未知环境下基于Dueling DQN的无人机路径规划研究
Research on UAV path planning based on Dueling DQN in unknown environment
赵恬恬 1孔建国 1梁海军 1刘晨宇1
作者信息
- 1. 中国民用航空飞行学院空中交通管理学院,广汉 618300
- 折叠
摘要
为有效解决无人机在未知环境下的路径规划问题,提出一种基于Dueling DQN的路径规划方法.首先,在DQN的基础上,引入对抗网络架构,从而更好地提高成功率;其次,设计状态空间并定义离散化的动作和适当的奖励函数以引导无人机学习最优路径;最后在仿真环境中对DQN和Dueling DQN展开训练,结果表明:①Dueling DQN能规划出未知环境下从初始点到目标点的无碰撞路径,且能获得更高的奖励值;②经过50000次训练,Dueling DQN的成功率比DQN提高17.71%,碰撞率减少1.57%,超过最长步长率降低16.14%.
Abstract
In order to effectively solve the path planning problem of UAV in unknown environment,a path planning method based on Dueling DQN was proposed.Firstly,on the basis of DQN,the adversarial network architecture is introduced to better im-prove the success rate.Secondly,the state space is designed and discrete actions and appropriate reward functions are defined to guide the UAV to learn Xi optimal path.Finally,the DQN and Dueling DQN are trained in the simulation environment,and the re-sults show that:①Dueling DQN can plan the collision-free path from the initial point to the target point in the unknown environ-ment,and can obtain a higher reward value;②After 50000 training sessions,the success rate of Dueling DQN is 17.71%higher than that of DQN,the collision rate is reduced by 1.57%,and the rate of exceeding the longest step size is reduced by 16.14%.
关键词
无人机/路径规划/深度强化学习/Dueling/DQN算法Key words
UAV/path planning/deep reinforcement learning/Dueling DQN algorithm引用本文复制引用
基金项目
四川省科技计划资助项目(2022YFG0210)
中央高校基本科研业务费专项资金资助(PHD2023-035)
中央高校基本科研业务费专项资金资助(ZHMH2022-009)
出版年
2024
NETL
NSTL
万方数据