sfujim / BCQ

Author's PyTorch implementation of BCQ for continuous and discrete actions
MIT License
597 stars 141 forks source link

Does discrete BCQ use vae? #13

Closed zichuan-liu closed 2 years ago

zichuan-liu commented 2 years ago

is this? """
i_loss = F.nll_loss(imt, action.reshape(-1)) """

thx!

sfujim commented 2 years ago

With discrete actions we can avoid using a VAE because a regular classifier can give multimodal outputs. See https://arxiv.org/abs/1910.01708 for details.