"Intel tbb lib, SIMD, hand isomorphism, O(n) showdown calculations and other optimization hacks - all this will be needed in final realization."
If you wanted to make a really optimized implementation you would do it all on gpu (only moving the new samples to cpu to add to the memory buffer) with large batched inference calls. So then why are you using intel tbb?
What is the purpose of hand isomorphism when the network generalizes over the card embeddings?
O(n) showdown calculations are not possible if you are sampling chance (as is done in external sampling)
Anyway the real bottleneck is in training the neural network each iteration, you may aswell just use python for everything and not worry too much about optimizing the mccfr iterations
Training will also be very expensive, glStatistics: Posted by llllllllll — Tue Sep 07, 2021 8:10 am
]]>