基于好奇心探索的深度强化学习算法研究
首发时间:2020-05-12
摘要:强化学习广泛应用于系统决策等人工智能领域,凭借强大的性能优势,能够解决大量复杂场景下的智能体决策问题,具有很高的研究价值和意义。但是奖励的稀疏和延迟阻碍了智能体的策略学习,尽管目前出现的好奇心探索有利于增强智能体的学习能力,但是好奇心奖励的构造方式有待进一步改善。本文基于于智能体的好奇心探索,提出基于方向好奇心的强化学习算法,设计方向探测器,利用先验知识指导智能体的探索方向,规避风险探索对奖励信号进行衰减处理,并且在Atari的游戏场景中进行实验,最后取得了更高的得分。
For information in English, please click here
Research on Deep Reinforcement Learning Algorithm Based on Curiosity Exploration
Abstract:Reinforcement learning is widely used in the field of artificial intelligence such as system decision-making. With its powerful performance advantages, it can solve the problem of agent decision-making in a large number of complex scenarios, and has high research value and significance. However, the sparseness and delay of rewards hinder the agent\'s strategy learning. Although curiosity exploration currently appears is conducive to enhancing the agent\'s learning ability, the construction of curiosity rewards needs to be further improved. Based on the agent \'s curiosity exploration, this paper proposes a direction-based curiosity-based reinforcement learning algorithm, designing a direction detector, using prior knowledge to guide the agent \'s exploration direction, avoiding risk exploration, attenuating the reward signal, and playing games in Atari Experimented in the scene, and finally achieved a higher score.?????
Keywords: Agent decision Reinforcement learning Reward shaping Curiosity exploration.?????
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于好奇心探索的深度强化学习算法研究
评论
全部评论0/1000