This could very well be a stupid question, but what happens if you put two of the same LHE bots against each other? Both just break even?

I assume they'd be 0BB/100 against each other long-term, but you'd still have variance, which'd make a break-even result unlikely.

Robot beating durrrr would be HORRIBLE for poker. You think recreational fish would want to play after seeing their hero get absolutely crushed by a robot? IF durrrr can't do it, who can?!

As if HULHE isn't dead already.

Not referring to HULHE

The fact that robots exist that can beat top pros will be discouraging to recreational players who hover around playing microstakes no limit (and discourage newer players who like watching durrrr/HSPetc)

Recreational players can be beaten by monkeys so they will probably be ok with the fact that they can be beaten by robots as well.

Sick deductive logic skills, do you coach?

Someone disagreed with Sklansky on this (I can't find the post since I'm on my phone) later in this thread and I think I have to side with Sklansky until proven wrong. I don't really understand this stuff so forgive my ignorance:

What if we had a GTO rock paper scissor bot (1/3 rock, 1/3, paper, 1/3 scissor) face a huge rock paper scissor whale that only throws rock. Wouldn't a human player adjust by throwing paper if he sees that this whale doesn't adjust while the bot continues to throw randomly?

This is a pretty extreme example but there are certainly many spots in poker where you must take a line that is non-optimal in a vacuum because of the opponent. This doesn't have to conflict with the superiority of a GTO bot vs a good player though right?

I doubt Durrr takes this up because I believe he was inferring that he could beat a nlhe bot that plays near perfect GTO. The version of polaris that plays closest to an equilibrium strategy is for limit (and was beat by stoxpoker coaches) and were not trying to test AI.

Here's the discussion about hu lhe and AI from a different thread. http://forumserver.twoplustwo.com/18...76/index9.html

a GTO play doesn't take into account what opponent is doing, it's a strategy which isn't exploitable, to be unexploitable i think you need to play sub-optimally in certain spots

a better example was a few pages back, if player 1 only ever bets the river with the nuts, a human would adjust by folding all non nutted hands, there for maximizing their win rate by losing less money, however a GTO bot would call some % of rivers to remain unexploitable.

For example, if player 1 began only betting the nuts on the river and player 2 quickly realised this by getting to showdown/looking at the stats then the obvious adjustment would be to fold all non-nutted hands. This adjustment is obviously exploitable because if player 1 had a brain they SHOULD then start betting the river with as many bluffs as possible (making player 2's adjustment exploitable)

although player 1 would essentially lose their money a little slower in this specific spot vs a GTO bot (due to the bot calling when p1 only ever bets river with the nuts), no matter what player 1 did they would always be making <0ev river bets because their isn't an adjustment they could make to start making the money back which they have been losing through not value betting/bluffing a more balanced range of hands (which the GTO bot would do)

i hope that makes sense (feel free to correct me if i'm wrong, i have only been looking into GTO over the past couple of days)

So you agree with me and Sklansky?

Whatever happened to Sonja anyway? Someone was asking about bots you could play against online awhile back and when I went searching for the link, I found out the pokerparadime site no longer exists.

BTW, if anyone wants to play around with some HU bots, a different UofA (Auckland) has the following posted online:

http://www.cs.auckland.ac.nz/poker/

One NL, one LHE. The interface sucks, but w/e.

I'm surprised so many ppl think it would be a big deal if Durrrr lost to a bot. He even made a comment like that himself in the hsnl thread iirc. It's probably a slightly better news blurb if he wins, but I don't see it making a huge dent either way. I guess I question how many ppl who aren't already passionate about the game would notice, let alone care.

Question for bot programmers out there:

Does anyone use an "evolutionary" approach to poker bot design? Meaning that you'd define variable design parameters and a way of measuring fitness (profit in this case) and let the bot "evolve" into the optimal design. I've read that this approach has been used to achieve strange but amazing designs in other areas (antennas I think) and wondered if the same approach would work in poker.

Sorry if this is an obvious/nonsensical question. I know very little about this.

Not a bot programmer and not sure about the term "evolutionary." I'm pretty sure there was someone posting in the IGT threads claiming to have been involved in developing their AIs, and that person stated that they used neural networks and the bots essentially develop their strategy by playing against other bots until they arrive at one that performs really well. I think that's the same as what you're suggesting?

I think the University of Alberta bots were designed in a similar way (see the HULHE thread I've already cited a few times) but I don't recall ever seeing mention of neural nets so the underlying approach to improvement might be different (more of a computational thing than an AI thing, not 100% sure though).

yeah although i didn't get the last question about conflicting with the superiority vs a good player (i assume you mean opposition like kanu/ike)

the GTO bot would play exactly the same and make more/less money depending on amount of mistakes (mistake = anything outside of GTO) made by any certain player?

ike posted earlier that he would guess a perfect GTO NLHE bot would beat his current game for 20bb/100 for the first 10 (or so) thousand hands. he also believed when he realised the bot would be crushing him he would study and attempt to emulate what the bot was doing to minimize his losses, ending with ike predicting a 10bb/100 win rate for the bot.

Great thread. One question I have though.......is Durrrr actually even a decent Limit player? It's a pretty different game to NLHE that he's known for crushing, so it seems pretty crazy to want to take on a specialist bot that 'knows' the limit game inside out AND plays pretty damn close to unexploitable AND has no emotional tilt issues etc. I mean surely there are far better high stakes LHE players out there than Durrrr?

Can't honestly see it happening, but would be very interested if it did.

I was using a different definition of medicre than you are. Few players that I would call mediocre seek out heads up online games.

Man, that NLHE bot sure makes some weird plays by current human standards:

==PREFLOP==
Spy Raise to $4.0
Sartre Raise to $53.0
Spy Raise to $400.0
Sartre Fold
==Fold==
Sartre folds.
Spy wins $53

BZZZT!

Thank you for playing.

The "O" in "GTO" stands for "optimal." A GTO strategy cannot be playing sub-optimally in some spots; otherwise it would not be a GTO strategy. A GTO strategy will be playing optimally in all spots. That's what "optimal" means: an optimal strategy is a strategy both is unexploitable and maximizes EV.

Playing GTO against an opponent who is playing exploitably will at least sometimes result in giving up EV. Optimal play won't win as much as exploitative play. But you cannot exploit the mistakes of an exploitable player without exposing yourself to exploitation.

EV is a thing that is used when human consider his exploitative options. When GTO bots are made, there isn't this EV function built into its computations. Think about it.

QFMFT...well said.

Even though I do not fully understand everything posted in this thread, it was an enjoyable read, however I do have a few questions regarding GTO and exploitable/exploitative play. If someone could answer that would be great.

The purpose of GTO play is such that we are not exploitable regardless of the adjustments our opponent makes, and any deviation from the equilibrium by our opponent is -ev for them, yes? Meaning their adjustment is a loss of ev for them and their play makes ours +ev.
Using the Roshambo example, we would always play 1/3 rock, 1/3 paper, 1/3 scissors regardless if our opponents plays 100% rock.

Is GTO a defensive strategy which assures we minimize our losses, but does not necessarily maximize our winnings; and so we make our winnings from our opponent hanging himself by deviating from the equilibrium?

And, in order to maximize our winnings we need to employ an exploitable/exploitative strategy which takes an offensive approach and we make the majority of our winnings from exploiting our opponents mistakes? Is this what makes humans more profitable against mediocre/bad players than a bot playing GTO? Again using the Roshambo example we would no longer play 1/3 rock, 1/3 paper, 1/3 scissors if our opponent plays 100% rock, but we would play 100% paper.

In regards to NLHE bots: is it, in theory, possible to create a bot which uses exploitative strategies based on population tendencies and adjusts to opponents as they play differently from these initially set tendencies?

In game theory, there are dominated strategies that are losing no matter what your opponent does,
http://en.wikipedia.org/wiki/Strategic_dominance
GTO bots mostly profit from those mistakes (and there are so many of them in poker).
And ofc it's possible to make exploitative bots, but the process is very different from making GTO bots.

we'll get this right some day nvg

The question also hinges on a variable that afaik is currently unknown; that is, the number and severity of dominated strategy decisions the opponent can make in comparison to the number of indifference points created by the GTO strategy. The reason the RPS GTO has a 0EV against all possible counter strategies is because the opponent can never make a dominated strategy decision. The more mistakes your opponent can make, however, the more potential the GTO strategy has to profit from mediocre and/or tough opponents.

A perfect exploitative strategy would ofc be more profitable, so it seems like another factor is how difficult it would be for the human expert player to identify and exploit his opponent's mistakes while keeping his own to a minimum.

Similarly, even if it could be shown that the GTO bot does have a higher winrate, there's still a question as to whether or not a human player's approximation of GTO would out-win the same player's attempt at an exploitative strategy. This could be something of a paradox since one's ability to do the latter is likely somewhat dependent on his ability to understand how to accomplish the former, but that too hinges somewhat on the quality of the opponent.

I guess to the extent that Hood says the iPoker bots are approximating GTO and out-performing humans, those results seem to suggest the answers to my first two questions. I think some ppl would argue that my third question can be answered by the degree to which many LHE experts have adopted a sort of goal to play in as balanced and unexploitable a way as possible, but even if we accept that, it doesn't seem clear to me that we can say the same for NLHE. My perception anyway is that while NLHE players incorporate balance, they still see a ton of value in finding ways to actively adapt to and exploit their opponents, even at the highest levels of play. So there's not that same sense of playing your best GTO-approximation while hoping to profit passively from your opponent's mistakes (that's perhaps an overstatement of the approach taken by LHE specialists but I think the distinction is real and helps to illustrate the point).