This article was first
published in
vol.
8, nr. 2, March 1997
| GIB:
Sensational breakthrough in bridge software (1) |
by Onno Eskes
Before saying more about bridge software, I'll first give you a tough
quiz to put your teeth into. If you pass this test, you measure up to the
very finest players. If you run into a few difficulties, you may find some
consolation in the fact that all hitherto existing bridge computer programs
get hopelessly stuck on these problems, too.
| 1. S/none |
7
(dummy)
K 5
K 9 8 6 5 3
K 10 6 4 |
|
| - |
 |
Q 9 5 3
J 10 9 6
Q
A Q 7 3 |
| e |
- |
|
|
---- |
The opponents bid 1NT(15-17) - 3NT. Partner,
West, leads 8 (attitude),
small, small to the jack. Declarer leads J
and plays small in dummy (partner 7,
high-low even). How do you, East, continue the defence? |
|
|
| 2. W/NS |
A 3
7 4
Q 8 7 4
A K Q 6 3 |
|
| - |
 |
- |
| e |
Q 2
A K Q 10 8 3
A K 10 3 2
- |
|
|
--- |
After
an opening bid of 2
on your left (at least 4-4 in the majors, and weak), you declare 7 .
The lead is J. Right-hand
opponent has doubled 4
on the way. What is your plan? |
|
|
| 3. E/none |
J 10 2
A 10 7 4 3
A Q 8
9 2 |
|
| - |
 |
- |
| e |
A K 9
7 6 4 3
6
10
A K 7 3 |
|
|
--- |
Against 7
the K is led. You
take A and ruff a
heart. The queen of spades does not drop under A,
but both opponents follow suit. How do you continue? |
|
|
| 4. S/both |
4 2
A Q 7 5
K J
Q 7 6 3 2 |
|
| - |
 |
- |
| e |
A K J
8 6 3
K 6 4 2
6 5
J |
|
|
--- |
And how do you tackle this 4 -contract
after a trump lead? |
|
|
Exactly a year ago, World Chess Champion
Gary Kasparov was defeated by a computer program named Deep Blue. The draughts-world
had already long been conquered by programs with simple names like Truus
and Flits. Bridgeplayers were eager to conclude from this how simple, or
even inferior, games like chess and draughts must be. Of course, they did
so preferably in the presence of fanatic draughts or chessplayers. Beaten
by a computer? Such a humiliation would never befall a bridgeplayer, would
it?
October 1996, Sunday night. I get home after a competition weekend in
the Dutch First Division (second National level). I am still feeling miserable.
My weekend was spoiled by this hand:
| 2. W/NS |
A 3
7 4
Q 8 7 4
A K Q 6 3 |
|
| - |
 |
- |
| e |
Q 2
A K Q 10 8 3
A K 10 3 2
- |
|
|
--- |
West opens 2 ,
showing a weak hand with both major suits. Unpleasant, but on the other
hand, it becomes a lot easier for us to stay out of a heart contract. We
confidently reach 7 .
West leads the J.
I greet dummy with approval. "Well bid," I remember thinking. I count five
trumps, twice Ace-King-Queen, A
and a ruff in dummy. What can go wrong?
Trumps four-zero. If left-hand opponent has them all, I will go down.
If oppo- nent on the right has them, I can finesse against his jack. But
then I cannot ruff a heart anymore. Well, then they'd better not be four-zero.
I take A and lead
a diamond to the ace. West discards a spade. I curse under my breath and
start thinking again. Are there any chances left? In clubs, maybe? If those
are four-four, I can discard a loser on the fifth club. |
|
|
But I have only one entry left. I should have started by ruffing
a club."One down," I concede, "how are the clubs divided?"
"Four-four," is the painful reply. Fortunately for us, exactly the
same thing happened at the other table!
The next day I am still disgusted with the hand. It is a nice problem,
however. I decide to present the hand to GIB, a new bridge program that
I received a week ago from the USA. I enter the hands and the auction.
I also enter the explanation of the bids (West at least 4-4 in the majors,
less than opening strength) and the opening lead. Then the computer starts
to bubble. After 30 seconds it produces
A.
I tell it which card East plays to the first trick. Again 30 seconds of
thinking. Ace of clubs! I let East and West follow small. The computer
discards
Q. Another
30 seconds. Small club, ruffed in hand! Forlorn, I watch the computer finish
the rest of the play in immaculate fashion. Ace of diamonds (discovering
the bad trump split), diamond to the queen,
K-Q
and a good club. East ruffs, South overruffs and he can now ruff his last
heart in dummy.
Beaten by the computer! The humiliation is complete when the machine
subtly announces that it just scored plus 2140. Humiliation is quickly
replaced by a feeling of excitement. Playing bridge seemed to be such an
insurmountable task for artificial intelligence. Many attempts have
been made to develop bridge software of some capability. But until now
not a single program has proven capable of offering more than mediocre
opposition to even the most inexperienced club player. And yet, GIB had
just played a hand better than two Dutch First Division declarers. This
has to be a revolutionary program. If this was not a coincidence, it would
mean a major breakthrough in the area of computer bridge.
I try another deal:
The final contract is 7
.
West leads
K and I
tell the computer this promises the
Q.
At the table
A was
taken and a heart ruffed, followed by
A,
small, small, small. I enter that as well. Then I set GIB to work. There
appear to be three chances: diamond finesse, two club ruffs in dummy (
Q
in the same hand with four clubs), or a squeeze. GIB requires little time.
After eight seconds it plays
A-K.
Then it thinks for five more seconds and plays
K.
The rest goes in tempo: club ruff, heart ruff, and all remaining trumps:
On the last trump West pitches a diamond, dummy can now spare the
10,
and East has to let go a diamond. GIB plays a diamond to the ace. Plus
1510, it announces cheerfully again. Impressive!
According to the report of the 1967 European Championships, former World
Champion Hans Kreijns needed three seconds for this double squeeze; GIB
fifteen. Kreijns doesn't have to worry,… yet.
What is GIB?
GIB stands for Goren-In-a-Box, and is named after the late Charles Goren,
one of the greatest bridge players of all time. GIB was developed by Matthew
Ginsberg, an American professor who specializes in artificial intelligence.
GIB can not bid yet, nor interpret an auction, but it can declare and defend
like a tiger.
Many IMP readers are familiar with the training program Bridge
Master , a set of 180 hands ranging from simple (level 1) to nearly
unsolvable (level 5). Until now, the best score was achieved by the American
program Bridge Baron, which brought 33 of the 180 problems (18.3%) to an
end successfully. GIB pulverized that score with a mighty 116 correct
(64.4%), including 24 of the 36 problems at level five (66.7%).
Click HERE
for details about the experiment.
In upcoming issues of Bridge Magazine IMP, many sides of GIB will be
highlighted, such as:
-
The Delphic oracle: the inner workings of GIB
-
Why is Zia afraid of Matt Ginsberg?
-
The adventures of GIB at a North American bridgeclub night
-
The future of GIB, and GIB's vision on computer bidding
-
Top experts outplayed by GIB
In order to stack up GIB against the world's finest cardplayers, I used
the following hand from the latest Macallan tournament in London. At all
eight tables the contract was 4
,
against which West led the
J.
| 4. S/both |
4 2
A Q 7 5
K J
Q 7 6 3 2 |
|
Q 10 9 5
J 10
A Q 10 8 7
5 4 |
 |
7----
9 8 3
K J 7 6
J 10 6 4 |
| e |
A K J
8 6 3
K 6 4 2
6 5
J |
|
|
|
|
|
|
Five out of eight declarers played the hand in exemplary fashion. GIB joins
this expert group (Robson, Chagas, Multon, Blakset and Bompis) by taking
Q
and leading a spade to ace. Heart to the ace and spade from the table.
East could ruff in front of declarer, but he'd be ruffing a losing spade.
Therefore, he discards a diamond. GIB wins the king, ruffs a spade and
is overruffed. East plays a diamond to the ace and and West exits with
a club. GIB ruffs, ruffs another spade and scores
K
for his tenth trick. Three world-class declarers went down.
Back to the weekend of the divisional competition. There was another
difficult hand. I modified it slightly to prevent 'noise'.
READER PARTICIPATION
Do you have interesting play problems that you'd like to present to
GIB? Send the full hand, with all details and the result at the table to
Bridge Magazine IMP or directly
by e-mail to: eskeso@xs4all.nl
To
IMP Bridge Index