README.md

Multi-Expert Distillation for Few-Shot Coordination (Student Abstract)

This repository contains the implementation of Multi-Expert Distillation (MED), based on PyTorch.

1. Getting started

Use the install script to install the python environment:

bash install.sh
conda activate med

2. Run an experiment

All the experiments can be run with the unified entrance file examples/train.py with customized arguments.

LIPO

The repository consists of a re-implementation of LIPO. For generating a population in Girdworld MoveBox or Overcooked, enter the examples folder and run the following command:

python train.py --algo lipo --env gridworld --task MoveBox --map multi_exits --exp_name test --use_wandb True --pop_size 8 --horizon 50 --n_iter 500 --eval_interval 10 --n_sp_ts 5000 --n_xp_ts 5000 --eval_interval 10

python train.py --algo lipo --env overcooked --map_name full_divider_salad_multi_ingred --exp_name test --use_wandb True --pop_size 8 --horizon 100 --n_iter 1000 --n_sp_ts 5000 --n_xp_ts 5000 --eval_interval 10

The results and models can be found in the examples/results folder.

MED

To run MED, the population model files should be placed in the harl/runners/generalist_runners/models folder. Users should make sure the file is named properly. For running MED, enter the examples folder and run the following commands:

python train.py --algo med --env matrix_game --exp_name performance --t_max 30000 --n_episodes 3 --use_wandb True

python train.py --algo med --env gridworld --task MoveBox --map multi_exits --exp_name performance --t_max 2000000 --horizon 50 --n_episodes 2 --use_wandb True

python train.py --algo med --env overcooked --map_name full_divider_salad_multi_ingred --exp_name performance --t_max 7500000 --horizon 100 --n_episodes 2 --use_wandb True

Training scripts are also provided in the examples folder.