Poker-AI.org

That's what makes me think that it stinks. I waste around a year in attempts to create a robust agent for multiplayer game using NN/reinforcement learning. Tried almost every and each algorithm and their augmentations, but it is impossible to ensure robustness (i.e. SNE) of resulting solution.

So, if their claims are true and pokersnowie's solution is any good, it'll still be heavily exploitable just because with neural networks you cannot guarantee convergence even to NE (not to mention SNE).

Heh, a guy named sn0w that tried the same thing as poker snowie. I'm curious what algorithms you tried, and how you tried them? I know with CFRM it's not such a straight forward task to train NNs.

It's a joke of universe. And a bad one.

Mostly I tried variations of RL. Also, NEAT was not bad at all. Just google around if you are interested in. In many cases the solution was good and pretty playable, but was always exploitable if two or more players are deviating from computed strategy.
Now I'm using multiplayer-modified version of CFRM and there are much less problems than with NNs. That's why I'm so sceptical about their NN methods whatever they are.

You used evolution to try and find NE? If so, what was your fitness function(s)? If you used reinforcement learning, what data were you reinforcing your learning with?

Snowie already has a successful backgammon business that has been running for years. He has been developing this for 10 years.

_________________
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!

Any info/links on this topic?

Don't recall any info or links. Suppose you deliberately played badly against someone and their strategy didn't change. Wouldn't you be a bit suspicious?

So, PS/FT do analyze each play in a manner "The fish hasn't been folding to any river bets 10 times in a row and the reg continues bluffing -> any good reg would stop bluffing -> this reg is a bot"? OR "SB shoves 10 times in a row preflop for 100BB HU and BB folds A2 on the 11-th time - > BOT!" Somewhat unbelieveable...

It seems to me they need to compute EQ for every hand and watch whether some players' actions correlate with EQ to a high degree in every hand he plays. But it also seems impossible to compute NE for thousands games/second that run on PS.

Chess seem much easier to detect bots since there are only few decent moves on every board and bots probably select some consistent lines on these boards. But in poker you can do several actions with different frequency with every hand in your range in every spot and the number of possible lines with your range is >>>>>> then in chess.

They do check the 'stability' of players, at least for some games (like HUL). Darse Billings (of the U of A), if I'm not mistaken, has consulted with casinos in this regard. This has been discussed on 2+2, etc. Also, I've had discussions with botters who have had discussions with PS security in this regard, where their accounts were flagged/suspended because they were too 'stable.' They, apparently, compare you to other players. It may not be enough to get you banned, but will probably get you suspended and seriously investigated until which time you can prove it was you that was playing (and this may take one of their guys coming to your home with a camera).

Where did I write "do analyze each play" (sic)? They could use the technique to confirm suspicions raised in other ways.



Of course it's hard to compute a NE in the first place, but relatively easy to recognize it once calculated. And again, it doesn't have to be every hand of every player.

I could offer more explanation but I'm not going to because I don't like your tone.

Sorry, I didn't mean to offend you in any way, just wanted to discuss it.

Yes, it's hard to compute NE, especially if there are thousands of e-EQs for different abstarction techniques/models and numbers of iterations run.

Could you please explain some more?

I'm tempted to make this my signature.