When to Switch: Planning and Learning for Partially Observable Multi-Agent Pathfinding.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Multi-agent pathfinding (MAPF) is crucial for coordinating multiple agents in shared environments.
• Traditional MAPF assumes full observability, which is often unrealistic.
• Partially observable MAPF (PO-MAPF) presents challenges due to limited agent perception and lack of communication.

Purpose of the Study:

• To develop and evaluate novel policies for the PO-MAPF problem.
• To investigate a mixed policy approach combining heuristic search and reinforcement learning (RL).
• To assess the performance and generalization capabilities of proposed policies against state-of-the-art methods.

Main Methods:

• Proposed two novel policies: one based on heuristic search and another on reinforcement learning (RL).
• Introduced a mixed policy that dynamically switches between the heuristic and RL approaches.
• Implemented three switching strategies: heuristic, deterministic, and learnable.

Main Results:

• The mixed policy demonstrated superior performance compared to individual heuristic and RL policies.
• The mixed policy generalized effectively to unseen maps and problem instances.
• Outperformed existing state-of-the-art PO-MAPF algorithms like PRIMAL2 and PICO.

Conclusions:

• A mixed policy approach is highly effective for solving PO-MAPF problems.
• The proposed switching strategies enhance adaptability and performance in dynamic environments.
• The developed methods offer a significant advancement in autonomous agent coordination under partial observability.