GIB wins the World Computer Bridge Championship
What was your overall impression in Bermuda? GIB clearly remains the strongest bridge program in the world. It went 6-1 to win the round robin and won easily in both the semifinals and finals. The margin in the finals (101 IMPs/58 boards) was substantially larger than GIB's winning margin in the 1998 event (63 IMPs/64 boards).
In our opinion, WBridge and Q-Plus Bridge are the next two strongest programs, with Micro Bridge close behind. There is another significant gap between these three programs and the others. Q-Plus didn't fare very well in the round robin, but it did beat GIB in our ten-board match. Q-Plus is the only program to have ever beaten GIB in match play, although an experimental pre-commercial version of GIB fared poorly in an IMP pair event back in 1997.
Are the other programs getting better? Absolutely. Programs that haven't changed much since the previous competition are clearly no longer competitive.
Why did GIB win in Bermuda? While GIB remains the strongest program in terms of card play, the gap is closing. All of the successful programs base their card play decisions on double dummy analysis, as GIB does. GIB's double dummy engine is still easily the fastest, but increasing CPU speeds are making this advantage less significant.
GIB won the championship primarily in the bidding. This is in part because it was playing a more effective bidding system, but mainly because GIB has bidding judgment that the other programs lack.
When faced with a bidding decision (e.g., should I raise 2NT to 3?), GIB deals out about 50 deals consistent with the auction thus far, and makes the decision that works out best on those deals. Here, a fast double dummy engine is essential, and the other programs have not (as yet!) been able to duplicate GIB's technique.
To help with this, we programmed GIB to think on its opponents' time during the auction. Especially in a computer-computer match, where a significant amount of time is spent entering the actions taken by each program, that bought us an extra 30 seconds or so of thinking time per deal.
Why did you play Moscito Byte in Bermuda? Our basic goal is not to win the world computer championships. The gap between GIB and the other programs remains substantial and is probably growing; we've set our sights on winning the human world championships. Moscito Byte may or may not be a better system than Standard American for human play, but we are completely convinced that its precision of definition makes it more attractive for machines to use than other methods. GIB playing Moscito is .25-.5 IMPs/deal better than GIB playing Standard.
As far as playing an artificial system in the computer championship, we believe that the conditions of contest in the world computer championships should mirror those in the human championships. This is especially true as computer bridge matures and the programs become stronger. Moscito Byte conforms to the ACBL's mid-chart restrictions, and therefore strikes us as completely suitable for a computer championship.
Did Moscito Byte give you an unfair advantage in Bermuda? No. All of the other programs entered the meanings of the Moscito bids manually. When it came time to play or defend, they had all of the inferences available that a human pair would have.
Moscito Byte is intended to be an action system, getting into and out of the auction quickly. It is common to play 4-3 fits at the two level, leaving the opponents guessing as to the best action. Our opponents found those guesses difficult, just as a human pair might. GIB's ability to simulate during the auction helped it make those guesses accurately, and it plays 4-3 fits very well.
Will Moscito Byte be part of the commercial GIB? Of course. Right now, we're pretty tired, and we have some vacations planned. Once we're back, we'll be releasing Moscito as part of GIB.
At this point, Moscito Byte is only about half implemented (the most important half, but still only about half). We'll be implementing the rest of it over the course of the year, and will be issuing regular updates to the Moscito database.
Did you have any relay disasters in Bermuda? Yes. On one deal, GIB passed a a 3D relay because its hand type wasn't in the database. Its opponent was kind enough to double, and the simulations then kicked in, with GIB staggering to the correct contract of 4S.
On another hand, GIB relayed to a poor 6C slam, off two aces (the style of denial cue bidding used didn't allow the captain to distinguish between several possible hand types; slam was cold on some, but on the one in question, it was hopeless). The slave bid 6C, and the captain was about to pass.
RHO doubled, and the captain did indeed pass. Unfortunately, GIB believes that the cheapest bid by the captain is always a relay unless it's above 5NT. Since "pass" is below 5NT, it was treated as a relay and the slave bid 6D! Double again, pass again, treated as a relay again, with the slave bidding 6H.
I fully expected the auction to end up at the 8 level, but 6H was a bad contract and the captain finally corrected to 7C. Down two instead of down one, which didn't actually cost that many IMPs.
Are those bugs fixed? The database gap is repaired, and GIB now passes the captain's bid if the current level is 6C or higher, as opposed to the captain's last call.
Fixing the problem with denial cue bidding is harder. Some time over the next year, we expect to be moving to a completely new scheme where the captain's first bid identifies the relay safety level, and a complex encoding scheme is developed automatically that makes optimal use of the available space.
This innovation strikes us as important for two reasons. First, it will be the first time that partnership agreements refer not to the meanings of bids, but instead to an algorithm that determines those meanings. Second, GIB's cue bidding methods will be unplayable by humans, who will not have the computational resources available to compute the encodings.
What about GIBson? Did it play in Bermuda? Unfortunately, no. It wasn't ready. The GIBson work is very difficult in a technical sense, and we're trying to get it working as quickly as possible. We're aiming for April or thereabouts.
What is GIBson, anyway? GIBson is a completely new style of cardplay engine. Instead of basing its conclusions on double dummy analysis, it is a true single dummy analyzer. We expect that GIBson will be as much better than GIB as GIB is better than the programs that preceded it.
GIBson can currently solve 8-card endings in about a second, but the branching factor is almost 100 per additional card, making it prohibitive to use it for full deals.
When are the next computer championships? Probably in August, in Maastricht, the Netherlands. That's the site of the next human world championship. The Bermuda contest that just ended was in some sense a delayed version of the 1999 competition; the Maastricht tournament will be the real 2000 event.
Will anything be different there? Probably. For a start, we expect that each program will be competing on two machines, with each machine only knowing one of the two hands. The informal days when each program was given the entire deal are probably over.
In 2001, we hope that the matches themselves will become automated, with play taking place via a table manager service of some kind, perhaps OKBridge or some other online bridge provider. This will reduce operator judgment and also allow us to play longer matches, making it more likely that the strongest program wins.
What was it like playing under the auspices of the WBF instead of the ACBL? It was great. The computer bridge community is truly an international one, with the four strongest programs coming from the United States, France, Germany and Japan. Although definitely weaker, the next two strongest programs are probably from Great Britain (Blue Chip Bridge) and Canada (Bridge Buff). Blue Chip was second in the bidding contest (behind GIB), and Bridge Buff, although 3-4 in match play, finished fourth in the Bermuda round robin. Given that the top six programs come from six different countries, it seems appropriate that the event have an international flavor.
It was also a real pleasure working with the World Bridge Federation. They are a very forward-looking organization, and we hope that our relationship with them will continue.
You said that your goal was to win the human world championships. When will that happen? In 2003.
We first made that predication in 1998 and so far, at least, we seem to be on schedule.