Sampled evaluation games

[TimYuenior]

Sampled evaluation games

[Narsil]

To be honest, I added evaluation games to training in order to use more data available.

It will decrease exploration but running this algorithm on a single machine with a singe GPU, it is kind of hard to replicate results anyway. I tried focusing on 9x9 which should be more easy but to be honest it does not achieve good performance even on this.

I suspect a bug somewhere more important than just this exploration problem but I might be wrong.
I haven't had time to focus on this lately, but my intuition was that it would be better to recheck the basic steps of the algorithm first.

Hope that helps.

[TimYuenior]

Yeah, it's hard to tell what actually is affecting the performance without 64 GPUs available.
Anyways, thanks for the response!

[TimYuenior]

As a side note. Appending the state samples and the target labels into 3 separate big tensors and giving it do model.fit once (with the number of epochs) is a fairly big speedup over calling model.fit for every sample state.

Edit: Ignore what I said, the tensors will be incomplete like this or you'll retrain on previous samples.