DiffLNS: Discrete Diffusion for Multi-Agent Path Finding
A novel hybrid framework called DiffLNS has been introduced by researchers, integrating a discrete denoising diffusion probabilistic model (D3PM) with the LNS2 repair-based solver for Multi-Agent Path Finding (MAPF). This task involves generating collision-free paths for several agents from their starting points to their destinations, a process that becomes combinatorially challenging in densely populated settings. Often, suboptimal initial plans lead to complex conflicts that complicate repairs. DiffLNS employs D3PM as an initializer, utilizing sparse social attention to derive a spatiotemporal prior from expert demonstrations, enabling the sampling of various joint plans. By directly engaging with the categorical action space, the discrete diffusion maintains the action structure of MAPF and samples from a multimodal joint-plan distribution, addressing the overlooked aspect of initial plan quality in repair-based solvers like LNS2.
Key facts
- DiffLNS integrates a discrete denoising diffusion probabilistic model (D3PM) with LNS2.
- D3PM serves as an initializer with sparse social attention.
- It learns a spatiotemporal prior over coordinated multi-agent action trajectories from expert demonstrations.
- The framework samples multiple joint plans.
- It operates directly on the categorical action space.
- The discrete diffusion preserves MAPF action structure.
- Initial plan quality critically affects downstream repair in LNS2.
- The approach targets complex and congested MAPF scenarios.
Entities
—