Evolution of path costs for efficient decentralized multi-agent pathfinding

Efficient multi-agent pathfinding (MAPF) is becoming increasingly relevant in real-world scenarios. MAPF aims to minimize the travel time for robots operating in a shared environment. To date, substantial research has been dedicated to offline centralized algorithms based on conflict avoidance. Recently, an entirely decentralized algorithm, IDCMAPF, was introduced, premised on online conflict resolution. Decentralized approaches offer potential advantages in scalability, flexibility, and inherent robustness compared to centralized methods, as conflicts are handled locally. However, in scenarios with high-robot densities, decentralized conflict handling is not always effective. In this paper, we therefore propose a combination of online conflict resolution and evolved local path costs that promote conflict avoidance by discouraging paths through highly congested areas. We represent the environment as a directed graph and evolve local path costs. Our study examines two encodings: edge weight, an explicit encoding of all edge weights in the environment, and node vector, a compact encoding where each cell in the environment is assigned a two-dimensional vector. We conduct a comprehensive set of experiments to evaluate the performance of decentralized MAPF with evolved path costs and compare the results with state-of-the-art centralized algorithms on benchmark maps. Our findings reveal that edge weight encoding outperforms node vector encoding in complex, high-density scenarios. Conversely, the node vector encoding shows promise when specific environmental features, such as long corridors, are present. We also find that decentralized conflict handling, combined with evolved path costs, is effective and often yields performance comparable to state-of-the-art centralized planners.