I ran CFRM with my large abstraction of around 100 million information sets for 2 million iterations as a test.
Looking at the preflop strategy, I noticed it is folding the right range when comparing with Hyperborean but unlike Hyperborean, it's limping with a wide range of hands. Does this mean my CFRM is wrong or was Hyperborean hardcoded to disallow limps preflop?
Another unrelated issue: I'm planning to parallelize by performing simultaneous iterations. I'm aware that collision issues will occurring in case 2 threads randomly end up with the same information set. I think generating new cards in the case of collision wouldn't be a good idea because common information sets would be undersampled. Would blocking the colliding thread solve this issue or would it still result in undersampling?
Looking forward to your responsesStatistics: Posted by algonoob — Thu May 23, 2013 11:52 pm
]]>