Poker-AI.org

Wondering, if anyone has ever tried programming logic for Open Face Chinese poker - there is no money in it as rooms do not dare offer it because it's so much easier to solve than NL HE. However, it still seems like an interesting research topic or just a project, has anyone tackled it? I am trying to program logic for one and would like to toss ideas around if anyone is interested.

Cheers

Hi !
(Sorry for posting without having my presentation done for now, I promess I'll do it soon !)
I had a project to do a web-app several monthes ago (a year maybe) and gave up when all others came out, but I programmed the logic a very unefficient (object oriented) ugly (fast programming), badly tested ("yeeha, it seems to work !") and not documented (indeed) way.
I also ran naive tests to estimate the max % of eligibility when playing only for it (the result was something like 95% for HU out of position).
AFAIR my test program can take any end state evaluator knowing what has been dealt to everyone and just tries to play optimal ignoring how other players played.

I thought using MCTS to start an AI, it seems appropriate to me. Your thoughts ?

Do you have some clues of how easy it is to solve ?
I really did my experiments with no other tools than Brecher's evaluator (modified for my needs) and no theorical knowledge at that time (found no resource at all), and sort of left all that behind since then to concentrate on NLHE experiments.

But I'm still interested in chatting about OFC

I've got a working MCTS AI for OFC. You need to make lots of tweaks and adjustments to MCTS to make it work though, since a standard approach produces something that is too conservative. My current version is probably about the same strength as a good human player (and much better than any bots I've seen on iOS), but I haven't played with it in a while.

My current guess is that it's a long way from being "solved" but should be not too much work to get something that's very good.

Tysen

Ok, it confirms what i thought. My old complexity evaluations gave me so huge numbers that I was surprised by the "solving" evocation

I was going by Pokerstars' statement about the huge steps that software took in solving OFC (http://calvinayre.com/2013/11/21/poker/ ... g-variant/).

To me, fantasy land is the main difference from holdem (other than no need to worry about your opponents cards). What is the best to precisely evaluate it, designing rollouts perhaps? If so, is dealing all 14 cards at once (as in OFC Pineapple) a good approximation or would it have a fundamental flaw that I am missing?

You should worry about your opponents cards (count your outs, see where you beat them...) to evaluate your EV. But they are public indeed
IMHO there is a lot of main differences. Like... a lot.


Not sure to understand (poor poor english) what you mean by "designing rollouts", and evaluate what ? If it means evaluating payoffs of all terminal situations (even playing only for fantasy land), building a strategy is strongly bound to the order you (and your opponents) draw the cards and fill the lines slots, so enumerating all possible final states ignoring this doesn't really make sense. Maybe this metric could help to build fast and unprecise evaluators however...

EDIT : 13 cards only, greedy you...

Pitt - yes, you got the idea of what I meant by opponents cards - just the fact that there is so much less uncertainty than in NL HE. By "evaluating fantasy land" I meant how much gain do we achieve in the following round by getting there on average. The only proper way of doing it I thought would be rollouts, which are basically tons of randomly dealt hands.

Does that make it any better?

Sorry Freddo, I was off topic (me must learn better english)
Now rollouts seem better, but I don't think ignoring opponents in your FL (next hand) simulation will be precise enough.
To me there is more work to do to evaluate this. Like having good IAs playing rollouts of FL hands for each position.

Tysen, what value did you give to FL in your IA ?

I don't like values ignoring position and such but people seem to estimate it between 12 and 15 points (in HU I guess).
http://forumserver.twoplustwo.com/21/dr ... n-1366971/
http://www.runitonce.com/other/value-of-fantasy-land/

Thanks for the links, Pitt, at least I know I am not alone in thinking FL is important:). The opponent's scores while in FL can be first set to 0 and then adjusted as the number of trials increases. The average will probably converge after the first few 1000's hands - so, no big deal.

But, once the value of FL is set, it's MCTS all the rest of the way, starting with the first 5 cards dealt - that was my plan.

I currently have it set to 12 * number of opponents, but I have not optimized it at all. You shouldn't need to do any rollouts of the next hand - it's unnecessary since all FL situations are equivalent. It just needs to be a fixed bonus number (that could vary by position I guess) that is added to a qualifying FL ending in the tree search. Then, you optimize that bonus value by using an EA or similar. That was my plan, but I moved on to other projects.

Tysen

Sorry for my bad english again, a little bit hard to use the right words

When I spoke about rollouts (misunderstood word) I meant running simulations of FL hands to adjust the values before any use of them in a static way by the general algorithm.
The problem here would be the autoreference to the values in the simulation.
Adjusting with EA looks more appropriate.

Btw I think the values should be more generally dependant on the distribution of players in FL. Even if I never saw two players reaching it the same time