Open luiscastro193 opened 5 years ago
Good idea. I don't think it would be complex for DP, but we would require some changes in AlphaGo.jl, in order to make it differentiable. step!
function of Go is not differentiable, since it makes use of setindex!
. We'll soon have zygote supporting that as well.
In the meantime, may I ask how well could we expect it to perform? I mean, given enough training (even if such a thing is not realistic), could it compete with AlphaGo Zero?
Since AlphaGo is already implemented here with Flux, it probably shouldn't be hard to implement a similar experiment with differentiable programming.
Would it make sense? Or is it too complex for differentiable programming?