My hex program

I created a hex program. Nothing that hasn't been done before, but it was fun. Only 10x10 for now. Will add more sizes.

Please try it and tell me what you think.

https://hexplorer-play.fly.dev

/Magnus Jäverberg

A clickable link.

/Magnus Jäverberg

Nice! But it's too strong for me on 10x10. Are larger grid sizes available?

Thanks. Adding more sizes is on the todo list, but I will first spend more time experimenting with 10x10. Right now I'm playing around with q derived policy. With time I intend to support 10 - 15, 17 and 19.

Isn’t Q-learning considered off-policy?
Can you tell more about your program, algorithms, language, and how it compares with existing ones?

Mostly standard alpha go style mcts so far. The training is in python, the rest in c#. Pytorch exports the model as ONNX, and I use that in c# to generate training data. I'm running about a 1000 trees in parallell to get a reasonable batch size, and I don't want to do that in python.

It annoys me that the dirichlet noise and other parameters heavily affect the policy target, so I'm experimenting with q based policy based on the paper "Policy Improvement by Planning With Gumbel".

I haven't tried it against other engines yet. I will, but it's a bit further down the to do list. So many other things I want to do, and not much time. Beside, I'm current spending part of my free time on another project.