Scholarlink

<p>Aiming at the problems of slow convergence speed, strategy oscillation and path redundancy of traditional deep reinforcement learning algorithm in UAV path planning task, a self-optimized soft actor-critic (SOSAC) algorithm is proposed in this paper. Firstly, the Q value prediction error is dynamically evaluated based on the critic network in each round of iteration, and the adaptive iteration loss is realized according to the probability threshold to reduce invalid exploration. Secondly, the learning rate of actor network and the learning rate of temperature coefficient are dynamically coupled with the prediction error of Q value to realize the dynamic adjustment of parameters, so as to reconcile the dilemma of exploration and utilization. Subsequently, the normalized multi-dimensional reward index realizes the centralization of reward and eliminates the strategy shock caused by scale differences, so as to enhance the stability of training. Finally, the SOSAC algorithm is compared with other deep reinforcement learning algorithms in three different simulation maps. Experimental results show that, compared with the existing TD3, DDPG, MNDDPG and other comparative algorithms, the average reward of SOSAC is increased by 75.3% -88.6%, and the shortest path length is shortened by 3.8% -9.8%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Self-optimized soft actor-critic: an efficient and stable approach for UAV path planning

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Zhaojun Zhang,
  • Shun Lu,
  • Guangyang Liu,
  • Peiye Cao

摘要

Aiming at the problems of slow convergence speed, strategy oscillation and path redundancy of traditional deep reinforcement learning algorithm in UAV path planning task, a self-optimized soft actor-critic (SOSAC) algorithm is proposed in this paper. Firstly, the Q value prediction error is dynamically evaluated based on the critic network in each round of iteration, and the adaptive iteration loss is realized according to the probability threshold to reduce invalid exploration. Secondly, the learning rate of actor network and the learning rate of temperature coefficient are dynamically coupled with the prediction error of Q value to realize the dynamic adjustment of parameters, so as to reconcile the dilemma of exploration and utilization. Subsequently, the normalized multi-dimensional reward index realizes the centralization of reward and eliminates the strategy shock caused by scale differences, so as to enhance the stability of training. Finally, the SOSAC algorithm is compared with other deep reinforcement learning algorithms in three different simulation maps. Experimental results show that, compared with the existing TD3, DDPG, MNDDPG and other comparative algorithms, the average reward of SOSAC is increased by 75.3% -88.6%, and the shortest path length is shortened by 3.8% -9.8%.