基于平均神经网络参数的DQN算法  

DQN Algorithm Based on Averaged Neural Network Parameters

作　　者：黄志勇 吴昊霖 王壮 李辉[1] HUANG Zhi-yong;WU Hao-lin;WANG Zhuang;LI Hui(College of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《计算机科学》2021年第4期223-228,共6页Computer Science

基　　金：教育部联合基金(6141A02011607)。

摘　　要：在深度强化学习领域,如何有效地探索环境是一个难题。深度Q网络(Deep Q-Network,DQN)使用ε-贪婪策略来探索环境,ε的大小和衰减需要人工进行调节,而调节不当会导致性能变差。这种探索策略不够高效,不能有效解决深度探索问题。针对DQN的ε-贪婪策略探索效率不够高的问题,提出一种基于平均神经网络参数的DQN算法(Averaged Parameters DQN,AP-DQN)。该算法在回合开始时,将智能体之前学习到的多个在线值网络参数进行平均,得到一个扰动神经网络参数,然后通过扰动神经网络进行动作选择,从而提高智能体的探索效率。实验结果表明,AP-DQN算法在面对深度探索问题时的探索效率优于DQN,在5个Atari游戏环境中相比DQN获得了更高的平均每回合奖励,归一化后的得分相比DQN最多提升了112.50%,最少提升了19.07%。In the field of deep reinforcement learning,how to efficiently explore environment is a hard problem.Deep Q-network algorithm explores environment with epsilon-greedy policy whose size and decay need manual tuning.Unsuitable tuning will cause a poor performance.The epsilon-greedy policy is ineffective and cannot resolve deep exploration problem.In this paper,in order to solve the problem,a deep reinforcement learning algorithm based on averaged neural network parameters(AP-DQN)is proposed.At the beginning of episode,the algorithm averages the multiple online network parameters learned by the agent to obtain a perturbed neural network parameter,and then selects an action through the perturbed neural network,which can improve the agent’s exploration efficiency.Experiment results show that the exploration efficiency of AP-DQN is better than that of DQN on deep exploration problem and AP-DQN get higher scores than DQN in five Atari games.The normalized score increases by 112.50%at most and 19.07%at least compared with DQN.

关 键 词：深度强化学习 深度Q网络 神经网络参数 深度探索 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]