[EricK]

Suppose a multibillionnaire with a passion for bridge decided to create the best bridge playing program i.e. hiring a load of top programmers and/or top bridge players, allowing them to work full time on this project, money being no object.

How good do you think they get a program to be within, say, a year? Better than any human player? Top ten? Not even close?

[cherdano]

One year is not enough, regardless of how many people are involved. I would think many aspects of improving a bridge program are a learning process (by the programmers) that just takes a while.

It might be better than the existing bridge programs after a year, but even that isn't so clear.

Arend

[Gerben42]

Well first they would buy the services and knowledge of the whole Jack team and start from there. After one year I'd think they could have a program that would beat the current Jack, I guess.

Í guess the program would be able to play the top national competitions with some success.

[hrothgar]

This is (obviously) a very complicated question... I have a couple thoughts:

1. The type of development effort that you do will very much depend on what kind of time frame you're looking at. Assume for the moment that you wanted to produce the best possible program within a relatively short period of time. Your development methodology is going to be very different than if you have five years or so to work on the problem.

2. If I had the time and resources necessary to do things the "right" way I'd start by looking at the problems with existing programs like Jack. Ideally, if would be interesting to Supercharge Jack by giving it access to a couple of order's of magnitude more processing power. After this, play a LOT of boards versus a top quality human team. The goal is to specifically identify the problem cases with the existing approach. Coupled with this, I'd work on a parallel effort focusing creating a syntax to be used for disclosure. (Human machine disclosure issues are very complicated)

Personally, I think that the whole issue is a matter of resources rather than difficulty.

[Gerben42]

Quote

Personally, I think that the whole issue is a matter of resources rather than difficulty.

I agree. Jack already played some matches against Dutch top pairs and this was extremely helpful for the programmers.

[hotShot]

If you play chess all there is to know about the game is visible on the board.
If you play bridge, you know what cards are still out, but you don't know how they are distributed. This is why a bridge programm will always be left with some uncertainty. So the best playing level a bridge programm can reach is that of the best statistical chance.

But i think the biggest improvements can be made in the biddding section.

[1eyedjack]

hotShot, on Oct 8 2006, 07:36 PM, said:

If you play bridge, you know what cards are still out, but you don't know how they are distributed. This is why a bridge programm will always be left with some uncertainty.

And yet, if there is one thing that computers do MUCH better than humans it is crunching numbers. So if the only (or main) problem in doing well at bridge were in the prediction of the placement of unseen cards, that is a problem that should be ideally suited to a computer, in which area it should shine in contrast with a human. Yes, it is left with some uncertainty, but no more uncertainty than a human presented with the same information, and it should quantify that uncertainty with greater precision in a shorter time.

[barmar]

hotShot, on Oct 8 2006, 02:36 PM, said:

If you play bridge, you know what cards are still out, but you don't know how they are distributed.

The same is true for human bridge players, isn't it?

Computer programs have all the same information that human players do, although it may be difficult to program them to take advantage of some of it (inferences are a tricky programming problem). But they have perfect memory (they'll never forget which spot cards have been played, and they can be preloaded with all the standard safety plays) and much better statistical capabilities.

If you want to compare with chess, chess programs are not very good at pattern matching (which is how we believe human chess experts play the game), but they can use raw computing power to examine thousands of moves to make up for this. Similar tradeoffs are possible in bridge.

[hotShot]

Each distribution has a different "best card" that should be played. By number crunching the computer can find the move that is best for most of the possible distributions.
So while a chess programms may be able to loose no game, a bridge program will loose 50% of all hand that depend on a finesse.

A chess program can enter a game at any point and take over from there.
A bridge program needs to know the bidding and the sequence in which the previous cards where played.
A bridge player/program needs to draw conclusions from the sequence the cards where played. While number crunching is easy to computers, drawing conclusions is not that easy.
The advantages of bridge programs are:
-almost no unforced errors
-almost perfect in percentage play
-reliable bidding style
Problematic areas:
-antipercentage play
-drawing conclusions from bidding and play

[mycroft]

One of the problems involved with Bridge computers - that has been very effectively removed by the World Computer Bridge Championship - is that in order for a bridge player - human or not - to succeed, it has to be able to understand and deal with differing bidding languages.

It's the same gripe in RL Bridge - "Why should I have to defend against that?" - it's like programming a chess computer to deal with the fact that in this game, the movement and capture styles of pawns are reversed; capture straight, move diagonally.

The computer has to both deal with defending against and using in analysis:

1C being
- 11-21, 2+ clubs, may easily be the shortest suit;
- 15+ any;
- 11-21 with clubs, or 11-14 or 18-19 balanced (2+ clubs) or 18+ unbalanced (0+clubs)
- 13+, (usually) 4+ clubs
- 0-7 any
- 8-14, 4+ *hearts*
...and many more.

You can make the same argument for any other bid - as Eric knows well (2C weak, eh?) This is a very solvable problem - in fact, probably easier for computers than for humans - but it just makes the single-dummy analysis part of the game harder.

For a while, the best way to foul up a chess computer was to do something silly early - to take the game "off book". The ability to do that in bridge is huge.

Again, as a computer programmer (which I'm out of, thankfully), "a year" isn't enough - see Brooks' Law for why. Five years? I think we'd have to see. It would be interesting to see also what they come up with in terms of bidding and cardplay systems, and where they're legal. Computers do have the advantage over humans that they don't "forget" anything playing "our real system" twice a year, vs "what we have to play in GCC events" the rest of the time - or "our real system" for two days of a Nationals vs "this other thing" the rest of the time.

Michael.

[hrothgar]

mycroft, on Oct 9 2006, 08:57 PM, said:

One of the problems involved with Bridge computers - that has been very effectively removed by the World Computer Bridge Championship - is that in order for a bridge player - human or not - to succeed, it has to be able to understand and deal with differing bidding languages.

I always thought that the decision by the administrator's of the World Computer Bridge Championship was extremely short sighted.

1. It drove some of the best programs out of the Championship. Ginsberg stopped entering GIB because of the new restrictive systems policy. (I'll note in passing that Ginsberg's decision to migrate over to MOSCITO prompted this whole brou-ha-ha)

2. If you are ever going to have an effective competition between humans and computers, disclosure issues are going to be key. Case in point: Zia psyches a LOT. You know it. I know it. The computer playing against him needs to know it. Limiting the range of permissable systems in the Championship ensures that the developers get to side-step a big ugly hairy issues. Personally, i'd prefer to address this head on and work at developing a reasonable solution.

[mike777]

A am bit lost on what the problem is. Zia is limited on how much he can psyche legally and ethically. Legal bidding systems in the Bermuda Bowl are wide ranging but limited. Even with screens which limits table action it does not remove it and inferences are allowed.

Computers should be allowed to learn and use whatever is legal and ethical just as humans do.

If computers cannot beat the best humans at the table in less than 20years I would be shocked.

Frankly I would be shocked if the best computers could not beat 50% of the current 25 million bridge players out there now.

[pigpenz]

on just a one hand basis i would think a computer would be able to do just fine, but when we bring in things like state of the match, emotions, etc. it would be harder for the computer to overcome those....it would be just crunching numbers for the computer. So on a just crunching numbers basis i would think computers should be just fine...its the other intangibles that they would have to overcome.

[mike777]

pigpenz, on Oct 10 2006, 10:24 AM, said:

on just a one hand basis i would think a computer would be able to do just fine, but when we bring in things like state of the match, emotions, etc. it would be harder for the computer to overcome those....it would be just crunching numbers for the computer. So on a just crunching numbers basis i would think computers should be just fine...its the other intangibles that they would have to overcome.

But that is just crunching numbers also, yes? Heck if World champs do not get these decisions correct, why should computers do worse? Can not computers learn to interpret emotions even with screens, body language, tone of voice, etc? I assume these computers can see, smell, hear, touch etc.

Heck computers can do all of the above now in a very limited and expensive fashion.

That just sounds like crunching the "right set of numbers". They just have to get them as right as often as humans do?

For sake of discussion lets assume computers in 20 years, at a very low cost, can process, access, etc at the speed, capacity, etc as the entire human race combined.