Experimental evaluations of model-based reinforcement learning combined with MPC
Qiming Zhang · Journal of Physics Conference Series · 2023
Abstract Model-based reinforcement learning has reached outstanding effects in control and decision areas with high sample efficiency compared with model-free algorithms. However, how to combine learning and planning, aiming at improving effect and reducing the cost of planning at the same time seems to be challenging in the model-based area. In this paper, several state-of-the-art model-based algorithms combined with MPC: TD-MPC, PlaNet and Dreamer, are discussed and compared. Theoretical background and features of these methods are analyzed based on the continuous control environments. This work also covers how to design model-based method for different tasks and future directions for improving performance for model-based reinforcement learning research.