Why are computer not better than they are at bridge?

[helene_t]

Surely not. You still have to place the opponents' cards.

SuitPlay goes (as I understand it) some way towards "solving" bridge by assuming that defenders follow an optimal mixed strategy. In priciple this could be generalized to the whole hand (not just one suit) and to assuming that all three players follow an optimal mixed strategy.

 

I don't say it is easy - declarer's play depends on entry restrictions, avoidance play etc and defender's play depend on the extend to which they know these things and even the extent to which they know whether their partner knows about it etc.

[Free]

What you fail to understand is that computers don't have to be perfect, or even close to perfect to be world class, you just have to play as well as the best human players. What difference does it make if you could single dummy analyze 5200300 hands for 1 trick? Not every hand is equally likely]. The world class player will eliminate a huge percentage of them based on the bidding or lack of, the choice of suit led, the card led, signals, discards, etc. Human players obviously can't do millions of computations but the best are viewed as world class. If humans don't need to do all those millions of computations to be world class, neither do computers. What they do need to do is reproduce or improve on the reasoning that humans do when planning the play whether as declarer or on defense. IMO, that is a reasonable and achievable goal but it will require some top programmers to spend a lot of time coding.

The 5200300 is just a number to give an indication of the magnitude of hands, and indeed not all hands are possible due to the bidding. Nevertheless, if there's a fraction of them left and you need to do single dummy analyses for 4 hands, you'll still require huge amounts of computing power. Limiting the amount of simulations is the only option with a lack of computing power and can still result in decent performance. Using algorythms to eliminate hands is an obvious choice to gain 'speed' and improve the number of quality-simulations, similar to chess ignoring ridiculous moves to focus calculations on serious moves. You will indeed get a decent improvement. But in case of computer bridge, more IS better, hence my argument against the claim that we have enough computing power.

 

Chess is actually a nice example where we lack computing power to "solve" the game, and that's also the reason why it can still be beaten from time to time! With enough computing power to actually solve the game, a chess computer would become better than world class chess players. Even a team of grand masters wouldn't be able to win (unless the game isn't neutral)! This can't be said by bridge, because bridge isn't an exact science with complete information. The optimal result you can get in bridge is a percentage play, and that will fail on many occasions. Playing perfect percentage lines all the time won't guarantee that you'll win.

[mike777]

"The 5200300 is just a number to give an indication of the magnitude of hands, and indeed not all hands are possible due to the bidding. Nevertheless, if there's a fraction of them left and you need to do single dummy analyses for 4 hands, you'll still require huge amounts of computing power. Limiting the amount of simulations is the only option with a lack of computing power and can still result in decent performance. Using algorythms to eliminate hands is an obvious choice to gain 'speed' and improve the number of quality-simulations, similar to chess ignoring ridiculous moves to focus calculations on serious moves. You will indeed get a decent improvement. But in case of computer bridge, more IS better, hence my argument against the claim that we have enough computing power."

 

 

 

"Given the current speed of progress, industry experts estimate that supercomputers will reach 1 EFLOPS (1018, one quintillion FLOPS) by 2018" 10(18)FLoating point Operations Per Second),

 

 

Just how much more powerful do computers need to become to do these simulations inside simulations to beat wc in bridge? If we do not have enough, what is enough, what is the number?

[PhantomSac]

They do not need to become more powrful, processing speed and power are not the issue in 2015 for why computers are not beating the best humans at bridge

[lackeman]

Hi free!

Lets say that a human based on what happened (ie a conclusion) can bee almoast sure on that for example IF having 11 trumf IT is 2-0 and therfore feel almoast sure on that it is right to finess. Then no computerforce Will be to Amy help for the computer.

[uday]

processing speed and power are not the issue

 

I have to agree with a number of points, including this one.

 

I loved Hrothgar's idea for explaining systems to another bot.

 

Even a tiny company like BBO has access to a lot of computational power. It just takes money, and the ability within the software to split up the problem (say, a simulation) so that you can have some large number of computers work on it at the same time.

 

I think there are several reasons why bots aren't better, tho I can only speak to our GIB. Just my opinions. I'm not an expert on the guts of this stuff, tho I tinker.

 

1. The lack of a meaningful incentive (or funding) to improve bots. I don't know why IBM developed Deep Blue. It couldn't have been cheap. What did they get out of it? In our world, who's interested/rich enough to fund this sort of development, and why would they?

 

2. The lack of a meaningful database or language to explain the meanings of calls ( or, if you prefer, the lack of a database that can explain what bid to make with some hand after some auction )

 

Sure, we add rules constantly. But it is like, as Georgi says, boiling the ocean. In GIB's particular case the language used to express the rules is quite difficult to work with and is limited in many ways. But making the developers of bot software also responsible for defining some sane system ( completely ) makes it harder.

There are no useful reference materials when trying to define stray sequences.

 

3. The difficulty of reading meaning into partners (opponents) actions on defence

 

4. Using double-dummy simulations instead of single dummy simulations

 

5. Limiting the computing resources ( time spent, if u prefer, on a single computer ) that the robot has access to per decision

 

6. The difficulty of trying to figure out what the trick target should be on any hand

 

There are others but those are the main ones that come to mind.

 

For GIB, the database of bids is the main technical obstacle, since flaws there cascade into other areas like card play and defence. But I think the main non-technical reason is mostly the lack of resources/incentive.

 

U

[mike777]

But again cannot learning overcome all of this. I assume computing power is cheap enough or will be soon.

 

For some reason members seem to think we need to teach the bots...no....let them learn.

 

 

computing power needs to be cheap, very cheap. I mean hardware and software

----------

 

 

If I understand Barmar and my words not his....computing power remains too expensive in 2015 But at least give us a number.....you guys do not give ua a number... a goal.

 

 

give it time..

 

 

I advocated hologram bbo bridge many years ago and still await it. I mean FULL hologram not cheap version. :)

[jogs]

 

 

1. The lack of a meaningful incentive (or funding) to improve bots. I don't know why IBM developed Deep Blue. It couldn't have been cheap. What did they get out of it? In our world, who's interested/rich enough to fund this sort of development, and why would they?

 

 

Watson beat Jennings in Jeopardy. Computers are learning to think like humans. Do these top bridge playing programs really use lots of simulations? Humans don't use simulations to play hands.

[Zelandakh]

Humans don't use simulations to play hands.

Humans also do not analyse chess in a brute force way, yet it was the switch to this from trying to emulate human thinking that caused the first big step forward in chess computer design. What works best for a computer is not necesarily the same as for a human.

[jogs]

In chess during the end game brute force is investigating every permutation. The chess program would always find the solution. A simulation examines a small portion of the possible permutations. Declarers count their sure winners, then their possible winners. A bridge program should also be counting sure winners. Maybe use simulation for possible winners. In many cases it may be in a table of percentages. Mainly probabilistic games like bridge aren't suitable for brute force analysis.

[Free]

Mainly probabilistic games like bridge aren't suitable for brute force analysis.

And why is that? Because we lack computing power to do it properly... :rolleyes:

[jogs]

And why is that? Because we lack computing power to do it properly... :rolleyes:

 

No. 90% slams still go down 10% of the time.

Deterministic games can be 'solved'. Probabilistic games cannot.

[Free]

No. 90% slams still go down 10% of the time.

Deterministic games can be 'solved'. Probabilistic games cannot.

That's what I've been saying the whole time, the best you can get from a computer is the percentage line. If we agree that finding the percentage line is actually making the world class play, then simulations can be beneficial.

 

I know there are various ways of calculating that percentage, but the easiest way is by doing lots of relevant simulations to reach a pretty accurate result.

[hrothgar]

No. 90% slams still go down 10% of the time.

Deterministic games can be 'solved'. Probabilistic games cannot.

 

You don't seem to be familiar with the definition of the word "solved"

 

"Matching pennies" has been "solved" for the better part of a century

[Free]

http://img.4plebs.org/boards/tg/image/1366/23/1366236683151.jpg

[billw55]

That's what I've been saying the whole time, the best you can get from a computer is the percentage line. If we agree that finding the percentage line is actually making the world class play, then simulations can be beneficial.

Sure .. if we define world class play that way.

 

I wonder however, to what extent human factors come in to play at that level. Things like picking up on an opponents fidget, or posture, or tension level, and finding key cards based on that. This is the kind of thing that it will be much more difficult to get computers to do well. The question is, how much that matters.

 

 

[Zelandakh]

Justin has given the impression that is not so important at the highest level, Bill, simply because the top players do not give anything away with such obvious tells. It is probably more important just below that level. Moreover, the computer has an advantage in not giving any tells at all, which nicely evens up any such disadvantage in head to head matches.

[billw55]

Justin has given the impression that is not so important at the highest level, Bill, simply because the top players do not give anything away with such obvious tells. It is probably more important just below that level. Moreover, the computer has an advantage in not giving any tells at all, which nicely evens up any such disadvantage in head to head matches.

So, it is a part of the advantage world class players enjoy over others, but not among each other? That makes sense.

 

And yeah .. those bots are some cool cats http://www.bridgebase.com/forums/public/style_emoticons/default/laugh.gif

[jogs]

Okay, I should say in deterministic games the outcome is predetermined.

 

The bidding portion of bridge certainly hasn't been solved.

And there may be a ro sham bo effect. System A beats system B.

System B beats system C. System C beats system A.

[antonylee]

Even in card play such non-transitivity exists, see e.g. https://en.wikipedia.org/wiki/Suit_combination#Mixed_strategies

It doesn't really matter; again there is always an optimal (in the sense of non-exploitable) strategy.

[jogs]

You don't seem to be familiar with the definition of the word "solved"

 

"Matching pennies" has been "solved" for the better part of a century

 

http://en.wikipedia.org/wiki/Solved_game

 

I found it. I was using "solved" correctly. You are confusing optimal strategy with solved.

[rhm]

Deterministic games can be 'solved'.

This may look evident to you, not to me.

What is the precise definition of a "deterministic" game?

How do you determine a (nontrivial) game (say chess or GO) is deterministic?

 

Rainer Herrmann

[Zelandakh]

It just means there are no random factors, Rainer. The rules of the game lead to the statement that chess is deterministic. Sports are typically not deterministic (physical factors are involved) abd bridge is clearly not due to the randomness introduced by the dealing process and presence of hidden hands.

[helene_t]

Maybe different textbooks use different definition of what a "deterministic" game is, but as I understand it the presence of hidden hands in itself doesn't make a game non-deterministic. See for example

 

http://www.cs.cmu.edu/afs/cs/academic/class/15381-s07/www/slides/021507gamesI07.pdf

[Zelandakh]

Maybe different textbooks use different definition of what a "deterministic" game is, but as I understand it the presence of hidden hands in itself doesn't make a game non-deterministic.

Can you think of any non-deterministic game that has random dealing + hidden hands?