jjbrr, on 2011-May-27, 13:55, said:
Human behavior and randomness have nothing to do with how the bots are programmed.
What is GTO? Donald Michie's match-box computer showed that optimum strategy depends on opponent's skill. In the multi-player example, against a tyro, a theoretically unsound strategy can win faster. Also, I think the computer could be good at detecting patterns in its opponent's play and exploiting play-history databases.
jjbrr, on 2011-May-27, 13:55, said:
The goal of programming the bots is to have the bot play as close to GTO as possible, otherwise it would be possible to find a way to exploit the bot's strategy. It absolutely does not depend on what anyone else is doing in the game. It does not assume the other player plays the same strategy. Indeed, it doesn't depend at all on what other strategy anyone else is playing. This is why you could publish your strategy or announce it before each action and it would not have any impact on the ability of people to beat you. Randomness has nothing to do with it in the long run.
I'm not sure about that. One goal of the programmer would be unpredictability. If for example, the computer bluffed on certain kinds of hand, in certain types of context, the programmer would not want to publish that strategy.
jjbrr, on 2011-May-27, 13:55, said:
The problem with programming a bot to play GTO is a matter of computing power. I think there are something like ~3.7*10^15 possible combinations of hands+betting action in HULHE (I didn't have 20 hours of college math, so if I'm way wrong, I apologize. 4 betting rounds, 16 ways a round of betting can go, 52!/(45!*3!*2) possible board + hole card combinations). It takes a long time for a computer to figure out how to play each one of these combinations correctly. It's not hard to see how that number jumps even higher when you add more players or allow bets in various sizes.
I've never played Holdem but I think jjbrii exagerates its complexity. I doubt that a programmer would even try to cater, individually, for each possibility and I think a computer-array would have more than enough power to handle generic cases. The difficulty is more in heuristics.
jjbrr, on 2011-May-27, 13:55, said:
So while it may be possible that we have a winning NLHE bot soon enough, don't hold your breath, and certainly don't expect more than heads up. It probably won't play the equilibrium strategy either, but close enough.
I won't hold my breath but I bet jjbrr a shilling that there will be a winning NLHU program within two years.