[calm01]

http://tinyurl.com/3kpc2r6

The same hand - this time a query on GIB leads.

Surely the best lead from this club holding is the 10 not the 6 both at trumps and NT.

It cannot be difficult to program standard leads based on a lead table having selected the best suit to lead - so maybe the standard lead tables are amiss or just missing.

This naturally raises the suit selected by GIB as the suit to lead:

- all too frequently GIB leads its shortest suit at NT even when it holds a likely majority of the likely defensive points.
- GIB rarely leads partners bid suit even when it has no decent lead of its own.

What rules, if any does GIB have for opening leads? If just a random selection of hands is generated to determine the best lead, then this would explain the poor choices GIB repeatedly makes.

Even if this random selection of hands must be the starting point, surely some allowance for long suit leads at NT might get a look in, using standard lead tables would not go amiss, and leading partners suit might be given some consideration.

Still, when improvements ensure GIB leads at last approach palooka status, the fun of discerning GIB's unfolding holdings from it opening leads would disappear.

Calm01

[barmar]

We know GIB's opening leads are poor.

We recently ran an experiment, having GIB play the boards from a round of this year's computer bridge finals, and compared its results against those of the finalists. We basically came to the same conclusion about its leads and not returning partner's suit.

This is obviously something we'd like to improve, but it's much harder than fixing bidding problems. GIB's play is mostly based on simulations, not rules. It does have the notion of standard leads, so this one may just be an omission there.

[cloa513]

Can you access GIB's simulations or does it delete them all as quickly as it creates them?

For that lead to be right, East would need to have some incredible hand such as Axx and say two other aces or winners and N-S 2-2 in clubs.

[barmar]

If South has K8xx(x), it doesn't matter which card East plays. GIB declarer play and defense are based on double dummy analysis, so East assumes declarer will take the correct finesse.

South is known to have either 4 or 5 diamonds, and 75% of 4-card holdings include the 8 (he pretty much has to have the K, so only the missing spot card is in question). So most of the time, it's not giving up anything by playing the J.

[Bbradley62]

 barmar, on 2011-November-07, 01:48, said:

South is known to have either 4 or 5 diamonds, and 75% of 4-card holdings include the 8 (he pretty much has to have the K, so only the missing spot card is in question). So most of the time, it's not giving up anything by playing the J.

So, if Play A and Play B yield equal results 90% of the time, and Play A yields a better result the other 10% of the time, GIB doesn't consider Play A to be better than Play B? Or are you saying that the number of simulations done may not identify the difference?

[barmar]

We ran a couple hundred tests. In total points mode, it played A 60% of the time, low 35%, and J 5%. At IMPs it was 50%, 46%, and 4%, pretty similar. In matchpoint mode it always played low.

When I looked at the hands where it played J, the simulated South hand didn't match the bidding (it wan't an opening hand, or it wasn't balanced -- I saw one hand with a singleton ♦). GIB's design currently allows some hands like this into the simulations; it deals a large number of hands, sorts them by how closely they match what it knows so far from the bidding and play, and then uses the top N. Most of these should match the bidding, but there will sometimes be a few outliers. This is needed to handle cases where GIB doesn't have any bidding rules that match the auction -- it would end up with no suitable hands in the simulations.

In the case of this hand, when South has his bids it doesn't often matter much what East does, as 3NT is cold. So those few mismatches end up as the tie-breakers. But notice that it doesn't happen in matchpoint mode, when every trick counts -- in this case, we don't have a close race among the matching hands, so there aren't enough outliers to push things the other way.

calm01, were using basic or advanced bots when this happened? This type of thing is more likely to happen with basic bots, since they start with fewer hands so there will be more mismatches when it takes the top N. In one of my tests with GIB in a very fast mode, it simulated 24 hands; 20 of them had J equivalent to a small card, 1 of them had J worse than small, and 3 of them had saved an overtrick by playing the J. South's hand didn't match the bidding on these 3 hands.

[Bbradley62]

 barmar, on 2011-November-07, 00:30, said:

It does have the notion of standard leads, so this one may just be an omission there.

While you're looking at standard leads, please remember low from three-small against NT.