Poker-AI.org

After dismissing the CFR and neural net approaches, I decided to create a multiplayer poker AI based on what's taught in Doug Polk's Upswing course, which is based on GTO principles, i.e. trying to stay balanced. I'm thinking of open sourcing and/or making it available for testing to gather feedback. Here's an outline of the algorithm on the flop:

We will keep track of our range. When it's our turn, we decide how to play all hands in our range, so the output will be for example "raise pot with these hands, call with these hands and fold others".

First we decide our raise (bet) size based on the board texture, number of opponents and stack-to-pot ratios. Typically, the bet size should be around 60% - 80%. To keep things simple, there will be just one bet size in a given situation.

Then we determine the required hand strength to value bet, for example: "top pair with J kicker or better" which will give us the value bet range.

Next, we will determine the call range based on the minimum defense frequency and pot odds.

And lastly, we will choose a bluff frequency, sort remaining hands in our range by bluff suitability and take the first x hands, where x is given by the bluff frequency.

What do you think, what are the possible problems I might run into with this approach?

My main worries are:

- I have the feeling that minimum defense frequency is not a good way to determine the call range.
- We're not tracking opponents' ranges. Maybe use simple heuristics such as: when the opponent was the preflop aggressor, be more careful on A high flops.

This is pretty hard to do very well imo.



Those formulas for calling and folding are very inacurate imo. The equity and playability of hands changes things a lot. For example if your opponent has equity with his bluffs, you have to call much more in order to make him indifferent. Also there is the problem, that the solution of subgames is not always a part of a solution of the full game, so for example you can not use that formular on the river without looking closer at what happened before.

I guess you can probably still win on low limits if you do it well though. I always find those rule based approaches a bit hard though, most players who read that book will probably deviate from very strict rules at some point, if it seems -ev, but it's hard to teach the computer to do the same thing that humans can do intuitively.

Thanks, I agree with your assessment. Haven't moved much since making the post because it's hard. What would your approach to 6-max be?