http://spectrum.ieee.org/tech-talk/c...for-first-time

Cliffs:

• Limit-holdem only for now
• New algorithm with a different "regret minimization technique" to select the best strategy at each step of the game
• Reduced temporal averaging of algorithm to most recent hands vs all hands, which reduces computation and memory required
• Used compression to reduced data set required to *just* 11 terabytes for counterfactual data and 6 terabytes for main strategy
• Uses 200 computer nodes each consisting of 24 2.1Ghz AMD cores, 32GB of ram, and 1TB disk
• Algorithm confirms that some basic tenants of optimal play hold up, including the advantage of position and aggression/raising, whereas some tenants were contradicted, like always bet-capping streets with the nuts and overly selective starting hand requirements out-of-position

.

wouldn't really call that conquering

http://news.yahoo.com/self-taught-co...192346149.html

i saw that one but it doesn't sound that great. soon enough tho.

Do they say what constitutes a game?

It really bothers me when reporters don't know enough to explain something. In the article they mention that the computer rarely caps preflop with aces. But that misleads non poker players who weren't told that the bet size doubles on the turn. I doubt it would make this play otherwise.

Here's the real article:

Science 9 January 2015:
Vol. 347 no. 6218 pp. 145-149
DOI: 10.1126/science.1259433

It's not public access.

They reraise with a pair of threes.

On balance I think the article did a good job summarizing the project. In writing just as in information theory, it's impossible to distill the essence of a something without losing some detail.

What about algorithms such as "deep learning" being applied to poker.

Whats the problem with these new algorithms is, computer scientist say they can apply the same algorithm to many fields, such as a program designed to learn how to beat pong and diagnos medical data.

Whats amazing about A.I is it really seems to only be mostly computational power and data.

Hasn't this been around for a while for limit?

http://poker.srv.ualberta.ca/

This is the university's project page. Its under some heavy load right now, but you can play against the bot and view its strategies.

Or at least that's what this article says:

http://www.washingtonpost.com/news/s...ying-computer/

Looks like it's backed by science so I would assume there is some legitimacy to it.

It looks like it's solved HU limit poker by the way of Snowie where it just played against itself billion billions (not a typo) of times to find the most optimal line in every single scenario. It's not an intuitive program but rather a database recall. Will be interesting what comes from this.

^ This sounds like the stuff where we have to ask China to invade Taiwan 1 000 000 times so we can calibrate our military planning AI. It's fine for games but its a bit of a dead end for more real world stuff.

I heard after a few million test runs the computer complains that the damn fish call any two cards.

No, it posts some sick graph up/to-the-right and claims to have solved poker.

You are being to kind. If they are going to make a point about not capping with aces they should be competent enough to elaborate so as not to mislead. (Although the guys that were interviewed might be culpable as well.)

so does this program adapt, to players or just play an unexploitable game?

I found a number of spots in my poker variant with certain ranges etc that are 100% unexploitable what i found though is that i beat the players but lost to the rake, i had to vary from that line to strategys that were exploitable but exploited the weakness in weak opponents.

a bot that is unexploitable may potentially just lose to the rake but im guessing it will have enough of an edge on randoms with no concept to avoid that fate i dont know.

Bloomberg article: http://www.bloomberg.com/news/2015-0...ong-moves.html

“We have a strategy that can guarantee a player won’t lose,” said Michael Bowling, a computer scientist from the University of Alberta, who led a team working on the program. “It’s going to be a break-even game. It’s only when someone makes a mistake that they could end up losing.”

In other words: they claim GTO


edit: It seems like they brute forced the solution, right? Just ran enough sims to the point where every random hand and random decision point had been up against every other random hand and decision point and from there on narrowed it down to the strategy with the most over-all +EV choice in every spot?

iirc they have been very close for a long time.

Heads up limit? How long until it becomes self-aware and starts flinging actual poo?

Hey! This is Mike Johanson, one of the authors on the paper. We're swamped right now keeping the website up, talking to the press, and running Twitter, so I only have a few seconds to spare, but I'll stop by 2+2 later to answer questions.

We're on twitter as @PolarisPoker . Check out my 30-tweet blitz from earlier today for a summary of what we've done.

We've been working towards this result for years, and have gotten closer and closer over time. Last year, we were down to 26 milli-big-blinds per game, or 1.3 BB/100, of exploitability: that's the most that a perfect counter-strategy / nemesis could win against our program. I've talked about that result on 2+2 a few times in the past, and the tools that we use to calculate that exploitability figure.

In October 2013, after I commented a few times in one 2+2 thread, our coauthor Oskari Tammelin got ahold of me. We'd worked with Oskari before, when he invented his PureCFR algorithm. He told us about two things he'd recently invented: CFR+, a new game solving algorithm based on our CFR algorithm that learns *incredibly* fast compared to what was then the state-of-the-art version of CFR, and a new compression technique that reduced the memory cost for solving holdem from 523 TB to 11 TB.

Oskari's two developments were hugely important, and opened the door to finally make solving HULHE possible. This paper is a joint effort between the U of A (me, Mike Bowling, Neil Burch) and Oskari Tammelin. It took us most of a year to write the code and tune it to be able to get the quality of result we wanted, at a memory limit that fit on our cluster, and at a CPU cost that was feasible. This was a tricky 3-way optimization. Actually running the computation used 4800 CPUs for about 70 days, with a total computation time of just under 1000 CPU-years.

The result is that our new program Cepheus is beatable for less than 1 milli-big-blind, or 0.05 BB/100. Even if you knew the perfect counter-strategy and could play it flawlessly, it'd take 60 million games to overcome the variance due to luck in order to actually have 95% confidence that you were winning. It's essentially solved: not quite perfect, but closer than any human could distinguish within a lifetime of play.

Our website is constantly going up and down right now due to traffic, but here's a link: http://poker.srv.ualberta.ca . The site will show you the preflop strategy (as a nice set of graphics), the strategy for any decision point in the game, and will also let you play against Cepheus. Although the "Play vs Cepheus" part is the part that is most swamped at the moment. It should get easier to get through in a few days.

I'll check back in on this thread tomorrow, when we're less hectic. If you have any questions about the result, I'll be happy to answer them then.

inb4 high stakes prop bets are made

i assume game = hand?

Who plays limit holdem anyway...

I'm one of the other authors (Neil Burch - webdocs.cs.ualberta.ca/~burch/two_plus_two.html for the skeptical) and I claim no responsibility for any poker errors in any media articles! I avoided any specific examples just because I figured I would mangle them. I'm not sure what Mike, Mike, or Oskari might have said, though...

Responding to others...
I will own up to the "200 games per hour" bit. Sorry. That probably should have been "200 hands per hour." From my point of view, as a game theorist and not a poker player, poker is a repeated game, but we tried to avoid that terminology specifically because it's confusing in the context of poker. We missed one, I guess. I _think_ I set it up so that it really is a single sided hand, not an average of playing both seats, but I'd have to double check the math to make sure.

I'm kind of happy to see a resounding "Meh." We hoped everyone with a bit of knowledge would be aware that we hadn't somehow ruined the game of poker by finding an (almost) game theoretically optimal strategy, but I wasn't sure exactly how much credit to give people... It's oddly reassuring to see a bunch of people say "who cares?"