lol, nicely done

i can remember 10 years ago threads about bots were always shouted down as ridiculous, and people said this kind of algorithm would never be possible in poker. an impressive accomplishment.

Unless there is a unique, non-mixed (it always picks a single action) Nash equilibrium, there are value-maximising ways to play against an equilbrium that aren't GTO themselves.

It's also worth keeping in mind that there are usually an infinite number of GTO strategies: there will be free choices where you can choose to make an action from X% to Y% of the time, and the rest of the strategy shifts around to accommodate this choice.

If you have a mixed Nash equilibrium, it's easy to get non-GTO play that maximises value against the equilbrium player: any decision it makes, with any non-zero probability, can be played with any probability to get the maximum value. For example, if the GTO strategy sometimes folds and sometimes raises in a particular situation, I could always fold or always raise in that same situation, and break even against the GTO player. This is because a Nash equilbrium - GTO play - maximises value against itself. If an action it played (even with a tiny probability) got less value than some other action, it wouldn't be maximising its value, and it wouldn't be GTO.

More generally, you can arbitrarily mix and match actions that any GTO strategy would ever play. All you have to do is avoid actions that no GTO strategy would make, and you're all set to play a GTO player.

So, any hand where GTO bluffs, you can always raise that hand, playing against GTO, and you won't lose. Or you can always fold. That's not recommended play, against anyone else.

It certainly seems inevitable that there will be some tourists that wander into tables, trying this out. I don't think that can be more than an (obnoxious) passing fad.

How slow does it have to be, to kill it as usable tool in a competition?

It's at least partly self limiting, in that it takes a while to answer a query - if no one else is trying to use it. That time is all limited by one set of hard drives, serving the entire world. The more popular it gets, the slower it gets.

Thanks for all the messages of congratulations! This week was an interesting experience.

And thanks for the interesting posts - there have been years of interesting reading in threads, off and on, watching game theory in the real world...

Let me add one more. As someone who has been following the CPRG work for years, this is quite an accomplishment. I look forward to reading the paper.
Congratulations to you and the rest of the team.

LOL.

It still blows my mind how many rather smart people and successful/winning poker players have zero concept of the existence of GTO or equilibrium strategies.

When the heads up bot first showed up at the Venetian, I had several relatively successful and intelligent poker players tell me how they couldn't wait to go crush it since it "doesn't know how to adapt". I would try to explain that the fact that it doesn't adapt doesn't make it beatable and they'd just reply with "IT DOESN'T ADAPT THO".

The following sentence from the paper makes me feel that you should have read some recently written books on HU-LHE before saying what human conventional wisdom is

You could even completely invert it & say how smart humans are for getting there years before the game was weakly solved.

So is limping just very marginal noise, or does an unexploitable strategy really need to limp 1 out of every 20000 or so hands?

Anybody else just get taken to two articles when they click the 'play cepheus' link on the website?

Almost any kind of pure research can be the target of this sort of argument. You realize if pure research would have been barred starting even 100 years ago you likely never would have been born?

Also this particular strategy should have a signature which probably surpasses fingerprints as a means of positive identification, right? This is just one of perhaps infinite weak GTO solutions, all of them starting at different random points and subject to different developmental vagaries, ultimately taking a shape which is unique to them alone, correct? Why wouldn't Stars and other rooms then just ban this strategy from use? If your play mimics the banned strategy with >95% accuracy or another appropriate threshold then you would be violating the T&C.

This is probably anti-botting 101 stuff but it seems like it might have to be adopted at some point, especially as very strong individual NL and PLO strategies become known. What I may be missing is the extent to which different strategies may resemble one another. But if I am understanding the CFR process correctly the particulars of any two strategies, such as weightings, are very likely to substantially differ at most decision points, at least substantially enough for the strategy to be differentiated from others within a reasonable sample of hands.


So you can play GTO, but you have to derive and implement it yourself imo.

hope it's plugs fall out

i guess problem is finding out what a strategy is ..the full set of moves that is. so reasonable sample size may be 60 million hands.

Regarding the near-death state of HS LHE:

HS 8-game is played HU pretty often. That's 1/8 LHE. Should players ask stars to remove LHE from their 8-game and make it a 7-game mix?

Dont let Durrrr hear this!

He will want $2k/$4k 100k hands no buy outs!

Nash who?

Hi

I am very curious why it is never correct to cap in HULHE. Looking on their pre-flop web page, if the bot is the dealer: the bot raises (they call it bet), opponent raises back. Now the bot will flat call with all hands.

This really surprises me. I can understand that it would be incorrect to only raise here with Aces, as you would turn your hand face-up for the gain of a small bet. A GTO strategy should be unbeatable, even to an opponent who knows the exact strategy, so giving away that we have aces, for the value of one small bet, on the preflop round, obviously isn't worth it. However, how come the GTO strategy cannot raise here with a range of hands that includes AA? Is it not possible to construct such a range that it is possible to raise here? Or would this give away too much information about the possible hands we have when we either raise or flat, even if we have a mixed strategy where we raise with aces, and some other hands, some non-zero percent of the time, but sometimes also call with them?

Could there be a different GTO strategy that will continue raising in this spot?

Could someone explain what is going on here?

Thanks

They can't tell, because the game is only "weakly solved" -- meaning you wouldn't be able to detect a statistically detectable margin of loss if this bot played the true GTO strategy for an entire lifetime.

What is true, is that for a human who wants to play GTO, it would be really foolish to try and mimic all these strategies that are used with a 0.2% frequency. There's no way you could get the right combination of balance and doing these things truly randomly.

It's much easier to just never take these actions -- or do what expert players have been doing for years. Things are slightly different postflop.

Because if you cap you can never force your opponent to fold any part of their range (their pot odds are good enough to call with any two). In NL you can raise any size & so choose a size that forces the opponent to fold at least a part of their range. So it's a result unlikely to generalise to more complex forms of poker.

(The same is true for 3-betting, the raiser will always at least call, but when you 3-bet it's against a much wider range than you can when considering a cap versus a 3-bet.)

Might be hard to get paid without taking special precautions if the money involved gets too big, especially if there is a side bet.

Look what Durrrr did to Jungleman.

http://www.huffingtonpost.ca/2015/01...n_6438140.html

Hi

yes I realise a raise will never get a fold in limit (except for the initial button raise against a very weak BB hand). However, by putting in another raise on the button, you can get more money in the pot when you are ahead of your opponent's range. So why is it never correct to do this? The only reason I can think of is that, if you try to raise with the part of your range that is ahead of the BB's raising range (ie the cards that got him to this spot), you are giving away too much information about your hand.

I am very surprised that the information you give away by having a raising range, in position, is worth more than the extra money you get in the pot. Is it the value of this information to your opponent that stops you raising again?

(I realise the logic in NL is completely different. I am not trying to draw any analogy to NL strategy.)

That can't be the correct answer. As I stated in the beginning of the thread I can't imagine that you wouldn't often cap if the bet size didn't double.

I love when things things feel like they are falling into synchronicity. I'm a huge fan of Sean Carroll and his quantum physics videos, so I was a little shocked to see his article on poker. I'm glad you found the link useful.

No, 8-game and limit HUFL are good as they are, if u cant beat them its your problem. But creating 7-game? Sure, why not. Maybe even 4 or 5-game, wtv.

Indeed. I'd guess that the key point is that folding to a flop C-bet in a 3-bet pot is rarely good. Thus BB, after a 3-bet, takes the cbet flop + bet turn line very often. There goes the extra Small Bet.

The same wouldn't apply to the previous preflop raise (BB's 3-bet), because pot odds are worse in a regular "just raised" pot, making the BB more likely to fold to a flop cbet, in turn making BTN barrel flop+turn less often than BB would in a 3-bet pot?