I'm wondering how people are using weight tables(probablity of what hand each opponent have).

Currently I am using this strategy:
First I save information on how often each opponent bets and calls each street and then change the WT based on that.

Example:
At the start of each hand every card in the WT has a probability of 1.
John raises 15% of the time preflop and calls 20%. This time John calls, so I change the weight of the top15% hands and the bottom65% to almost 0.
If John would have raised instead I would have changed all but the top15% hands to almost 0.

This strategy is easily exploitable though, if a player only raises with AA and 32o the bot will think that this player only raises KK/AA which makes it easy to bluff.
Anyone know how to fix this problem?

Hi,

The trouble with that approach is maybe that you get too much contrast (ratio of probabilities between most likely hands and least likely hands) - particularly when the opponent has a tight range. In general having too much (inappropriate) contrast in an opponent model is more dangerous than not having enough I think.

At the moment you are trying to directly estimate the probability that the opponent has a certain hand given their action. Instead maybe try estimating the probability that they do each action - given that they have a certain hand. If that looks easier then check out bayes theorem.

- PeppaPig

I agree with the OP that using a probability weighting on opponent hands based on their pre-flop betting behaviour is an extremely useful approach (the probability of winning calculated in the OpenHoldem bot platform is based on this). On average it gives good results. Bizarre opponent betting can 'fool' it, but such strategies are not very effective long term.

Regards
M.

There are some good articles on how they calculate their weight tables from the University of Alberta Poker Research group.

http://scholar.google.com

Considering only PreFlop:

In theory you are doing right, but some things must be taken into consideration.
For example, how are you ranking the hands in preflop, to know which ones are the best 15%? Keep in mind that this list might not be the same for your opponents or might not be static (changes from time to time).
Another thing, there is much more in the preflop than just the rank of the hole cards:
The opponent might raise AK if he's in late position, but might never do it in early position against a full ring of players. Or may do it sometimes only.
Also, he might slow-play some hands.

First of all I think you shouldn't be so strict with the limits, it might be better to use linear interpolation (or another function) between the threshold for the action frequency and some variance. This is described in UoA papers.

Secondly, from my experience the re-weighting in preflop might have some drawbacks.
In one of my assessments, I concluded that decreasing the weights for the top% hands, when an opponent calls was not efficient, as many times the opponents would not raise preflop, because of some other factors (Early position, slow-playing, number of players, randomly...).

So, maybe you should add some noise to it. And do lots of testing.

Also, PeppaPig' opinion is valid, you can change the way of reweighting, by iterating through the opponent's possible hands and calculating the probability he would do such action. This is much like the Probability Triple's approach by UoA. In essence, if you use just the threshold of an action frequency to compute the probability triple, you'd be doing the same as before. But with this approach there's room to improve, as you can more easily account for other strategies (depending on a hand: like a flush draw) and for pot odds (more interesting in postflop phase).

Anyone thought about taking into account particular game context when calculating preflop WT? For example, players tend to limp in more frequent after winning previous round (at least I observed it at 0.25/0.50 stakes).

I don't think anyone who have this approach has taken any psychical aspects into account. There's much info as there is without these and the psychical aspects are often hard to log.