Poker-AI.org

I'm currently evaluating a new abstraction technique and want to share how I tested it. For me, the quality of an abstraction is good, if it puts hands that have similar strategic implications (on a certain board) into the same bucket. Obviously we don't have an exact notion of whats similar in a GTO strategy, but I'm just using common sense and basic poker knowledge.
I created two 600 bucket abstractions, one based on EHS2, another one on my new method to evaluate. Next, I randomly generate 5 cards x-times and determine which buckets according to both strategy they belong. For each of those hand/board combos, we create 1000 other hand board combos and determine their buckets. If the buckets to both bucketing strategies are different, I just continue. If the bucket of strategy1 is similar, I store it in a list for s1 and similar for strategy2.
What we get is a random hand and 2 sets of other hand/boards, that would reflect a similar bucket according to strategy1 or 2. If we print them out, we can compare the results manually. For instance, here is an excerpt:

Hand & Board: 8h4h6dAh7d
NEW method: 4d8s2s6s7h; 8h7s4d6dJc; 3h6h4c7sTh; 5h4dAc2hAh; 7s4s6h8sAh
EHS method: Jh3cKs4cKd; 8h3hTh7cJc

We can see that we have a no SD value but a gutshot and a backdoor FD. The NEW method puts boards with also a guthot into the same bucket, while EHS has one hand which has the same property, but another hand (the first one) which has strategically nothing in common.

Hand & Board: 6h8d6c4c3h
NEW method: 7s9c6d7d4c; 7s2s7h4h6s; 8cKs6h7s8h; 5c2h3c6h5d; As6s6h5d7s
EHS method: 3hQsJc8cQh; 7hTs8h9c7c; AdKd5s5cQd; 5s9h3s9sAs; Qs9c9hAd6s; 8d8sTd5h9c; 3d8sTh5d8d; 5hQs8dQcJd; 7d2hTh7s3c; 8d3s6dJc8h; 4cQs4sTs3d; 2dTdTcJcJh; 2s8d8sJd5c; 9c3s8d9d5s; JhJsKsQs8s

Here we have a weak TP and the NEW method groups similar hands together (low board cards, weak top pair or better second pair). The EHS method however, merges very different hands together. From TP on a Q-high board to third pair.


Hand & Board: JhKhTcAh9h
NEW method: AsJsKdQs6s; Jc2c2hTcKc; 5hAh7h3h2d; 5c6cJcAc5s; 6c5c9c6hKc
EHS method: AdTsThQh7c; AsKc6s4h6d; Qd6s4s6c3s; 8sJsJhQsTs; 9hQsTc9d8s; 7sQdTs4d7c; 8c8d9h4h4c; 7c7h2cTd3h; 3s4h3c2c5s; 8dJh4dTd8h

Finally, we have a Monster draw (Nut-FD + gutshot) which means that we can play it very agressive. The NEW method puts similar hands into the same bucket, while the EHS method puts very different hands into the same bucket.


One further step to investigate abstractions is to see, where hands in the same bin of S1 are clustered to in S2 (and with which other hands) and vice versa. This should give you an idea, if the abstraction really separates hands according to their strategic impact or not.

I currently use a combination of EHS^2 and HP for my bucketing. Are you using a single variable? Despite being quite outdated, Ian Fellows used a combination of EHS and HP (clustered, I think) that seemed to work quite well in practice, even though his abstractions were fairly small.

I'm using an abstraction which is taking into account handstrength and potential versus different ranges on different board textures. One experience I have is that abstractions that work in practice well for FL don't perform necessarily as good in NL. The reason is - at least thats what I'm thinking - the exponential bet-sizing which changes the EV of drawing hands significantly.

This is a pretty nice technique. I'm interested to hear more about that new abstraction method as well, classifying hands into the same buckets like a poker pro would actually do has always been my ultimate goal. Looks like yours does that pretty well.

_________________
Cheers.