Enhancing MADDPG for Multi Agent Path Planning: Based on LSTM and Progressive Course Learning

This study focuses on the field of multi-agent path planning and proposes an optimized MADDPG algorithm that integrates LSTM networks with a curriculum learning mechanism. It aims to address three major technical bottlenecks commonly faced by traditional multi-agent reinforcement learning approaches in path planning applications: the challenge of non-stationarity caused by dynamic environments, the difficulty of coordinated policy optimization among agents, and the inefficiency of the training process. By systematically analyzing the core features of the MADDPG algorithm-such as its distinctive centralized training with decentralized execution framework, gradient variance suppression mechanism, and adaptability to diverse scenarios-this research establishes two key optimization pathways. In terms of technological innovation, the study achieves dual breakthroughs. First, by incorporating the LSTM neural network architecture, it significantly enhances agents' ability to capture temporal dynamics, enabling a more accurate understanding of environmental evolution. Second, it introduces a progressive curriculum learning paradigm, implementing a staged task difficulty strategy that effectively refines the agents' learning trajectory. Empirical results demonstrate that the optimized algorithm exhibits outstanding performance in multi-agent path planning tasks. A high-fidelity simulation testbed with three heterogeneous agents was developed, and extensive comparative experiments confirmed the superiority of the proposed method. Experimental data show notable improvements across several key performance metrics: the task planning success rate increased to 95.12%, the average path length was reduced by 0.8 meters compared to baseline methods, and the average navigation time was shortened to 10.4 seconds-all significantly outperforming existing approaches. To further validate the algorithm's performance, the research team conducted a quantitative analysis using agent position heatmaps, which vividly illustrated the enhanced rationality of path planning and the improved environmental adaptability offered by the proposed algorithm.