It is difficult to train BEN doing the right things - follow an interesting board.

[ThorvaldAagaard]

From yesterdays match, where BEN lost heavily South has this hand

All vulnerable, IMPS, and partner open 1C.

BEN has no doubt

With a 4333 hand, mostly quacks and a bad suit, I think many will bid 3N. showing a balanced hand with 13-16 HCP. BBA does not allow a 4-card unless clubs, and at least a 4-card diamonds should be allowed, but also bad 4-card majors with 4333.

Now that could be OK, but I know BEN might meet opponents, that bid 3N on this hand, so BEN has a problem, when simulating samples, as a hand like this will be removed as it would NEVER bid 3N.

So we need to find a way how we can compensate for this kind of evaluation.

OK, the bidding continues:

And we have to play 4H on this layout:

West lead Ace of spades, and continue spades. What now?

Again we generate some samples for EW to find the best play

This is all hands that would bid as shown above, but on many of these I don't think X is the right bid from West. Typical having 4333 should not double.

As an example

I am not sure, why BBA now suggest 3N on a very bad stopper, and 5D could be right,

But I do not think X is right, and even worse is A9x.AKx.KJ8x.T8x

Have you made up your mind how to play it?

The only risk is a spade ruff, so it could be right to overtake Jack of spades and discard dummys spade and lead trump 9.

Perhaps this is the risking going down if trump are 4-1, but I hope the simulation would find that for us.

BUT: BBA seems unable to double on a 5-card spade even with 5-2-4-2, yes it could be a strong hand with 5 spades, but here NS has a majority of points, so any sample will only give West 4 spades, so there is no risk of a ruff.

Trying to move around the cards I found:

1: Giving West 5 spades, and BBA will bid 1S 2: Giving West 5 spades and 5 diamonds BBA will use Michaels (Reported as an error in another issue) 3: Giving West 4 spades and 5 diamonds BBA will use Unusual 1N (Also reported as an error as West is unpassed, so 1N is natural)

So we can only find samples with 4-3 or 4-4 in the unbid suits.

So no matter how many samples we generate we will never find the actual hand as West Actually held this hand

I have added to BEN, that we can have different models for the opponents as we would normally trust partner is following the same rules as we do, but to get this to work we need a model, where the bidding can go as bid by other players.

For this board it will probably be solved by fixing the two issues with 1N and 2C, that both should be natural, so West might have 5-5 in the unbid suits, so the DD will notice the risk of a ruff.

But any ideas how we can create examples for training opponent neural network is welcome.

Currently my idea is that when sampling we should not focus on "Will bid" but much more on "Might bid".

Any ideas how this can be solved using BBA?

BEN's opponent just bid 3N and had no problem making it.

[EdwardPiwowar]

Maybe we need to generate some samples with other bots: BBA-GIB and BBA-WB5.

This will also reveal bugs that have not been detected yet.

[ThorvaldAagaard]

Yes, for the opponents model we should include samples with different bidding on the same hands. The problem is to create those deals, ie finding the deals, where the bidding is different.

Having a separate model for opponents is probably the right step