关巍, 曲胜, 张显库, 等. 基于改进DQN算法的船舶全局路径规划研究[J]. 中国舰船研究. DOI: 10.19693/j.issn.1673-3185.03866.
引用本文: 关巍, 曲胜, 张显库, 等. 基于改进DQN算法的船舶全局路径规划研究[J]. 中国舰船研究. DOI: 10.19693/j.issn.1673-3185.03866.
shu
GUAN W, QU S, ZHANG X K, et al. Research on global path planning of ships based on improved DQN algorithm[J]. Chinese Journal of Ship Research(in Chinese). DOI: 10.19693/j.issn.1673-3185.03866.
Citation: GUAN W, QU S, ZHANG X K, et al. Research on global path planning of ships based on improved DQN algorithm[J]. Chinese Journal of Ship Research(in Chinese). DOI: 10.19693/j.issn.1673-3185.03866.
shu

基于改进DQN算法的船舶全局路径规划研究

Research on global path planning of ships based on improved DQN algorithm

    摘要
  • 摘要:
    目的 为提升实际海域环境下船舶航行路径的经济性与安全性，提出一种改进DQN算法的船舶全局路径规划方法。
    方法 首先，引入优先经验回放机制赋予重要样本更高的权重，提升学习效率；然后，再通过决斗网络和噪声网络改进DQN的网络结构，使其对特定状态及其动作的价值评估更加准确的同时具备一定的探索性和泛化性。
    结果 实验结果表明，在Maynila附近海域环境下，相比于A*算法和DQN算法，改进算法在路径长度上分别缩短了1.9%和1.0%，拐点数量上分别减少了62.5%和25%。
    结论 实验验证了改进DQN算法能够更经济、更合理的规划出有效路径。

     

    Abstract:
    Objective In order to improve the economy and safety of ship navigation path in the actual sea environment, a ship global path planning method with improved Deep Q Network (DQN) algorithm is proposed
    Method First, the preferential experience replay mechanism is introduced to give higher weights to important samples to improve learning efficiency. Then, the network structure of DQN is improved through duel network and noise network, so that it can evaluate the value of specific states and actions more accurately, and at the same time have certain exploration and generalization.
    Result The experimental results show that compared with the A* algorithm and DQN algorithm, the improved algorithm reduces the path length by 1.9% and 1.0%, and the number of turning points by 62.5% and 25%, respectively, in the Marine environment near Maynila.
    Conclusion It is verified that the improved DQN algorithm can plan the effective path more economically and rationally.

     

    • HTML全文
    • 参考文献(17)
    • 相关文章
    • 施引文献
  • 资源附件(0)

/

返回文章
返回