Reinforcement Learning Guided Partial-Update Extended Kalman Filter

The partial-update extended Kalman filter seeks to restrain the influence of measurement updates to state and covariance estimates using trust factors. To date, there is little work addressing how to use data inference techniques like machine learning to choose the optimal trust factors. This work presents a simple reinforcement learning structure for developing policies which use readily-available estimation variables and simple rewards to discover these trust factors for estimation scenarios, without relying on non-observable information of the estimated dynamic systems. A learned policy is validated against the re-entry vehicle problem where a well-performing, manually-tuned policy with static trust factors is known, and an advantage in convergence times of state estimates is exhibited by the policy determined through reinforcement learning. Policies trained through the proposed method are also shown to have convergence times that are less sensitive to measurement frequency changes.