Abstract: Traditional deep reinforcement learning methods rely on a large number of samples and are difficult to adapt to new tasks.By extracting prior knowledge from previous training tasks,meta reinforcement learning provides a fast and effective me-thod for agents to adapt to new tasks.Meta deep reinforcement learning based on maximum entropy reinforcement learning framework optimizes strategies by maximizing expected reward and strategy entropy.However,the current meta reinforcement learning algorithms based on the maximum entropy reinforcement learning framework generally adopt fixed temperature parameters,which is unreasonable in the multi-task scenario of meta reinforcement learning.To solve this problem,an adaptive adjustment strategy entropy algorithm is proposed.Firstly,by limiting the entropy of the strategy,the original objective function optimization problem is transformed into a constrained optimization problem.Then,the dual variable in the constrained optimization problem is taken as the temperature parameters,and the updated formula is obtained by solving the dual variable by Lagrangedualmethod.According to the updated formula,the temperature parameters will be adjusted adaptively after each round of meta trai-ning.Experimental data show that the average score of the proposed algorithm on Ant -Fwd-back and Walker-2D increases by 200,the meta training efficiency improves by 82%,the strategy convergence on Human-Direc-2D requires 230 000 training steps,and the convergence speed increases by 127%.Experimental results show that the proposed algorithm has higher meta training efficiency and better stability.

Key words: Maximum entropy, Meta learning, Reinforcement learning