| --- |
| license: mit |
| tags: |
| - pytorch |
| - multi-agent-path-finding |
| - mapf |
| - decentralized |
| --- |
| |
| # DMM: Decentralized Master-Mind |
|
|
| DMM is a decentralized multi-agent pathfinding policy that refines agents' action |
| intents over several local communication rounds before committing to actions. |
| The models are pretrained on expert solutions with imitation learning and |
| optionally fine-tuned with MICPO, a critic-free reinforcement learning method. |
|
|
| [Paper](https://arxiv.org/abs/2609.32019) · [Code](https://github.com/CognitiveAISystems/DMM) |
|
|
| | Checkpoint | Imitation pretraining iterations | MICPO optimizer updates | |
| | --- | ---: | ---: | |
| | [DMM-08M.pt](DMM-08M.pt) | 1,000,000 | — | |
| | [DMM-3M.pt](DMM-3M.pt) | 1,000,000 | — | |
| | [DMM-MICPO-08M.pt](DMM-MICPO-08M.pt) | 1,000,000 | 96,000 | |
| | [DMM-MICPO-3M.pt](DMM-MICPO-3M.pt) | 1,000,000 | 96,000 | |
|
|
| MICPO fine-tuning comprises 500 outer iterations (96,000 optimizer updates). |
| Both model sizes use four communication rounds. Training details are reported in |
| [the paper](https://arxiv.org/html/2609.32019v1#S5.SS1). |
|
|
| See the [GitHub repository](https://github.com/CognitiveAISystems/DMM) for code, usage instructions, training, and evaluation. |
|
|
