GL for your PhD thesis Mickael !
I'm sure it will be as well as your Master thesis, and I will read it conscientiously

Thanks for that. People who really understand something can explain it simply and you did.

It seems to beg an obvious question: Where is the perfect poker machine against which people are developing the strategy?
-

thanks to others who posted the links.. I'm reading the Hawrilenko interview that I passed on when that other thread was running. And I see the math guys are back so I'm back to observing... have fun.

First, no problem with you being "that guy", you are obviously shmart and edumacated and I respect your opinion.

I kind of agree with your first paragraph (depending on how you define "pretty badly") but don't really get the second (or get it and disagree). Can you elaborate?
I am not sure if/how your point about future hands makes sense, probably have to think that through, but wouldn't that contradict the definition of gto?

Also, if you look at heads up as two seperate games (one with yourself in the sb and one with yourself in the bb), what do you think about your ev in the sb where you actually can make it a push/fold game and the bb can only react? Wouldn't using the chart be preferable to any other strategy against a significantly stronger player?

It's relatively easy to compute the best max exploitive strategy vs given strategy. Idea of algorithms like CFR is to train vs such perfect opponent.
It should be intuitive: if you tell me how you play exactly with every hand in every spot you would expect me to come up with smart way to play vs you pretty soon. Computers can do that faster and more accurately.

so... they probably have computers hammering away at each other 24/7.
On a scale of 1 to 10, ten being HU game solved, where does it now stand?

That scale doesn't make sense. If you mean HULHE, then they are pretty close If you mean HUNLHE, then they are not very close

glad i'm not the only person who admits to using wiki as an occasional reference...

I get it.. and admit to grunching a few TLDR posts in this thread.

another dumb question: When a GTO strategy is someday formulated (for limit), will a human be able to play it or will only machines be capable of following it?

please stop

The question upset you... I'll take that as a "No, humans cannot do it."

it's not that i would be upset if you asked one dumb question, dumb questions are fine. but you are derailing this thread when it should be more than obvious to you by now that you are a beginner in game theory and should study some instead of asking your entire way through it. it also seems like you haven't read the thread, or haven't tried hard enough to understand its content.

read all the posts regarding how a gto start in hu limit looks like, you will read stuff about the size of the game, how many decision points it has, what a mixed strat is etc. then ask yourself whether a human can apply this strat correctly and fully, and i have full faith in you that you'll end up with the right answer by independently thinking about it.

I don't know who in this thread knows what they are talking about and who doesn't.
It's not like I haven't done any research.. been reading about Nash.. struggling though a PDF about counterfactual regret minimization..

but like I said earlier, humans tend to ignore extraneous pathways when solving a problem. And to get right to the heart of the matter, I want to know if this thing holds some promise for me in a real game, or if it must be boiled down to a few imprecise rules of thumb to have any value at all.

I appreciate the book tip, gonna look it up.

You're suggesting that the bots have a balanced range for each betsize. 2 things:

1. They don't. At least based on the 100 hand sample provided here, they appear to minbet mainly weak hands and potbet mainly nuts.
2. Let's assume for the sake of argument that their 1/4pot bet on a drawy flop has a balanced range. How is that the optimal betsizing? I already pointed out all the things wrong with this sizing so I'm not gonna repeat myself. Your reply covered another problem I presented, but not this one.

Whether GTO bots will eventually crush strong players for >20BB/100 is a matter of opinion, the state of the current NL bots is not. They're horrible. They are so bad I would be happy to play them for real money and I'm just a lowly bumhunter. If you think the bots are better then you can put your money on the bots and I'll play them. Should be easy money for you, don't even need to put the hours in.

You're missing the point. The bots are most likely far from optimal, and I wouldn't find it hard to believe that they're horrible by real-world standards.

The point is your analysis demonstrates a fundamental lack of understanding of the means by which a ridiculous-looking strategy can be good. You also misapplied several GT concepts.

Your overall argument (that the bots aren't good) is reasonable, though.

That isn't what I suggested. I said it's possible for a player (whether in human, bot, or dog form) to avoid the problems you mentioned by having balanced ranges for each sizing (i.e., they aren't constrained to the flawed either-or proposition you provided). As evidenced by your repeated use of the word 'if', we were dealing with hypothetical strategies at that point--not these two specific bots' actual strategies.

Hmm.

It's probably not. But the relevant question is "How can it (or something similar) be the optimal sizing?" Or more precisely, "How can it (or something similar) be one of the sizings you'd use in this spot as part of an optimal strategy?"

You'll probably be able to answer that question, or at least make educated guesses, after reading MoP.

Yes, you are correct and I was wrong on that point.

On another point, I find it really hard to believe small bets (like minbets) from position on the flop and turn could ever be part of an optimal solution. We give our opponent the odds to call with almost anything on almost any type of board. We almost never see a fold. If we are called we get 1BB more into the pot. If villain decides to CR, we have to fold like 40-60% of the time, when we could have just checked behind. The only time we gain something is when we have a really strong hand and we induce action that we wouldn't be able to get with a normal sized bet or a check. There's just no way that's enough to offset the huge # of times we
- put 1BB more into the pot as an underdog
- put 1BB more into the pot and then fold
- put 1BB more into the pot when we have a strong hand and our opponent would have called/raised a larger bet

@FullyCompletely

Tyvm for your elucidating replies in this thread. I've learned a lot from what you've had to say.

I was wondering if you might comment on the software "Snowie" which seems to be getting a lot of attention lately. It's my impression that most of this attention stems more from the way the software has been marketed than from its true merit. You're certainly more qualified to judge it however. What concrete achievements has the software made? What would your response be to their following claim? -

"While for the specific variant of fixed limit heads-up the computer is known to be just as good as the best professional Poker players, nobody could create strong artificial intelligence-based software for the most popular (and most complex) variant: no-limit full ring game. This goal has now been achieved with the release of PokerSnowie."

hmm that's a brainfart, forget about it...

see:
http://forumserver.twoplustwo.com/sh...&postcount=120

by all practical standards, I would say that what you have inferred is incorrect. Maybe not in terms of 'solved' but being able to beat any human over a meaningful sample is close enough imo.

It's a matter of how you define the game.

You're definitely right that the push/fold chart gets crushed by a strong player that is not restricted to push/fold (human or computer), down to fairly short stack sizes. That chart is a very close approximation of the solution to the toy game in which you are only allowed to shove or fold. When franxic says you can't lose if you play according to the chart, he means you can't lose to an opponent who is also restricted to push/fold.

If you were analyzing something like a HUSNG, then yeah, a really thorough solution would need to account for the value of future hands as a function of stack size and position. In that case, "the game," is the entire tournament.

However, academics who have worked on computer poker treat a single hand as "the game." This doesn't mean the solutions they produce are imperfect, it means they're solving a slightly easier problem.

Practically speaking, I can't imagine this matters in something like a deep stacked cash game with fluctuating stack sizes. In a fast structure tournament format, I could see it being a significant factor.

not sure if i get that right, the concept is avoiding small +ev spots in order not to miss future bigger +ev spots?

Or taking small -EV spots now to avoid larger -EV spots in the future.

Suppose you're playing a push/fold HUSNG with a total of 20bb in play. Further suppose that playing the button with 9bb effective stacks is a sufficiently losing proposition that when you start the hand with 9bb and the button you're actually only 44% to win the tournament (rather than the 45% of the chips in play that you have).

If you're in the big blind with equal stacks, you should call a jam with 44% equity or better, rather than the 45% you would expect from looking at things in terms of maximizing chip-EV on a hand by hand basis.

the advance of AI is unstoppable & insuperable.

thx god poker is played between humans . Sad fact: bots, if done right, won't be detectable so poker pros might need to move back to the casinos in the future. Good fact: There is a high change micro/low stakes players won't be affected as much .

@FullyCompletely: How much are u guys banking at online limit games ?

Read http://www.computerpokercompetition....owall=&start=2 before analyzing the hand histories. The bots are not trying to play "GTO", they are trying to win the competition.

Two great things happened in this thread

1) i learned a crapload of stuff from a very smart phd student who could probably be a good poker player if he tried, thanks alot FullyCompletely!

2) I blocked joeschmoe

What's clear is this: players are not equal and if AI bots are not ready yet to take on such mature and accomplished pros like kagome, ragen70 or sallywoo, but even given knights-odds to such overated players like that swedish kid with banana mashed puree in the lieu of the cerebral matter who would be beaten to a pulp without further AI development in this case.