Poker-AI.org

Hello everyone

I have to give a little lecture to close a Game Theory course; I want my students to realize that there is much more to game theory than prisoner's dilemma and small case studies.

I'd like to let them know about what can be done for "real" games and why NE isn't the ultimate solution. As far as i know, CFRM is used to find eps-NE in LHE games, but I am unaware of what algorithms are used for opponent modelling nowadays.

Has there been any advance in multiplayer bots?

Would you say that "academia" is up to date with bots, or in your opinion private bots are one generation ahead?

I don't need any specific implementation (although every info is valuable!), it's just to give the students the "big picture"

thanks!

I am probably the least knowledgeable when it comes to academics and how they think there is a NE solution for NL. I personally feel that there will not a NE solution for 100BB+ poker in quite some time. I am not even sure if a NE solution can even exist is that makes sense. For any strategy there is a counter strategy. Most bots do not adapt quickly enough and by the time they adjust they have already been exploited and the villain has adopted a new strategy. So its just a never ending battle of adjustment and whoever makes the adjustments quick enough will win the money in my opinion.

I recently saw a NE chart designed for a new game and I developed a counter strategy for anybody using it in just a few hours..so any published NE chart/solution will not last long.

Opponent modeling is rather straight forward and any bot not adapting to a players tendencies will have difficulty winning anything (if at all). However..with that being said you can have all the stats in the world on a player and still miscalculate EV and make mistakes all over the board. The reason for this is that players can have identical stats but completely different ranges leaving some pretty serious grey areas on what to do. I am not sure if this could ever be solved.

So..as far as what kind of bots are in the private sector and what is in the academic world. I think the academic world is probably on the cutting edge of simulation based bots. They have the math guys, tools and resources in order to accomplish that stuff so I do not see them being a generation behind in that field.

I personally do not think NE or simulation is the right approach however. I feel its best to write a core basic strategy that plays a solid game of poker with rules to get near break even with rake. This type of bot is also very easy to debug and change on the fly. Try debugging a simulation based bot in which it relies on number crunched data that took months to compute running 24/7. You better hope there is not a flaw in the calculations eh? That river bug could cost you thousands...

So..the rule bot can really only be done with the help of an expert poker player with a high degree of GTO knowledge. Then use opponent modelling using stats from programs such as PokerTracker or just write your own database to deviate from the strategy to exploit tendencies which is where the money is.

I also think the basic strategy should be aggressive right out of the gate..meaning force the villains to adapt to you rather then you adapting to them. Aggression wins. This means raising 100% of buttons until he is forced to call or 3 bet. And when he does you adapt until equilibrium is reached and then you switch gears again playing tighter for awhile..

That is the vicious adjustment cycle I was talking about...

There has been some work on opponent modelling in academia but it hasn't got very far. eg
http://www.poker-ai.org/archive/www.pok ... =64&t=1634
http://poker-ai.org/archive/www.pokerai ... =64&t=4037
http://www.poker-ai.org/archive/pokerai ... &sk=t&sd=a

If you know something about your opponent strategy then Data Biased Response is useful
https://www.cs.auckland.ac.nz/~jrub001/ ... intAIJ.pdf

Judging from various conversations here and my own work, I'd be very surprised if private development hasn't gone much further. The now defunct Sonia was an exploitative bot (HUFL and FRFL) that beat the best Alberta Nash Bot and nearly everything else back in the day.

http://poker.cs.ualberta.ca/publication ... sponse.pdf shows that equilibrium bots created up to 2011 were still quite exploitable.

What happened to Sonia? Seems strange nobody continued on with it for NL. I remember you mentioned it was also cognizant of its own range and how it looked to the villain.

Sonia died. Somehow I got the idea that they were not trying to make money out of Sonia herself: rather they wanted to create a business that used the general AI/stats techniques in Sonia. I think they believed unleashing Sonia on the world was of dubious morality/legality. When I last checked one of the developers is a quant in Bournemouth UK, and the other an independent trader in Switzerland.

Well that is too bad. All that work for nothing although they probably applied it to their current jobs. I think stock traders have a more dubious reputation as of late. Would you rather have lunch with Bernie Madoff or Phil Ivey?