Two Plus Two Newer Archives - Computer to play 2 Poker Pros for 50k

[ QUOTE ]
four 500 hand !duplicate! matches = 4000 hands. This may yield a statistically significant result.

[/ QUOTE ]

The "duplicate" thing will sure help but it has much to do with the styles too; if the opponent is pushing small edges by heavy aggression or by roping down a lot, then the fluctuations when one player only has a marginal edge (like 1.5 big bets per 100 hands) can give results like five big bets won per 100 hands during the first 5k of hands and five big bets lost per 100 hands during the next 5k hands. Then there are players who do a lot of e.g. (check-)raising rather than roping and the opponent plays it solid; in that combined case the fluctuations are much smaller, one being a heavy favorite to be ahead after 5000 hands even with a small edge.

If the computer or the humans get beaten in a big way like five big bets per 100 hands, it could be because of the fluctuations, but in case the humans say it was an easy game that they really did have that much of an edge, then there's about no way the computer is going to do anything but lose heavily. What those players will say about it is more important than the fluctuations, no matter what the result. As far as my experience against bots goes, they have been at least as bad as weak-passive real life opponents, and that's more true the less players on the table, e.g. at full ring they can do pretty well but as soon as it gets shorthanded they start losing much more, and heads up they get beaten even more.

One edge that the computer can have over humans is that humans have experience against humans, not against bots, and the match is short enough that there's not all that much time to learn, though during 2000 hands one can learn a lot, though the computer is made unpredictable enough so it will have some edge for a start, except that both players do rate to have experience against bots as they say themselves, and only if the bot will be very different the humans in the future may start to ask to practise against it before the match, like they do at chess, though in that chess case the computer can study their opponent's style before the match while that may not be so much a factor in this case.

I am not too happy about the factor that the computer is not adjusting its optimal strategy during the match, though it's also possible to use that adjustment against the computer - there's no problem in the real life as the opponents do not know they are playing against a bot. Then I also believe that if the computer is using the right style, it can't be exploited much even when one gets a complete line, but it would be better if the computer would vary its style during the game more or less randomly, though there is something like that in its strategy, so it's not playing just the same "optimal" strategy all the time as far as I understood. Additionally it could trick humans that when they start to adjust it picks a strategy that does well against that adjustment, and the humans will not know that.