Action-Guided Adversarial Imitation Learning

building

Algorithm framework

agail

How to use

1. Download expert data

Download expert data from this dropbox url or google drive url. Unzip and place it in the scripts folder (i.e., scripts/expert_data/)

2. Running training scripts

This repository consists of two version of AGAIL in folder scripts: one for discrete actions and another for continuous actions.

Discrete action control:

# run a single process
python3 agail_trpo_cartpole.py --loss_percent 0.25 --seed 0 --algo agail # running agail
python3 agail_trpo_cartpole.py --loss_percent 0.25 --seed 0 --algo state # running state-GAIL
python3 agail_trpo_cartpole.py --loss_percent 0.25 --seed 0 --algo trpo # running TRPO

# run multiple process, e.g., run agail
sh run_cartpole.sh 

(Checkpoints and logs will be written into checkpoint and log_trpo_cartpole folder)

Continuous action control:

# run a single process
python3 agail_trpo.py --loss_percent 0.25 --env_id Hopper --expert_path expert_data/mujoco/stochastic.trpo.Hopper.0.00.npz --algo agail # running agail
python3 agail_trpo.py --loss_percent 0.25 --env_id Hopper --expert_path expert_data/mujoco/stochastic.trpo.Hopper.0.00.npz --algo state # running state-GAIL
python3 agail_trpo.py --loss_percent 0.25 --env_id Hopper --expert_path expert_data/mujoco/stochastic.trpo.Hopper.0.00.npz --algo trpo # running TRPO

# run multiple process
sh run_mujoco.sh

(Checkpoints and logs will be written into checkpoint and log_trpo_mujoco folder)

3. Plotting curves

python3 plot_curve.py --env_id Hopper --timesteps 5000000

Experiment outcomes

Overall
Effectiveness
Robustness

Citation & Question

If you use the repository for your research, please cite our work:

Adversarial Imitation Learning from Incomplete Demonstrations

Mingfei Sun, Xiaojuan Ma

International Joint Conference on Artificial Intelligence (IJCAI 2019)

If you encountered any problems when using the codes, please feel free to contact Mingfei (mingfei.sun@ust.hk). Or you can create an issue in this repo.

Visit www.mingfeisun.com for more research projects on the relevant topics.