General Information about BBO Robots

This section gives general information about BBO robots, under the following headings:

1. Origin and History To Date of Computer Bridge

2. GIB Robots

2.1 GIB System Notes

3. Advanced GIB Robot System Notes

1. Origin and History To Date of Computer Bridge

This section is derived from the Wikipedia entry on Computer Bridge,

(https://en.wikipedia.org/wiki/Computer_bridge), and may be of interest to all levels of CBC

members.

Computer bridge is the playing of the game contract bridge using computer software. After

years of limited progress, since around the end of the 20th century the field of computer bridge

has made major advances.

The GIB robot system won the world computer bridge championship in 1997 and 1998 but not

since, as other bidding and play systems have emerged.

The Wikipedia entry (click here) includes written material and links under the following

categories:

• World Computer-Bridge Championship

• Computers versus humans

• Cardplay algorithms

• Comparison to other strategy games

• The future

2. GIB Robots

This section describes the general characteristics of BBO’s GIB robots, their bidding system and

approach to both declarer play and defense.

Most of the material was sourced from the BBO website. The links to specific items referenced

in the material (eg. the standard GIB Convention card, and, understanding GIB bid explanations)

can be found in the ‘how-to section which follows.‘

Bridge Base Online offers GIB to its users, either for free or at a small cost depending on when

and how it is used. GIB is a bridge-playing robot, or computer program. Since GIB is a

computer, there are certain things it does really well as a bridge player.

For example, GIB never miscounts a suit. It never misses a spot card. It never forgets its own

system. It never fails to notice… well, anything. . Robots play at least as well as the average

BBO member. In simulations, GIB robots currently average around 55% in a duplicate

tournament setting.

However, GIB is far from perfect. When GIB does make an error, it tends to be a dramatic error

of the type that humans would rarely make. Why does this happen? There are quite a lot of

reasons. GIB may be programmed incorrectly for the auction. GIB may have misunderstood

what someone else meant by their bid, and reached a faulty conclusion. Sometimes when GIB

doesn’t know what to do, it examines a bunch of simulated hands to reach a decision, and

although those simulations tend to be quite accurate, they can from time to time reach a silly

conclusion for a variety of reasons.

The robots are quiet and polite. They may never thank you for a well-played hand, but they will

never yell at you for not making a contract. Interaction between the human opponents is not

allowed so there is no chance of cross table abuse. Their personal hygiene is impeccable.

The robots used on BBO are called GIB (Ginsberg's Intelligent Bridgeplayer.) GIB is widely

considered to be one of the best computer bridge programs ever created. It is occasionally

capable of brilliance. It is also occasionally capable of some really poor bids and plays (just like

all human players).

Some players may find it frustrating if a particular robot partner plays especially poorly (or if a

particular robot opponent plays especially well), on a given hand, but, these things will even

themselves over time. Sometimes the robot does something totally crazy (don't we all.) But

everyone has the same robot as partner and opponent so everyone is on a level playing field. If

your GIB does something crazy, BBO wants to know about it so we can see if it is fixable. (ed.

Note: the GIB robots, developed in the late 1990’s are now on version 40, last time I looked, so

there have been lots of improvements and debugs)

The GIBs used on BBO play a relatively simple and natural 2/1 bidding system. You can

find out the meaning of any bid by clicking on that bid as it appears in the bidding

diagram. Furthermore, when it is your turn to bid, moving your mouse over the buttons for

the various possible bids will cause an explanation of the bid you are considering (as your

GIB partner will understand it), to be displayed. These explanations can be somewhat cryptic,

but reading them carefully before you bid will help you to avoid misunderstandings with your

GIB partner.

2.1 GIB System Notes

BBO has created a standard convention card for GIB. Click here to see GIB's convention card,

and click here for the Basic Robot System Notes, and click here for the Advanced Robot System

Notes

In general, the GIB robots on BBO use the 2/1 system described below. You can click on any of

GIB's bids for an explanation, and pause your mouse over a bid you plan on making to see how it

will understand it. Click here for help understanding these explanations.

3. GIB Advanced System

When you purchase robots on BBO, you have a choice of Basic or Advanced. Advanced is more

expensive, and offers a slightly different, more reliable system. Click here to see the Advanced

system notes.

(the next level of click access material starts here)

2.1 GIB System Notes

Overview

2/1 Game Force with 5 card majors, strong NT, strong (17+) jump shift, weak 2 bids and a strong

artificial 2♣.

HCP vs Total Points

Gib uses both old fashioned HCP (A=4, K=3, Q=2, J=1)) and “Total points” (HCP+3 for void, 2

for singleton, 1 for doubleton, but short suits containing an honor are reduced by 1 point). It will

usually force to game if it thinks it has 25 Total Points between the two hands.

How GIB Defends

It's difficult to describe precisely how GIB defends. It doesn't use rules and guidelines, like

humans often do. It simulates hands based on the auction, using double dummy analysis to

determine the average result of each defensive play, and chooses the one with the best average.

Sometimes this simulation comes up with the same choice that a human would make (there's a

good reason for some of the guidelines -- they actually work well), but not always (some of our

rules of thumb have become popular simply because they're easy to remember and "good

enough"). When it has a choice of equivalent cards, it will choose based on leading and

signalling conventions.

GIB doesn't interpret your signals or make many inferences from the play, it uses simulations

based on the auction. However, it's usually able to figure out that when you lead an honor, it's

part of a sequence.

GIB usually leads passively against NT (read the book Winning Notrump Leads to understand

why). Don't assume it's leading its longest suit. When you lead, it doesn't assume you're

leading your best suit, which is why it doesn't always return the suit like a human would.

In suit contracts, GIB's opening lead is frequently a side singleton or doubleton, to try to get a

ruff. When it leads a suit bid by the opponents, this is almost always the reason. Read the

book Winning Suit Contract Leads for insight on the way GIB leads against suits.

If it leads an honor that's part of a sequence, it uses standard honor leads (K from AKx, A from

AK doubleton). If it leads from a long suit, it leads 4th best (but see above: it doesn't always lead

its long suit). When leading from 3 small, it leads low against both suit and NT contracts.

It doesn't use any signals when making discards, it just tries to make safe discards. In a suit

contract it will frequently discard from a short suit while it has trumps left. Otherwise, it tends to

discard from a long suit that's safe to shorten.

When it's following to partner's opening lead, it will usually give an attitude signal:

• High spot card with an Ace or King

• High spot card with a Queen behind dummy's Ace or King

• Low in any other situation

Note that it doesn't give count in this situation, so it's hard to know when you can give it a ruff.

When it's trying to win the trick in third hand, it will play the lowest of equals. Otherwise, when

following suit it usually gives standard count signals (high = even); an exception is when it's

forced to play equivalent cards in a doubleton, it will randomize them because of "restricted

choice".

(end of intro)

GIB Basic Robot System Notes

Bidding

Basic Approach

Opening bids

1♣ could be 3 if 4333,3433 or 4423. 2♣ response is forcing, inverted

1♦ usually 4 unless 4432. Opens 1♦ with 4-4 in the minors. 2♦ response is forcing,

inverted. 2♣ response is game forcing.

1♥ 1♠

normally show 5 in all seats. Opens 1♠ with 5-5 in spades and clubs. 1M-2M

direct raise shows 7-10 points. 1N response is forcing. Jacoby 2NT. Splinters.

Two-way game tries.

1NT balanced 15-17 HCP, may have a 5-card major (GIB treats 17 with 5-card major

as 18). Followups

2♣ strong, artificial. 22+ HCP

2♦ 2♥ 2♠ weak 2 bid. Disciplined, with honors in the suit

2NT balanced 20-21 HCP, may have a 5-card major. Followups

Responses and Rebids

• Opening jump rebid (1♣-1any-3♣ for example) promises 6+ card, 17-20 HCP

• Opening major rebid after 2/1 response does not promise 6 cards in the suit.

• Jacoby 2NT

• Raising responder's suit usually promises 4 cards, but will occasionally raise with only 3

Competitive Auctions

• 1-level overcall shows 5+; 8-17 HCP; 9-19 TP. However, might overcall 1-major with

decent hand and a strong 4-card suit.

• GIB uses the law of total tricks.

• Takeout doubles to 4♥

• Negative X and Responsive X up to 3♠, Support X up to 2♥ (GIB may do support X with

Kx).

• Weak jump overcalls (aggressive, 3-9 HCP, no allowance for vulnerability)

• Cappelletti when opponents open 1NT (aka Hamilton)

• Lebensohl after 1NT, double of weak 2, and after reverse.

• Michaels cue bid (over a weak 2 in a major, 4 of the same major shows a good hand with

both minors, while 4NT shows a weaker hand with both minors).

• Unusual NT (when non-vulnerable GIB only does Unusual NT with intermediate hands,

xx-KQxxxx KQxxx for example, but it needs a better hand when vulnerable)

• Unusual versus Unusual, lower cue = limit raise or better (GIB does UVU, but doesn't

alert it with this name)

• Sandwich 1NT (by passed hand only)

• Truscott after partner's opening doubled (aka Jordan)

Other conventions and treatments

• Fourth Suit Forcing (1♣-1♦-1♥-1♠ is artificial game force, 1♣-1♦-1♥-2♠ is game forcing

with spades)

• New minor forcing (one-way)

• Roman Key Card Blackwood. GIB doesn't like to use Gerber, but it will respond

appropriately

• Strong (Soloway) jump shifts by unpassed hand, fit jumps by passed hand (except jumps

to 3♣, which are natural and invitational).

• Reverse Drury by passed hand in uncontested auctions.

Notes

• Vulnerability does not affect most opening/overcall decisions. GIB tends to consider

vulnerability and form of scoring only when thinking about leaving in a double for

penalties, and other high-level competitive decisions (they get used when performing

simulations and estimating the expected value of different outcomes). Some bidding

rules, such as the decision whether to use Michaels or Unusual NT, take vulnerability

explicitly into account; these were decided on a case-by-case basis, there's no general

rule.

Conventions that GIB does not play

• Gambling 3NT

• Namyats

• Bergen or Reverse Bergen Raises

• DONT

• Puppet Stayman

Two-way Game Tries

After a single raise of a major suit, GIB plays two-way game tries.

• The next step (1♥-2♥-2♠, 1♠-2♠-2NT) is a short-suit game try, showing unspecified

shortness. Responder can bid the next step above that (2NT or 3♣) to ask where the

shortness is. Opener bids the short suit, or bids his major if the short suit is one of the step

suits.

• Any other bid below 3 of the major is a long suit game try, at least a 3-card suit with

some honors. After 1♥-2♥, 2NT is a LSGT in spades (since 2♠ would be a short-suit

game try).

• 3 of the major is a general strength game try, showing about 17 points with no singleton

or void (GIB rarely makes this bid, since this would probably be a 1NT opener).

Roman Keycard Blackwood (RKCB)

RKCB is a 4NT bid that, unlike regular Blackwood, asks for "keycards" instead of Aces. There

are always 5 keycards - the 4 Aces plus the King of the agreed trump suit. If no trump suit has

been clearly agreed, the King of the most recently bid suit is typically counted as the 5th

keycard.

Responses to 4NT RKCB 0314:

• 5♣ 0 or 3 keycards

• 5♦ 1 or 4 keycards

• 5♥ 2 or 5 keycards, but no Queen of the agreed suit

• 5♠ 2 or 5 keycards plus Queen of the agreed suit

• 5NT An even number of keycards plus an unspecified void

• 6x An odd number of keycards with a void. If 6x is below 6 of the agreed suit then the

void is in the suit bid. If 6x is a bid in the agreed suit then the void is in an unspecified

higher-ranking suit.

After the 5♣ and 5♦ responses, the 4NT bidder can bid the next step that is not a signoff in order

to ask for the Queen of the agreed suit. Then:

• Bidding the agreed suit at the cheapest level denies the Queen of the agreed suit.

• Bidding a new suit promises the Queen of the agreed suit plus the King of the suit bid.

• Bidding 5NT promises the Queen of the agreed suit and denies a side King that can be

shown below 6 of the agreed suit.

A subsequent 5NT bid by the 4NT bidder (regardless of whether or not an ask for the Queen of

the agreed suit has taken place) asks for specific Kings. The 5NT bid promises that all of the 5

keycards and the Queen of the agreed suit are accounted for. Then:

• The responder to RKCB is entitled to bid a grand slam if he thinks that 13 tricks rate to

be available.

• If the responder to RKCB has a King that is lower-ranking than the agreed suit, he should

bid that suit at the 6-level. If he has more than one such King, he should bid his lowest-

ranking King.

• Otherwise the responder to RKCB should bid 6 of the agreed suit.

DOPI after interferences:

• Interference after 4NT (whether RKCB or regular Blackwood) is handled by the DOPI

convention.

• Double with 0 keycards (or ace, playing regular blackwood), pass with 1 keycard (or ace,

playing regular blackwood).

After a 1N opening bid

If the opponents overcall 2♣ (Cappelletti any 1-suited hand), Double is Stayman and all other

bids as below. If the opponents double, all systems are on; Redouble is used to run out to a minor

(opener should bid 2♣, responder passes or corrects to 2♦). After any other

interference, Lebensohl is used.

• 2♣ Stayman (promises at least one 4-card major unless inviting 3NT).

o 2♦ No 4-card major

▪ 2♥/2♠ Invitational with 5 of the suit bid and 4 of the other major

▪ 2NT Invitational (does not promise a 4-card major)

▪ 3♣/3♦ 5+ card suit. Forcing to game

▪ 3♥/3♠ Smolen (forcing to game with 4 of the suit bid and 5 of the other

major)

▪ 4NT Invitational to 6NT.

o 2♥ 4 hearts (could also have 4 spades)

▪ 2♠ Invitational with 4 spades

▪ 2NT Invitational, denying 4 spades

▪ 3♣/3♦ 5+ card suit. Forcing to game

▪ 3♥ Invitational

▪ 3♠ Artificial slam try with at 4+ hearts, usually balanced

▪ 4♣/4♦ Splinters (singleton or void in the suit bid, 4+ hearts, interest in

slam)

▪ 4NT Invitational to 6NT

o 2♠ 4 spades (denies 4 hearts)

▪ 2NT Invitational (does not promise 4 hearts)

▪ 3♣/3♦ 5+ card suit. Forcing to game

▪ 3♥ Artificial slam try with 4+ spades, usually balanced.

▪ 3♠ Invitational

▪ 4♣/4♦/4♥ Splinters (singleton or void in the suit bid, 4+ spades, interest

in slam)

▪ 4NT Invitational to 6NT

• 2♦ Jacoby Transfer Bid (promises 5+ hearts). Opener would normally bid 2♥, but can

superaccept with a maximum and 4-card heart support. After 2♥:

o 2♠ 5+ hearts, 5+ spades, invitational to game.

o 2NT Exactly 5 hearts, invitational to game.

o 3♣/3♦ 5+ hearts, 4+ card suit. Forcing to game.

o 3♥ Invitational with 6+ hearts

o 3NT Exactly 5 hearts. Choice of games (4♥ or 3NT).

o 3♠/4♣/4♦ Splinters (6+ hearts, singleton or void in the suit bid, interest in slam)

o 4♥ 6+ hearts, no singleton or void, mild slam interest

o 4NT Exactly 5 hearts. Invitational to 6♥ or 6NT

o 5NT Choice of slams (6♥ or 6NT)

• 2♥ Jacoby Transfer Bid (promises 5+ spades). Opener would normally bid 2♠, but can

superaccept with a maximum and 4-card spade support. After 2♠:

o 2NT Exactly 5 spades, invitational to game.

o 3♣/3♦ 5+ spades, 4+ card suit. Forcing to game.

o 3♥ 5+ spades, 5+ hearts. Forcing to game.

o 3♠ Invitational with 6+ spades

o 3NT Exactly 5 spades. Choice of games (4♠ or 3NT).

o 4♣/4♦/4♥ Splinters (6+ spades, singleton or void in the suit bid, interest in

slam)

o 4♠ 6+ spades, no singleton or void, mild slam interest

o 4NT Exactly 5 spades. Invitational to 6♠ or 6NT

o 5NT Choice of slams (6♠ or 6NT).

• 2♠ Minor Suit Stayman (Usually at least 54 in the minors, forcing to game)

o Opener would normally bid a 4+ card minor if he had one, but can bid 2NT with

3343 or 3334 distribution.

o If responder's next bid is 3 of a major, he is promising a singleton or void in that

suit, but not necessarily slam interest

• 2NT Minor Suit Transfer (Promises 6+ clubs. Opener must bid 3♣)

o If responder's next bid is 3 of a new suit, he is promising a singleton or void in

that suit, but not necessarily slam interest

o Responder's 3NT rebid is a mild slam try (usually balanced).

o Responder's 4NT rebid is RKCB

• 3♣ Minor Suit Transfer (Promises 6+ diamonds. Opener must bid 3♦)

o If responder's next bid is 3 of a new suit, he is promising a singleton or void in

that suit, but not necessarily slam interest

o Responder's 3NT rebid is a mild slam try (usually balanced).

o Responder's 4NT rebid is RKCB

• 3♦/3♥/3♠ Singleton or void in the suit bid, at least 4 cards in the other 3 suits, no 5card

major, forcing to game.

• 3NT Signoff

• 4♣ Gerber

• 4♦ Texas Transfer (Promises 6+ hearts, opener must bid 4♥)

o New suit rebid by responder is a cuebid.

o 4NT rebid by responder is RKCB.

• 4♥ Texas Transfer (Promises 6+ spades, opener must bid 4♠)

o New suit rebid by responder is a cuebid.

o 4NT rebid by responder is RKCB.

• 4NT Invitational to 6NT

• 5NT Invitational to 7NT

After a 2N opening bid

• 3♣ Stayman (promises at least one 4-card major)

o 3♦ No 4-card major

▪ 3♥/3♠ Smolen (forcing to game with 4 of the suit bid and 5 of the other

major)

▪ 4♣/4♦ 5+ card suit. Interest in slam

▪ 4♥/4♠ Signoff bids

▪ 4NT Invitational to 6NT

o 3♥ 4 hearts (could also have 4 spades)

▪ 3♠ Artificial slam try with 4+ hearts

▪ 3NT Choice of games (4♠ or 3NT). Promises 4 spades.

▪ 4♣/4♦ 5+ card suit. Interest in slam

▪ 4NT Invitational to 6NT

o 3♠ 4 spades (denies 4 hearts)

▪ 3NT Signoff (promises 4 hearts)

▪ 4♣/4♦ 5+ card suit. Interest in slam.

▪ 4♥ Artificial slam try with 4+ spades

▪ 4NT Invitational to 6NT

• 3♦ Jacoby Transfer Bid (promises 5+ hearts). Opener would normally bid 3♥, but can

superaccept with a maximum and 4-card heart support. After 3♥:

o 3♠ 5+ hearts, 5+ spades, interest in slam

o 3NT Exactly 5 hearts. Choice of games (4♥ or 3NT)

o 4♣/4♦ 5+ hearts, 4+ cards in suit bid, forcing to game

o 4♥ Mild slam try with 6+ hearts

o 4NT Exactly 5 hearts. Invitational to 6♥ or 6NT

o 5NT Choice of slams (6♥ or 6NT)

• 3♥ Jacoby Transfer Bid (promises 5+ spades). Opener would normally bid 3♠, but can

superaccept with a maximum and 4-card spade support. After 3♠:

o 3NT Exactly 5 spades. Choice of games (4♠ or 3NT)

o 4♣/4♦ 5+ spades, 4+ cards in suit bid, forcing to game

o 4♥ 5+ spades, 5+ hearts, choice of games (4♥ or 4♠)

o 4♠ Mild slam try with 6+ spades

o 4NT Exactly 5 spades. Invitational to 6♠ or 6NT

o 5NT Choice of slams (6♠ or 6NT)

• 3♠ Minor Suit Stayman (usually at least 54 in the minors, forcing to game)

o Opener would normally bid a 4+ card minor if he had one. Otherwise he would

bid 3NT.

o If responder's next bid is 4 of a major, he is promising a singleton or void in that

suit

• 3NT Signoff

• 4♣ Gerber

• 4♦ Texas Transfer (Promises 6+ hearts, opener must bid 4♥)

o New suit rebid by responder is a cuebid.

o 4NT rebid by responder is RKCB.

• 4♥ Texas Transfer (Promises 6+ spades, opener must bid 4♠)

o New suit rebid by responder is a cuebid.

o 4NT rebid by responder is RKCB.

• 4NT Invitational to 6NT

• 5NT Invitational to 7NT

Soloway Jump Shifts

GIB plays Soloway Strong Jump Shifts by an unpassed hand in uncontested auctions. A jump

shift shows one of the following types of hands:

1. Strong rebiddable suit, 17+ total points, 4+ controls (A=2, K=1), no side 4-card suit

2. Solid suit, 17+ total points, 4+ controls, may have a side 4-card suit

3. Rebiddable suit, 18+ HCP, 4+ controls, 5332 or 6322 shape.

4. Rebiddable suit, 17+ total points, 4+ controls, 4-card support for opener's suit

Strong jump shifts are only from the 1 level to a higher suit on the 2 level. Jumps to a lower suit

on the 3 level are natural and invitational.

Opener can rebid his suit to show 6+, raise responder or bid RKC Blackwood with 3+ support,

bid a side suit to deny support and show least KQ in the suit, or bid NT at the cheapest level to

show any other hand.

Jump shifter shows which type of hand it had with its next bid:

• With types 1 or 2, it rebids its suit, jumping to game with a minimum and solid suit (note

that it never shows the side suit in type 2).

• With type 3, it bids NT or raises NT to game.

• With type 4, it raises opener's suit with no side shortness, or bids its short suit (this is why

it can never show its own side suit – a new suit is a splinter in support of opener).

Reverse Drury

GIB plays one-way Reverse Drury when partner opens a Major in 3rd or 4th seat. A 2♣ response

shows at least 3-card support and invitational values (11-12 total points). This is not used if there

is any interference; Jordan/Truscott 2NT is still used to show a limit raise over a double, a cue

bid is used after an overcall, and 2♣ is natural (weak after a double, one-round force after an

overcall).

Opener's rebids are as follows:

• 2♦ Full opener, inviting game, no extra shape.

• 2M Sub-minimum opener, no interest in game.

• New suit without jumping 4+ cards in the suit, less than 18 total points.

• New suit single jump 18+ total points, singleton in suit bid.

• New suit double jump 18+ total points, void in suit bid.

• 2NT 5332 shape, one-round force, less than 18 total points.

• 3NT 6322 shape, one-round force, less than 18 total points.

• 3M 18+ total points, balanced.

• 4M To play, nothing extra to show.

© 2018-2020 Bridge Base On Line LLC • About • Privacy • Rules • Term

(end of Basic system section)

(beginning of Advanced GIB Robot System Notes

BBO Advanced (2/1=Game Force)

Notrump Openings

1NT=15-17: Stayman, Jacoby, 2♠->3♣, 2♠->3♣, 2NT->3♦, 3♣=5-5 minors weak, 3♦=5-5 minors GF, 3♥=31(54), 3♠=13(54), Gerber, Texas 2NT=20-21: Stayman, Jacoby, Texas, Gerber, 3♠->3NT (minor 1 or 2-suiter) 3NT=Solid minor no outside A/K

Major Suit Openings Minor Suit Openings

1♥/♠=5+: 1NT=forcing, Jacoby 2NT, 3NT=15-17, Splinters, Weak jump raises, 3♣=limit raise (4+), 3♦=good raise (4+), 1♠-3♥=weak

1♣/♦=3+, 2NT=11-12; 3NT=13-15, 1♣-1NT=8-10, Inverted raises, 1♦-3♣=invitational

2-Level Openings

2♣=strong: 2♦=waiting, 2♥=negative, 2NT=♥ positive 2♦/♥/♠=weak: 2NT=asks feature; new suits forcing

Other important notes

4th suit game force, Unusual vs. Unusual, Ingberman over reverses (2NT=negative), 2-way NMF, Wolff over 2NT rebids

Doubles Notrump Overcalls

Negative->4♦, Responsive->4♦, Support x and xx ->2-level, Maximal

1NT=15-18 (system on) 2NT=unusual (2 lowest unbid)

Simple Overcalls Over 1NT Openings

8-16+ HCP; new suit=non-forcing; jump raise=weak; cuebid=limit raise+

DBL=penalty 1NT=strong (Cappelletti): 2♣=any suit, 2♦=majors, 2♥=♥+minor, 2♠=♠+minor 1NT=weak (Landy): 2♣=majors, 2♦/♥/♠=natural

Jump Overcalls Over Takeout Doubles

Weak New suit forcing at 1-level; RDBL=10+ HCP; 2NT=limit raise+; jump raise=weak; jump shift=weak;

Direct Cuebid Slam Bidding

Michaels RKCB 1430; Gerber; Grand slam force; DOPI; DEPO

Defensive Carding

Attitude: high=encouraging; low=discouraging Count: high-low=even; low-high=odd; 4th best leads; Ace from AKx versus suits

(start of Wikipedia article on computer bridge)

Computer bridge

From Wikipedia, the free encyclopedia

Computer bridge is the playing of the game contract bridge using computer software. After

years of limited progress, since around the end of the 20th century the field of computer bridge

has made major advances. In 1996 the American Contract Bridge League (ACBL) established an

official World Computer-Bridge Championship, to be held annually along with a major bridge

event. The first championship took place in 1997 at the North American Bridge Championships

in Albuquerque. Since 1999 the event has been conducted as a joint activity of the American

Contract Bridge League and the World Bridge Federation. Alvin Levy, ACBL Board member,

initiated this championship and has coordinated the event annually since its inception. The event

history, articles and publications, analysis, and playing records can be found at the official ebsite.

Contents

• World Computer-Bridge Championship

• Computers versus humans

• Cardplay algorithms

• Comparison to other strategy games

• The future

• See also

• References

• External links

World Computer-Bridge Championship

The World Computer-Bridge Championship is typically played as a round robin followed by a

knock-out between the top four contestants.[1][2] Winners of the annual event are:

• 1997 Bridge Baron

• 1998 GIB

• 1999 GIB

• 2000 Meadowlark Bridge

• 2001 Jack

• 2002 Jack

• 2003 Jack

• 2004 Jack

• 2005 Wbridge5

• 2006 Jack

• 2007 Wbridge5

• 2008 Wbridge5

• 2009 Jack

• 2010 Jack

• 2011 Shark Bridge

• 2012 Jack

• 2013 Jack

• 2014 Shark Bridge

• 2015 Jack

• 2016 Wbridge5[3]

• 2017 Wbridge5[4]

• 2018 Wbridge5[5]

• 2019 Micro Bridge[6]

• 2020 championship has been cancelled

Computers versus humans

In Zia Mahmood's book, Bridge, My Way (1992), Zia offered a £1 million bet that no four-person

team of his choosing would be beaten by a computer. A few years later the bridge

program GIB (which can stand for either "Ginsberg’s Intelligent Bridgeplayer" or "Goren In a

Box"),[7] brainchild of American computer scientist Matthew Ginsberg,[8] proved capable of

expert declarer plays like winkle squeezes in play tests. In 1996, Zia withdrew his bet. Two years

later, GIB became the world champion in computer bridge, and also had a 12th place score

(11210) in declarer play compared to 34 of the top humans in the 1998 Par Contest (including

Zia Mahmood).[9] However, such a par contest measures technical bridge analysis skills

only[clarification needed], and in 1999 Zia beat various computer programs, including GIB, in an

individual round robin match.[10]

Further progress in the field of computer bridge has resulted in stronger bridge playing programs,

including Jack [11] and Wbridge5.[12] These programs have been ranked highly in national bridge

rankings. A series of articles published in 2005 and 2006 in the Dutch bridge

magazine IMP describes matches between five-time computer bridge world champion Jack and

seven top Dutch pairs including a Bermuda Bowl winner and two reigning European champions.

A total of 196 boards were played. Jack defeated three out of the seven pairs (including the

European champions). Overall, the program lost by a small margin (359 versus

385 IMPs).[volume & issue needed]

In 2009, Phillip Martin, an expert player, began a four-year project in which he played against

the champion bridge program, Jack. He played one hand at one table, with Jack playing the other

three; at another table, Jack played the same cards at all four seats, producing a comparison

result. He posted his results and analysis in a blog he titled The Gargoyle Chronicles.[13] The

program was no match for Martin, who won every contest by large margins.

Cardplay algorithms

Bridge poses challenges to its players that are different from board games such as chess and go.

Most notably, bridge is a stochastic game of incomplete information. At the start of a deal, the

information available to each player is limited to just his/her own cards. During the bidding and

the subsequent play, more information becomes available via the bidding of the other three

players at the table, the cards of the partner of the declarer (the dummy) being put open on the

table, and the cards played at each trick. However, it is only at the end of the play that full

information is obtained.

Today's top-level bridge programs deal with this probabilistic nature by generating many

samples representing the unknown hands. Each sample is generated at random, but constrained

to be compatible with all information available so far from the bidding and the play. Next, the

result of different lines of play are tested against optimal defense for each sample. This testing is

done using a so-called "double-dummy solver" that uses extensive search algorithms to

determine the optimum line of play for both parties. The line of play that generates the best score

averaged over all samples is selected as the optimal play.

Efficient double-dummy solvers are key to successful bridge-playing programs. Also, as the

amount of computation increases with sample size, techniques such as importance sampling are

used to generate sets of samples that are of minimum size but still representative.

Comparison to other strategy games[edit]

While bridge is a game of incomplete information, a double-dummy solver analyses a simplified

version of the game where there is perfect information; the bidding is ignored, the contract

(trump suit and declarer) is given, and all players are assumed to know all cards from the very

start. The solver can therefore use many of the game tree search techniques typically used in

solving two-player perfect-information win/lose/draw games such as chess, go and reversi.

However, there are some significant differences.

• Although double-dummy bridge is in

practice a competition between two generalised

players, each "player" controls two hands and the

cards must be played in a correct order that

reflects four players. (It makes a difference which

of the four hands wins a trick and must lead the

next trick.)

• Double-dummy bridge is not simply

win/lose/draw and not exactly zero-sum, but

constant-sum since two playing sides compete for

13 tricks. It is trivial to transform a constant-sum

game into a zero-sum game. Moreover, the goal

(and the risk management strategy) in general contract bridge depends not only on the

contract but also on the form of tournament. However, since the double-dummy version is

deterministic, the goal is simple: one can without loss of generality aim to maximize the

number of tricks taken.

• Bridge is incrementally scored; each played trick contributes irreversibly to the final "score"

in terms of tricks won or lost. This is in contrast to games where the final outcome is more or

less open until the game ends. In bridge, the already determined tricks provide natural lower

and upper bounds for alpha-beta pruning, and the interval shrinks naturally as the search

goes deeper. Other games typically need an artificial evaluation function to enable alpha-beta

pruning at limited depth, or must search to a leaf node before pruning is possible.

• It is relatively inexpensive to compute "sure winners" in various positions in a double-

dummy solver. This information improves the pruning. It can be regarded as a kind

of evaluation function, however while the latter in other games is an approximation of the

value of the position, the former is a definitive lower bound on the value of the position.

• During the course of double-dummy game tree search, one can establish equivalence classes

consisting of cards with apparently equal value in a particular position. Only one card from

each equivalence class needs to be considered in the subtree search, and furthermore, when

using a transposition table, equivalence classes can be exploited to improve the hit rate. This

has been described as partition search by Matthew Ginsberg.

• Numerous strategy games have been proven hard in a complexity class, meaning that any

problem in that complexity class can be reduced in polynomial time to that problem. For

example, generalized x × x chess has been proven EXPSPACE-complete (both

in EXPSPACE and EXPSPACE-hard), effectively meaning that it is among the hardest

problems in EXPSPACE. However, since there is no natural structure to exploit in double-

dummy bridge towards a hardness proof or disproof, unlike in a board game, the question of

hardness remains.[14]

The Future

Unsolved problem in

computer science:

Is the problem of deciding the

winner in double-dummy

bridge hard in any complexity

class?

(more unsolved problems in

computer science)

In comparison to computer chess, computer bridge has not reached world-class level, but the top

robots have demonstrated a consistently high level of play. (See analysis of the last few years of

play at www.computerbridge.com.) However see below the Philippe Pionchon's article (1984).

Yet, whereas computer chess has taught programmers little about building machines that offer

human-like intelligence, more intuitive and probabilistic games such as bridge might provide a

better testing ground.

The question of whether bridge-playing programs will reach world-class level in the foreseeable

future is not easy to answer. Computer bridge has not attracted an amount of interest anywhere

near to that of computer chess. On the other hand, much progress has been made in the last

decade by researchers working in the field.

Regardless of bridge robots' level of play, computer bridge already has changed the analysis of

the game. Commercially available double-dummy programs can solve bridge problems in which

all four hands are known, typically within a fraction of a second. These days, few editors

of books and magazines will solely rely on humans to analyse bridge problems before

publications. Also, more and more bridge players and coaches utilize computer analysis in

the post-mortem of a match.

See also[edit]

• List of computer science awards

• Computer Olympiad

• Monte Carlo method

• Monte Carlo tree search

• Importance sampling

• Hanabi (card game)

References[edit]

1. ^ ACBL/WBF World Computer-Bridge Championship Official Site

2. ^ World Computer Bridge Championship - List of contestants and their links

3. ^ "20th Ourgame World Computer-Bridge Championship". World Computer-Bridge

Championship. Retrieved 3 November 2019.

4. ^ "21st World Computer-Bridge Championship". World Computer-Bridge

Championship. Retrieved 3 November 2019.

5. ^ "22nd World Computer-Bridge Championship". World Computer-Bridge

Championship. Retrieved 3 November 2019.

6. ^ "23rd World Computer-Bridge Championship December 2-7, 2019 San Francisco,

CA, USA at the ACBL's NABC". World Computer-Bridge Championship. Retrieved 14

December2019.

7. ^ Bridge Base Online Help

8. ^ Ginsberg profile

9. ^ "Rosenberg Wins Par Contest" (PDF).

10. ^ Foreword to Man vs Machine - The Bridge Match of the Millennium Archived 2006-

05-14 at the Wayback Machine

11. ^ Jack homepage

12. ^ WBridge5 homepage (in French)

13. ^ [1]

14. ^ Hearn, Robert Aubrey (2006). Games, puzzles, and computation (Doctoral).

External links[edit]

• World Computer-Bridge Championship – ACBL/WBF bridge-bot world

championship (official website)

• BridgeGuys. "Participating Computer Bridge Software for the World Computer Bridge

Championship and the History of these Software Programs" (PDF). Archived from the

original (PDF) on November 8, 2016.

• Ginsberg, Matthew L. "GIB: Steps Toward an Expert-Level Bridge-Playing

Program". CiteSeerX 10.1.1.52.2188.

• Bethe, Paul M (January 14, 2010). "The State of Automated Bridge Play" (PDF).

Philippe Pionchon (1984). "Artificial intelligence and Bridge game". Le Bridgeur

Understanding GIB Bid Descriptions

Hand features

GIB internally describes hands using the following features:

• High card points, using 4321 count • Total points = HCP + short-suit points (void=3, singleton=2, doubleton=1 – subtract 1 for

each short suit with HCP) • Length of each suit • Quality of each suit • Total losers • Stoppers in each suit

• High honors shown or denied in each suit

Suit Quality

3-card exactly 3 cards

4-card exactly 4 cards

biddable 5+ cards, or 4 cards + 3 of 5 honors or Ace or King

rebiddable biddable + 1 card

twice rebiddable rebiddable + 1 card

strong rebiddable twice rebiddable with 4 of 5 honors or AKQ

solid 6-card AKQTxx or AKQJxx

solid 7-card AKQxxxx

solid 8-card AKJxxxxx or AKxxxxxxx or better

A suit gets the highest quality that describes it.

Stoppers

unstopped none of the below holdings

partial stop length+HCP = 4 (Qx or Jxx)

likely stop length+HCP at least 5 (Kx or Qxx)

stop A, QJx, 5+ HCP, or length+HCP at least 7

two stops length+HCP at least 8

Honors

Cue bids, responses to Blackwood, and some other bids (e.g. help suit game tries) indicate or deny various combinations of honors in a suit. GIB explains this by stating the minimum honor holding it can have and/or the honors that it has denied. For instance, if it's cue bidding and bypasses 4♣ to bid 5♦, it will say ♦A, no ♣A.

When it shows some honor combination, it could be better. GIB uses an 8-4-2-1 system to calculate the value of honors: Ace = 8, King = 4, Queen = 2, Jack = 1. If it says that it has KQ of a suit, that's 4+2 = 6, but this is the minimum; it could have the Ace instead, since that's worth 8 by itself. This is somewhat like the old "quick tricks" evaluation of honor combinations (A and KQ When it shows

some honor combination, it could be betterare both considered 1 quick trick).

How these features are used

These features are used in bidding rules in two ways:

1. Does the hand the robot holds, plus what other players have shown, fit the requirements for the bid? This is the criteria.

2. What type of hands does the bid show? We call this the specification.

You might expect that these would be the same thing (if it requires a 5-card suit, it must also show 5+), but that's not always the case. For example, Stayman can be used with a variety of hands, so it doesn't show anything specific. And when the criteria incorporate what partner has already shown, it's difficult for the software to infer the specification from this, so we need to enter it explicitly. When you mouse over a bid in a robot game, the description is produced from the specifications of all the rules that fit the auction and produce that bid.

Descriptions from multiple rules

Sometimes two bidding rules will result in the same bid. For instance, if your partner opens a weak 2, you would raise to game with a weak hand and 4-card support (raising the preempt based on the Law of Total Tricks), or with a strong hand and 2-card support (because you think you can make it). When combining rules like this, we use the minimum of each feature shown.

In the above example, the rule for the weak hand shows total points = 4+, length = 4+, while the rule for the strong hand shows total points = 16+, length = 2+. GIB is not able to describe a hand as showing "this or that", so it shows the least common denominator of each feature: 4+ total points, 2+ cards. This means that the hand has at least 4 points and at least 2-card support, although it can't have the minimum of both at the same time. Unfortunately, there's no way for you to tell that this is what it means.

Descriptions from multiple bids

When a player has made multiple bids in the auction, each of them contributes a specification. These are combined by taking the maximum of each feature. So if a player makes a bid that shows 5+ hearts, and later makes a bid that only shows 4, we will still display that he has shown 5+.

Descriptions containing related features

Sometimes you'll see a confusing description like 5+ ♦, 3-card ♦. How can it be both a 3-card suit and have 5+ cards, you ask. This is because these are two different features: 5+ comes from the Length feature, while 3-card comes from the Quality feature. Although these are related (a 5-card suit is always at least biddable), GIB doesn't take these relationships into account when describing the hand. If one bid only shows suit length, while another bid only shows suit quality, the final description will show each of these as they were shown by their respective bids. As mentioned above, we maximize each feature across multiple bids that show that feature, but we don't reconcile

different features, even if they're related.

Why doesn't the robot's hand always match the description?

When humans play bridge, they don't just follow rote rules for bidding; they often use their judgement to find better bids, or fill in holes in their system. We would love it if we could program judgement into GIB, but that would be pretty advanced artificial intelligence. As with many game-playing computer programs (e.g. chess programs that routinely beat grandmasters), we substitute brute computational power for thinking. Many of GIB's rules allow it to perform simulations.

GIB starts by finding the matching bid in its bidding rules (we call this the "book bid"). If simulations are allowed, it then makes some adjustments to its hand (adding a card to each suit, adding/subtracting a few total points) and finds the book bids for those similar hands. Then it deals out cards to the other hands at the table consistent with the rest of the auction. For each hand and possible bid, it does 2 things:

1. Determine how the auction will probably continue if it makes that bid (to avoid exponential complexity, it only considers book bids for this continuation, not what would happen with subsequent simulations), and

2. Calculate the double-dummy result of the final contract determined in step 1.

• It then selects the bid whose expected value is highest across all the hands. This takes into account the form of scoring; this is how we emulate rules of thumb like "bid games more aggressively when vulnerable at IMPs". The description that's displayed comes from the specifications of the bid that was chosen; this corresponds to the standard rule about disclosure

in bridge: you must describe your agreements, not your actual hand.Review. Classic analysis of

AI applied to bridge.