Dual head network like AlphaZero

[b0n541]

I would like to use KotlinDL to create a AlphaZero like dual head network for my game.

Unfortunately I haven't found any hint on how to accomplish that.

Is this already possible or would it be a new feature to be implemented?

[zaleslaw]

Hi @b0n541 if you want to create an AlphaZero, could you please share some notes about Keras/PyTorch implementation for that model?

Probably after that I could say more about is it possible or not

[b0n541]

I have found the following information:

https://medium.com/applied-data-science/alphago-zero-explained-in-one-diagram-365f5abf67e0
https://github.com/deepmind/open_spiel/blob/master/docs/alpha_zero.md#model
https://github.com/deepmind/open_spiel/blob/master/open_spiel/python/algorithms/alpha_zero/model.py

What we would need is a way to split the network after the first couple of layers into two stacks of following layers with dedicated outputs for the game value prediction and the move prediction.

[EgoLLC]

Do you mean reinforcement learning?

Can I implement reinforcement learning in Kotlin DL to play chess or any other 1v1 board game where the model will play against itself?