Artificial Intelligence is Catching Up
I'm a big fan of any piece of software that can give me a better understanding of game theory optimal play. Recently, I've been experimenting with a Russian developed, cloud-based program called Simple Postflop. The program can take the pre-flop ranges of any heads up confrontation and compute perfect GTO outcomes across all 1,755 possible flop types in about an hour. Here's an extremely informative tutorial from Alex "asuth" Sutherland.
The most useful part of this software is its ability to differentiate between maximally exploitative and minimally exploitative play. Maximally exploitative play requires perfect knowledge of your opponent's tendencies. Minimally exploitative play only requires a basic understanding of your opponent's frequencies.
Poker is a game of imperfect information. How often do you really have a perfect understanding of how your opponent is structuring his ranges? Alex goes on to explain how the "maximal exploit" feature in CREV isn't especially useful in real life situations for this exact reason.
He uses a rock, paper, scissors example to explain these concepts. The GTO solution to RPS is to throw each hand gesture 1/3rd of the time at completely random intervals. Clearly, this would make it impossible for anyone to exploit you. But what if your opponent starts throwing rock half the time and you want to exploit this tendency?
It seems that the maximal exploit would be to always throw paper, but your opponent can compensate by throwing rock half the time and scissors half the time, thus neutralizing your advantage. He can do this because you don't have a perfect understanding of his tendencies. So how can you exploit him if you don't know what he's going to do the other half of the time? The solution is in this video.
Alex brings up a lot of really good points. I think the most interesting part of this discussion is the fact that Simple Postflop suggests that range advantages can be strong enough that Villain has no incentive to meet minimum defense frequency on certain flops. This is a concept that game theorists have hypothesized for a long time. It's exciting that our computing power is finally strong enough to verify this.
It's even more exciting to see that this software can be used to boost the profitability of pre-flop flatting ranges, especially if you're using a mixed strategy. Still, it's extremely time consuming. Even the cloud isn't powerful enough to do this in a reasonable amount of time.
Chess is a game of perfect information and it still took IBM years to solve it using Deep Blue. Backgammon is a frequency based game of perfect information and it still took BGSnowie a really long time to solve it. Poker is a frequency based game of imperfect information with ever-changing bet sizes and stack sizes. It's exponentially more difficult to solve. PokerSnowie has run trillions of simulations in the last ten years and it still makes mistakes.
Watch this video to get a better idea of the complexity. It took Simple Postflop several hours to run these pre-flop flatting range simulations. The results are extremely interesting.
We've finally figured out how to solve flop to river range versus range simulations. The last step is gathering enough computing power to finally solve pre-flop play. We're getting close. Take a look.
Conclusion: we're all really bad at poker. The big winners are just less bad than everyone else. I'm not the best -- far from it. But I'm learning at an accelerated rate and even if this site never earned me a dollar I would continue to invest my time developing it because it forces me to improve.
Every time I coach a student I have an opportunity to look at poker from a different perspective and my game has improved tremendously as a result of those opportunities. If you care to improve with me, simply fill out the form at the bottom of the home page. Until then, good luck at the tables!
