arigreen

Not playing maximal doubles, GIB's ranges here seem pretty reasonable to me. 3S is competitive (less than invitational values) and 4S is to play (game going values). (Yes, 4S could be an advance sacrifice but most of the time it isn't.) With invitational values GIB simply has to choose which bid to make, so it will probably simulate. Sometimes it will bid 3S and sometimes it will bid 4S. However, it is important that the description for 3S not include hands with up to 17 points -- if it did, then partner-GIB would often raise 3S to 4. It would not be terrible if the description of 4S were changed to "16-22 total points" even though not all hands with 16 total points bid 4S. Still, I think this is a bit of a misdescription since many hands with 16 total points will not bid 4S. In any case, GIB is instructed never to bid again over opener's 4S rebid in this particular auction. As a result, it doesn't matter so much what the precise description of 4S is.

Whatever it is you thought that 3♠ showed, GIB obviously didn't play that. I've updated the rule definition so that the cuebid now promises values.

I've located and fixed the bug. When we release a new GIB I'll include the description of the fix in the release notes.

I like 4♥ with this hand. I've modified the range for splinters to include this sort of hand. We'll let you know when the next version is released.

You are saying that the longest pauses occur when the human is dummy, including the case when there is a singleton to be played? This seems unlikely to me. Has anyone else experienced unusually long pauses when human is dummy?

I agree that opener has at least 16. But it is unsafe for responder to play opener for 21, and that is just what responder was doing before I made the last change. Adding a rule along the lines of "3NT ends all auctions unless..." might be interesting.

Thanks for the report. This shouldn't happen again once we release the next version of GIB.

I modified GIB to use HCP in place of total points and ran a 100-board match. As I had expected, the version with total points destroyed the version that used only HCP.

It's not easy to drop total points altogether: - Many rules rely on total points, so removing these rules will leave many gaps in the rule database. - Even without a fit, total points can be useful for estimating playing strength of a hand. It might be interesting, however, to convert all references to "total points" to HCP. I suspect that if we did this, however, GIB's judgement would be worse.

One problem is that 6-9 total points is not really a playable range for 3D. 3D must be either NF or F. If NF, it should be limited to 7 total points. If F, it should be 8+ total points and forcing to 3N. With that said, 3N shows 25 combined HCP. Partner has only promised 4HCP, so GIB thinks that its 3N bid promises 21 HCP. A common pattern that we see is confusion between HCP and TP. Perhaps a better way of looking at this would be: - Responder could have up to 9 HCP for his 3♦ bid. - With 25 combined HCP, opener will want to bid 3N - Hence, responder should not play opener for more than 16HCP.

Here is the hand in question: http://tinyurl.com/2ffj4s7 I very much like GIB's bidding here. However, the explanation of 3♠ is incorrect. It should read "4+ D, 10-13 TP, values in spades" GIB does not always bid its longest suit when responding to a takeout double.

I fully agree with you that improving the bridge logic behind GIB's decisions is paramount in improving the quality of the robots. The problem isn't that the bridge logic itself is too difficult, but rather that there are thousands of auctions that each need to be implemented separately. Even the definition of "biddable suit" differs based on the context of the auction. We try to focus our time on improving the more frequent auctions.

Hi Stephen, Thanks for your system suggestion -- it looks great to me. I agree that the continuations in this auction are rather horrible. My point was that GIB can't simply choose the bid that best matches the descriptions, because the descriptions themselves are poor. The best thing we can do is to add rules for as many specific auctions as possible. I'll start by adding Stephen's suggestions over 2♣-P-2♦-P-3♣.

Good point. In order to persuade GIB to pursue diamonds we need to further restrain opener's 3♦ rebid, to show a more distributional hand. In several of GIB's simulations, partner was 54 with mediocre diamonds and 3N was actually better than 5♦. Another fix would be to make 4♦ forcing so that GIB could bid 4♦ with this hand and land in 4♥ if partner is 64.

GIB simulated among 3S,3N,4D,4H and found 3S to be the long term winner. When opener was x64x, all bids except 4D led to a contract of 4H, usually making. Opener passed a 4D rebid. When opener was 1543, opener bid 3N over 3S but 4H over 3N. 3N was usually the better contract. The simulations here performed poorly for several reasons: - GIB's understanding of the types of hands that bid 3D was poor. For example, GIB chose 3D with K.AQxxx.Axxx.QTx where most people would bid 3N - GIB's predictions for how the auctions would continue was often incorrect. For example, correcting a 3N rebid to 4H but correcting a 3S rebid to 3N. Furthermore, in GIB's simulations partner never raised 3S to 4S, even when 3541. Thus the 3S bid was not penalized by the possibility of partner raising to 4. As usual, the best fix for this sort of anomaly is improve the rules. In particular, the rule that dictates that GIB should bid 4♥ after P-P-1♥-P-2N-P-3♦-P-3N-P with 1543 should be removed. When I deleted this rule and reran the simulation, GIB bid 3N instead of 3♠ with the hand.