Poker-AI.org

The two great bots in action ...

http://www.youtube.com/watch?v=W4CqUw_VP_Q
http://www.youtube.com/watch?v=PlfEfKu5tZA

I am quite disappointed of what is happening there (especially regarding bet-sizing) ... But maybe it is just my understanding of poker that lacks behind.

EDIT: The secion below this line can be ignored. I thought about a general weakness in the CFRM-solution-application-procedure. But anything turned out to be working correctly. I leave my thoughts anyway ... Because I already wrote them down
----------------------------------------------------------------------------------------------------------------

Nevertheless I kept thinking of how CFRM-constructed bots are working and (sadly) figured out that something I considered being trivial all the time maybe isn't that trivial at all: The usage of an obtained 'EQ-Strategy'.

I will try to give an example:

Assume you obtained a strategy using Monte Carlo CFRM variant ASS for T'Holdem HUNL. Assume you are playing in the BB. Furthermore assume that in your model/game tree the SB has the options to fold, limp, minraise, bet 6BB, all-in. Assume hat for all hands the probability of playing (bet 6BB) is 0÷ (as is the case for a few strategic options in my strategy).

So, let your opponent now bet 6BB. Now you will lookup frequencies for a spot in which the nemesis has an empty range.

Any defending range would do the job against an empty opponent's range.

Let us now give our range some interpretation: Our range is the average of all 'regret-ranges' that were solid defending ranges before the nemesis realized that this spot is less profitable than other spots he could reach. The implementation of ASS ensures (due to the exploration guarantee EPS) that (in infinity) the average strategy equals the latest regret-strategy.

Hmmm, sounds fitting now. You just have to make sure your bot never runs into a branch where his action probability is 0÷ which could easily happen during the action translation procedure.

Sauce123 has some new videos commenting on the play of those two bots which are quite interesting. I believe he had something to do with slumbot.

Betsizing "holes" seem to be a big problem for simulation based bots from what I understand of them. Just because a guy bets half pot does not mean he is betting for value nor does a 2x pot bet on the river mean he is bluffing. Good pros will mix that up so you have no idea what to do..is he looking for a call or is he bluffing?

I think that its better to make the decision entirely on the villains range and betting actions throughout the hand in real time and not use any kind of LUT that was determined from months of number crunching. If you can range the villain accurately then its going to get much easier to determine the correct bet size to extract maximum value and vise versa (meaning determining what his bet size could mean)

Lets say the bot has TPTK and thru hand combinatorics/betting actions you have determined his range to be mostly air and bluff catchers (like 2nd pair or worse) on the river. Making a small bet here is probably not the best line as your going to get mostly folds and some bluffs. The problem with that is you do not want to have to make a tough decision with just top pair if he does bluff. Cole South once said "Its better to make the last bet or make the big bet yourself". So by making a big 2x pot bet yourself you accomplish the same thing..your gonna get the folds either way but your going to get more calls on his bluff catchers just because it looks like a bluff. He is going to be less likely to call the small bet with his bluff catchers (not always the case but generally) because it looks like value. It is also doubtful he would choose to shove or reraise based on his range.

Who knows. You will never know for sure, if he only has bluffcatchers and air, combinatorics will only help you, if you have very solid reads and your opponent plays a static strategy, if he is good, he will slowplay in the right frequency.

The same goes for the betsizing. IMO a solid strategy based on the nash equilibrim is more promising than a rule based approach. It doesn't work perfectly for NL yet, but it's just a question of technology, FL showed that this approach is very effective.

Well I think the slowplaying frequency should be the same as the bluffing frequency which is something between 25-33% which is what I have mine doing. Correct me if I am wrong here. I like the the whole concept of the "delayed c bet" to prevent the villain from always leading the turn if you checkback the flop.



Well I guess its true that its just a question of technology. I heard solving 100bb+ NL could be a decade or two away. One prominent NL player said "not in my lifetime" which is interesting.

Has 10 player FL been solved or ? I know HU was pretty close. I made a living from a 10 player FL bot for a number of years (rule based of course). It took about 14 months of work to get it right but it was not that difficult really. The time consuming part was having the pro find the mistakes. Sifting thru 5-10k hands at a time takes awhile.

- Computers are getting faster by about 80% every year. This has not been slowed by the end of Moores law because parallel techniques are taking over. The power of modern GPU's is awesome.
- Algorithms for finding NE are improving too. The best HUFL algorithm in 2006 produced a solution which was exploitable 0.32 bb/h The best FLHU algorithm in 2011 was exploitable 0.1 bb/h
- In real poker, as opposed to bot competitions, it's important to recognise what your opponent is doing, and find a strategy to exploit it. Pattern recognition is progressing apace now. The best algorithms at recognizing street signs in a real noisy environment are now better than humans. The algorithms for finding counter strategies are not so different to the ones for finding NE
- Very few poker players have any clue about what can be achieved by computers.

An exact solution might be many years away, but good approximate solutions already exist and they are getting even better.

You can still find a solution with techniques like bucketing, even though you have to treat a few hands in the same bucket the same way, it will usually be still more accurate, than what humans do.
Very few players use more than very general categories. TPTK, TPGK, MP, strong draw, weak draw, then a few boards. They might use more stuff, stats, history, board patterns, but it's usually by feel, a human will never have the mathematical precision a computer has.



On the topic of slowplaying: Yes, it's important and it's good you do it, but I was talking about your assumptions of your opponents. You said, that you can determine, that your opponent is weak on the river, but how can you be sure? He might still be slowplaying. Of course, if he is very bad, you can just assume he will always play the same way, but many players won't.
If he does slowplay, your strategy of making a big bet on the river will be exploited, because he will pick you off with his monsters.

That's just one example, I think a strategy, that is not based on GTO will never be extremelly strong. I am not saying to play GTO, but have it as a core and adapt from there. You will most likelly get better adaptations with a good algorithm, than with rules.

I have been playing FL and NL for years, but a good bot still owns me and when I am exploiting opponents, my reads are sometimes extremelly inacurate. I don't even do this botting, because I want to make some change on micros, I believe, that I can build a bot, that is way better than me.

That and what spears said about faster computers then I guess its not that far away. I better get cracking on my PLO bot if NL is about over...



Well the bot can without a doubt play better then a human in many many areas which is why all of us are working on it I guess. I think the non tilt factor and the ability to play unlimited tables without error are the big ones for me.



Oh right..I totally missed what you where saying. Well as with anything in poker you can never really be sure about anything so it has to be an educated guess. Ok..about the big bet on the river. I also have it making the big bet against his entire range when I have value as well so its somewhat polarized but not totally. Some value will be lost by not making the perfect betsize vs his range but I think it evens out over time. Polarizing the river has some issues but I think it overcomes some of the ranging errors that can occur when the villain is slowplaying. I have thought about weighting the possible value range when he calls down to the river versus totally eliminating those hands from his range. It is definitely a problem area. Like I think the villain would be more likely to slowplay on a dry board so perhaps I could weight those hands as 25% likely in those spots or something. On a wetboard I think most players will try to protect the hand but not always. I have my bot employing a probe bet in some of those situations in order to get a better read on his range.



Well I am sure you are right about that. I mean you cant beat GTO.



Hand reading is not easy but I am getting better at it. When the villain is leading the betting its way easier then if he is not obviously just because of the slowplaying factor. But yeah..I dont think I can play better then my bot either which is why am doing it too

I didnt think of this until just a bit a go but I think I should be weighting the value range of the villain at the GTO bluffing frequency which is roughly 33% if he is not leading. This would change the range equity by quite a bit and account for the slowplaying problem. As far as hand combinatorics I could just include a third of the hand combos in order to get the betsize correct. By doing this I could de-polarize the river in order to get maximum value in every spot.

Does this sound right or ? Any insight would be appreciated.

Well I made the change to include weighted ranges when the villain is calling down. I did two things. First, when the villain calls my raise OOP I accounted for 15% slowplaying big pairs. Secondly, I went to a weighted range on all hand combos that are a str8 or better at 20%. I am not sure about that number and I am trying to get more intel on what that should be. 20% could be too high for a general average. It could also be very villain specific but I think its better to always account for it.