3/22/2006

03-23-06 [poker] - 1

03-23-06 [poker]

Poki-type AI's have the very bad property of exaggerating their mistakes. You have some AI function A() which returns {fold,call,raise} probabilities for a given hole and state. You use the same A() in your opponent model - that is, you assume your opponents think the same way you do, and this A() is used in the Bayesian way to guess the chance they have each hole, which you then use to measure how often you're ahead or behind in the hand.

If A() is too tight/weak, it will only bet with very good hands, and check or fold a lot. If you then use the same A() for your opponents, then you assume they are only betting with very good hands, which makes you even more scared!!

Similarly, if A() is too loose/aggressive, it bluffs way too much and plays a lot of junk cards, then you will assume that your opponents also are betting a lot of junk. This will make you want to bluff them even more, because you give them no respect for having hands, you'll try to bluff their bets & call their raises with really marginal stuff.

If you had an accurate opponent model, these flaws in A() would still be flaws, but by using the same A() for the opponent model it amplifies the problem, which is a horrible feedback loop.

Of course bad human players make the same mistake - they assume that the opponent plays the same way they do.

No comments:

old rants