Deep reinforcement learning for static noisy state feedback control with reward estimation

Deep reinforcement learning (DRL) has demonstrated extraordinary capabilities in learning optimal policies for Markov decision processes (MDPs). However, when measurement noise affects the observation of the state, the problem transforms into a partially observable MDP (POMDP), which becomes nearly intractable when the system dynamics are unknown. To this end, we establish a reward estimation-based DRL algorithm to evaluate the long-term reward and learn a static (memoryless) noisy state feedback (SNSF) policy under additional assumptions. Numerical simulations validate the algorithm's effectiveness, and the related code is open-sourced at https://github.com/RanKyoto/RE-POMDP.