dimanche 2 décembre 2018

GTO 202 : building a flop subset with increased accuracy

After two years playing with Game Theory Optimal poker solvers, I have decided to compute a set of advanced features, which I will describe in this GTO 202 set of articles.

Solving perfect heads-up play starts with a detailled computation of Expected Value (EV) and Equity (Eq) of every hand of poker, against every hand that the opponent may have, on every postflop situation (after a flop, after a turn card, after a river card). This topic started three years ago and poker solvers like PioSolver worked on finding a list of weighed flops that would represent the whole postflop game. You may read the whole story from November 2015 here.

Still, the year 2018 is nearly close to an end : from three years ago, we now have increased computing power and this is an appropriate moment to make the same computations again, dedicating more computing time in order to achieve increased accuracy in the results.

The above image shows how exploitability (the blue line) decreases with time : a computer will quickly (about 150 seconds, when the red line goes up) take us from "bad quality" to "medium quality" accuracy; however it takes a much longer time to get from "good quality" to "even better quality" (nearly 700 seconds for most flops, up to 1150 seconds for the most complex flops). In simple words, it takes more than five times to compute a result with a five-time better accuracy (that is, dividing the exploitability by five).

This computation is currently running on my computer, and the estimated time to compute all hands, all flops, all turns, all rivers, is somewhere around 18 days. When this is done, I will take you to the next step : analyzing the characteristics of the computed data and creating simpler ways to describe this data sample. This will lead to the generation of new and improved flop subsets.