20

*Jan*#### Mean-field Multi-agent Reinforcement Learning Platform

Zida Wu reinforcement learning; multi-agent; mean-field;**Paper Target**: IROS 2022

Current open source mean-field reinforcement learning code, which is based on MAgent

**Blue:** Group 1 (Proposed Algorithm)

**Red**: Group 2 (Opponent)

Algorihtm | Information | Collision | Task |
---|---|---|---|

Current | using all agents as neighborhood | One state (position) only one agent | Homogenous |

Our proposed | only using the closest neighborhood | Allowing overlapping | Heterogenous |

**Next Step**

- Step 1: Deal with information problem: using only closest neighborhood information
- Step 2: Deal with collision problem: allowing overlapping
- Step 3: Deal with task types problem: add heterogenous attributes.

**Term illustration**:

- A
**round**is the one batch training in multi-agent neural network. All algorithms are trained by 1599 rounds. - MF-AC: Mean-Field Actor-Critic algorithm
- MF-Q: Mean-Field Q-learning algorithm
- IL: Individual Q-learning algorithm
- AC: Individual Actor-Critic algirhtm

All experimental results fit the results in original paper: Mean field multi-agent reinforcement learning, which should be MFQ>IL>MFAC>AC. Explanation could be founded in original paper.

MFAC Group **VS** AC Group

MFAC Group **VS** IL Group

MFAC Group **VS** MFQ Group

MFQ Group **VS** IL Group