[Eubulides]

My impression is that the BBO robots are programmed by human programmers. ChatGBT, facial recognition systems, etc. are self-trained robots that learn from their mistakes. There are billions of deals in BBO archives. Robots trained on that data would have judgment based on vast experience. Their bidding system would have the conventions that work in the long run. Just as in chess, the world champion would be a program.

[LBengtsson]

Nope. It's a bit more complicated than that. The robots would have to understand why they are making the bids. Bridge is a far more subtle game than chess, which is effectively number crunching various positions in a linear mode. The Russian and world chess champion Mikhail Tal used to make positions deliberately complicated for his opponents, but if his opponent had been a computer program, the program would just number crunch faster and more accurately than him. His chess method was sometimes psychological, but can computers learn a similar psychology at bridge? I doubt it.

There are far too many variables for a bridge computer program to assimilate: chess effectively is a series of patterns, with moves that are conditioned by how a piece moves. Even if a new bridge program could look at the billions of boards that have been played on BBO, or at least a few million played by expert players, it also needs to understand so many factors other than bidding and play. I am not saying it will not happen within the next 20 years or so, but it certainly will will not be happening by 2025 in my opinion.

[johnu]

 Eubulides, on 2023-June-12, 09:18, said:

My impression is that the BBO robots are programmed by human programmers.

And you are correct. There is no AI in BBO's robots. They are not capable of "learning" from previous hands.

[pescetom]

 johnu, on 2023-June-12, 11:21, said:

And you are correct. There is no AI in BBO's robots. They are not capable of "learning" from previous hands.

Nor were the human programmers entrusted with fixing the robot capable of learning from bug reports, even when they (once) tried.
Carefully read, this forum is an entire book about how to improve BBO's robots.

Not that I share the pessimism about AI being capable of learning to beat us all at bridge. I suspect it is just that there is not enough financial return on effort.

[pilowsky]

 LBengtsson, on 2023-June-12, 10:23, said:

The Russian and world chess champion Mikhail Tal ...

Tal, Jewish from Latvia, would be offended!
In any event, the oft-repeated idea that chess is 'easy', compared to bridge, because all the information is available and the answer is only limited by number-crunching is false.
A single Bridge hand has a 'solution'; so in that way it is comparable to an end-game in chess.
A chess game has an infinite number of possible outcomes and is therefore much harder to comprehend and play - no matter how good the calculating ability of the carbon or silicon based player.

[smerriman]

 pilowsky, on 2023-June-12, 17:34, said:

In any event, the oft-repeated idea that chess is 'easy', compared to bridge, because all the information is available and the answer is only limited by number-crunching is false.
A single Bridge hand has a 'solution'; so in that way it is comparable to an end-game in chess.
A chess game has an infinite number of possible outcomes and is therefore much harder to comprehend and play - no matter how good the calculating ability of the carbon or silicon based player.

You have these backwards. A chess game has a finite number of possible states, and each position is a theoretical win, draw, or loss. (And endgames with up to 7 pieces have been 'solved' via a tablebase, so you instantly know what the right move is).

It's bridge which has an infinite number of states, since you have to take into account not just the current state of the cards, but also the potential reasoning that was behind every past bid and play from everyone else at the table.

[pilowsky]

Bent Larsen would probably disagree with L. Bengtsson.
But perhaps this discussion deserves a thread of its own.

[barmar]

 Eubulides, on 2023-June-12, 09:18, said:

My impression is that the BBO robots are programmed by human programmers. ChatGBT, facial recognition systems, etc. are self-trained robots that learn from their mistakes.

There has been research (not by us) on using neural networks in bridge robots. One of the problems with this is that they're inscrutible -- they don't have specific agreements that can be disclosed to opponents. They just do what seems to work.

Furthermore, as far as I know, no one has yet implemented something that can handle competitive auctions. That adds the additional complexity that you need to be able to inform the robot of the meaning of the opponents' bids. In our GIB application, we just have the robots programmed to assume the opponents are playing the same system as the bots. But as you've likely seen, its ability to describe the meaning of bids is limited -- it has no way to represent multi-meaning bids (consider 2♠-4♠ -- 4♠ could either be weak, increasing the preempt, or strong, expecting to make, so GIB explains it as the intersection, which is effectively meaningless).

[pescetom]

 barmar, on 2023-June-15, 08:27, said:

There has been research (not by us) on using neural networks in bridge robots. One of the problems with this is that they're inscrutible -- they don't have specific agreements that can be disclosed to opponents. They just do what seems to work.

Bridge between humans is already hugely challenged by the peculiar situation that agreements must reflect partnership experience as well as what was written and they must be disclosed, yet it is legal to deviate from them all the same. If we allow (as the current laws would imply) AI to develop and play agreements "that just work", then it looks probable that the robots' ability to "disclose" their chosen agreement (in ways that are already very challenging to define, we are effectively stuck at the system card of two pages) will far exceed the human ability to understand it and to formulate effective defences.
Future Law could attempt to limit this by mandating or constricting allowable systems for a mixed event, an option which has its own critical aspects. I suspect the AI would still run rings round us even if forced to bid a standard system and conventions, but at least our results against our human peers (be it human-human or human-sameAI) would have sense.

[steve2005]

 pilowsky, on 2023-June-12, 17:34, said:

A chess game has an infinite number of possible outcomes and is therefore much harder to comprehend and play - no matter how good the calculating ability of the carbon or silicon based player.

Chess is a finite game, it is just a very large number