Revealed: How Libratus bot felted poker pros – and now it has cyber-security in its sights • The Register Forums

Tuesday 19th December 2017 09:04 GMT John Smith 19

"The techniques that we developed are largely domain independent "

And at 1.35 PFLOPS quite computationally expensive.

But this is pretty impressive

No doubt the Strategic Systems have found some refinements that will be proprietary and lower that or improve the decision making at that level. Of course the human processor does it in 2Kg package that uses about 200W.

Incidentally wasn't improvements to playing "imperfect information" games like poker what the scientist in "War Games" was famous for?

7 1 Reply
1. Tuesday 19th December 2017 10:11 GMT steelpillow
  
  Re: "The techniques that we developed are largely domain independent "
  
  "...and can thus be applied to other strategic imperfect-information interactions, including non-recreational applications."
  
  I assume they emailed a copy to the Pentagon.
  
  4 0 Reply
  1. Tuesday 19th December 2017 22:12 GMT Bob Dole (tm)
    
    Re: "The techniques that we developed are largely domain independent "
    
    I assume they emailed a copy to the Pentagon.
    
    ... I assume it emailed a copy of itself to the Pentagon..
    
    2 0 Reply
2. Tuesday 19th December 2017 10:18 GMT jmch
  
  Re: "The techniques that we developed are largely domain independent "
  
  "Of course the human processor does it in 2Kg package that uses about 200W."
  
  More like 1.5kg and 20W if only the brain is being considered
  
  http://faculty.washington.edu/chudler/facts.html
  
  https://www.popsci.com/technology/article/2009-11/neuron-computer-chips-could-overcome-power-limitations-digital
  
  Of course, the human nervous system extends al over the body and various other systems (eg endocrine) also regulate the performance of neural functions. It's not at all the case that a 'brain-in-a-jar' would be computationally the same as a brain 'in situ', as it were. So it's maybe a fairer comparison to use numbers for the whole package - approx 70 kilos* and 100W. That's still incredibly inpressive.
  
  Of course, human poker players have been training for years and years, so for fairness you could train the AI with higher power to get an equivalent training time and run it in-tournament at 100W for a fair comparison.
  
  *Yeah, I know
  
  3 0 Reply
  1. Tuesday 19th December 2017 11:16 GMT Muscleguy
    
    Re: "The techniques that we developed are largely domain independent "
    
    Though high spinal injured patients who have lost everything below the neck still think perfectly well. The hormone mediate effects on the brain are slow wave effects on set point and not information processes. Similarly while as headless chooks demonstrate much of the fine control of movement is done by the spinal cord when you don't need to move that subroutine does not contribute to central computation in the brain. The brain does not send off compute tasks to the Spinal Cord neurons and receive them back.
    
    So to the brain in a vat you need add a sensor filled cranium with a mouth for energy consumption and an associated gut and heart for circulation, also a liver and a kidney or two to handle water balance. The liver is not just there to detox the food, it is the food store for the brain. The brain is not only an obligate glucose user* that cannot burn fats but it has no energy stores. Cut off the blood supply to your limb and you can still use it, albeit painfully. But do that to the brain and unconsciousness results. Ditto if your blood sugar drops too low. The prime energy store for the brain is the liver. Brain in a vat needs more than glucose solution feed.
    
    *the neurons themselves actually eat lactate, the glia break down the glucose to that point then feed the lactate to the neurons. Think of the neurons as like a nitrous fuelled race engine living life fast and clean.
    
    The no fat burning is probably at least in part an adaptation to the fact that the brain is largely made of fat in terms of membranes, of axons and dendrites and myelin sheaths. Burning fats in that environment would be akin to lighting a library with candles.
    
    4 0 Reply
3. Tuesday 19th December 2017 12:07 GMT joed
  
  Re: "The techniques that we developed are largely domain independent "
  
  I'd add that I'd lose against computer controlled game bots most of the time (at levels > easy) but I'd not call it AI. Simply I and majority of human player are novices at these games (and have no intent or time to change that). I bet the same was true for thousands of poker hands played by Libratus, especially if entry fee was low.
  
  3 0 Reply
  1. Tuesday 19th December 2017 12:43 GMT jmch
    
    Re: "The techniques that we developed are largely domain independent "
    
    "Simply I and majority of human player are novices at these games (and have no intent or time to change that). I bet the same was true for thousands of poker hands played by Libratus"
    
    IIRC from the original article a few months ago, it was playing against tournament-level professional players, not just any mugs who could afford an entry fee
    
    2 0 Reply
4. Tuesday 19th December 2017 17:03 GMT Tom 7
  
  Re: "The techniques that we developed are largely domain independent "
  
  You'd only need 30 of those new google chips for this tho. We're talking half a rack or so.
  
  1 0 Reply
Tuesday 19th December 2017 09:15 GMT Anonymous Coward

OK so 1.35 PFLOPS is a lot

But how much does it really need while playing? That is, if you train it on the super expensive supercomputer, could it run on say a high end PC and/or GPU and still play well?

They could fund their development costs (and supercomputer expenses) by turning it loose in online poker games and winning millions.

7 0 Reply
1. Tuesday 19th December 2017 17:34 GMT Mage
  
  Re: OK so 1.35 PFLOPS is a lot
  
  And it's not really AI at all. It's easier for a computer than a person to remember all the cards played and calculate odds.
  
  2 0 Reply
  1. Wednesday 20th December 2017 10:18 GMT John Smith 19
    
    "And it's not really AI at all. "
    
    It never is, once you can explain how it works.
    
    1 1 Reply
Tuesday 19th December 2017 09:16 GMT mark l 2

It might need a super computer today, but in another 5-10 years time it could be running on kit that is within reach of normal consumers, and at which point the online gambling business could be loosing a lot of money to bots playing games on their systems. That is assuming the online gambling site aren't already rigged, as it is much easier for software to be written in favour of the casino winning than it is to rig a real life poker game.

5 0 Reply
1. Tuesday 19th December 2017 10:23 GMT jmch
  
  "That is assuming the online gambling site aren't already rigged, as it is much easier for software to be written in favour of the casino winning than it is to rig a real life poker game."
  
  I sort of assume they already are. After all, in a real-life casino everything is so much rigged in favour of the house that if you get really lucky you're liable to be thrown out for suspected cheating.
  
  8 0 Reply
  1. Tuesday 19th December 2017 15:23 GMT Anonymous Coward
    
    well
    
    "I sort of assume they already are. After all, in a real-life casino everything is so much rigged in favour of the house that if you get really lucky you're liable to be thrown out for suspected cheating."
    
    Yea, but in poker, the house doesn't have a need to manipulate the game. The house has no interest in if you win or lose (unlike craps, blackjack, etc.). The house survives on the rake (a percentage that comes out of the pot).
    
    I'm thinking that many casinos also bank on the poker player either wandering over to the blackjack table for a while, or bringing a friend/SO who spends time at the slots.
    
    2 0 Reply
2. Tuesday 19th December 2017 18:05 GMT Anonymous Coward
  
  I think it is safe to say that 1.35 petaflops will not be available to normal consumers in 5-10 years. Probably NEVER will, or at least not in either of our lifetimes. Moore's Law can't continue past fundamental physical limits, and will probably stall out for economic or technological reasons before that.
  
  0 1 Reply
  1. Tuesday 19th December 2017 19:15 GMT Brewster's Angle Grinder
    
    @DougS Betting that Moore's law will end is a bit like betting that the Bitcoin bubble will burst.
    
    That said, the argument we will all have peta-FLOPS systems only depends upon the price/peta FLOPS decreasing; it doesn't require on improvement in chip performance. And if we do hit the hard limits, we might still be able better optimise the hardware for the specific application or find software improvements.
    
    1 0 Reply
  2. Wednesday 20th December 2017 10:25 GMT John Smith 19
    
    " 1.35 petaflops will not be available to normal consumers in 5-10 years. "
    
    Funny you should say that.
    
    It's estimate that there are 1 x 10^15 connections in a human brain.
    
    If each is involved in a calculation then you could say ordinary consumers already do have such access.
    
    And what is AWS or Azure for but to provide such access on demand?
    
    Current transistors are about 140 atoms wide, so the end of Moore's law is in sight. But in principle they can still shrink linearly 14x, till the conductors are 10 atoms wide and the insulator thickness (historically the real smallest dimension) 1 atom thick.
    
    1 1 Reply
Tuesday 19th December 2017 09:49 GMT K

Great article... ahead of its time..

So much so, that it has terse relevance to cyber-security..

But, still I'd have a beer with it.

1 0 Reply
Tuesday 19th December 2017 10:03 GMT Anonymous Coward

quadrillion floating-point math calculations per second.

wtf is a quadrillion? I had to look it up, apparently its about a million billions.

per second.

Do they really need that to work out the probabilities in a game of cards?

0 2 Reply
1. Tuesday 19th December 2017 10:10 GMT Charles 9
  
  Re: quadrillion floating-point math calculations per second.
  
  That only gives you a rough estimate, even. And yes, the kinds of game's provide scenarios that easily run away from you. Remember, chess and go have similar issues.
  
  3 0 Reply
  1. Tuesday 19th December 2017 18:37 GMT Anonymous Coward
    
    Re: quadrillion floating-point math calculations per second.
    
    About that:
    
    "... containing about 10^13 nodes – much smaller than the 10^161 nodes..."
    
    "The planck length - much smaller than a galactic supercluster"
    
    2 0 Reply
Tuesday 19th December 2017 12:09 GMT Stratman

What would be the result

if two instances of Libratus were to play each other? Would the best hand always win?

5 0 Reply
1. Tuesday 19th December 2017 12:49 GMT jmch
  
  Re: What would be the result
  
  "Would the best hand always win?"
  
  I would guess its more complex than that because "best hand" isn't only initially drawn cards but can vary with the luck of subsequently drawn cards. These probabilities will be taken into account by Libratus, so might vary over a couple of games. Of course over a large number of games the probability that each machine gets about equal cards rises towards 1.
  
  However the benefits of having a big pot of money available are also there, so maybe the machine that gets some early luck can beat the other long-term even if the luck of the draw evens out
  
  2 0 Reply
2. Tuesday 19th December 2017 15:35 GMT Anonymous Coward
  
  Re: What would be the result
  
  The "best hand" isn't all that great in poker, that's what's hard for people to grasp, and what makes people assume there's more chance involved in the long run as a poker player. Pocket aces versus 7-2 offsuit, heads-up starting hands in Texas holdem are only about a 10:1 advantage preflop.
  
  Beginning players see AA in their hand, get excited, and bitch about online poker being rigged when they lose. Pros understand they'll lose at least one out of every ten times they draw rockets, and bet accordingly so that they come out ahead in the long run.
  
  0 0 Reply
3. Tuesday 19th December 2017 15:54 GMT scrubber
  
  Re: What would be the result
  
  Best PLAY would usually win. Position is huge in poker, acting last is a major advantage. Starting hands are important too, but you rarely win a hand pre-flop.
  
  If two identical strategies came up against each other then it would be the luck of the draw, but you'd still see very few showdowns.
  
  1 0 Reply
  1. Wednesday 20th December 2017 07:34 GMT Charles 9
    
    Re: What would be the result
    
    "If two identical strategies came up against each other then it would be the luck of the draw, but you'd still see very few showdowns."
    
    Unless FORCED, like in a head-to-head situation. This also minimizes the advantage of the Button since it alternates between hands.
    
    0 0 Reply
4. Tuesday 19th December 2017 22:12 GMT englishr
  
  Re: What would be the result
  
  > if two instances of Libratus were to play each other?
  
  A STRANGE GAME
  
  THE ONLY WINNING MOVE IS
  
  NOT TO PLAY
  
  1 0 Reply
Wednesday 20th December 2017 00:21 GMT a_yank_lurker

AI of?

AI or just a big database with lots of computing power behind it?

0 0 Reply
1. Wednesday 20th December 2017 13:31 GMT phuzz
  
  Re: AI of?
  
  Isn't "a big database with lots of computing power" a pretty good definition of AI anyway?
  
  0 0 Reply
Wednesday 20th December 2017 04:01 GMT Anonymous Coward

Bad beats (cracked aces) / Bluffing etc

Hard to get a picture on how the algo handled these cases, anyone know?

0 0 Reply
1. Wednesday 20th December 2017 15:05 GMT Charles 9
  
  Re: Bad beats (cracked aces) / Bluffing etc
  
  I think it would just play along since it's hard to bluff consistently and come out ahead. Reason being there's a risk your bluff backfires because you're up against someone with a genuinely good hand (ex. trying to bluff with a 7-2 off-suit when the opponent holds an actual pocket pair or at least something competitive like a suited connector). The last thing a bluffer wants is to get called, which is why bluffing is more of an art. Plus, in an online setting, there's no source of facial tells to get a read on what players are hiding.
  
  0 0 Reply