Abstract
Multi-agent reinforcement learning for multi autonomous underwater vehicle path planning faces significant technical challenges in data collection tasks within the Internet of Underwater Things. These challenges are particularly pronounced under dynamic underwater conditions, where the policy network tends to prematurely converge on suboptimal policies, thereby undermining its adaptability to environmental variations. As adaptability continues to decline during extended exploration, training stagnates, and learning efficiency decreases. To address these challenges, we propose an adaptive plasticity injection based twin delayed deep deterministic policy gradient (PITD3) framework that integrates an adaptive trigger and a dual-mode plasticity injection method. The adaptive trigger continuously monitors fluctuations in reward and temporal difference loss, activating plasticity restoration when learning stagnates or exploration becomes inefficient. The dual-mode plasticity injection method reinitializes the actor head and critic heads through selective parameter injection, restoring learning capacity and preventing long-term performance degradation. Experimental results demonstrate that the proposed PITD3 framework out performs baseline frameworks, achieving substantial improvements in collision avoidance, data collection rate, and energy efficiency.
Funding Information
  • National Natural Science Foundation of China (U22A2011)
  • Open Fund of State Key Laboratory of Acoustics and Marine Information (SKLA202501)
  • National Natural Science Foundation of China (62572107)

This publication has 36 references indexed in Scilit: