Can LLMs solve large-scale pathfinding problems?
MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale
September 4, 2024
https://arxiv.org/pdf/2409.00134This paper proposes MAPF-GPT, a novel, decentralized solution for multi-agent pathfinding (MAPF) problems, where multiple agents must navigate a shared space without collisions, reaching their individual goals.
Key highlights for LLM-based multi-agent systems:
- Imitation Learning: MAPF-GPT leverages imitation learning, trained on a dataset of one billion observation-action pairs generated using an expert MAPF solver.
- Transformer Architecture: A decoder-only transformer processes tokenized representations of agent observations (local map data, agent positions, goals, and past actions) to predict optimal actions.
- Decentralized and Scalable: Each agent operates independently based on local observations, enhancing scalability for large-scale multi-agent applications.
- Dataset and Benchmark: The research provides the largest MAPF dataset to date, valuable for training similar systems, and a benchmark for comparing different approaches.
- Zero-Shot Learning: MAPF-GPT demonstrates zero-shot learning, effectively solving previously unseen MAPF scenarios.