Can you show some papers, or theoretic materials for this algorithm?

[alanyuwenche]

Thanks for your sharing.

This "Deterministic" method, which policy only selects actions for states sampled from expert data, is quite interesting. Can you show some papers, or theoretic materials for this algorithm?

[alanyuwenche]

I suppose this "Deterministic" method is similar to "Behavior Cloning" in imitation learning. Therefore, it suffers from the well known limitation including sample- inefficiency, training instability, and dataset bias, which you recognized as high variance in results.

However, you use GAN-based method to learn the policy. Its performance should be better than traditional supervised learning. Finally, it is correct to set random_seed = 0, which means more variety in training.