Poker-AI.org

Hi, I'm here again with more doubt.

Here:
http://poker.cs.ualberta.ca/publication ... on.msc.pdf

In page 31, talk about Selective sampling and Simulation betting strategy...
1)Is CFR based on this principle? or is it other approach altogether different?

And here..
http://www.poker-ai.org/phpbb/viewtopic ... 2520#p4411

2)"Vanilla: nothing is sampled".. this means that hole cards and next board cards are fully simulated, right?
3) which means "Strategy" here? mean Fold,Call or Raise? or tight, bluffer, etc?
4) I got the amax's code, it is beautiful. Amax create these methods? you use these? or which?
5) Where I can read more about External Sampling and Outcome Sampling? (if you use these)

Thanks

1. CFR is quite different to the strategy described by Davidson, but both use sampling.
2. Vanilla means nothing is sampled. But groups of cards are collapsed into buckets
3. In the context of CFR, strategy means "for each situation, how frequently do I fold, call or raise". Situation = cards held, board cards, opponent previous actions, our previous actions.
4. IIRC Amax wrote the code based on the UofA papers. I don't use that approach any more
5. Not sure. UoA publication? Marc Lanctot video? Poker-AI? Google with the site option is your friend.

It's out of date now but still worth reading https://www.cs.auckland.ac.nz/research/ ... _AIJ11.pdf

Thank you spears, why you don't use that anymore? wich of these approach you recommended me?

Hard question.

CFRM and related approaches use a lot of computing resources. You might get a decent HULHE player that is difficult to beat, but I doubt it will make much money either. The resources required for any bigger game are insane.

Exploitative approaches like Davidson's don't seem to adapt fast enough to the opposition and don't know how to play defensively.

An approach like Sartre might work but I doubt it will ever beat the world because it is just copying other players.

If you know how to play poker then an expert system might be a good way to go. I have some ideas about how that could be simplified.

I'm working on an approach that is distantly related to CFRM and fictitious play, but uses far less memory. Once that works, I will make it adapt to different opponents. You cannot hope to get sufficient information on your opponent's strategy from a few hands, so you have to hope that when opponent appears to play a certain way he is using a similar strategy to all the other players that played like he did.

Have I put you off yet?

I maybe put off when I come to the finish line, and I tried 30 times more after see that it is impossible.

You say that limit 6-max is so bigger for these approach?

About opponent model I thought persist opponents data of each hand played by they usernames, and collect these data, put in a NN, data like stack transition of the last 10 hands, and other results of expert systems, etc. I have not fully defined this yet.

If I have the best predictor, CFR still is worse than expert system in effectiveness?

- 6 max is a lot bigger than HU for CFR
- Even if you can build a good opponent model, you still have the problem of knowing how to exploit it. If you use best response you are susceptible to counter attack.
- CFR produces a static Nash Equilibrium strategy. NE is the same regardless of opponent. It doesn't exploit flaws in the opponent.
- You should read the Watson Rubin paper.

Thanks..
About your own approach... I'm interested in helping with that and I've a lot of time, you can trust in me , really