Would there be any merit to compiling detailed stats of the reigning equilibrium bot champion and then learning to play against it for every branch of the game tree? Its game tree should represent a (near) Nash equilibrium, no? So why couldn't one simply find a counter strategy given sufficient data for each decision that maximizes the gain? I imagine such calculations would be fairly trivial once all of its later street tendencies are known with a high degree of accuracy, just like it becomes easy to know the right play when someone only bets the turn with >1p over 500 hands. With extensive hh's we know its complete strategy a priori. So we can devise an optimal response. And that response should be a (near) Nash equilibrium limited only by how close the bot's strategy itself is to that status and secondarily, as our information on its strategy approaches perfect, right?
My thoughts on this are for practical applications. So that if you found yourself getting raised on a j55 flop you could ask, what would Sonia (or whatever) do here? The bots ofc take board texture into account but if you could filter for similar board textures then get frequencies of what the bot did in that spot, you could get a good idea of what the gto play is there. And that sort of method could be used anywhere on the game tree. So you get bet rr'ed on the turn with top fd+bp, what would the bot do? Could this method be useful or is it just chasing shadows?
If you find the thread that the University of Alberta guy was posting in, he stated that they have a method of determing the perfect exploitative strategy versus any of their bots, and that they use this to determine how far away the bot is from GTO. His own opinion was that the resulting strategy is likely useless outside of that context.
I think you can learn a lot from studying how the bots play, and I haven't done as much of that as I would like to. I still think for all the talk of GTO in the short-handed/HU LHE community, the key to success versus other humans is still in the game of adjustment and re-adjusment (i.e., identifying specific ways of exploiting your opponet's play while simultaneously balancing your own plays -- the latter of which is a GT concept but does not imply GTO in this context).
You could do that but since we're idiots there's no way we could remember the full strategy for every spot. There are also so many nodes that we'd never be able to infer the bot's whole strategy based on the limited sample.
Ignoring suits there's something like 1,755 different flops. The best you could do is build a "model" of how various flops differ to get this number down to a more manageable number. Then you'd try to figure out the bot's approximate strategy for each different "flop-type" in your model. Then repeat for all later streets.
Thanks for the replies. Sounds like just studying the bots play and trying to make general inferences would be a better route, I was impressed with your analyses and explanations in The Intelligent Poker Player and would like to incorporate that sort of thing into my learning process as well.
Muppets, do you have a link to that thread with the bot guys explaining their methods? Thanks.
Two Plus Two
Linear Mode
