed alleviate this issue, bias still poses great challenge for MBRL. In work, we propose leverage prior knowledge underlying physics environment, where governing laws are (partially) known. particular, developed physics-informed framework, equations physical constraints used inform search. By incorporating information can be notably improved, while required interactions environment significantly reduced, leading better performance. effectiveness merit demonstrated over handful classic control problems, environments governed canonical ordinary/partial differential equations.