[helene_t]

Lorseker asked for ideas for a signaling system for Ben. There are a number of tasks that could be useful and/or interesting to look into:

1) Comparing the efficacy of different systems, for example (high means positive OR even) versus (high means positive in some situations and even in others), vinje versus suit preference etc.
2) Implementing a standard signaling, with a system to decide how good a suit has to be for positive attitude or suit preference etc.
3) When count signals are in principle in effect: when to give honest count, dishonest count or random count. Kit Woolsey wrote a series on BW, comparing the declarer play of WB5 to human experts, and it transpired that WB5 often won because it took advantage of opps giving honest count all the time, even when a good human defender would realize that they were just helping declarer.
4) Letting the AI develop its own Sci-Fi signaling system

I have been thinking about how one might approach 4), but the technique I envision might be useful for the three other tasks also:

Step 1: Using a DD solver on a single hand, identify a situation in which a defender (the "signaller") should play a spot card but has two or more to choose between.
Step 2: From signaller's partner's vantage point, sim a bunch of ways the unseen 26 cards could be distributed between declarer and signaller, consistent with the auction and the play so far (including the condition that signaller must have 2+ spot cards and that DD favours a spot card in this situation).
Step 3: DD-solve each of the hands generated in step 2 and colour the hands red and blue whenever signaller's partner's optimal play depends on which of the hands signaller holds.


Repeat step 1-3 bunch of times to generate a bunch of red/blue features of the hands which signaller might have. Now cluster, into two clusters, signaller's hands based on the red/blue labelings. Those two clusters are now what the signal should distinguish. Finally, characterize those clusters in terms of suit lengths and possession of specific high cards.

Any thoughts?

[pescetom]

Was he really looking for ways to build a better mouse-trap, or just ways for BEN to be more useful on BBO?

I would have expected the priority to be something that a good number of human partners would find familiar and all would find useful to some extent.
So start from "Bridge World Standard Defense" or something similarly well defined and then simplify and clarify a bit.
Ideally there would be also be UK, Italian, Polish etc. systems to choose from, but that requires a lot more work.
I don't think allowing users to define their own signal system to BEN would be very useful, most BBO users are simply not up to it.

But maybe I missed the point.

[helene_t]

pescetom, on 2024-January-20, 11:17, said:

Was he really looking for ways to build a better mouse-trap, or just ways for BEN to be more useful on BBO?

Yeah I get your point. An "optimal" design system will inevidably be too weird and complex for use with human partners.

But I think the same approach could be used to optimize the use of existing signalling systems, such as deciding when to give honest count, and when you actually want to show positive attitude.