This is the source code of RPBT, which is proposed in the paper "Learning Diverse Risk Preferences in Population-based Self-play"(http://arxiv.org/abs/2305.11476). This repository provide a single file implementation of RPPO(Risk-sensitive PPO) in toyexample/rppo.py, and a lightweight, scalable implementation of RPBT(Population based self-play with RPPO). All the experiments were conducted with one AMD EPYC 7702 64-Core Processor and one GeForce RTX 3090 GPU.

• single-agent setting

  • Toy example in the paper
  • classic gym environment

• multi-agent competitive setting

  • Slimevolley
  • Sumoants

  The videos are available at https://sites.google.com/view/rpbt.

We use python 3.8

pip install -r requirements.txt

We provide a single file implementation of RPPO for toyexample.

run

python toyexample/rppo.py --env-id ToyEnv-v0 --num-steps 200 --tau 0.2

--tau is the value of risk level $\tau$ in the paper. If we set --risk False, we recover the original PPO.

The hyperparameter configs are in config.py. We provide 2 training scripts：

For Slimevolley, bash train_vb.sh

For Sumoants, bash train_sumo.sh

If not to use PBT, set --population-size 1 , we recover the RPPO.

We appreciate the following repos for their valuable code base implementations:

• CleanRL

• MAPPO