I am posting this on NVG and The Heads UP Limit Forum.

I just found out for sure that it switches "personalities". It does not switch based on how it is doing or how you are playing. It does not adjust to your play in any way. But because it switches after random periods of time you may get that impression. You thus be misplayying against it because you have recognized patterns in its play that will no longer be true shortly after you recognize them. For example you may notice that it reraises in the big blind preflop with 95% of its ace rags so you can be almost sure it hasn't paired the ace on the flop when it just calls. But by the time you spot this it has switched to a personality that reraises only 70%. This new personality induces the bluffs you have recently been making so it looks like it has adjusted to you. Actually it hasn't. But even so its multiple personality technique will cause you trouble, especially if you didn't know about it . Now you do.

wow. with some careful legwork work this thing should be massively exploitable.

How so?

I think you're begging the question. If a given personality is exploitable, then having that on its menu of possible strategies would be exploitable once you figure that out. If it has multiple unexploitable personalities, switching among them would make it no more or less exploitable, except that it is harder to spot tendencies in each personality and for some period of time after the switch your perfect exploit is pointless.

David, most of us aren't in Vegas. I haven't really seen anything that makes me terribly excited to play the thing. Are there people who think playing the machines is a profitable activity. Using the methods the programmers used, it seems likely that you could make a bot that wouldn't be worth playing.

I've seen a number of people post in these threads that talk about "strategy X" from the machine in terms of how we as players understand poker. As I understand it, looking at a piece of software this way isn't that useful. It didn't come to understand poker in any way we did. It tried to optimize its results and wound up here. There isn't a explicit "reraises in the big blind preflop with 95%" strategy that is different from a "a personality that reraises only 70%". One set of learning came up with one and one came up with the other.

It is possible that there were pre-programmed shortcuts in the SW, but to date none of the information I've seen supports that. Thus, trying to interpret motive or other play based on any stat is misleading. Among the many strange things this algorithm chooses to do it does this one thing preflop.

Do you have any indication that playing the machine is a good idea? As a guy who writes code, it seems interesting. As a poker player, the more I know the more I want to just walk on by and play the people beyond.

Interesting stuff, could have been inferred really from what's been reported but hard to for certain.
So the obvious best way to maximally exploit it is to play standard GTO in a vacuum type play, ignoring all previous HH. And given the leaks that have been reported under whatever personality, I think it would be well worth playing for decent profit at the higher stakes.

fwiw Polaris had multiple "colors" (personailities), which were selected by the computer based on its opponents' tendencies. In fact, my understanding was that most of the "thinking" the computer did during the matches was just evaluating which color to use.

why? can't each personality still be a close GTO approximation?

I think huflhe gto only has one personality. I know this isn't a fundamental necessity for this class of game, but based on my understanding of the game's structure, I believe it to be the case. I think it will be shown (if it hasn't already) that all optimal equilibria to huflhe are in the same family, and that family has a single dominant pair.

If this is correct, then any noticeable personalities should represent significant deviations.

This is a very different situation from Polaris - Polaris adjusted her personalities in an attempt to exploit opponent's tendencies. This thing randomly plays chunks of hands with a given, pre-defined personality.

I also may be vastly under-estimating the minimum magnitudes of the deviations. If we visualize the flhe game space as game values rising out of a 'strategy plane', it is probably very lumpy and odd. There are likely countless local maxima and minima in untold manner of weirdness.

What if we set a strategic stake in the sand, such as '3bet more than 40%' or 'never cap btn', and then find the most balanced strategy that satisfies that constraint? So we're setting a boundary in our strategy plane, and climbing up the highest hill we can find within it. These constrained peaks could be quite close in height to the true optimal peak. In this case, duder could exhibit multiple personalities without being very exploitable.

There's no way to really know. I guess I was basing my assertion that this thing should be 'massively exploitable' on an assumption that the strange and unarguably terrible plays that people have posted do not represent errors in abstraction solving (as I had originally assumed), but actually represent facets of duder's various personalities.

But, these two causes aren't mutually exclusive, so idk.

From the control system world, the knock on this type of programs is that they cheat when optimizing. As I understand it, they are geniuses at finding holes in your statement of the problem. In the NVG thread, you see a series of insane river folds from the machine. I wonder if this is the artifact of some sort of river play optimization that it has done. I wonder if the real exploit on this sort of machine would be to understand how the software likely learned the game. It isn't looking to balance stats. It doesn't care about percentages on street or unexploitable play, unless someone put in some shortcuts to make it do so.
This depends. Any claim of "GTO" for the machine assumes that there is some GTO strategy and that in optimizing, the machine found a close approximation to it. I'm not sure if that is a given. From the claims in threads, the machine was in no way given standard LHE HU concepts as a starting point.

We can also guess the machine spent a lot of time playing itself. Other than that, it played HU experts, as could be found by the developers. I wonder how broad a set of strategies it saw while training. In the end, the potential exploit could be to play in a way it was never trained. Thus, working on slight variations of the current state of the art is playing exactly how it has the most experience.

good post as always.

i agree with you that it's quite possible there is only 1 GTO optimal solution. However, I think there would be a host of strategies that are globally close to that solution yet have some dramatic local differences. Thus the multiple personalities would imply some minimum level of exploitability, but by no means guarantee that it could be exploited for much. And indeed I would be surprised if it could.

Btw how are you visualizing this strategy plane? What do the axes represent? I was intrigued by that concept

i'm guessing it's just your sentence that is unclear, but you know the first half of your assumption is not an assumption, right?

Every time a discussion of the machine comes up, there is a resulting GTO conversation that goes with it. I know little about GTO vs. HU LHE. I assume that it is not certain that there is a single GTO strategy to find. Many of the posts I have seen about the machine assume that there is one and that the programmers tried to find it.

I have some idea of what the programmers did to produce the machine. Based on what I think they did, it converging to this GTO strategy would be an unintended result. They claimed they didn't start out with the current best possible balanced strategy and then let it search ways to play when near that. I have a friend building an egg-bot. He didn't tell it anything about how to balance on two wheels. He told it a bit about physics, and it mucked around trying to figure out how to win, to stand upright. As a control system engineer, many of the choices the egg bot made seem odd. I suspect as a poker player, you'll make the same observations about this machine playing poker. The egg bot SW looks nothing like a classical control system. The poker bot may not seem much like a good poker player.

it is certain that there *at least* one GTO strategy. whether there is exactly 1 or more than 1 isn't known, afaik.

Isn't there a "can be found" bit, that is in doubt as well?

nope. hu lhe meets the criteria of the minimax theorem:

http://en.wikipedia.org/wiki/Minimax#Minimax_theorem

Is knowable in question? Just because it meets an existence theorem doesn't mean that it has to be possible to discover, does it? That was my point (perhaps not clearly stated) in my previous post. I'm not sure that we'll ever discover the strategy, and thus be able to compare it to the machine's play.