facebookresearch / facebookresearch/BenchMARL
MAPPO fails to converge on VMAS Balance task with default configuration
- Dominant language
- Python
- Stars
- 661
- Forks
- 135
- PR merge metrics
- No merged PRs in 30d
Description
**Description**
I recently cloned the repository and attempted to run the default example following the README instructions. However, the training process fails to converge. Actually, it does not converge when using QMix.
**Modifications**
I have not modified any core logic.
The only change made was adding TensorBoard logging for visualization.
**Steps to Reproduce**
Clone the repository.
Install dependencies.
Run the following command: python benchmarl/run.py algorithm=mappo task=vmas/balance
**Expected Behavior**
The model should show learning progress and converge on the vmas/balance task, distance to goal should become shorter and shorter.
**Actual Behavior**
The training does not converge.Please see the attached TensorBoard screenshots below:
Contributor guide
Assessment
This issue has not been assessed yet.