Poker-AI.org

Hey,

how can I parallelise my chance-sampled cfrm algo over multiple machines? It's pretty obvious that data-transfer between the machines is the bottleneck. Would it work if I just run the algo for the same gametree independently on two machines and merge the regrets and strategies after n iterations?

Thx.

combining the regrets/strategies from two runs of n iterations is definitely not equivalent to one run of 2n iterations. i'm guessing the improvement would be marginal.

Ideally you would want to partition it to use the least amount of network. i.e. Splitting it at earlier branches in the game tree.

The Polaris paper talks about parallel implementation. You will have to combine the results every iteration. The biggest advantage IMO is the Ram of 2 machines, not a speedup.