Abstract: Reinforcement learning is an important branch of machine learning.With the development of deep learning,deep reinforcement learning research has gradually developed into the focus of reinforcement learning research.Model-free off-policy deep reinforcement learning algorithms for continuous control attract everyone’s attention because of their strong practicality.Like Q-learning,algorithms based on actor-critic suffer from the problem of overestimations.To a certain extent,clipped double Q-lear-ning method solves the effect of the overestimation in actor-critic algorithms,but it also introduces underestimation to the lear-ning process.In order to further solve the problems of overestimation and underestimation in the actor-critic algorithms,a new learning method,randomly weighted triple Q-learning method is proposed.In addition,combining the new method with the soft actor critic algorithm,a new soft actor critic algorithm based on randomly weighted triple Q-learning is proposed.This algorithm not only limits the Q estimation value near the real Q value,but also increases the randomness of the Q estimation value through randomly weighted method,so as to solve the problems of overestimation and underestimation of action value in the learning process.Experiment results show that,compared to the SAC algorithm and other currently popular deep reinforcement learning algorithms such as DDPG,PPO and TD3,the SAC-RWTQ algorithm has better performance on several Mujoco tasks on the gym simulation platform.

Key words: Continuous action space, Deep learning, Maximum entropy, Off-policy reinforcement learning, Q-learning, Soft actor critic algorithm