Model Free Safe Control for Reinforcement Learning in a Clustered Dynamic Environment
Guiliang Zheng, Minhao Yang, Yuxuan Wu · 2022
While reinforcement learning (RL) has been wildly used in continuous control tasks with impressive performance, there are two main challenges in RL development: continuously satisfying the safety constraints in a clustered dynamic environment and lacking the explicit analytical models of dynamic systems for typical safeguard algorithms in RL training. This paper proposed a model-free safe control strategy to safeguard the RL agent in a clustered dynamic environment. By extending and applying the adaptive momentum boundary approximating (AdamBA) method to monitor and modify the RL nominal controls in a clustered dynamic environment, experimental results have shown the better safeguard performance of the proposed algorithm than other safe RL methods. The proposed algorithm could be easily extended to other RL algorithm with black-box implicit analytical model of dynamic systems.