Data-Efficient Uncertainty-Guided Model-Based Reinforcement Learning with Unscented Kalman Bayesian Neural Networks
Abstract: In recent years, reinforcement learning (RL) has made substantial progress by providing feasible solutions to many planning and control problems. However, the majority of RL algorithms, particularly model-free RL, suffer from low learning efficiency. To address this issue, this paper proposes the utilization of the Kalman Bayesian neural network (KBNN) to learn a tractable dynamics model of a system from data that captures uncertainties of the system state. Additionally, we employ the unscented Kalman filter to propagate these uncertainties over the control horizon, and we exploit the propagated uncertainties explicitly in the cost function. This approach presents a novel solution to improve data-efficiency in RL. This is validated on classic control problems in a comparative analysis against state-of-the-art methods.
Loading