11-30-05 [poker]These are my lifetime stats that GoldBullion has seen :
chukb hands:9023 plays: 32%(K8o) raises: 45%(A8s) steal:447/599 bbf:136/224 cont:308/433 big:803= 8%
There are various little todos in Goldbullion to improve the stat tracking, but the big thing is to try to use the AI to watch the way people have played hands to build a better Bayes-based read on them. That is, given how they've played various situations in the past, what do we think they have based on their play so far in the current hand.
The basic way of doing this, as I've outlined in the GoldBullion AI notes, is to have a parameterized AI model for opponents. eg. the model has various parameters a,b,c,d, etc. which influence play. They start at some basic setting which will give "standard" play. Based on what you see of a person's play, you adjust the parameters so that the AI model will make actions that match their play in those situations. There are two primary ways to adjust the model. One is statistically - that is, if you see the % of hands someone calls, folds, raises, etc. you can tweak the model so that it will generate those same statistics over all. The other way is with specific hands. When you see someone's hole cards, you see how they played a specific hand, and then you can tweak the model so that it plays that hand in a similar way.
This is all very complicated by the fact thats that - 1) poker is a game of very sparse statistics, so you will only get very scattered data on their play, not enough to build your model without a lot of assumptions, and 2) there are lots of extra circumstances that affect people's play, eg. are they on tilt, what are the stacks and payouts, etc., and 3) people will occasionally make very strange plays that can throw off the model if you weight them too strongly.
Of course none of this is very exotic, this is just what I'm doing myself when I play.