Taking nothing away from PN and his study of bots, but what he did was throw out some numbers and say, "This is how I think the bot likely plays pf and for a variety of reasons I don't believe reports to the contrary." I'd be very interested to hear RR's thoughts on this as he's actually spent some time playing the thing. I'm also hoping Sklansky will reappear and further develop his comment.

I don't think PN ever really commented on the notion that weird pf lines might cause the bot to play in an exploitable or straightforward fashion postflop.

Although I agree this must be true, I don't think there's sufficient evidence of PN's opinion that this is a result primarily of perfect pf play.

Again, I'm kind of playing devil's advocate here. I don't want to be painted as the guy who went around saying that a limping strat was a good one. I do find it interesting that it seems like a growing number of people who have actually played the thing have somehow stumbled upon this idea, however, and while Phil's opinion is meaningful, I don't think it, in and of itself, establishes anything.

I just said I don't really believe the guy who claims to beat it for 38BB/100 with a max. downswing of 28BB. He could have at least changed those numbers by at least a factor of ten and then maybe I'd believe him more.

I guess my point was that I don't really think that raising limps at some frequency above 50% is necessarily exploitable. I laid out some maths earlier. Now a limp-raising frequency of 75% might genuinely be somewhat exploitable but a frequency just 10% or so below that might not. Has anyone whose played it collected enough data to exclude a 65% limp-raising frequency from a 95% confidence interval?

Lots of people limp-re-raising against the thing might just be because they see the thing rarely folds to raises in the BB and often raises limps (which in themselves aren't bad things). I think they could probably do as well by playing a normal pf strat from the sb then exploit the postflop mistakes its been noted to make. I don't have an opinion on whether "weird" preflop lines make it play worse or better. Bots probably have less of a sense for "weird" than we do and you might be better off putting yourself in a situation you've played more frequently.

Of course I haven't actually played the thing. If people who have could give more info. on its preflop game I could probably come to a more informed opinion.

Trifecta, beeaatch!

[IMG][/IMG]

I bet the machine changed personalities after that hand.

The machine is interesting to say the least. My original thought was that being able to see 100% of hole cards would be a huge edge for me, then the idiot light dimmed and I realized the AI was most surely recording mine as well.. :0

I value raised 2nd pair on river, Q high board, it folded AQ (wtf)
I bluffed turn, bet river on scary board, got called by Q high..


I only have about 1k hands vs the machine, Without any evidence only my own experience playing HUHU I believe it changes personalities, and pretty quickly when your seemingly beating it.. I doubt there would be one strategy that would continue to be profitable over time with how many variations this thing has, if anyone is truly beating this thing it should require adjusting your playing style.. thats just my 2 cents from my limited experience playing this machine.

I'm coming to the belief that it doesn't really change or have any personalities at all, that's just what extreme gto looks like to a pattern recognising human.

Kidkash it knowing your cards doesn't matter - it (as far as we know) doesn't adjust to your play. Similarly, you knowing his won't help too much if you assume it's 'perfect' gto.

I tried following this strategy of being the agressor in the hand by raising the flop or check raising the flop. I was taking down a lot of hands initially, but then it stopped being aggressive and that strategy no longer worked. I think it is impossible to beat this machine.

The more I play it, the more I believe this.

That said, tonight I dropped a buy-in +$50 to live fish in 1/3NL (if you're counting, that's roughly tree fiddy), then turned around at beat a GTO HU machine for 115 big bets.

Poker -- how does it work?

I'm not a HULHE pro, but I'm at least reasonably competent at the game and have played 1k+ hands vs the machine. A few thoughts:

1. I thought at first it had to be beatable due to mistakes it was making. In one hand I posted about in my blog, it folds bottom pair + the nut flush draw on an unpaired board getting 7:1. Obviously it shouldn't fold getting the immediate odds to draw to the nuts, but clear mistakes like that are (from my experience) pretty rare. Much more common are just weird plays like c/r'ing J5o on a Q93 rainbow flop, calling my 3bet, then c/f'ing a blank turn. While plays like that and a lot of the similar strange hands that people post about aren't things I'd personally do much, it's often tough to say objectively that they're mistakes.

2. You can get unlimited rakeless action from the machine at pretty high stakes. If there aren't top HULHE pros grinding the bot fulltime, that makes me skeptical that I'm beating it.

3. I for one greatly enjoy OP's posts, with my favorite OP moment being:

Is that just weird or is it a clear mistake? If you have any kind of reasonable 3bet range about the only turn cards he can call are J and T, making the flop call -EV. At this point there is practically zero fold equity for him for the rest of the hand. Off the top of my head his equity is probably 15% given you have a sane xr range. That seems pretty bad to me. Maybe it's somehow GTO. I guess if I understand it right when it (he,whatever) is getting 9:1 to call it should be folding only the bottom 10% of it's range. What range? It's xring range, right? So maybe it's xring that flop so frequently with weakish hands that J5o falls outside that threshold. So it feels compelled to call because folding would be immediately exploitable. It checks the turn because sometimes it might be checkraising again so you can't exploit its checking by thin value betting the lower part of your 3 betting range. It still seems like a bad play.

Then it folds TPTK on the river to a raise. I think it has a Russ Hamilton personality.

pretty sure u could find the GTO superstar bots (hyperboraen f.ex) making plays like these from time to time. So my gutfeeling is that this is hardly a sign of weakness at all.

The bottom pair+FD fold tho is ofc funny and likely a sw glitch or something

I just crunched some numbers on that and it said that assuming you raise tp+ and 30% mp on that dry flop, the flop call is actually more than 1 unit negative EV. The RIO on hitting jacks on the turn kill you. The strategy can be made significantly worse by calling turned 5's.

I am very interested in learning how to implement GTO concepts into my game and I feel I don't understand them well. But how can making a call that actually costs you significantly more than the 1 unit you're putting in be part of an optimal strategy?

Let's not get carried away with these assumptions that the bot is playing GTO. There are several documented cases of the bot making clear mistakes. Isn't there a photo where it folded like AJo to a 3-bet pf? There's no way that's part of a GTO solution for HULHE. Nor is it the result of an abstraction. As PN has pointed out, pf is the easiest street for the bot to play perfectly. It would make sense for the bot to make some trivial abstractions such as playing any AJo combination the same way, but that's about it.

I think some of you guys hear about concepts like indifference points and balance and you think this means every play in every spot is varied in a GTO model. There are tons of spots where there is a dominating strategy, which means that there is only one correct play to make. There are other spots where it might be correct to play the hand in one of two ways, but the third option is not valid.

The range you describe is fine for a lot of TAGs and loose-passives, but I doubt it's optimal. The bot's optimal c/r'ing range should include some air, oesds, gutters, and perhaps some bdsd+bdfd combos, so an optimal 3-betting range almost certainly includes re-bluffs and possibly some free card plays. Whether or not it makes enough of a difference to justify a call is another question.

I just changed the 3 bet range to 100% and it's still -.37 units. So in terms of EV its a clear mistake. You can't really tweak the assumptions after that point either. You either call TJ or 5 turns or you don't. Calling 5's is bad. Not calling T's or J's is bad.

PN mentioned this type of thing in his book too. IIRC he said he didn't believe it was optimal to give up too much EV for the sake of balance, using 3betting pf as an example where withholding information from your opponent cannot overcome the lost ev from pushing a 20ish% edge with a hand like AT.

Edit: I had originally given hero a pfr range of 72%. I changed that to 85% and the ev increased to -.16 units, still in the red though slightly less. But this still assumes the nut maniac 3 betting range of exactly the flop betting range.

The other thing to consider, although it's not quite as definitive, is that the ip hero is probably meant to smooth call some % of his made hands.