uidilr / gail_ppo_tf

Tensorflow implementation of Generative Adversarial Imitation Learning(GAIL) with discrete action

Generative Adversarial Imitation Learning

Implementation of Generative Adversarial Imitation Learning(GAIL) using tensorflow

Dependencies

python>=3.5
tensorflow>=1.4
gym>=0.9.3

Gym environment

Env==CartPole-v0
State==Continuous
Action==Discrete

Usage

Train experts

python3 run_ppo.py     

Sample trajectory using expert

python3 sample_trajectory.py

Run GAIL

python3 run_gail.py  

Run supervised learning

python3 run_behavior_clone.py 

Test trained policy

python3 test_policy.py  

Default policy is trained with gail
--alg=bc or ppo allows you to change test policy

If you want to test bc policy, specify the number of model.ckpt-number in the directory trained_models/bc
Example

python3 test_policy.py --alg=bc --model=1000

Tensorboard

tensorboard --logdir=log

Results

Fig.1 Training results	legend

LICENSE

MIT LICENSE