Poker-AI.org

I'm writing a poker bot for up to 9-max games. I'm still undecided which approach to take, which is increasingly frustrating.

Right now, I'm thinking of using CFR with bucketing, where the nodes are grouped into groups such as: "flop, we're the the preflop raiser, we have over pair, board is dry, 2 players behind one ahead, etc.".

Is this a good approach, if something like that was tried before, what were the results?

And in general, how should I approach this, i.e. up to full-ring bot that plays reasonably well?

CFR consumes huge amounts of memory and processing time. The larger the problem the larger the memory and processing time, and you have selected the biggest possible problem. These problems can be reduced by an abstraction, but with that comes error. I don't really play much poker and am unfamiliar with much of the jargon so I don't understand the abstraction scheme you are proposing: you should be able to figure out the memory requirement yourself. If the abstraction is too coarse the error becomes so high that the solution is easily exploitable. I don't think I have heard anyone proposing it before, but then I haven't read and understood all the literature. If you haven't done so already I would recommend you look at Noam Brown's you tube video and also Michael Bowling's Deep Stack video.

Sorry if this sounds negative. With a deep understanding of the problem you might be able to find an abstraction that is both accurate and reduces the problem size to something manageable. Alternatively, you might be able to hand craft an expert system as I mentioned before.

One of the huge problems with static strategies is that once your opponent understands it they can play ridiculously exploitable strategies themselves to exploit you with impunity. So you might be able to get away with a weak initial strategy if you can continuously modify it to exploit your opponent.

Off-topic but I'm quite surprised to hear that

@OP: Full-ring 6max is pretty much the biggest challenge in poker AI at the moment and you're starting all the way from pre-flop aswell. Try out rule based population explotative bots instead if you're trying to make a profit