Poker-AI.org

I found a ton of information about GTO bots, but few research has been published about good exploitive or adjustive bots. For instance, I would like to know which techniques bots like Neo, PULPO or MANZANA are employing. Unfortunately, all I could find are "one-liners" about those bots or longer papers about bots that basically failed... Anyone knows how to find more on this topic?

PULPO or MANZANA weren't adaptive.

Neo has/had a website.

I read a post from Johanson in the old forum where they discussed the competition and he argued that so far it was basically a contest of having the best equilibrium, but due to manzana and his dominance in one year he assumed that people would change from pure equilibrium, so I assumed he was at least exploitive. Do you have any infos that he is an equilibrium solution too?

Regarding Neo: I saw the website, but the information there is basically just a few keywords without saying anything

Manzana was an NN trained on actions from EQ(s) then purified. The U of A did testing on it's exploitability and found that it was the most exploitable of all the contestants, despite it's dominance. But, since all the contestants were EQ-based, none of them adapted to Manzana's (obvious) strategy. Pulpo was the predecessor to Manzana -- same concept, same author.

If you really want a challenge, try figuring out how Sonia works/worked. Now offline: pokerparadime.com

Yeah Sonia is really the only one I know that supposedly is highly adaptive and a great player. Not much information out there about it though and as Nasher said they seem to be out of business now too... whatever their "business" was before.

_________________
Cheers.

I (thought) I actually found one of the authors of Sonia on Facebook (Marc Cohen). He didn't respond to my message. There's a Marc Cohen that has/had a company in the UK called "All in Equity" (get it?) that declared $30k in revenue at some point. He seems to have dropped out of the poker world. Probably sold his idea for a few million and retired. I'd show you his Google+ profile, but I'm not sure he'd want it posted on a public forum.

I chatted briefly with Spears about Sonia on the old forum, we think she's actually modeling herself from her opponent's perspective, then solving some linear equation between her own (external) range and her opponent's. The authors mentioned she's technically a Bayesian model, but don't go into detail about how she functions. She did beat Polaris, which at the time was the strongest EQ (publicly known).

IIRC Polaris 2008 had several "tilted" equilibrium strategies from which it chose the best at runtime according to opponent. http://en.wikipedia.org/wiki/Polaris_(poker_bot)

If I'm not mistaken, I think they only played Orange (the default EQ) versus Sonia.

There's not much info about them but I think OneDayItllWork and PeppaPig have successful exploitative bots

I do too, just not for HUNL.

Searching for information on poker agents using Neural Networks/Machine Learning techniques will give you a few results.
Most seem based (very loosely) on learning opponents via history, then simulating the current game state to find the action with the highest EV, then using this action - possibly using a mixed strategy.
However, this almost becomes a task in ML solely as you'll try to aim for the highest % on your classifier for modelling opponents.
I imagine it gets more technical and interesting when trying to adapt to shifting opponents, meeting new opponents and finally attempting to not be an easily exploitable agent.

I'll be working on a Best Response agent next year for my thesis (NL HU or 6-max) - have a basic bot at the moment, but only at around 71% accurate on predicting opponent actions.

I can have a dig around for some links if you'd like? If it's what you're looking for anyhow?

I found a lot of papers, but they lacked in either one of those areas:
1. they either require an exact understanding of opponents private strategy (this would require in depth knowledge we typically don't have)
2. they are only able to adjust after a very long history
3. they are not performing pretty good

If you know a paper that doesn't have these drawbacks, I'd be happy if you can give me a reference. The use case I'm interested in is finding ways to exploit in short-run games like SNGs where we have only a very limited number of observations.

The recent U of A paper on using clones and Exp4? Or did you not want to use a DBR approach?

even though its one of the better papers, the clones could only be created in an articifial setup where they had access to the hands from last years competition (many hands) *and* hole card information was available in each game, even if villain did fold... And the same information is also used (as far as I understood) in the selection process of the best agent. So I think its hard to generalize the findings in a real-world setting.

I don't think the clones were the important aspect of the paper, rather their use of Exp4 and findings that smaller agent profiles perform better. You could replace "clone" with "dbr player" throughout the paper.

maybe I got it wrong. I understood the portfolio thing and that the smaller was better. However, based on which criteria did the exp4 select the bandit to play? as far as i understood, they did EV calcs, but aren't those based on the private information too? For instance, you have 3 models A, B and C. How do you - based on a small history of say 50 hands with 3 hands that went to showdown select the best model?

Importance Sampling? ... doesn't require knowing the op's strategy.