Dear all,

I am a student in artificial intelligence and a passionate poker player. For my bachelor's thesis I wrote an algorithm which is able to yield pre-flop decisions. The aim is to later expand towards post-flop play.

In short: there is no pre-defined strategy (not even bet sizes and such kind of things), no 'training' (there is no Machine Learning involved at all), no big data sets (an 'average amount' of hand histories suffice - most of people on this forum defenitely have played enough hands); while the algorithm actively exploits players without having to pre-define ways in which to exploit players, and takes into account situational factors (such as positions, stack sizes, VPIP and PFR values of opponents). These things are good, since the 'bottleneck' in No-Limit Hold'em is generally the abundance of different situations - to 'compute them all' is generally impossible.

Although there is still much to improve, I think I proposed an at least interesting approach towards creating poker playing algorithms.

Now, I would like to ask you guys for an offer. It would help me extremely much if some of you could comment on some of the pre-flop descisions made by my algorithm. Preferably I am looking for players being aware of at least some poker strategy.

The questions take about 5 minutes in total to answer. Please send my a private message if you are willing to answer. Again: I would appreciate it extremely much!

Note: I will send the questionnaires by sunday evening or monday afternoon.

Why not post the survey link here instead?

Preflop algorithms based on opponent frequencies are not new in poker a.i. (but they might be in academia). These algorithms all have the same problem though- they cannot adjust quickly enough to counter-exploitation.

They treat opponent frequencies as a static variable. If your opponent suddenly changes their strategy, and you're only looking at overall VPIP/PFR, then it could take hundreds or thousands of hands before your algorithm adjusts. For that reason you need to develop some kind of edge-detection mechanism to pick up situations where villain has suddenly changed strategy. And all of this assumes you actually have information on your opponents, which often times you won't.

That's why the recommended approach is to develop a baseline strategy using game theory, then adjust slightly based on recent data.

I'd be willing to check out your survey if you send it to me.

Hi! Thank you very much for your response. I really appreciate that you would like to fill it in for me. I will send all surveys by sunday evening. I might also put a link in the topic by then.

The algorithm is not based on player frequencies but rather measures the 'similarity' between historical poker situations and a currently presented poker situation (though the VPIP and PFR values for all involved players do contribute to some of the 'characteristics' of a 'poker situation'). If a historical situation has a relatively high value for its similarity with a currently presented situation, it will have more 'signifiance' when using that situation to (for example) put together a range for an opponent. This it the main idea behind it. The frequencies are not directly used to make predictions for an opponent's action.

I've finished your survey. Interesting stuff. It seems to have 3bet sizing down. Range assumptions seem reasonable. It's doing some very strange stuff though like flatting KK in the SB vs an UTG open raise.

I imagine that your algorithm will be quite sensitive to the hand histories it uses to define "historical situations". Poker evolves so quickly that data from several years ago might not be very representative of poker today.

Could you tell us a little more on how it works? Is it pure Bayesian analysis?

where's the link