基于改进Q-Learning算法的智能体路径规划研究

打开文本图片集

关键词：Q-Learning；路径规划；动态学习率；蒙特卡洛树搜索；启发式奖励图分类号：TP391.4；G434 文献标识码：A 文章编号：1006-8228（2025）11-01-04

Path Planning for Agents Based on Improved Q-Learning Algorithm

Liu Shuo'，Dong Xisong²，Zhao Wei1

（1.School ofRail Transportation，Shandong Jiaotong University，Jinan，Shandong25oooo，China;

StateKeyboratoryfultimodalArificialntellgenceSstemsstitutefutomatoineseAcademyofiece）

Abstract：Withthegrowingdemandforpathplanningbyagentsincomplexanddynamicenvironments，thelimitationsof traditionalQ-Learningalgorithms—suchasslowconvergence，lowobstacleavoidanceeficiencyandlimitedglobaloptiization capability—havebecomeincreasinglyapprent.ToaddresstheseshortcomingsofQ-Learninginpathplanning，thispaperproposes animprovedmethodthatintegratesadynamiclearningrate，anadaptiveexplorationrate，andMonteCarloTree Search （MCTS）. First，anexponentiallydecayingdynamiclearningrateandexplorationrateareitroducedtobalanceexplorationablityintheearly trainingstagewithpolicystabiltyinlaterstages.Second，MCTSiscombinedwithQ-Learningtoleverageitsglobalsearch capablityforoptimizingtheQ-valueupdateprocess.Furthermore，aheurisicfunctionisincorporatedtoimprovethereward mechanism，guidingtheagentmoreeficientlytowardthetarget.Experimentalresultsdemonstratethattheenhancedalgorithm significantlyoutperformstheconventionalapproachintermsofaveragepathsteps，convergencespeed，andstabilityhisstudy offersan eficient and robust solution for agent path planning in complex environments.

eywords：Q-Learning;Path Planning;DynamicLearningRate;Monte Carlo Tree Search；Heuristic Rewai

0引言

在智能化浪潮奔涌的当下，制造业智能化转型迫在眉睫。（剩余7533字）