Brief Survey of Model-Based Reinforcement Learning Techniques

摘要

Model-free reinforcement learning (MFRL) usually has better asymptotic performance than the model-based reinforcement (MBRL) learning algorithms, especially in complex environments. But MBRL algorithms are very often much more sample-efficient, and sometimes are able to learn control tasks in just a handful of trials. In addition, in some domains, the MBRL algorithms can reach the MFRL performance with better sample efficiency. In recent years, MBRL research has increased in various application domains, such as robot control tasks, or game environments with complex observations. In this paper, we review the most popular techniques used in MBRL and look at some useful classification of algorithms in this area.