---
license: apache-2.0
library_name: epymarl
tags:
- multi-agent-reinforcement-learning
- marl
- mappo
- smacv2
- starcraft
---

# MAPPO on SMACv2 protoss_5_vs_5

Multi-Agent PPO (MAPPO) agents trained with [EPyMARL](https://github.com/uoe-agents/epymarl)
on the SMACv2 `protoss_5_vs_5` scenario (StarCraft II).

- **Algorithm:** MAPPO (shared parameters, RNN policy, centralised value function)
- **Environment:** SMACv2 `protoss_5_vs_5`
- **Checkpoint step:** 10051667 environment timesteps
- **Greedy test win rate at upload:** ~0.50 greedy (converged, 10M steps; peak 0.54-0.65)

## Files
- `agent.th` — actor network weights
- `critic.th` — centralised critic weights
- `*_opt.th` — optimiser states
- `config.json` — full training configuration

## Usage
Load into EPyMARL by pointing `checkpoint_path` at a directory containing a
`10051667/` subfolder with these files:

```sh
python src/main.py --config=mappo --env-config=sc2v2 \
    with env_args.map_name=protoss_5_vs_5 checkpoint_path=<dir> evaluate=True render=False
```