Interesting. So if we assume human opponent plays his relatively normal GTO'ish approach, not going for hyper-exploitative approach, then computer would lose about -65bb/100 if it always folded? (-25bb/100 in SB, and let's say -40bb/100 in BB if human in SB only open-raises 80%)

And current AI's underperform this by even more? This sounds like a lot. Especially since couldn't computer easily improve by always shoving hands that are +EV to shove with the current stack size, not even balancing in bluffs to those hands? Would you bet on the side that Libratus will be losing at least -10bb/100?

Kinda surprised to see Doug not taking part to a $200k ($50k a piece?) freeroll. Too baller or too leery that his alleged #1 HU position would be on the line? This would've made for some good YouTube Polker hands content, too.

Sam,

Should we have any Bayesian priors for this competition based on past competitions, or are the strategies of both the humans and computer so different that we should reset?

Correct, I have no involvement with this event.

i wont million

old thread:

http://forumserver.twoplustwo.com/29...u-bot-1526750/

I'm pretty amazed reading this right now; we used to have a lot of discussions about exploitability and specifically CMU's decision to not publish any explo numbers for Claudico in the past. Now this



in addition to this



and this



kind of confirms (we still don't know for sure) a lot of the suspicions others and I had. I won't go into it any deeper, it's all in the old thread.

Bonus:



#bull****science

More bs from CM:

https://www.cs.cmu.edu/news/upping-a...intelligence-0



For one thing, I don't know how I feel about the statement that Poker poses a more difficult challenge than Go -- surely incomplete information makes things more difficult, but there are many factors that add to the complexity of a game and how easy it is to solve.

But that's besides the point -- which is that Prof. Sandholm sells it like letting a hunl poker bot play vs human players is the only/best way to evaluate its strength. In 1vs1 poker, there is a scientific measure that is precise and comparable, exploitability. That's a direct property of a Nash equilibirum for 2 player zero-sum games. And while it may be sometimes tough to calculate it based on how the bot is working/how the strategy is saved, there surely is a way to compute a lower bound of exploitability, just as the two authors in the paper linked above did. Note how these authors aimed at comparing the strength of various hunl bots and used exploitability as the natural measure -- how would you feel if they instead played every bot for 30k hands each and published their results?

How will CMU compare Claudico and Libratus strategies for instance? By comparing the results of both brains vs ai competitions? The article said that CMU is running the supercomputer for around 15 million core hours (compared to 2-3 million for Claudico) -- are they never checking whether running it longer is actually making progress in the right direction when using that many resources? If yes, how if not by comparing exploitability? Is there any theoretical proof that their algorithms converge to equilibirum?

Bottom line is, whenever someone presents you a poker strategy that is supposed to be good/aiming for optimal, and they dont't present you a (lower bound of) exploitability number, then that means they either don't know what they are doing since they aren't measuring what they are doing, or they know that whatever they have done isn't working too well and consciously decide to not to let people know but rather sell it through another way.

When is this happening and when will it be streamed

in b4 CMU was in the lab all year coming up with max exploit strat against Doug & friends.


btw, how exploitable the AI is doesn't seem too relevant in this case. I'm certain even the best humans are exploitable to some ungodly number against their nemesis. It might be exploitable in ways humans won't even be able to figure out.

You don't need a perfect nash bot to beat humans, you just need one that's slightly closer to optimal than we are, and I don't think we're very close to nash at all.

Looking to bet a decent amount on the bot. Anyone?

They are all a part of the "empire" so doug will be there in a sense.

Right, but CMU are not a private company with the goal of building an AI that can beat humans at poker, they are trying to develop techniques (and show that they work!) to solve real-world problems:



It doesn't really matter whether the bot will win the competition because that doesn't really prove anything; it says nothing about how close to optimal the solution is, how well the algorithms do, it doesn't even reliably tell you whether Libratus is better than Claudico because of variance and the fact that human play is not a constant benchmark, etc. Competitions are a very noisy measure at best.

It's one thing to have bot vs human competitions in games where no such measure is feasible, but if it is for this game, it's obvious that one should use it to test/measure progress and evalute the result. More importantly, exploitability is relatively easy to measure (esp if you are good with computers anyways!), so why not publish it? ...

^^^^^^
Economist Martin Armstrong has already done this with his Socrates computer.

No one more prescient on world events. Check out his blog.

Hi Sam. The conclusions in that paper are really interesting! I am no expert in mathematical programming nor poker playing, but I would like to throw my opinion here.

I had the conviction that the abstraction techniques typically used (Imperfect Recall, K-means Earth Mover for example) should give very poor results. But finally got to believe that heads up nl bots were so strong that the abstraction was good enough if sufficient buckets were made. This paper demonstrates current abstraction technique is not good enough.

May be I don't fully understand the abstraction of programs like Claudico but I think they would benefit strongly from a mixed Imperfect Recall/Perfect Recall abstraction. May be you don't need to remember the previous round hand strength, but you need to remember the public board texture sequence. May be you do not need to remember the exact betting sequence, but you need to remember the stack to pot ratios at the beginning of each round.

I believe that imperfect Recall is really good to get a detailed enough card abstraction especially on first rounds, but perfect recall of some critical information would be better than a super fine card abstraction.

I'm also looking to bet on the AI, 5-1 odds, my 1k to your 5k. With escrow to someone known.

Looking to bet between 500 to 1k on the AI to win, 5-1 odds with escrow to someone known.

Cant edit this. Looking to bet 1000 on the AI to win, 4.9-1 odds. My 1k to your 4.9k. With escrow to someone known.

I haven't heard from the poster who wanted 5-1 on comp after we PMd back an fourth a bit; me being in the US makes it a lil harder to send/receive funds but PM details and if that falls through I'll likely take ur action. Escrow will be necessary unless u got a crazy good rep. I have zero history with trades/bets on here personally

Pm sent.

I'm surprise at the amount of people willing to bet on the Bot. Has the development of AI improve that much in a short period of time?

Shhhhhhh. EZ money.

Same here. One of the major reasons why I'm on the human side is that the match takes place over 20 days. Team Brains has ability to share exploitative reads over the bot for every team member each night (and they're probably well incentivised to do so given the $200k freeroll) and adjust accordingly. However Team AI comes into the match with a fixed strategy, and the AI team will not be able to adjust AI during the course of the match because of lack of computing power. (Unlike Deep Blue was apparently adjusted during the infamous '97 matches, I think.)

At least this is my understanding of the current state of bot development, and if I'm wrong, my bet probably was not that good.

he said he was going to release a video about the subject tomorrow.

Anyone know if there is a time limit for how long the computer gets?

Well...think about it like this: What edge would the humans need to have (in bb/100) to be a 5:1 favorite in a 120k sample given the standard deviation from the previous competition. Or if that STD is not publicly available, use typical HU NL STD with a loose, aggressive opponent.

Also, anyone making bets should specify whether "win" means to win money after 120k hands or to have a statistically significant win (as per scientific publishing standards). Last competition, humans only won by the former definition, so the organizers called it a draw.

Looking to bet a few hundred on the bot @ 4.2 to 1. Send me a pm