THE 9th ANNUAL WORLD COMPUTER-BRIDGE CHAMPIONSHIP

Estoril, Portugal

October 25-30, 2005

 

Alvin Levy[1]

 

The 9th annual World Computer-Bridge Championship, sponsored by the American Contract Bridge League (ACBL) and the World Bridge Federation (WBF), was held in Estoril, Portugal alongside the WBF’s World Team Championships.  Six of the best bridge-playing software programs, or robots, competed for the 2005 computer-bridge world champion title.

 

History

The ACBL inaugurated an annual computer-bridge championship in 1997 as a way of encouraging computer-bridge software developers to accelerate their robots’ development to expert class, with the hopeful result of more software development for educational purposes and generally promote bridge.  The WBF supports this event and it has become a joint venture.  This year’s championship was dedicated to the memory of Alan Truscott, bridge historian, New York Times Bridge editor, and world-class player, who encouraged me to organize this championship ten years ago.

 

Since its inception in 1997, the championship has been held every year at an important human international bridge event.  The first eight championships were held three times at ACBL North American Bridge Championships, four times at WBF World Championships and once at the European Bridge League’s first European Open Championship.  For a complete history and details of previous championships go to ny-bridge.com/allevy/computerbridge or computerbridge.com.

 

Technical remarks

A bridge “table” consists of a central server, or Table Manager (TM), and four connecting computers that “seat” the robots.  The TM distributes the four hands of each board to the robots.  Play proceeds automatically with the TM receiving and passing information to the robots and recording the play.  This year P4 3 GHz/512 MB PCs were used, running under Windows XP.  The speed of play was set at two minutes per pair per deal, approximately half that of human play.

 

Without getting into the details, most of the robots are programmed with a combination of knowledge-based AI (sets of rules) and search-based AI (simulations).

 

Preparation

This is very much a group effort between the ACBL, the WBF, robot developers and myself as organizer.  An administrative committee sets the Conditions of Contest (CoC) with input from the competitors and bridge experts.

 

It is important that the developers understand the opponent robots’ methods well in advance of the competition so that they can prepare and program defenses.  In computer-bridge play the developers need time to enter detailed information about their opponents methods and program bidding defenses.  Contestants are required to submit a Convention Card (CC) with their bidding and carding understandings one month before the competition.  Since this information is stored in the robots’ memories before play begins, few explanations (alerts) are necessary during play.  In the instances where explanations are required by the CoC, the pertinent information is manually input into memory and play continues.


Play format

The competition is in the form of team matches, with a team’s four identical robots seated N and S at one table and E and W at the other table.  A match, or part of a match such as a 16-board segment, is played sequentially, first at one table and then at the other table.

 

This six-day event started with a 40-board round-robin, played in halves, with the top two robots advancing to a 64-board knockout (KO) final with carryover.  The round-robin was scored on an international 30-VP scale, with the winner receiving a maximum of 25 VPs for a 69 or more IMP victory.  To earn a carryover in the final KO match, a robot must both win its head-to-head round-robin match against its KO opponent and end higher in the overall standings.  The carryover is the lesser of the two VP differences.

 

The competitors

With six teams entered, the smallest field to date, this was still one of the strongest fields of robots ever competing.  Four-time defending champion, Jack (The Netherlands), was the favorite, with Bridge Baron (USA), the 1997 champion and 2003 and 2004 finalist and WBridge5 (France), the 1999 and 2002 finalist, losing to Jack in 2002 by 1 IMP, expected to threaten Jack’s recent domination.  Other competitors were Q-Plus Bridge (Germany), three-time finalist, losing by 2 IMPs to Meadowlark (USA) in 2000, Micro Bridge (Japan), 2001 runner-up, and Blue Chip Bridge (U.K.).

 

The competition

The round-robin ended with Jack topping all competitors with 111 VPs.  WBridge5, 20 VPs behind at 91VPs, would play Jack for he title.  Q-Plus Bridge made a strong finish but fell short with 83 VPs.  Micro Bridge was fourth with 70 VPs, followed by Bridge Baron with 48 VPs and Blue Chip Bridge with 37 VPs.  To show the tightness of the competition, or possibly the randomness of short matches, even though Blue Chip Bridge finished last, it won two matches, 96-93 against last year’s finalist, Bridge Baron, and 116-110 against this year’s eventual champion, WBridge5.

 

The robots, their developers and the final round-robin standing are shown in Table 1.  Table 2 shows the IMP/VP results of each round-robin match and the round that each match was played.  Gérard Joyez operated Q-Plus Bridge.

 

Table 1.  Robots, developers and round-robin standing

Robot

Developers

Country

VPs

Jack

Hans Kuijf, Wim Heemskerk and Martin Pattenier

The Netherlands

111

WBridge5

Yves Costel

France

91

Q-Plus Bridge

Hans Leber

Germany

83

Micro Bridge

Tomio and Yumiko Uchida

Japan

70

Bridge Baron

Stephen Smith, George Yanakiev, Jason Rosenfeld and Tom Throop

USA

48

Blue Chip Bridge

Ian Trackman and Mike Whittaker

UK

37

 


Table 2.  Round-robin match results

 

Jack

Bridge Baron

WBridge5

Micro Bridge

Q-Plus Bridge

Blue Chip Bridge

Position

VPs

Jack    IMPs

 

VPs

 

117-53

 

24-6

50-57

 

14-16

102-48

 

23-7

116-40

 

25-5

163-37

 

25-0

1

 

111

Bridge Baron

53-117

 

6-24

 

69-130

 

6-24

67-88

 

12-18

56-98

 

9-21

93-96

 

15-15

5

 

48

WBridge5

57-50

 

16-14

130-69

 

24-6

 

89-73

 

20-10

89-73

 

17-13

110-116

 

14-16

2

 

91

Micro Bridge

148-102

 

7-23

88-67

 

18-12

88-127

 

10-20

 

80-112

 

11-198

104-40

1

24-1

4

 

70

Q-Plus Bridge

40-116

 

5-25

98-56

 

21-9

73-89

 

13-17

112-80

 

19-11

 

176-23

 

25-0

3

 

83

Blue Chip Bridge

37-163

 

0-25

96-93

 

15-15

116-110

 

16-14

40-104

 

6-24

41-176

 

0-25

 

6

 

37

 

Some of the boards played in the second half of the round-robin were the played earlier in the Bermuda Bowl/Venice Cup/Senior Bowl world championship round-robin.  Four boards from the fifth human round-robin follow.  As the Bermuda Bowl played 20-board matches and the Venice Cup and Senior Bowl played 16-board matches, board 18 was only played in the Bermuda Bowl.

 

Board 18

Dlr: East

Vul: N-S

North

ª Q 9 5 3
© A 6 3
¨ A 3 2
§ A Q 4

 

 

 

West

ª ---

© 10 9 5 4

¨ K J 10 9

§ K 10 9 5 3

 

East

ª K 10 8 7
© K 7 2
¨ 6 5 4
§ J 7 2

 

South

ª A J 6 4 2
© Q J 8

¨ Q 8 7

§ 8 6

 

 

 

 

The robots played in 4ª at all six tables.  In the Bermuda Bowl, 4ª was played 14 times, 3NT by North 7 times, and 3ª once.

 

The play at two tables is mentioned here.  If the defense slips, 4ª can be made on a non-diamond lead by East.  At one table Q-Plus Bridge won a heart lead by East with the ©Q, took a club finesse, a spade to the ªJ, §A, club ruff, ©A and a heart to East’s King, leaving this position:

 

 

North

ª Q 9 5
© ---
¨ A 6 3
§ ---

 

 

 

West

ª ---

© ---

¨ K J 10 9

§ K 10

 

East

ª K 10 8
© ---
¨ 6 5 4
§ ---

 

South

ª A 6 4
© ---

¨ Q 8 7

§ ---

 

 

 

 

East exited with a diamond.  4ª is now cold if East has the ¨K and cannot be made, with best defense, if West has the ¨K.  If East has the ¨K, declarer’s winning line would be, ¨Q, followed by the ¨A, exiting with a diamond, then ducking a spade to East.  However, Q-Plus Bridge played West for the ¨K and ducked to the ¨9.  West went wrong by returning the ¨10 rather than a club.  Declarer ducked to the ¨Q, then won the ¨A and ducked a spade, endplaying East.

 

In another match, on a club opening lead, WBridge5 found itself in the same end position as shown above but with North on lead.  WBridge5, not allowing for an opponent error, played best by playing the ¨A and a diamond toward the ¨Q, going down on the lie of the cards, but only one-down when it endplayed East in trumps.

 

All six declarers played in 4ª, with Q-Plus Bridge making, as shown, and four others going one-down.

 

Interestingly, of the 15 times that the contract was played in spades in the Bermuda Bowl, declarer took only eight tricks 7 times, going –200 in 4ª and –100 in 3ª.  On this board the robots were less imaginative than the experts, never playing in 3NT, and declared better than the experts, only taking less than nine tricks once in six contracts.

 

Board 5 from the same set follows:

 

Board 5

Dlr: North

Vul: N-S

North

ª K 8 5 4
© J 8 7
¨ J 10 8
§ K 6 2

 

 

 

West

ª Q 10 7 6

© K Q 4

¨ 9 7 3

§ 9 8 5

 

East

ª J
© 10 9 5 3 2
¨ A K
§ A Q 10 7 4

 

 

South

ª A 9 3 2
© A 6

¨ Q 6 5 4 2

§ J 3

 

 

 

 

West

North

East

South

 

WBridge5

Micro Bridge

WBridge5

Micro Bridge

 

---

---

1©

Pass

 

2©

Pass

3§

Pass

 

3©

Pass

4©

All Pass

 

Micro Bridge’s §J lead resulted in +450 for WBridge5.

 

A simulation shows that, after 1© – 2© with the opponents silent, game is better than 50% and if partner doesn’t accept, 3© will almost always make (game will still make over 25% of the time), and if partner accepts, game will make most of the time.  The risk of a game-try is nil as going down in 3© is offset by South possibly entering the auction, if East passes 2©, and going plus in a spade contract.

 

WBridge5 bid to 4© as shown.  This appears to be a faulty simulation.  On a similar auction, Jack, against Bridge Baron, reasonable (and correctly) passed 3©.  2© was played at the other four tables.  Micro Bridge and Blue Chip Bridge, playing four-card majors, bid 1©-1ª/2§-2©/Pass (with West holding 4 or more spades and only three hearts, game is less than 50% with the risk of going overboard greater than 25%, therefore pass is the percentage bid), as did Bridge Baron, playing five-card majors (with West holding 2 or three hearts pass is the percentage bid).  Q-Plus Bridge, playing 5-card majors, bid 1©-2©.  Given their methods, four judged well with Micro Bridge, Blue Chip Bridge and Bridge Baron passing, and Jack stopping in 3©), Q-Plus Bridge was too conservative and WBridge5 was too aggressive.

 

It is interesting to compare the robots’ play to that of the experts in the world championship events.  In the Bermuda Bowl, 4© was bid twelve times and made ten times (miss-guessing the ©J at two tables, I presume), 3© was played nine times, and 3§ was played once (too aggressive, assuming a variety of systems).  In the Venice Cup, 4© was bid and made nine times, 3© was played three times, 2© was played five times, and there were five other inferior contracts, 2NT, 3NT, 2§, 2ª by N (-2), and 3§ (average aggressiveness with too many accidents).  In the Senior Bowl, 4© was bid and made eight times, 3© was played five times, and 2© was played nine times (average aggressiveness, assuming a variety of systems).  In our championship, 4© was bid and made once, 3© was played once, and 2© was played four times.

 

One trademark of the Bermuda Bowl experts is reaching slim games even when non-vulnerable.  Possibly this is an emotional flaw, never wanting to miss a makeable game, or, based on experience, allowing for the defense to be less than best.  The robots could have the edge here by unemotionally sticking to the odds as they see it.

 

Board 19

Dlr: South

Vul: E-W

North

ª J 9
© A 9 8 7 4 2
¨ 4 2
§ J 10 7

 

 

 

West

ª Q 8 7 4 3

© 6

¨ A K J 6

§ 8 6 3

 

East

ª A K 10 5 2
© K Q J
¨ Q

§ A 9 5 2

 

South

ª 6
© 10 5 3

¨ 10 9 8 7 5 3

§ K Q 4

 

 

 

 

Bridge Baron and Micro Bridge reached 6ª.  Bridge Baron was the only robot to open 1ª.  In the Bermuda Bowl, 6ª was reached 7 of 22 times, approximately the same percentage as the robots.  This is difficult hand to reach a slam.  At some tables West splintering in hearts, which didn’t encourage East.

 


Board 12

Dlr: West

Vul: N-S

North

ª A K 10 9 8 4

© Q 10 3
¨ 10 9 6
§ 6 3

 

 

 

West

ª Q 5

© J 9 8 6 5 4

¨ J 5

§ Q 10 8

 

East

ª 6
© A K 7 2
¨ K Q 10 6 3 2
§ J 6

 

South

ª J 7 3 2
© ---

¨ A 9 4

§ A K 7 5 4 3

 

 

 

WBridge5 reached 7ª and Q-Plus Bridge reached 6ª.  The only accident was Blue Chip Bridge, + 150, allowing Q-Plus Bridge to play 5© undoubled.  The others were in 4ª.

 

In the Bermuda Bowl, 6ª was played 4 times, and 7ª twice, in the Venice Cup, 6ª was reached six times and in the Senior Bowl, 6ª was reached six times and 7ª once.  Interestingly, there were many accidents on this board, including 3¨ (-50), twice the result in the Bermuda Bowl, four times in the Venice Cup and five times in the Senior Bowl.  6§ was played once in the Senior Bowl.

 

The robots duplicated the experts in reaching slam and had fewer accidents.

 

In their final match, WBridge5 had a one IMP lead at the half, 45-44, and then far outscored Jack in the second half, 91-23, to take the crown.

 

 

Carryover

1-16

17-32

33-48

49-64

Total

Jack

0

27

17

4

19

67

WBridge5

0

31

14

41

50

136

 

The final quarter was broadcast on BridgeBase Online to an audience of over 1,300.  The finalists demonstrated fine play, not being hampered by the nervousness that less than expert players feel when on viewgraph.  The play of all the final boards, as well as the BridgeBase Online broadcast of the fourth quarter can be found at computerbridge.com.

 

The audience was treated to some good play.  Of the thirty-two results for each team of robots, sixteen at each table, WBridge5 reached its par result fifteen times (including 7NT on board 9 with thirteen top tricks, gaining 13 IMPs), had three bad results and fourteen neutral results.  Par being either reaching a game or slam that makes, staying out of a game or slam that fails or getting a good score because the opponents erred.  A bad result is not obtaining a reasonably obtainable par result.  A neutral result is either reaching a normal contract that luckily makes more, as for example, playing in 4ª on board 10 which makes six, but is only even money to bid at IMPs, or the opponents reaching a normal contract with no error on your part.  Par results, as used here, are not necessarily best bridge as, for example, an inferior contract might make on the lie of the cards, but par results, in general, demonstrate reasonable bridge.  Bad results usually determine a match.  Jack had eleven par results, six bad results and fifteen neutral results.  Of the three bad results by WBridge5, two were reasonable bridge.  One occurred on board 5 when, with Kx opposite Qx in spades, 5 of a minor was reached at both tables and failed when diamonds broke 3-1, while 3NT had nine top tricks.  Another WBridge5 bad result came on board 16 when, holding KQ104 of trumps, WBridge5 played low at trick two on a trump lead from a dummy holding 9832.  This resulted in Jack making 4ª.  Only one bad result was clearly an error, doubling 5ª on board 10, which could never make less than 11 tricks, was cold for 12 tricks on the lie of the cards and actually made 13 tricks.  Of Jack’s six bad results, four were errors, including one at each table on board 11.

 

Board 11

Dlr: South

Vul: None

North

ª K Q 10 2

© A K Q J 10 5
¨ A 10
§ Q

 

 

 

West

ª J 6 3

© 8

¨ K Q 9 6 3

§ A J 8 2

 

East

ª A
© 7
¨ J 8 5 2
§ K 10 9 6 5 4 3

 

 

South

ª 9 8 7 5 4
© 9 6 4 3 2

¨ 7 4

§ 7

 

 

 

 

West

North

East

South

 

Jack

WBridge5

Jack

WBridge5

 

-

-

-

Pass

 

1¨

Dbl

3§(1)

Pass

 

3¨

4©

Pass

Pass

 

5¨

Pass

Pass

5©

 

Pass

Pass

Dbl

All Pass

 

(1)    diamonds and clubs

 

Opening lead ªA, making 5, and +650 to WBridge5.

 

West

North

East

South

WBridge5

Jack

WBridge5

Jack

---

---

---

Pass

1¨

Dbl

3§

Pass

Pass

4©

5§

Pass

Pass

Dbl

All Pass

 

 

3§ showed 8-10 hcp, not forcing

 

Opening lead ©2, making 5, +550 and 15 IMPs to WBridge5.

 

Yves Costel explains that the 5§ bid by WBridge5 resulted after a simulation that gave better results (including going down less) to 5§ than to 5¨.  Interestingly, Jack, not knowing about E-W’s double fit, was less likely to save in 5©.

 

In 5©x, the ªA was led.  Jack does not play suit preference signals in this instance.  East returned a diamond, as West was more likely to hold the ¨A than the §A.  Hans Kuijf explains that the programming of a suit preference signal in certain situations, such as this one, is very difficult and, if not done near perfectly, can easily lead to mistakes in more frequent situations.

 

General remarks

The fourth quarter of the final match was broadcast on BridgeBase Online at 0900 Portugal time.  While all of North America slept, there was an audience of over 1,300.  A much higher number than anticipated which may be showing an increased interest in computer-bridge play, especially outside North America.  With little advertising, computerbridge.com averaged over 500 hits-per-day during a period of five days before the event to five days after the event.

 

The level of play of the robots continues to improve, with the top robots showing greater consistency.  The bad results that sometimes occur are, in my opinion, overstated.  This may be because there are no human excuses.  Those who watched the human championships on viewgraph and online witnessed some bad play and considerable less than expert play.  This is often passed off as being due to fatigue and pressure, at the table and between sessions (for the best players) or just not being expert players (for some players in all three championships that are less than world-class but still represent their country or zone).  Note the number of accidents on board 12 of the third round of the human round-robin.

 

While bidding and defense are still the weakest part of robot play, as they are in human play, the top robots are greatly improving in these areas.  Specific meanings in competitive auctions are being introduced to advantage.  Defense will continue to be the weak link, for a while, because programming computer simulations for a defender, where one of the two unknown hands belongs to partner, is more difficult than programming for declarer, where partner’s hand is known.

 

Eventually, the great advantages in memory, partner understanding and omission of emotion and fatigue, will lead to robot play that will rival the experts.

 

For more information on this year’s championship, past championships, published articles, photos and the robots’ websites, go to computerbridge.com or ny-bridge.com/allevy/computerbridge.



[1] 22 Hamlet Drive, Commack, NY 11725, USA. Email: allevy@aol.com.