WCGRIDER said that he believes Claudico would beat stakes up to HU NL400.
This is a program that made its strategy using supercomputer.
Is it really possible to make winning midstakes bot running on regular computers?
How does it work?

Sure, we'll just tell you exactly how to make a bot. No problem

This is a misconception. It doesn't take a super computer to run the bot. It took a super computer to program the bot and teach it. Once it's optimized, it could be run on a standard laptop.

also there's so betsizing standardization at midstakes that a bot there can get away with being significantly less complex than claudico

^this, I think it would do just fine with just 3 betsizes.

This is so stupid it practially drools

What's so stupid about it?

very big dick style of donger to be winning vs the bot while bjorn is crushing it so hard

Well these guys are quite a bit better than the average midstakes reg (especially on the sites which have more bots) so that accounts for something.

Also, these guys know they're playing a bot, which makes a difference, as not all regs will know player X is a bot, especially since most bots are on those sites where you can change SN's often etc..., or the bots just make multiple accounts regularly anyway.

Also, considering Claudico's maker's lack of knowledge about poker theory, and some other things, I believe that the bot is definitely not as good as it could be. If you gave a team of top players and top programmers access to a supercomputer, especially with today's poker software, I think a better bot could be made.

So it's mostly a combination of 1) and 2). Some people may overestimate the botting problem, but for the most part, it's understimated imo.

Also, the bots that are apparently doing well at midstakes are probably built to exploit the particular tendencies of the players in these games. i.e. they're exploitative bots whereas Claudico seems to be trying to play more GTO-ish.

I think we need to be careful when comparing their effectiveness as they don't really pursue the same objective.

For us poker players, some of the stuff that the AI chief says may come off as ignorant regarding "poker theory" (sure enough, anyone who sits at a table not knowing what the term "donk bet" means, will make the other guys at the table salivate), but the way such a bot is built, it may have a negligible effect if the programmers are up to date on the latest theory, or if top players are in the development team. The bot is trained by playing against itself for a huge amount of hands, and probably hardly any poker theory (in the way a poker player understands "poker theory"), a lot of which is a way for humans to grasp the game, becomes any sort of input to the program. For all I know, many of the successful chess programs were built by players who are fairly good but by no means outstanding.

Yes but it's clear they made a lot of bad decisions in the design (unless the goal was different than making the strongest poker playing entity and it might well be different, for example applying as generic framework and proving it can learn everything by itself) and those decisions wouldn't have been made by programmers who have a lot of experience with poker.
This is a bit scary because sheer computing power they have was enough to overcome:

-lack of poker knowledge (for example they could've cut the tree very significantly by reducing options which are worthless or close to worthless like limping preflop or 6x overshoving turns/rivers)
-not very efficient code (taking this kind of time to recalculate rivers on this kind of hardware means their code is very slow)

Now imagine what happens when people who are good at writing efficient code and know a lot about poker start making AIs.

You know of course way more about this kind of stuff than I do (it's your job, not mine ), but wouldn't it seem like they rather did not build enough options into their tree, instead of too many (like if it had explored more of the min-donk lines, it wouldn't have been thrown off by this tactic so much ?).

The river delay seems like a design decision to incorporate Sam Ganzfried's "endgame solver" http://www.cs.cmu.edu/~sganzfri/Endgame_AAMAS15.pdf ... again, whether, this is any good or not, I am not in a position to argue with you

How are you so sure that those options are worthless?

I am one of the developers of Claudico and I know a lot about poker. I was SNE twice on PS and have played for 9 years.

Our 2014 agent Tartanian7 beat everyone with statistical significance in the computer poker competition, including teams that had programmers with expert poker knowledge. Claudico is way better than Tartanian7.

This was a pretty massive undertaking; this Claudico project. A lot of money and hours from top professionals went in. I would not assume they were lacking for developers or poker playing consultants. I wouldn't be surprised if this bot is as good as it can get right now or near too it. Nobody in their basement has a super computer. This bot was built by a major university. I don't see many efforts more massive than this even being put into NLHE. We will see. Punter you kind of make it sound like had this Claudico team been more poker based (I.e. Like the GTORB team?) they wouldn't have 'wasted' all these resources and built a much better bot.

Thank you. Good read.
I have to assume that all or some of the humans in the competition also read this before the first cards were dealt.

It's obvious once you run enough cases and see they don't change EV by much.
As to the limp preflop... the almost perfect solution produced by Alberta team doesn't have limping in HU limit Holdem. There is even less incentive to limp in NLHE with deep stacks.
I mean, when you are far far away from NE it doesn't make sense to spend your time on those lines (unless your goal is different than simply producing the strongest poker playing entity). You need to know how to defend vs those plays but it's easier than having those in your arsenal as well.
Once you are close to NE maybe it's worth investigating but as it is it's just about making choices about where to spend your energy/computing power.

Nice to know, still my point about bad decisions stands.
Winning against other academic teams doesn't mean much
People who have the strongest AIs don't play in those and the strongest public AI won't come from academic world either. This has long tradition in programming as well: the strongest playing programs in various games never came from academia, the reasons I think are obvious.

What about addressing specific things: calculating rivers having 64 core machine for that and taking 30 seconds in there doesn't sound too convincing. Including options which can be safely discarded without much (or any) EV loss is there as well.

The impression about lack of knowledge about poker comes from interviews of the chief architect btw, not only the ones made on Twitch streams but also the ones published before and from one thread we had about it before in NVG (which ended once the most straightforward question: how exploitable is that AI was asked). I had the impression he is more about applying general techniques and seeing how they do in poker because the impression is he doesn't understand much about the game itself (as documented by many quotes here and conversations with Doug).

This is really an interesting thought.
My gut-reaction was that they weren't proficient enough to optimize their method but after this quote I remembered that the big-picture of this stuff is developing algorithms that can solve complex real-world problems. of course optimizing is a huge part of that, but maybe they have their reasons for wanting to approach it this way even if it isn't leading to the purely strongest poker playing entity in the quickest possible fashion. Maybe I've been too harsh, or maybe I haven't - I'm definitely more open minded about it at this moment.

Either way, the bot is still (to me) impressive even though it isn't perfect and think we can all agree that the stream has been extremely entertaining to watch.

i think it's better that claudico is coming from an academic background and less a pure poker background

computing power says that eventually NL will be GTO and if the "solve" is unorthodox then it shows just how much we have to learn in approaching GTO and range weighting

i think chess is comparable in this regard; i don't play much chess but i've heard that the bots who crush grandmasters do really weird things

a lot of the standard sizing and plays from these days would get you crucified 5 years ago

Haha, yeah, maybe soon ppl will be like "What you mean, you don't ever shove 6x pot ???"

The computer poker competition isn't just for academics. Many of the strong entrants are from hobbyists (some who are also poker players), and some from professional botters as well. Anyone can submit. For HU limit Texas Hold'em the current strongest agent is from academia -- the University of Alberta just "essentially solved" it to exploitability of 1 mbb/hand and would beat anyone. One year a team submitted claiming to beat up to 5/10 hunl online, and we beat them. I agree though that it's possible some amazing bots exist that chose not to participate.

Our team had three members. Each brought different skills, and had a variety of background in poker.

There are some limitations in academia that wouldn't apply to people outside of academia. We can only work on approaches that can be published in strong conferences. Often reviewers know close to nothing about poker, and a paper that is considered "poker specific" will just be auto rejected. (Earlier versions of) my endgame solving paper that was linked above for example was rejected four times, before it was finally accepted with scores right around the minimum threshold for acceptance. People outside of academia are free to work on poker-specific approaches without regard to whether they would lead to publication.

A big goal of this research is to develop techniques that will apply to domains outside of poker, e.g., national security and medicine, where problems are modeled as imperfect-information games. So we prefer approaches that are more broadly applicable to ones that are purely poker specific.

I'd be happy to discuss specific details of our algorithms and potential improvements after the competition is over.

why would it be easier to defend vs. a line than it would be to employ the line yourself? I don't see how you can know how to defend vs. a line without knowing how you'd competently play that line yourself

also a programming team that's ignoring strategic options from the getgo is going to take a much longer time to reach a NE (that's the ultimate goal here right?) than a team that starts from the ground up, yeah?

That should be interesting, I'm sure that you'll be taken up on this, looking very much forward ...