Poker-AI.org

Did some of you guys tried to compute online CFRM for postflop decisions?
As you know the flop, the remaining game tree is way smaller.
Is it enough to compute a good NE approximation of LHE HU in a short time or is it still intractable even with super computers?

The problem is that you can't view the postflop state isolated: CFRM needs to know with what ranges you and your opponent are going to the flop...

Realtime equilibrium computation is an easy way to face the problems of Sparbot/GS3: strategy knitting just don't work as good as is expected to. And you still need reach probabilities for every information set (it'll be a hard time to get a good estimation of them).
However, if I recall correctly, some of Tartanians solve their LP in realtime, so theoretically you may try to run CFRM the same way but it'll be hard to get good results in 2-3 seconds.

Isn't it the same idea that imperfect recall?
Both player will have all possible hole cards with the same probability, I know it isn't correct but on the other hand we have the benefit of a smaller tree.
Does it introduce a huge bias?
You can also have ranges from your expert model or precalculated with CFR-D http://poker-ai.org/phpbb/viewtopic.php?f=25&t=2433

haven't looked into CFRD into detail, but your first approach will not work: you are basically assuming any2 ranges in limped pots as well as in 4 bet pots...

Not if I consider preflop ranges from an expert model for example.

But then you calculate only a subgame with strong restrictions (the given ranges for your and your opponents). Within the sub-game, you'll get an optimal solution, but it might be quite a bit off from the unrestricted nash solution. The problem is you cannot judge a priory how big this gap is - it might be that its a minor issue and the improvement in speed outweights the accuracy, but it might also be the other way... Moreover, you cannot calculate the unabstracted game and the abstracted game and generalize the difference to other games (e.g. other stack sizes, other board textures, ...)
Speed-wise it should be doable online as the game is much smaller and you can change the card generation such that only hands part of the ranges are generated.

no, you won't get good results in time. maybe with heavy postflop abstraction.

you could still use the method of predefined preflop ranges for offline calculation though.

how many iterations did you go through per second? I never tried it but the arguments for quick convergence are:
- the board is fixed, so you only have to cope with 47*46 different turn/rivers
- the game tree is way smaller as you cut off a lot of the preflop branches
- you don't deal any2 ranges anymore

let's do some rough estimates:

140 preflop hands.
200 river sequences.

140*47*46*200 = 60536000 information sets

and so we need about 600 million iterations. we won't get that in a few seconds.

obviously it's not as many when removing isomorphs, but still not doable without buckets.

obviously I was thinking about using buckets... No one is using a non-bucketing approach even offline for the full game.

well, we probably can't reduce the abstraction, so what's the point?

in no limit this idea would hold some merit however. if we are facing an unexpected bet size, we may solve with that bet size included in the abstraction to remove the issue.

the problem is that to get results in time we have to abstract, a lot. probably too much.

which abstraction are you talink about? say we use 100 buckets with imperfect recall and LUTs, I'd guess we reach a decent strategy at least in a short time.

yes. but what would be the point when we can use 10000 buckets offline?

If you don't have the resources UoA has available I'm pretty sure you are not using close 10k buckets per street even offline... Furthermore, a ton of buckets don't necessarily lead to better results. And finally, a human implicitly creates maybe 20 buckets in his head with way less accuracy.

The online calculation can be a good idea, but river only .
It allows to use a wide abstraction because the game tree is so small.

The problem is you cannot judge a priory how big this gap is - it might be that its a minor issue and the improvement in speed outweights the accuracy

_________________
I have been a FunForFreedom survivor since 7/2/12!Runescape 2007 Gold
diablo 3 items Runescape Gold
Supporting Leon X Zara! <3
I have a blog!