Poker-AI.org

Thoughts about this http://www.pokersnowie.com/ ?

sounds like a scam to me...

Interview with the CEO of Pokersnowie
http://www.youtube.com/watch?v=T9l4nXqydlg

It seems like players are concerned it would kill online poker. Many reference to "smart guys" and pros stating it would be unbeatable by human players, many others give examples, trying to prove it would be total scam.
For sure the guys that created Snowie are familiar with AI (also see http://www.bgsnowie.com/) I assume their product to be better than any bot created by one person only.

On the other hand they are bounded by the same restrictions as we are so unless they have access to a quantum computer I assume (at least their 6max-version) to be far away from world class level.

I mean: UoA is now researching in the domain of poker for more than 15 years; and their statements seem to be way more realistic "[...] amateur level [..]". The point simply is that poker players don't have access to anything closer to EQ so it will be recommended almost everywhere and the guys from Snowie will make a lot of money until better programs appear.

Anyway, this appears nothing but arrogant and wrong to me http://www.pokersnowie.com/about/pokersnowie-the-best.html



I would love to see their product at the yearly http://www.aaai.org/home.html contest (http://www.computerpokercompetition.org/ in particular) - but I am almost sure they would be afraid their claims would not hold The competitions seems to be held this month. Maybe that's why they not release their software before July

PS: You were wrong http://poker-ai.org/archive/www.pokerai.org/pf3/viewtopic59af.html?f=3&t=3123&view=previous

http://www.computerpokercompetition.org/ requires competitors to submit executables and doesn't permit an internet connection. Nobody who wants to make a business out of this is ever going to agree to those terms.

I seriously doubt this is a scam. Snowie already has a successful backgammon business that has been running for years. He has been developing this for 10 years. These are not the trademarks of get rich quick scammers.

And I don't see why somebody shouldn't be able to do better than U of A. Sonia did it already. Personally, I think there is a big flaw in the U of A methods: I just wish I could find the time to develop my own better idea.

Neo poker?

Where did you see it's not released until July?

it was mentioned in the interview (link above)

It's all relative. Remember when Poker Academy and Sparbot first came out? They had The Count, a pro player, test it's HUL HE play... If I'm not mistaken he thought it was good. Then Ian Fellows released Fell Omen 2, which at the time was a strong(er) HUL strategy, but it too eventually became regularly beatable. The guys that created Sonia (a Bayesian BR player) beat the U of A's best EQ in an exhibition match, but their online "coaching" service went out of business after a few years.

EQs are easy to detect online -- Poker Stars and Full Tilt actively look at the stability of a player's strategy to discern if he/she is a bot. Billing predicted this would happen, much like it does with Chess bots.

It might make those that use it better players, but there will always be fish -- the occasional hobbyist that can't or doesn't want to spend the effort to improve his game.

Looking at the pokersnowie coaching software, it seems to be offering pure solutions preflop (i.e. saying an action should be taken 100% of the time) and giving only f/c/r options with a single recommended raise amount (not a distribution). I'm very curious about the exploitability of it's play, given that we've seen NNs that play pure strategies win competitions (Pulpo, Calamari, etc.) but remain extremely exploitable.

if they play by gto, as written at their website, it will not be exploitable...

Neural networks found a NE solution via self play in multiplayer game? That stinks. They missed "quantum" somewhere on the site.
I think, it's all about money and nothing more. Really, who of us wants to sell a money printer?

... within their abstraction which seems to be limited to one bet amount and might therefore very well be exploitable in the unabstracted game.

IIRC the game tree of HUNL-100 consists of about 10^70 nodes; the game tree of HUL, which - as well as their abstraction - is limited to one bet amount each action, consists of about 10^12, which is approx. 0.00000000000000000000000000000000000000000000000000000001% of HUNL-100's complexity, so by human means the game has been simplified by a factor of 100%. And then you only have a model which is about the size of HUL, which ... I think? ... is still not perfectly solved nowadays. Take these numbers with a ton of salt, give each exponent a range of +/- 5.

So I would agree that there might remain some exploitability in the unabstracted, REAL HUNL

Actually they are not selling the AI. IIRC the CEO mentioned in the interview it is going to be subscription-based. You may upload a couple of hands and then Snowie tells you how to improve your game.

Just curious how many players are going to expose their strategy (in terms of hand histories) to some strangers giving them the opportunity to perfectly exploit them

Maybe he thinks he can repeat the multi million pound success he made with his backgammon product. Maltese company accounts available if you want to look.

Maybe... Well, they could say whatever they want, but it will take at most three-four weeks to integrate AI into poker client. And if their solution is so good, they will make much more money (in order of magnitude) then any subscription service ever will.
All of this looks really strange.

- Sell 100 subscriptions and charge 10% winnings you make 10 times what you could make yourself.
- I doubt psuedo NE play actually makes big money.

As you all know the game is just astronomical http://webdocs.cs.ualberta.ca/~johanson ... l-size.pdf.
And this is just HU...
I am very skeptical of their claims.

It's exploitable, unless it's a perfect EQ. If a pure strategy, it will be easily identifiable.

If you notice on the their main front page with the "Snowie Tip" it mentions floating with that gutshot. This does not sound like an equity based bot or NE as floating would be -EV on any math calculation. This sounds more like an expert system to me. I am not sure how floating could be figured out mathematically but it makes sense on certain boards (mostly wet) and player actions. I have several floats incorporated into my bot but I need to add in many more. Technically I think any kind of backdoor draw/gut could be considered a float but I think having nothing but air and no outs would be a "true float" if that makes any sense.

In fact after thinking about that Snowie Tip I think its totally wrong. You cannot rep many hands on that 95A rainbow flop. So apparently he called that flop with no draws present so all he could rep is top pair or a set which is rather unlikely.

Playing devil's advocate a bit here - that is human thinking not EQ thinking. Every non-terminal action in an EQ solution needs to contain strong and weak hands. If you never call on flat boards with weak hands then that is exploitable. Floating is not -EV by a math definition, it is part of a balanced strategy

Tysen

That is a very good point. I will have to think about that as I am probably missing out on opportunities on dry boards. I have it calling many other types of weak hands on dry boards so perhaps I have already compensated for that. I have to gone to big lengths in order to make it balanced. I guess my definition of floating is probably not right. When I think of floating I think of repping a hand I do not have with total air. So I have no draw, no pair and no outs. So then I take my preflop action and board texture and determine if I could credibly rep a draw (like if I called a preflop raise which would generally indicate a typical flatting hand like a broadway, suited connector, etc) or a made hand in order to take the pot away. On a dry board I could really only rep a made hand which is harder to do especially against a good hand reader.

You have me thinking now on other possible floats when I know his range to be weak. On a dry board his range will be even weaker...

I give them credit for getting an NN to converge to any kind of descent solution, even if it doesn't live up to the hype. Any thoughts on the "iterative algorithm" they used?