Instead of training move played to 100% prior should be an option to say have it be X% (say 50%) and use leela network at Y nodes (say 800) to determine the rest. Would save data in format that has Q and node counts so could be added to later for ensembling, kinda like cyan's patch for training:
LeelaChessZero/lczero-training@adc3ddd
And described in this blog post https://blog.lczero.org/2018/10/understanding-training-against-q-as.html#more
But I believe he does it in realtime instead of ahead of time, and doesnt have such flexibility...
Instead of training move played to 100% prior should be an option to say have it be X% (say 50%) and use leela network at Y nodes (say 800) to determine the rest. Would save data in format that has Q and node counts so could be added to later for ensembling, kinda like cyan's patch for training:
LeelaChessZero/lczero-training@adc3ddd
And described in this blog post https://blog.lczero.org/2018/10/understanding-training-against-q-as.html#more
But I believe he does it in realtime instead of ahead of time, and doesnt have such flexibility...