Curriculum-guided decentralized deep Q-network for multi-agent pathfinding under intermittent obstacles

Abstract

Multi-agent pathfinding (MAPF) requires multiple agents to reach assigned goals while avoiding vertex, edge, and intermittent-obstacle conflicts in a shared workspace. Learning-based MAPF is attractive for online decision-making, but fully decentralized reinforcement learning remains vulnerable to sparse rewards, non-stationary interactions, and weak cooperative behavior under local observations. This paper presents EnhancedDQN, a fully decentralized deep Q-network framework that integrates an environment-level curriculum, a reward-responsive non-monotonic epsilon schedule, and a local cooperative observation-reward design. Experiments on 8 × 8 grids with 3 and 5 agents show that EnhancedDQN improves success rate, makespan, total time, path length, and average conflicts over an independent Q-learning baseline. The method adds no inference-time communication or centralized planning module. The reported values are point estimates; repeated-seed statistics, stronger baselines, and component-level ablations remain future work. These mechanisms are designed to improve learning stability while preserving the simplicity of independent execution.