[lorserker]

Thank you for the suggestions!

I did some improvements which I plan to push to Ben & Friends sometime next week.

Fixed the bug when it was running from double hoping to escape double.

Toned down the crazy bidding by making it fear getting doubled more. This reduces boldness and insanity a bit.

Grafted gib onto ben. Gib has priorities on its bidding rules. When a very high priority rule applies (e.g when answering blackwood) then we just obey gib's suggestion. This should remove annoyance of wrong blackwood answer, passing forcing bids and similar.

Next I plan to improve the card play, maybe some signalling etc. Let's see

[smerriman]

lorserker, on 2024-January-19, 12:10, said:

Grafted gib onto ben. Gib has priorities on its bidding rules. When a very high priority rule applies (e.g when answering blackwood) then we just obey gib's suggestion. This should remove annoyance of wrong blackwood answer, passing forcing bids and similar.

That sounds much more promising an approach. As well as the priority system, GIB has flags for when it's allowed to deviate from the book bid, so that's another potential way of detecting forcing situations. (Some of GIBs highest priority rules are leaping to slam, when slower routes would have been more advisable; simulations don't help it understand the information gathering advantages, but maybe it would help Ben. Or maybe this hybrid version should be called Gen.)

The good thing is that you can empirically test the difference; generate a large set of hands, let GIB handle the cardplay at both tables, but Ben the bidding at one. If you can average a positive result, you're definitely making forward progress.

[Joe_Old]

Perplexing auction.

With AKxx AJx Qxx 98x Ben opens 1C. With x Kxx xx AKQJTxx Responder bids 2C. Then 2NT from Ben and 3H from Responder (4NT is identified as quantitative). Ben bids 3NT and 4NT is now identified as Ace asking in Clubs (bid despite a useless xx in D). Ben bids 5C (0 or 3, etc).

Why? Is Ben afraid to go past 5C? Some other esoteric reasoning?

[lorserker]

Joe_Old, on 2024-January-20, 07:49, said:

Perplexing auction.

With AKxx AJx Qxx 98x Ben opens 1C. With x Kxx xx AKQJTxx Responder bids 2C. Then 2NT from Ben and 3H from Responder (4NT is identified as quantitative). Ben bids 3NT and 4NT is now identified as Ace asking in Clubs (bid despite a useless xx in D). Ben bids 5C (0 or 3, etc).

Why? Is Ben afraid to go past 5C? Some other esoteric reasoning?

Just clumsy bidding, no deeper reason. We'll try to improve this.

[Joe_Old]

lorserker, on 2024-January-20, 08:17, said:

Just clumsy bidding, no deeper reason. We'll try to improve this.

Is it? From the bidding Ben could easily imply lack of a D honor in Responder's hand. But, an intelligent Responder would have Diamond shortness, and still want to be in a Club slam opposite Ben's actual hand. Do we really know the extent of Ben's understanding on that issue?

[lorserker]

Joe_Old, on 2024-January-20, 08:24, said:

Is it? From the bidding Ben could easily imply lack of a D honor in Responder's hand. But, an intelligent Responder would have Diamond shortness, and still want to be in a Club slam opposite Ben's actual hand. Do we really know the extent of Ben's understanding on that issue?

If you post the link to the hand, I could run it through Ben to see what exactly it had in "mind" and what bids it considered.

[Joe_Old]

lorserker, on 2024-January-20, 11:02, said:

If you post the link to the hand, I could run it through Ben to see what exactly it had in "mind" and what bids it considered.

Sorry, I haven't the faintest idea how. The hand was from today's daylong (1/20/24), assuming that you have access to BBO's files.

[guinnypoo]

http://tinyurl.com/ytjcum8p
Club on trick 6 seems like an obviously bad play, his partner has preempted diamonds and I have the rest of the club tricks showing in dummy

[rhm]

http://tinyurl.com/yqdhqkkn

On 4NT Ben bid 5C which according to the explanation showed 0 or 3 key cards
I would expect that a robot trained on millions of hands is capable of counting its key cards

What's the point of such an undertaking if Ben can not count?

Please explain

[lorserker]

rhm, on 2024-January-21, 07:28, said:

http://tinyurl.com/yqdhqkkn

On 4NT Ben bid 5C which according to the explanation showed 0 or 3 key cards
I would expect that a robot trained on millions of hands is capable of counting its key cards

What's the point of such an undertaking if Ben can not count?

Please explain

sorry about this. i don't have anything smart to say except "hang on, we are fixing it in the next version"
thanks for having given it a try.

[acfromatl]

Can Ben learn to adjust his bidding to counter opps who play systems other than SAYC or 2/1 (Acol or Precision for example)

[etha]

yes. You can make a neural net versus specific opponents even, that is how nukkai managed to beat a number of expert human declarers. They trained their model specifically for the auction 1NT 2NT 3NT played against wbridge. No surprise then that their bot was pretty good at this pretty restricted game. If you want to play precision with ben you have to train it to do so. Ben is open source and available at hub.com/lorserker/ben. Here you can find a precision net for example that I trained.
However it plays Meckwell lite so it might not be what you expect.

[thorvald]

guinnypoo, on 2024-January-21, 03:50, said:

http://tinyurl.com/ytjcum8p
Club on trick 6 seems like an obviously bad play, his partner has preempted diamonds and I have the rest of the club tricks showing in dummy

This is an interesting aspect of robot play.

As humans we "know" that partner don't have a trump, as that would be very bad declarer play, but as a robot it simulates the deal, and finds that partner might have a trump, so that is the only chance for setting declarer.

[hrothgar]

Speaking strictly for myself, I'd very much prefer seeing BBO invest resources in providing API's and the like to integrate third party solutions rather than trying to develop their own.]

There's a bunch of really good third party bridge programs - think WinBridge or Jack or ....

You'll get the most bang for the buck by licensing these (or allowing folks to rent these) rather than yet another stand alone project

[helene_t]

hrothgar, on 2024-January-22, 07:27, said:

I'd very much prefer seeing BBO invest resources in providing API's and the like to integrate third party solutions rather than trying to develop their own.

Intobridge does this.

As for Jack, they have their own playing site already so they would need strong incentives to integrate it in other sites.

[lorserker]

Hi, from tomorrow the "Ben & Friends Daily" tournament will use an updated Ben in which we tried to fix some of the biggest issues which were reported in the past two weeks:
- made bidding more reliable
- tamed crazy bidding in competition
- opening leads are a bit more normal (human)

Many thanks for playing and reporting issues.

[pescetom]

hrothgar, on 2024-January-22, 07:27, said:

Speaking strictly for myself, I'd very much prefer seeing BBO invest resources in providing API's and the like to integrate third party solutions rather than trying to develop their own.]

helene_t, on 2024-January-22, 07:31, said:

Intobridge does this.

I agree and would like to see BBO and it's competitors working together to define generic standard robot interfaces that require little or preferably no integration specific to each platform. And with the involvement of WBF, although I can imagine what hrothgar has to say about that

[pilowsky]

Is this (4S bid) normal behaviour for opener in a lebensohl sequence?
http://tinyurl.com/ylsngvus

[lorserker]

pilowsky, on 2024-January-25, 05:06, said:

Is this (4S bid) normal behaviour for opener in a lebensohl sequence?
http://tinyurl.com/ylsngvus

4S is a bit creative, but I don't consider it a blunder.
5C does have equity, and you may give some info on the way.
would you prefer if N passed 4C? or if N bid 5C directly?

[etha]

Any luck on 1430 fixing ? I am currently making the info file for a 420k net using only hands with 1430 in them. Previous attempts suggest this won't help but we'll see.

Have you tried just normal blackwood does the lack of a 5th ace make it learn correctly ?

We might have to end up just hard coding the RKCB responses, but it does seem weird that it can't learn it.


And indeed wrong answers still common to RKCB, really odd so more data may not be the answer, moving on to leaping michaels, and then the impossible to get right doubles.