Robot Tournaments

[mizzoukid]

I live on the West coast of the U.S., so the tourneys available are more limited. I have played in the robot tourneys often, but am ready to stop. This is due to the fact the robots play trump contracts so poorly. It's amazing how often in a trump game contract where the robot is playing the hand that the robot fails to pull trump when it's obvious that's the best line. As such, far too often, my side goes down on contracts that are very makeable. I find myself bidding wrong just to avoid allowing the robot play the hand. Surely the programmers can tweak the tourney to increase the tendency to pull outstanding trump when it's clearly the best line to make.

[Stephen Tu]

Post some example hands?

 

Gib usually declares fairly well, if the time controls are set reasonably. Maybe the tourneys are having them play too fast. But in any case there's no heuristic for "drawing trumps", that's not how the software thinks. It calculates what line will work best percentage wise at each turn, which may or may not involve playing trumps.

[zeit]

I agree with mizzoukid. I think the "robot" does not play appropriately in many instances. Further, some of the bids are often questionable- esp. things such as inviational bids and sacrifices. (So far, it has never penalty doubled).

 

Zeit

[JoAnneM]

I have been playing in the Robot Tournaments and if there is a variable to be set for GIB to think it must be set at the shortest time. I usually have to review every trick because the cards flash by so quickly. I have noticed the same thing in just the seven tournaments I have played about GIB not pulling trump when it would have been the right line.

 

 

My disclaimer is that I am really enjoying these tournaments.

[helene_t]

I think GIB plays ok. If you rent GIBs, you can set it to playing slow and it will play very well. I think GIB will do better in a year or two even without software improvement, as CPU power becomes cheaper.

 

If I were in charge of GIB development, I would at the moment give higher priority to improving bidding decisions than to improve play. But I am sure Fred has his priorities straight, and GIB's bidding has improved a lot over the last year. Fred always listens to constructive criticism, so if you have some concrete hands that illustrate what appears to be a general problem, you may send them to Fred.

[JoAnneM]

I was wondering if the speed of my own computer controls how quickly the tricks are played and flashed across the screen. At least it is easy to mouse over the trick.

[jdonn]

If Gib often doesn't draw trumps when it should, it's because in its sims no hands came up where it mattered. That is something you have to live with unless they completely overhaul the algorithm for how Gib decides on its plays. Complaining about this probably means someone doesn't understand how Gib works.

[mtvesuvius]

An interesting thing I noticed when playing one of the ACBL Robot Duplicate tourneys... GIB plays the auction 1♠-1NT-3♣-3♠ as 10-11 HCP and 3♠. What does GIB do with other hands here?

[jdonn]

An interesting thing I noticed when playing one of the ACBL Robot Duplicate tourneys... GIB plays the auction 1♠-1NT-3♣-3♠ as 10-11 HCP and 3♠. What does GIB do with other hands here?

Bids 3NT and 4♠ way too often.

[peachy]

What I have noticed is that very often GIB doesn't have what the explanation (during auction when I mouse over to GIB's bid to see the meaning) says it has. Sometimes the difference between the explanation and the actual is HUGE. A second thing is that GIB leaves in takeout doubles that should never have been left in, and similarly, it never makes a penalty double.

 

As to its declarer play, sometimes it takes lines that I would not have taken but at least it goes by fast and I can go on to the next hand:)

[mizzoukid]

I HAD the specific tourney and board written down at home, but tossed it. Wish I could retrieve it but cannot, but this wasn't a one-time thing. Guess I'll have to play another robot tourney and write down any similar experiences. Again, it seems to occur only in TRUMP games, not so much in NT contracts, but I truly DO try to steer the final contract away from any trump bid if it means the robot is going to play it since I've been zapped too many times. Thanks, Randy

[cherdanno]

If Gib often doesn't draw trumps when it should, it's because in its sims no hands came up where it mattered. That is something you have to live with unless they completely overhaul the algorithm for how Gib decides on its plays. Complaining about this probably means someone doesn't understand how Gib works.

That sound like: A: "I find it unacceptable that we are spending so much money on health care in the US and still have so many people without insurance not getting the help they need." B: "That just shows you just don't understand how health care in the USA works."

[mtvesuvius]

<edit>

A second thing is that GIB leaves in takeout doubles that should never have been left in, and similarly, it never makes a penalty double.

<edit>

I believe GiB's failure to make certain penalty doubles is that almost all doubles are defined as "takeout". Maybe this is a system card error, but when I make a Double of the oppenent's game or slam contracts, the double is explained as "takeout". GiB always leaves the doubles in anyway, but I'm wondering if this may have something to do with it...

[jdonn]

If Gib often doesn't draw trumps when it should, it's because in its sims no hands came up where it mattered. That is something you have to live with unless they completely overhaul the algorithm for how Gib decides on its plays. Complaining about this probably means someone doesn't understand how Gib works.

That sound like: A: "I find it unacceptable that we are spending so much money on health care in the US and still have so many people without insurance not getting the help they need." B: "That just shows you just don't understand how health care in the USA works."

I don't think that's a good analogy, and not just because Gib is a toy. When they fix the bidding they make corrections or slight changes to the code. Fixing the play would involve rewriting the code entirely in a different way, in other words writing a new program altogether.

 

In other words, the healthcare system could be changed in major ways and still be the healthcare system. Gib could not be changed in this way because then it wouldn't be Gib. So the only constructive element to complaints like this are really for someone to write a new bridge-playing robot and for BBO to use that instead of Gib.

[peachy]

<edit>

A second thing is that GIB leaves in takeout doubles that should never have been left in, and similarly, it never makes a penalty double.

<edit>

I believe GiB's failure to make certain penalty doubles is that almost all doubles are defined as "takeout". Maybe this is a system card error, but when I make a Double of the oppenent's game or slam contracts, the double is explained as "takeout". GiB always leaves the doubles in anyway, but I'm wondering if this may have something to do with it...

A common type of Dbl that it leaves in is like (1H) P (2H) P P DBL I have stopped balancing in robot tourneys... and if I do, it will be a bid in a suit because GIB does not understand balancing at all. However, after balancing in a suit, Gib has no problem raising...

[cherdanno]

If Gib often doesn't draw trumps when it should, it's because in its sims no hands came up where it mattered. That is something you have to live with unless they completely overhaul the algorithm for how Gib decides on its plays. Complaining about this probably means someone doesn't understand how Gib works.

That sound like: A: "I find it unacceptable that we are spending so much money on health care in the US and still have so many people without insurance not getting the help they need." B: "That just shows you just don't understand how health care in the USA works."

I don't think that's a good analogy, and not just because Gib is a toy. When they fix the bidding they make corrections or slight changes to the code. Fixing the play would involve rewriting the code entirely in a different way, in other words writing a new program altogether.

 

In other words, the healthcare system could be changed in major ways and still be the healthcare system. Gib could not be changed in this way because then it wouldn't be Gib. So the only constructive element to complaints like this are really for someone to write a new bridge-playing robot and for BBO to use that instead of Gib.

You seem to know a lot about GIB. For example, apparently you know that it is impossible to automatically generate more simulations with bad breaks when random simulations do not yield significant differences between various lines. And you know that many similar ideas that I didn't come up with while thinking about it for 2 mins are impossible to implement either?

 

In any case, there is a difference between saying "here is a problem" and saying "XYZ should fix this problem!". I can complain about the disastrous state of public transport in the USA even if I know that this is almost impossible to fix.

 

Anyway, welcome back :)

[jdonn]

You seem to know a lot about GIB.

I have roomed with Ari several times lately and talked a lot to him about it over the past few months. I'd go out on a limb and say other than the programmers themselves I know more about how Gib works than just about anyone.

 

For example, apparently you know that it is impossible to automatically generate more simulations with bad breaks when random simulations do not yield significant differences between various lines. And you know that many similar ideas that I didn't come up with while thinking about it for 2 mins are impossible to implement either?

That wouldn't be impossible. Here are the problems (just that I can think of off my head right now.)

- It would make Gib extremely slow. Already the general complaint is that it runs too many simulations when it doesn't need to, and it would make that problem even worse.

- You would need to either run the sims for bad breaks in each suit (even slower) or decide somehow which suits are important.

- You would have to rerun the sims at each trick that goes by if the trump isn't immediately drawn (even slower).

- There is always the chance that those sims lead Gib to make the wrong play anyway so the problem would be diminished but not eliminated.

 

I don't know that there are or aren't other ideas, but they probably involve huge changes that would either slow Gib down way too much, or be equivalent to rewriting the program anyway like I said.

 

In any case, there is a difference between saying "here is a problem" and saying "XYZ should fix this problem!". I can complain about the disastrous state of public transport in the USA even if I know that this is almost impossible to fix.

 

Anyway, welcome back :)

Complain about whatever you want. I didn't say not to complain, I was just saying complaining about this particular problem (with the expectation that there is a fix any time remotely soon) shows someone probably doesn't understand how Gib decides on its plays.

 

Good to be back. :)

[Vampyr]

What are robot tournaments?

[barmar]

What are robot tournaments?

Tournaments on BBO where each human plays by himself at a table with 3 GIB robots, competing against other humans at similar tables. There are two kinds:

 

1. Robot Race or Robot Rewards: You all get randomly dealt hands, different at each table, you play as many hands as you can in the time allotted, and at the end you're compared on a total points basis with all the other tourney entrants. Robot Race tourneys are less expensive ($0.25 US) and the leaders only win BBO masterpoints. Robot Rewards are more expensive ($1 and $5), and the leaders get a share of the pot.

 

2. Robot Duplicate: This is like a normal Matchpoint duplicate tourney, except with robots at the table instead of other humans. All tables play the same boards, and they're matchpointed.

[porh]

i think there has been an improvement in gib in the recent months.. I've noticed that nowadays they actually balance sometimes amongst other things.. I think one wld get used to it's bidding style after many games.. Just my 2 cents

[CarlRitner]

GIB becomes a much better partner when YOU get on the same page as it. Since the other way around doesn't happen.

 

Two partners playing the same so-so system will consistently out-perform two partners playing a superior system, but differently.

 

That's nothing specific about GIB, as you all know. It's just a change in the normal diplomatic approach of 50/50 compromise to more like 100/0.

[Wayne_LV]

Personally I see a frightful trend in BBO. Why are there so many robot tournaments and why are they so popular?

 

There will not be a computer program that will play good bridge for the next 1000 years, if ever.

 

How do I know, you ask. 20 years ago I wrote the first simulation of the World Series of Poker and created an expert system that played Texas Hold'em, 7 card stud, Omaha, and draw poker, well enough to best several past world champions and a number of other expert poker professionals.

 

Was it AI? Absolutely NOT!

 

There is no such thing as artificial intelligence. Real human intelligence is pretty rare. What passes as AI is really nothing more than expert systems programmed to react to predetermined circumstances. A computer cannot and will not ever be able to learn in the human sense of the word.

 

There are games that are more adaptable to expert systems; Poker being one of them and Chess being the classic. A computer approaches chess by trying every possible move, given a board position, evaluating them all until the most satisfactory or least unsatisfactory move is found. A computer can do this at the speed of light ... a human cannot and besides most humans would be bored out of their goard long before a fraction of the possibilities are explored.

 

I used to love the game of chess until I realized that a machine is far better at it than I could ever become. I now play chess extremely rarely.

 

A game of chess against a computer is like a contest between a human adding numbers in their head vs a calculator. The human will never win.

 

And then there is Bridge. The game has way too many variables for any programmer or team of programmers to anticipate every combination. At best a small percentage of hands can be programmed to bid and play well by a computer.

 

The best chess programs can and will beat 99% of the chess playing human population. The best bridge programs will beat only a small percentage of the bridge playing population.

 

So why bother with a computer when there are far more proficient human players available with which to enjoy the game?

 

Or have we become so introspective that we cannot deal with each other on a human level, accepting our flaws and rejoicing in our talents?

 

I think we have .... and that fact makes me very very sad :(

[hrothgar]

 

Personally I see a frightful trend in BBO. Why are there so many robot tournaments and why are they so popular?

 

 

I had written a long post which ultimately collapsed into the following realization

 

"Robots are the McDonalds of the bridge world"

 

1. Robot's aren't as good as the best human partner's but they're a damn sight better than the worst. Moreover, you know exactly what you're getting.

2. Robots are convenient. I don't need to spend 20 minutes discussing system with the robot. We can sit down and just play bridge.

 

I'll always prefer dining at a fine restaurant to Micky Ds. However, McDonalds has its time and its place (and it serves up a hell of a lot more food then Le Bernadin)

 

As to some of your other comments:

 

At best a small percentage of hands can be programmed to bid and play well by a computer.

 

As I recall, there was a par contest 10-15 years ago back where GIB beat most of the human participants.

I think that Michael Rosenberg was one of the few humans to beat the computer...

[awm]

I don't see any reason robots shouldn't be able to play/defend better than humans simply by trying all possibilities, especially after fairly simple auctions that don't carry a lot of inferences (i.e. 1NT-3NT). It is a harder computational problem than chess because of the unknown information, but the combination of better algorithms and faster machines should still be able to carry the day.

 

Bidding is more difficult because the number of possible assignments of meanings to bids is so large. However, if you hold a "simple systems" event where everyone has to play more or less the same methods, I don't see any reason that a good rule-based system combined with simulation can't do well here too.

 

Surely there is a "human aspect" to bridge, where the top players try to read their opponents tempo and such to gain an advantage. But one would think there is even more such effect in poker, and yet machines can compete with human poker experts. Anyway, I think the "1000 years" statement is very much an overbid; if a concerted effort (with serious funding) was made to build a machine that was good at bridge (much as IBM made such an effort for chess) I think 5-10 years a very realistic timeline. Since no such effort appears forthcoming, we are probably looking at more like 15-20. No, it will not "reason like a person" -- it will do much more number-crunching and simulation, but it will play just as well.

[mike777]

It has been discussed in other threads, the bots already are better than 50% of ACBL players. That is a good start, not perfect but a good start.

 

No reason to not think in the coming years as speed/power gets cheaper the bots will improve.