Computer Faces Poker Pros in No-Limit Texas Hold’Em Competition | Hacker News Reader

Computer Faces Poker Pros in No-Limit Texas Hold’Em Competition

cmu.edu

33 pointsempressplay17 comments

The scientist behind the AI, Tuomas Sandholm, has a nice 115 page resume: http://www.cs.cmu.edu/~sandholm/cv.pdf

That's crazy.

The most important piece of information seems to be missing. What's the stack depth? In the AAAI HUNL competition it's 200 big blinds which is significantly harder to be good at than 100 big blinds which is the standard for online HU play.

I think you can actually build a fairly competitive HU bot for 100bb games with very simple means (rules based with a bit of statistics thrown in, no opponent modelling or advanced features or learning). At 200bb it's a lot harder and researchers that approach it from an "understanding of the game" POV rather than a "pure computation" POV have a tougher time.

Edit: I only know Doug (aka WCGRider). Wish they would play against Ike Haxton, jungleman12, sauce123, Galfond or Durrr

>it's 200 big blinds which is significantly harder to be good at than 100 big blinds which is the standard for online HU play.

I'm quite surprised by this!

100bb has higher variance and therefore more luck, where 200bb has lower variance

So, I would say the 200bb tournament rewards skilled (human) players more effectively. As a skilled player, I would seek out the 200bb tournaments.

My immediate thought would be that a computer player would be rewarded better in the 200bb tournament.

But, even in the 200bb situation, players will have to play short stacked some of the time. Perhaps the difficulty for bots is the requirement of changeable modes here?

They're not playing tournament style. The blinds are fixed and I think the players stacks are reset before each hand.

The smaller the effective stacks, the better it would be for a bot. For instance, below 15bb stacks, a bot could just play nash equilibrium and be unexploitable. By definition, the human would have to do the same, or become -ev by deviating one way or another.

The fact that every hand starts at 200bb's in this challenge is actually very good for humanity.

On another note, it's very interesting to see that the bot found independently that huge overbets on the river are close to optimal. This is something that only recently some top HU pro's started incorporating in their games.

I don't see where they got 10^161 as the number of situations in no-limit. You can bet ANY amount, making it potentially unlimited. You also have to consider the ratio of the size of your stack to the size of the blinds, which means the number of situations is unlimited (arbitrarily deep stacks).

I.e., if the blinds are $1-$2, you follow a completely different strategy when you have $10 as when you have $100 or $10000.

> You can bet ANY amount, making it potentially unlimited.

Not really.

The minimum bet size is fixed by the blinds/antes.

The maximum bet is fixed by the total number of chips in the tournament (or that the player holds).

The bet amounts are limited by the size of the smallest chip - in bricks and mortar tournaments tournament managers will 'chip up' the small chips whenever it becomes sensible to do so. In digital tournaments the smallest chip might remain at $0.01, but it still exists.

Suppose again the blinds are $1-$2 and the stack is $10000.

I can bet $2, $4, $6, $8, $10, $12, $14, any amount up to $10000. Each of those is now a separate game tree. A couple of bets, and now we're way past 10^161.

It may never be optimal for me to bet $132, but a perfect computer player should be prepared for that bet.

While you are theoretically correct that the game tree is massive, the difference in tree between betting 1bb vs 2bb in a 50bb pot is marginal; the algorithm they use considers an abstraction of the game.

So they're just doing an approximate game tree.

Small differences in bet size can matter, because professional players will sometimes try to give you the pot odds to make a bad call.

Nope.

All that matters is the ratio of bet size to the blinds.

Basically you normalize everything into "BBs". Actual $ amounts aren't relevant.

False. The strategy is different based on the size of your stack relative to the blinds.

If the blinds are $1-$2 and you have $10, you're probably going to go all-in or fold.

If you have $10000, then you can think more about trying to trap an opponent and get all their chips, instead of just trying to take the blinds.

You're both saying the same thing except you're saying "if the blinds are $1-$2 and you have $10" and he's saying "if you have 5bb". You're saying "if you have $10000" and he's saying "if you have 5000bb".

What I'm saying is that I don't see how "all possible bet sizes" leads to a game tree of 10^161. It's got to be more unless they're taking some shortcuts, such as saying that a pot size of 24xBB and 25xBB are treated as equivalent.

Also, it seems that they are only considering one stack size, where you have 100x or 200x the big blind. If the poker solver does all possible stack sizes, then it can't possibly be 10^161.

In practical play the small blind is often the basic "unit", so usually the smallest delta between successive bet sizes are the size of the small blind, or half that (e.g. at $1/$2, allowed bets might be $2, $2.50, etc)

Nope.

Having a $200 stack at $1/$2 is the same having a $2000 stack at $10/$20. It's 100BB in either case, and the appropriate strategy is the same.

Reply on news.ycombinator.com