Adaptive Filtering Algorithm based on Reinforcement Learning

Real-time state estimation of dynamic systems is a basic task in control theory and signal processing. Kalman filter is a minimum complex optimal state estimator for linear state space models with fully known system parameters, which is widely used in industrial production. However, in practical problems, accurate models and linear state-space models are often difficult to satisfy. In this paper, by combining reinforcement learning and Kalman filter, an adaptive strategy of state space model is designed based on PPO algorithm to improve the performance of state estimation under the inaccuracy conditions of the system model. The algorithm is called Reinforcement Learning Kalman filter algorithm (RLKF). Finally, we prove by numerical results that the Reinforcement Learning Kalman filtering algorithm overcomes the model mismatch and nonlinearity, and the accuracy is better than the traditional filtering algorithm.