计算机应用研究2024，Vol.41Issue(6)：1699-1703,5.DOI:10.19734/j.issn.1001-3695.2023.10.0440
面向动态三维迷宫的综合奖励设计
Integrated reward design for dynamic 3D mazes
摘要
Abstract
Dynamic 3D mazes present more challenging environments for reinforcement learning due to their uncertainty and incomplete information.Conventional reward functions can lead to slow and ineffective task training.This paper proposed an event-triggered integrated rewards scheme to solve the problem of finding multiple targets in a dynamic maze using reinforce-ment learning.The scheme expressed the various behavioral states in the 3D maze as events,which in turn derived the rewards.This paper divided rewards into environmental rewards and internal rewards.Environmental rewards directly related to the 3D maze mission and included node rewards reflecting the mission objectives and constraint rewards reflecting the mission con-straints.Internal rewards linked to the agent's emotional state during the learning process and encompassed both judgement and mood rewards.The average performance of the integrated reward shows a 54.66％improvement compared to the upgraded reward.The results suggest that the integrated reward scheme offers benefits by increasing satisfaction with task completion,promoting exploration,and boosting training efficiency.关键词
三维迷宫/奖励设计/强化学习/事件触发Key words
3D maze/reward design/reinforcement learning/event trigger分类
信息技术与安全科学引用本文复制引用
焦昌成,王少威..面向动态三维迷宫的综合奖励设计[J].计算机应用研究,2024,41(6):1699-1703,5.基金项目
国家自然科学基金资助项目(62073249) （62073249）
湖北省技术创新专项重大资助项目(2019AAA071) （2019AAA071）
