摘要
随着深度强化学习的不断发展,深度Q网络(DQN)在机器人路径规划中得到广泛关注和研究.首先,简要介绍DQN以及Nature DQN、Double DQN、Dueling DQN和D3QN等算法的基本原理和改进思想.针对算法存在的样本获取成本高和交互效率低的问题,系统梳理并总结了从奖励函数、探索能力、样本利用率等方面进行优化的研究成果和思路.最后,讨论了DQN在现代物流中进行机器人路径规划的优势,对每个场景提出了算法的优化方向,涵盖状态空间、动作空间以及奖励函数等多个关键方面.
Abstract
With the continuous development of deep reinforcement learning,deep Q-learning network(DQN)has received extensive attention and research in robot path planning.Firstly,the basic principles and improvement ideas of DQN and algorithms such as Nature DQN,Double DQN,Dueling DQN and D3QN is briefly introduced.In view of the problems of high sample acquisition cost and low interaction efficiency in the algorithm,the research results and ideas of optimization from reward function,exploration ability,sample utilization rate,etc are systematically sorted and summarized.Finally,the advantages of DQN in robot path planning in modern logistics is discussed,and optimization directions for each scenario is proposed covering key aspects such as state space,action space,and reward function.
维普
万方数据