leo27945875 / Adversarial-Soft-Advantage-Fitting-1

Adversarial Soft Advantage Fitting - 1 (ASAF-1)

Imitation learning without policy optimization !

Introduction (In Traditional Chinese):

https://www.notion.so/Adversarial-Soft-Advantage-Fitting-441698eb0ccb40eab4f59275d637466a

Description

Preparing expert demos:

Every expert demo (state-action pairs) file must be a pickle file and in this form: [[np.array([state0]), np.array([action0])], [np.array([state1]), np.array([action1])], ...]

Training:

Adjust the parameters in ./src/train.py and then run it.

Testing:

Adjust the parameters in ./src/test.py and then run it.

Experiment

Reference

https://arxiv.org/abs/2006.13258