Poker-AI.org

I have created a few bots, but the only succesful ones so far were HU bots.
This thread is about good approaches for creating 6-Max bots, not so much about implementation details, but about general ideas and algorithms.

I know some people are using rule bots, which basically mirror their own understanding of poker. This would be pretty hard for me, as I am not a great poker player and it seems, that you need a ton of rules.
There is no real GTO solution for 3-Max+ games, so simply running CFRM would not produce an optimal strategy, also its infeasible for anything above 3-Max and even there the trees get huge.
I have heard, that some people decompose the game into a lot of HU situations and solve those with CFRM. I tried that approach, used Snowie preflop ranges as a start for those HU spots, but there are so many spots, that it seems infeasible.

Then there are techniques like neural networks, but even though I know a bit about NNs, simply having NNs doesn't really work as well, I still need to combine it with some features, or CFRM or something like that.

So what are some ideas for 6-Max bots?

Hi, I made rule base bot, as you said I made it according to my understanding of poker. It utilizes most of the gto approaches, like "unexploitable defending", and a lot of math like which hands can be value bet, which bluffs are +EV, using many player stats and population stats together with some predefined rules in order to define villain ranges (which is the toughest in poker). My bot beats Nl10, but it will be better over time, I don't have much time to invest in bot because I have full time job.
I am not sure if neural network approach would work 6 max considering that if you want to play multiple tables and have only 15+ seconds to react.

I did a similar approach as mlatinjo for 6 max PLO. Its a combo of rules and equity using population data only to form ranges and uses standard known GTO frequencies. You would be surprised at the differences in population data from limit to limit. I am probably losing quite a bit of EV as its not very exploitative. I have kind of done this on purpose as I think bum hunting is coming to an end and more anon games are going to become more prevalent in the future. You can make some of it up by being very aggressive as most villains over fold in general.

It has been a difficult project and I have made many mistakes. I personally feel that simulation based bots are overkill and are not necessary. Had I gone with only rules from day one I think I would have been much farther along to be honest. Sometimes equity is not a good indicator of playability and I have struggled with this quite a bit. Equity also does not consider positional advantage.

I think some of the most difficult aspects of 6 max are dealing with multiway all-ins which require quite a bit of processing power to figure out correctly due to all the various stack sizes, side pots etc. HU is pretty simple but 3 and 4 way low SPR situations are very complicated. My river bluff catcher routine is also complicated but I am not sure if its even needed for games under PLO25. I think if you can get your RCE (River Call Efficiency) around 1.3 that your doing good enough. I think the holy grail is around 1.75. Playing perfect on the river is not as easy as it sounds.

Its also easy for bots to over fold post flop. mlatinjo has helped me with this and he really knows his stuff. I have really loosened up the floats and it has made a big difference.

Also..one other thing that simulation bots struggle with is blockers. My bot uses both nut and draw blockers quite a bit. The only way you can do this is with rules.

Ok I see, so most seem to use rules. I would assume it's really hard to write all those rules?

Are you saying there is a theoretical problem? If so, could you post a reference?

Creating rules is easy and you would be surprised that you can cover 90% of the situations rather quickly. My approach was to create a rule set for each major scenario then run simulations over all board types to see where the frequency of that rule set settled at. At the same time you can see the distribution of value and bluff combos and adjust accordingly for GTO purposes. So if your striving for a 65% flop HU c bet you just create a rule set to achieve it by widening or tightening the rules. It sounds complicated but it really isn't.

Not a great reference, but you can watch Bowlings video starting at 56:00
https://www.deepstack.ai/

I think there was some more in some Alberta papers, but I didn't find it.

Thanks

i have half of a 6max bot ready to go. the poker logic part. Based on pokersnowie preflop ranges, and postflop defending 1-a and, betting flop,turn , river in a GTO balanced way.

If you have the code to get information from the pokerclient and get input to the pokerclient. I am happy
to share code.

In a 2-player zero-sum game a NE is also a min-max equilibrium, every player can assure a guaranteed minimum winnings/maximum loss (a constant called game value). A Nash Equilibrium still extists for any game of finite number player and finite number strategy, but in 3+ player there is no notion of game-value. So there is no strategy that can assure to not loss. A NE strategy is the best option only if all other player are playing the same NE, if one or more player deviates from nash (or playing strategy from another NE) you could loss also playing nash.