融合环境模型与深度强化学习的游戏算法

The game′s algorithm based on integrated environment model and intensive learning

针对基于深度强化学习的游戏算法在训练Agent过程存在的训练速度慢的问题,提出一种新的融合了环境模型与深度强化学习的游戏算法.该算法考虑到Agent以前执行过相关任务的经验的重要性,并将这些任务经验转化为环境模型.首先Agent从环境模型中学习,待Agent从环境模型中学习后,然后再与环境交互.这样可以让Agent少犯错误,加快Agent的学习速度.为了验证算法有效性,在ALE游戏平台上选取两部游戏进行两种算法的仿真实验.实验结果表明新的算法在选取的两部游戏中都可以有效的提高Agent的训练速度.

A new game＇s algorithm in integrating environment model intensive learning is put forward based on the problem of the games algorithm of deep reinforcement learning in the process of Agent training, which includes the problem of low training speed. It has considered the importance of related tasks taken by the agent, and translated these tasks into environment model. Firstly, Agent learns from environmental model, then it interacts with the environment. In this way, there will be less errors made by agent, and the Agent＇s learning speed will be faster. In order to test the algorithm efficiency, the two games using two different algorithms have been selected in the ALE game platform to conduct the simulation experiments. The result demonstrates that the new algorithm selected from two games can improve the Agent＇s training speed efficiently.