From: AAAI Technical Report SS-02-02. Compilation copyright © 2002, AAAI (www.aaai.org). All rights reserved. Large Scale Q-tables Kaile Prorok Departmentof ComputingScience Ume~University SE-901 87 UMEA Sweden kalle.prorok@tfe.umu.se Abstract Thisarticle studiesthe possibilitiesfor Q-learning to learnthe biddingin the card-game bridge.Thelearnt biddingsystems are quiteusefulbutbelowthe level of an average club-player. Some studiesregardingthe relationbetween the learningrate, the exploration rate andinitial Q-values are alsodone. The Application Bridge is a card-gamenormallyplayed by four humans.It has someinteresting properties making.itsuitable as a test platform regarding Collaborative Learning Agents. Each player is on his ownbut playing together pair-wise; a dynamic, collaborative and adversial game. Someagreements are madebetweenthe players on beforehandbut mostsituations have to be taken care of by commonsense logic. This makesit difficult to pre-programa computerized player. Instead it makesit interesting to let the computer-player learn from exhaustive experiences. The optimal behavior is also unknown,which makethe results from such an artificial player valuable. Myhopeis that these experimentswith Bridgewill be applicable for other domainslike networkmanagement (informationtransfer), compressionalgorithms(finding descriptions of i.e. a face automatically),electronic commerce (how to performan auction) and robotics (learning to interact). Another area might be language learning; howdid we invent a language"in the beginning". Blocks world (Winston1977) has been used as a smallscale modelsystemfor decadesin the Artificial Intelligence community.Today’sproblemsare of a different kind; cooperation, stochastic and/or non-linear systems, emergent properties etc. Mysuggestion is to use the international card-gameBridge as a newmodel world with someinteresting properties fromthe real worldbut still on a reasonably small scale. Someof the things that can be studied using the Bridge platform are: Collaboration/Cooperation within pairs and teams; Competitionbetweenpairs, teams and countries; Communication - information transfer with limited bandwidth;Insecure domain- not all information is visible for all parts; Stochasticdomains- there is a randomfactor (bowthe cards were dealt); The role of concertCopyright (g) 2002,American Associationfor Artificial Intelligence(www.aaai.org). All rights reserved. 69 tration; mentaltraining; Planning(of biddingand playing); Emergent properties; single suit versus multi-suit plays, endplays; Democratic(descriptive) or Captain (enquiry) based methods;Tactics, Strategy, Risk taking, "the only chance" or safety play "guarding"; Constructiveversus destructive waysof doing things; Psychologicalaspects - to be unpredictable, whento fool partner or opponent;Philosophical aspects - try to find out whatpartner or opponentsthink; Memory;a complexbidding systems versus an easy to remember,understandand use bidding system; Getting information - what and when; Reducingamountof information to opponents whomaytake advantage; Learning from mistakes and successes. The gameis described in manystandard textbooks and the rules can be learnt within a fewdays but requires several years of practice to become a master. Thegoal with the gameis to score points related to the numberof tricks your partnershipis able to take and this is dependentof howgood youare to reach a suitable contract; the gameis thus separated into twophases: the bidding (auction) into a contract and then the play in such a contract. Anexampledeal is presentedin Fig. 1 and a biddingin Fig. 2. Background Bridgeis fairly international withsimilar rules all overthe world and there are more than one million players in the world. Nowit is also an Olympicgame. There are bridge programsperformingwell and it is also a matter of time before the computerbeats most humanbridge players. The mainreason for this is the technical advantagea computer has (rememberingall cards during the play and the bidding system,possibilities of statistical simulationsand calculations makingit possible to play well) compensating for the lack of intuition and other psychologicalaspects by making fewer mistakes. Today’ssituation (December2001) is that the play with the cards is very goodbut bidding lacks behind. The reason for this is that entering humanjudgement as rules in a bid-systemdatabaseis verytedious, Fig. 3. There are somecomputerprogramsaround (reviewed in (Frank 1998)) able to play Bridge but most only at a (beginners)level. Oneof the never and better candidates AI-researcher professor MatthewL. Ginsberg’s GIB(Ginsberg 1999), whichhe guess will be better than humansby the year of 2003(Prorok1999).It is able to do declarer play North/Deal 4iv T9632 ¯ T72 6 QT873 West East ¯ AJT842 AK7 * Q963 v 85 ¯ AJ64 4. J86 South * K75 v QJ4 ¯ K92 4. K953 ¯5 4. AQ4 Figure 1: A typical deal in schematicform. It is dealt by player North whothen starts the bidding with East next in turn. Northis void in (have no) spades but has a weakhand and passes. East has a strong hand and normallyopens the bidding by giving a bid, maybe1 6. South has an average hand but not enoughto enter the bidding with a bid other than Pass. N E SW P 18P 24 P 4*P P P Figure 2: The schemaof one possible bidding with the hands in Fig 1. Threepasses ends the auction. 4 Ik requires EWto take 6 ("the book")+ 4 = 10 tricks out of the possible 13. 1"#08115" {5{[CDHS]:" !.! %0% %.% <0x1083>"2153"#08116" 5[÷#b]" !DOPI! %6% %:H[~:]*~8&[^:]*-S&% "2160"#08117" 5[+÷#b]" !DOPI! %7% t:H[^:|*-8&[^:]*~8&{^:J*~8&t "2159"#08118" 5[+++#b]" !DOPI[ 18% %:H~8&.’S&.’S&.~8&% "2156"#08119" 4N:$7" !.! " %0% t.t *1501"#08120 {SNI[67][CDHSN]):" !.! " %0% %.% <0x1083>"2153"#08121 X" !DOPI! " %3% I.% "2154"#08122 P" !DOPI! !DOPI! %4% %:H[^:]*’S&%"2155"#08123" X" %5% %:H[^:]*~8&[^:]*~8&% "2154"#08124" P" !DOPI! %6% %:H[^:]*~8~[^:]*’8&|^:]*-8~% "2155"#00126" .*{|{{{[I||{{[[CDHS]<#j=3><#k=6>:P:4N" !.! %0% %.% *2037*#00127" :(PIX):" !.! " [BLACRWOOD! %0~ I.%#00128 @ACE BLACKWOOD-" Ii% %.% "1501"#00129" 8ACE RKCB-" [RKCB! %1% %.%#00130" [1231{{{{{|{{T{(({.:P:5N$12" !GSF! " %0% t.% "2229"#00131 :,:" !,[ %0% %.%~00132" 7#a" !.! " !.! %5% %:H[~:]*-9>#a% *2034*#00133 7#a" !.! %5% %:H[*:]*-6=#a% *2034*#00134" 6#a" Figure 3: A small part of the GIB-database, showinghow the Ace-asking bid Blackwood(4NT) should be used. 7O on a very high level. This also makesthe biddingbetter becauseit simulatesthe play to find the best or mostprobable contract. GIBusestheBorelsimulations algorithm (Ginsberg 1999): Toselect a bidfrom a candidate setB,given a databaseZ that suggest bids invarious situations: I.Construct a setD ofdeals consistent withthe bidding thus far. 2,Foreachbidb E B andeachdeald E D usethe database Z toproject howtheauction willcontinue if thebidb ismade. (Ifnobidissuggested bythedatabase, theplayer inquestion isassumed topass.) Computethedouble-dummy result oftheeventual contract, denoting its(b, d). 3.Return that b forwhich Y’~a s(b, d)ismaximal. There aredozens of other programs forplaying, generatingdeals(Andrews 1999)etc,fanFrankhasmade end-play analyzer/planner FINESSE whichseemstowork wellbuttooslowforpractical use(written ininterpreted PROLOG) andhehasalsowritten anexcellent doctoral’s thesis "Search and Planning UnderIncompleteInformation" (Frank 1998). PYTHON, (Sterling & Nygate1990) as cited in (Bj6mGamback &Pell. 1993) is an end-play analyzer for findingthe best play but not on howto reach these positions. The former World Bridge ChampionZia Mahmood has offered one million poundsto the designer of a computer systemcapable of defeating him but has withdrawnhis offer after the last matchagainst GIBwhichwasa narrowwinfor him. Most programsbasically use a rule-based approach. The first one (Carley 1962)with four (I) bidding rules with performance like "Theability level is about that of a person whohas played dozenor so hands of bridge and has little interest in the game".Wasserman (cited from(Frank 1998)) divides the rules into collection of classes that handledifferent types of bid, such as openingbids, respondingbids or conventional bids. The classes are organized by procedures into sequences."Siightlymoreskillful than the average duplicate Bridge player at competitive bidding". MacLeod (MacLeod1991)(cited from (Frank 1998)) tries to build picture of cards a player holds but cannotrepresent disjunction; whena bid has morethan one possible interpretation. Staniers (Startler 1977)(citedfrom(Frank1998)) inlroduces look-ahead search as an element of Bayesian planning and LindeR$f’sCOBRA (Lindelof 1983)uses quantitative adjustmerits. Lindel6fclaimsthat COBRA bidding is of world expert standard. Ginsberghas invented a relatively compact but cumbersome (Fig. 3) way of representing bidding systemsin his GIB-soi~vare.By using such representations he reduceda rule-base of about30,000rules into one with about 5,000. Suchrepresentationsare of course proneto errors. The Rules of Bridge Bridge is a card gameplayed with a deck of 52 cards. The deck is composedof 4 suits (spades 6, hearts ~, diamonds <> and clubs &). Eachsuit contains 13 cards with the Ace as the highest value and then the King, Queen,Jack, Ten, 9, .... 2. Thefirst five are often abbreviatedto A, K, Q, J, T. The gamebegins with someshuffling of the deck and then the cards are dealt to the four players, often denotedNorth, South, East and West (N, S, E, W)Fig. 1. N and S play together in a team against E and W. Before the card play begins an auction (bidding) takes place. Oneof the teamst wins a contract to makeat least a certain numberof tricks. Somelevels give bonuspoints. Topreventopponentsfromtaking a lot of tricks in a long suit of theirs, a trump-suitis lookedfor in the bidding. If no common trumpsuit is found, a gamewithout trumps can maybebe played; No Trump(NT), which is ranked higher in the biddingthan the suits whichare ranked6 (lowest), ~, to 6 (highest). Duringthese twophases, the bidding and the play, there is cooperationin pairs playing versus each other with four players at each table. The pairs communicatevia bidding wherethe bids are limited to a few (15) "words", but many sequences"sentences/dialogs" are possible. Opponentsare 2 about the meaning(semantics) of the bid kept informed sequence. This can be seen as a mini-languagewith words like 4, 1, Spades, Double, Pass etc. and these words can be combinedvia simple grammaticalrules into short "sentences" like 3 Spades.The sentences are put in a sequence "conversation"called the bidding. It can look like N:Pass - E:I Spade - S:Pass - W:2Spades, N:Pass - E:4 Spades - S:Pass - W:Pass, N:Passwith semantic meaningslike "I havea handwith spades as the longest suit and aboveaverage numberof Aces, Kingsand Queens"- "I have support for your spades but not so goodhand", "I have someextra strength"- "1 havenothingto add". Thefirst non-passbid is denotedopening.A full bidding schemais presented in Fig. 2. When the biddingis finished, a pair handlesthe final contract (4 spades here) and the play with the cards can begin by taking tricks. After the play, the pairs are given points accordingto the result. In tournamentplay the points are comparedto the other pairs playing a duplicate3 version of the deal, almost compensating the factor of luck with "good cards". In our workwe have studied undisturbed bidding which meansthe pair is free to bid on their ownwithout disturbing interventions from the opponents. Thereis no need for preemptivebids and sacrificing or to keep in mindthe importance of reducingthe amountof informationtransferred to the opponents.By havingthese limitations weare able to learn better in shorter time. Less memory is also needed. Representation A bridge-hand as a humanviews it is seen in Fig. 4. A shorter, standard, representation with almost the samemeaning (losing the graphic"personality"of the cards) Figure 4: A typical hand Card Ace King Queen Jack Highcard points (h.c.p.) 4 3 2 1 Table1: Theestimatedstrength of the high cards. 6AK87 cYQJ52 QQ7 6T96 or, simply, AK87,Q J52, Q7, T96with the order of the suits understood. The total numberof possible hands are 6 ¯ 1011. There are 8192different suit holdings 92, AK82,etc.), and only 39 hand patterns (5431,5422, etc. disregardingsuit-order) (Andrews1999)and 560shapes (5431,4513, 5422,.) in total. To estimate the strength in hand, Milton Worksinvented the high card points (h.c.p.) (table 1). The total numberof h.c.p, in a deck is 40 and the average strength per hand is exactly 10. A hand with AKQJ, AKQ,AKQ,AKQcontains 37 points which is the maximum.Suchevaluations are neededto estimate the numberof tricks a pair of handsare able to take. RobinHillyard(Hillyard 2001) has calculated the expected numberof ~cks at no trump, presented in table 2 basedon G1B-simulations and the results are close to basic rules of thumbused by bridgeplayers. ThomasAndrews(Andrews1999) has made some analysis showingthat the value of the Aceis slightly underestimated in the MiltonWorks432 l-scale. Of coursethis scale is just to makean roughestimateof a handand every bridgeplayeradjust (reestimate) the value the hand, maybedynamicallyduring the bidding due to fit ~ etc. Oneadd-onadjustment invented by Charles Gorenis the distributional strength (Table3) wherethe total value of our hand is 12 (h.c.p) + 1 (doubleton ~) = 13 points. H.c.p 20 Tricks 6.1 21 6.7 22 7.2 23 7.6 24 8.2 25 8.7 H.c.p 26 27 28 29 30 Tricks 9.1 9.7 10.1 10.6 11.1 t Withthe highestbid. 21ftheyask. 3Or replayed later bystoringthe dealin a wallet. Table2: Theaveragenumberof tricks with a given strength. 71 Number of cards in a suit no card (void) one (singleton) two (doubleton) Distributional value 3 2 I Whatvalue of the rewardr should the evaluator give the learners? a) Calculatethe differencebetweenthe actual score to the score achievableon this combinationof hands: Table3: Distributionalpoints. r = score - achievable idea behindis that shortage in a suit makesit possible to. ruff in a trumpcontract andtherebyincreases the strength of thehand.Thesedistributional pointsare normally notadded during a search for a no trump(hiT) contract. Thestrength is used to define a bidding systemin terms of shape-groups and strength intervals. Our hand can be represented as a 4423-shape(4 spades, 4 hearts, 2 diamondsand 3 clubs) with the strength of 12 h.c.p. Thenumberof possiblehandsin this representationis 560 ¯ 38 = 21180. In our approachwe reduce it further by just keepingthe distribution of the cards (the shape), andthe rangeof strength: 4432, medium where mediummight show12-14 (Goren)points. Allowedsystems In most real tournaments the possible systems are limited by assigning penalty points (dots) non-natural opening bids (Svenska BridgefOrbundet, SBF and WorldBridgeFederation, WBF).This is so becauseit is difficult for a humanopponentto find a reasonabledefence within a fewminutesto a highly artificial biddingsystem.A certain numberof points are alloweddependingon the type of tournament. : ---- qsituat,bid q- lr * (7" q- "7 I~ida~ qsituat’,bid’ -- qsltuat,bid) It worksbadly due to high exploration. Assuming a start value of Q withzero "optimisticstart", almostall tried bids will start with negative scores (strange contracts) making themworsethan hitherto untested alternatives. This will lead to a huge, time consumingexploration phaseof all alternatives. b) Differenceas abovebut added200 points: r = score - achievable + 200 (3) A contract not worsethan -200 is denotedbetter than the untested cases. The Pass alwayssystemevolved guaranteeing a score of 200. Maybe a long run will improvethis policy and the pass-systemis a kind of eye of a needle? c) Absolutescore. Simplyreinforce the learner by the actual score: r = score * 10 (4) It workedbest, at least on our short (oneday or two)runs. Thescore was sealed with 10 to try to avoid integer roundofferrors whenmultipliedby the learning rates. d) Heuristic reinforcement. r’ = r - 100 * isToHighBid (5) To avoid high-level bidding with bad cards there wasan extra negative reinforcementon those bids that were above the optimalcontract. The Learning The Q-learning (Watkins &Dayan1992) updating rule qsituat,bid (2) (1) wasusedfor all three (response,rebid, 2ndresponse)bidding positions (whenapplicable) after each deal. Theshape, strength, position (E,W)and bid-sequence is encoded part of the situation and the q-valueswere stored in a multidimensionalinteger array, q is the averageresulting score havingthis handand followingthe policy, lr is the learning rate, 7 the discount rate and r is the reward. The bid with the highestq wasselected in a particular situation unless exploration was actual, on which any bid was selected with equal probability. If several bids had the sameq-value, one of themwas randomlyselected. In practice this meantone out of 2048opener werecombined with one of the 2048respondersaccordingto the current examplein the databaseand they had to learn together with their partners; thence multi-agentreinforcementlearning. The pair’s q-values wereupdated whenthe bidding was finished. This wasrepeatedmultiple times. Reinforcement signal Somepreliminary tests with the rewardwasdone. ?2 Discount/Decay How far-sighted should the bidders be’?. Byselecting different values for -y in the Q-learningdifferent behaviorswas obtained. A lowvalue makethe learners prefer a short bidding sequenceand resulted in guessinginto 3NTmoreoften. Maybea value higher (not tested) than the normallimit one will encouragea moreinformative, long sequence?Eligibility traces wasnot usedbut the result after the bidding wasusedfor all pohitions. Results Someof the ideas above were implementedin a program called "Reese", after the famous bridge player Terence Reese who died 1996. He was knownfor his excellent booksand simple and logical bidding style. Regardingour bidding programthere was hopefor finding a super-natural bidding system but so was not the case. This programrequires a computer with at least 348 MBof memorywith the described configuration. The experimentscan be seen as what can be achieved with these methodsin a complex two-learner task and also gives somehints of howto chose learning parameters. Introduction By using the huge databaseof GIB-results whenplaying againstitself for 717.102deals, wefounda reasonable evaluative function. Thedatabasewasgeneratedusingan early versionof GIBbut both the declarerandthe opponents were not so strong so maybe the result is quite reasonableanyhow. Wehave used the 128 most commonshapes which includesall handswithat mosta six-cardsuit exceptfor the rare 5440, 5530, 6511, 6430 and6520shapes. This covers about91%of the handsand, neglectingthe inter-handshape influence, 0.91.0.91 = 83%of the hand-pairs.Thestrength wasclassified into the following16 intervals(althoughthis is easy to modify): Class 0 1 2 3 h.c.p. 0-3 4-6 7-8 9 Class 5 6 7 4 10 11 12 13 h.c.p. Class 8 9 10 11 h.c.p. 14 15 16 - 17 18 - 19 Class 12 13 14 15 h.c.p. 20 - 21 22 - 24 25 - 27 28 - 37 Theintervals wereheuristicallyselected in accordance to frequencyof appearing. To reducememory requirements,learning time and simplify the restriction of learninginto allowedbiddingsystems, the openingbids wasdefined by the user as a small twopagetable. Partof the definitionof the openingbids in bidsyst.txt is (the strengths9, 13, 14 and28 wereremoved here ~r ease of showing: Shape 6313 6241 6232 6223 6214 6142 6133 6124 5521 5512 5431 5422 5413 5341 5332 0 4 P 3S P P 3S P 3S P P P 3S P P P P P P P P 7 i0 2D 2D 2D 2D 11 12 1S 1S 1S lS 1S 1S 1S 1S 1S IS IS 15 IS IS IS IS IN 16 18 20 22 25 2S 2S 2S 2S 2S 2S 2S 2S 2S 2S 2S 2S 2S 2S IS 2C 2N 2S Thebidsyst.txt canbe rearranged (preferablywitha program)fromthis matrixrepresentation into an inputfile with examples.Eachstrength 0..37 is used to generatean example, resultingin 128 ¯ 38 = 4864examples,whichlook like this (high cardpoints, number of Spades,Hearts,Diamonds, Clubs, recommended opening-bid): 0, i, 2, 3, 4, 6, 4,2, I, pass 6, 4,2, i, pass 6,4,2,l,pass 6, 4,2, l,pass 6, 4,2, i, pass 73 It canbe represented as a decisiontree or rulesetbut this introduceserrors (due to somepruning;about2%is erroneous),is larger (five andten pagesrespectively)andconsiderablymorecomplexto read. The See5, inductive-logic, sotbvare (available via http://www.rulequest.com) generatesthe following(slightly modified)decisiontree: See5 [Release 1.13] Wed Mar Options: Generating rules Fuzzy thresholds Pruning confidence 4864 cases Decision (5 attributes) 14 2001 level from 50% open.data tree: pts in [20-38] : :.s in [5-6] : : .pts in [25-38]: 2S (390) : pts in [0-24] : : :.pts in [0-21] : : : : :.s in [1-5] : : : : :.d in [5-6] : 2S (4) : : : : d in [1-4] : :.h in [5-6]: 2S (4) h in [1-4] : :.h in [1-2] : :.d in [1-2] : : :.d = i: 2S (2) : : d in [2-6]: : : :.h = I: 2S (2) : : : : : ¯ h in [2-6] : IS (2) . . . . . d in [3-6] : : : : : : :.h = I: IS (4) : : : : : h in [2-6] : : : : : : :.d in [1-3] : 2C (2) : : : : : d in [4-6] : IS (2) : : : : h in [3-6] : : : : :.h in [4-6]: IS (6) ... (5 pages) And the ~llowings~ ofrulm(mycommen~ on fight of the *): Rule i: (512, lift 3.9) pts in [0-3] -> class pass [0.998] * Pass with a very weak hand Rule 2: (132/2, lift 3.8) pts in [0-II] s in [1-3] h in [1-4] d in [5-6] d in [1-5] -> class pass [0.978] * Pass even up to average and 5 diamonds Rule 3: (39, lift 3.8) pts in [0-12] s in [4-6] s in [1-4] h in [2-6] h in [1-2] d in [1-3] -> class pass [0.976] * Pass with up to average and 4 spades ... (i0 pages] Theprogramuses the original bidsyst.txt andit is doubtful if the tree or the rules canbe useful to a human reader. Noopponentswere taken into accountexcept for the specified openingbids whichincludedpreemptivebids like 2 diamonds(showing a strong hand with diamondsor a weak hand with a six-card major) and 3 in a suit showinga weak handwith a six-card suit. Thedatabase contains a vector with the actual numberof tricks taken in any suit or no trumph.Wehave assumedEast to be declarer in all cases due to performancereasons. The maximum numberof bids in a row was set to four (a pass wasautomaticallyaddedas a fifth bid if applicable) and the numberof possible bids waslimited to sixteen. Learningrates wereencodedas integers and shifted right ten steps, effectively scaling them downwith a factor of 1024. Explorationrates wasinteger codedin parts per thousand. A T-value of zero wasused and the initial values of the Q-tables for east and west respectively was a command line argument. The exampleswere separate! in a training and a test set and training exampleswere selected in randomorder to avoid biasing effects. The desired numberof exampleswere selected in order from the database except for not representable uncommon shapes which were simply skipped. Exploration was not used during test-evaluations but used duringtraining evaluations. Thescore wascalculated from the numberof tricks taken and final contract. The vulnerability was "none" and the dealer was always East. Undoubledpart-scores, games, small-slams and grand slams was calculated and contracts going more than two downwas assumedto be doubled. Wetried to find the best possible biddingsystemby doing three experiments;first a surveyrun with different parameters, a secondsurvey with somenew,better, parametervalues and then a longer run with the parametersestimatedto be best by the humanobserver. The parametersto be selected werethe learning rate (common for both hands), exploration rate (also common) and initial values for the Q-tables. In the first run weusedthe learningrates 5, 10 and20, explorationrates 5, 10 and20 andinitial Q-values-100for East and -200, 0 and +200for West. Eachrun was repeated three times with different randomnumberseeds (13579, 5799and 8642respectively). This resulted in a total of 3,3.3,3 = 81 cases, Ten million training exampleswereselected in each epochand 200 epochswas run in each case. Eachcase took about 41 minutes and the total CPU-time was about 55 hours on a 1.5 GHzIntel Pentium IV running Windows2000 and equipped with 512 MBytememory.Each training occasion required about one microsecondincluding evaluation. Every 74 Lr = 5/1024 Lr = 10/1024 Lr = 20/1024 e = 0.005 27.2.4- 8.7 24.9.4- 1.8 33.8.4- 3.6 ¯ = 0.010 27’.6.4- 0.2 40.4.4- 6.1 50.1.410.0 ¯ = 0.020 30.7+ 9.9 41.7+ 3.7 63.6 .4-3.1 Table4: Theaverageof the last ten test-evaluations, averagedover three runs. q0 is zero " :,- " =’ Figure 5: Runswith Q0= -200. The smoothcurve is for the training set and the ruggedcurve the test-set. 200epochs, each10 miijon tries, wererun and repeatedthree times to be averagedinto these curves. 99th epoch,a biddingsystemwasemittedinto twofiles; responses.txt (115 kB) and rebid.txt (525 kB), the second sponsewasnot stored. Thegenerationof the rebid-file takes the openingrestrictions into considerationto reducethe size. A log-file (6 kB) with the time, average training-sum and averagetest-sum wasgenerated and the standard output was redirected into a file (1.4 MB)with the actual bidding sequencesfrom all the test-cases for future reference. About 2MB*gI=I62MB of disk space were u~ed. 550,000 training examplesand 200 test examples(positioned from 710,000 and forward) were used. 109,940 (16.7%)were skipped, reasonably close to the assumption above(83%kept). Selecting Learning Rate, Exploration and Initial Q-value Thefirst run was madewith a qo of 0. The results are shown in Table 4. Thehighest value for both of themresulted in the best performance; a averagevalue of 63.6 -4- 3.1 with an learning rate of 0.019 and an exploration of 0.020, denoting a 2%risk of selecting a randombid. The experimentswere rerun with a qo of +200, meaninga untested alternative is worth 200 points. Becausemost bridge-results are in the -100..140region, this is the optimistic approachresulting in an (more)exhaustivesearch and slightly worseresults. A third run with qo = -200 where done with better resuits. Thisvalue (-200)results in less searchof alternatives becausea reasonableresult like -100is preferred to untested alternatives. The training and test-curves (Fig.5) seems Lr = 5/1024 Lr = 10/1024 Lr = 20/1024 ¯ = 0.005 29.3 :h 5.1 28.4 + 5.6 27.2-4- 2.4 ¢ = 0.010 36.7 -4- 13.6 40.9 + 6.2 48.3 ± 8.3 e = 0.020 51.4 ± 6.9 54.6 ± 3.2 68.7 ± 4.9 Table 5: The average of the last ten test-evaluations, aged over three runs. q0 is -200 aver- 150 100 ! ! 5O Figure 6: A 5-dayCPU run of learning. Lr is 20/1024, e = 0.050 and qo = -200.The maximumtraining case has a score of 104.89, max test-score is 121.7 and average on the last ten is 94.65. follow each other but at a different level. Thetest-curve is almost always below the training curve although no exploration is used. The training-curve is also more smooth due to the large number of examples (10 million, some identical). The best result (Table 5; 68.7 -4- 4.9) was also obtained with the highest parameter values; Learning_rate = 20/1024 ,~ 0.019 and e = 0.020. A long, single run with lower learning rate but high exploration is shownin Figure 6. The levels seems to more or less stabilize around 100. Some of the interesting examples (many removed) from the bidding are presented in table 6. At a later position (1000) under the same training conditions, a bidding system was emitted. Only a very small part of it (part of the responses whenpartner has started with Pass) is presentedin table 7. It is illustrative becauseit indicates what suit to open with and which strengths are associated with a particular shape. Manycases are innovative and maybe unknown to most human bridge-bidders. Some examples are: opening with only seven h.c.p, whenhaving six spades, two hearts and a four card minor (club or diamond) suit, opening with the lower of adjacent suit if weak(then pass) or strong (continue bidding) and higher suit if medium strength, pass if halanced and less than fourteen h.c.p., open one or three NTwith strong hands and any shape, 2 hearts and spades(stronger) shows a good hand with eight+ cards the majors, open with nine h.c.p, and unbalanced with sixcard major or diamondsor when having a five-five shape,.. 75 Deal 2 5 6 7 13 16 24 26 27 28 29 33 34 42. 52. 53. 54. 56. 60. 64. East 2614 2434 4522 3145 1543 3514 3334 5152 4243 2434 3613 5134 3514 3334 2623 2254 4342 3163 4144 4441 Hp West Hp 16 2353 15 IH 9 5323 14 P i0 1255 13 P 7 2434 15 P 9 5332 16 P 12 3253 14 IH 6 3343 14 P 4 3613 i0 P 2 2641 22 P 12 4243 12 IH 13 3262 15 IH 6 3622 Ii P ii 4135 13 P I0 5512 13 P 10 2236 7 2D 13 2632 I0 ID 2 3514 ii P 12 5422 15 ID 0 3325 17 P 9 5431 ii P E W E 3N 4N IS 2C IC IS IH IN IS 2C 3N P ID P IH P 3N P 3N P 3N P IH IN IC IN IH IS 2H P 3N P IC ID 3N P IN 2H IS IN W 7H 2H 2C P 3N 2H P 4S IH 2S 2S Score 1510 110 ii0 90 400 430 -50 110 -100 -50 520 -50 150 480 110 400 -50 430 -100 110 Table 6: Someinteresting examples of bidding from a reasonably good run with Lr = 0.02 and exploration set to 0.02, q0 is set to -200. Anaverage score of 87 on the training set and 100 on the test set is achieved after 562 epochs. Response 10 3613 IH 3541 P 3532 P 3523 P 3514 P 3451 P 3442 P 3433 P 3424 P 3415 P 3361 P 3352 P 3343 P 3334 P 3325 P 3316 IC 3262 P to P: ii 12 13 14 15 16 18 20 22 25 IH IH IH IH IH IH IH 2C 2D ? IH IH IH IH IH IH 3D 2S 2H 2S P P P IH IH IH IN 3N 3N 2C P P IH IH IH IH IH 3N 2C 2H IC IH IH iC IH IH IH 3N 3N 3C ID ID ID ID ID ID IN 2H 3D ? P P P ID ID IH IN 3N 3N 3N P P P iC IH IH IN IN 3N 2H P P iC IC IC IH IN 3N 3N 3N IC IC IC IH IC IC IH IN 3N ? ID ID ID ID ID ID IN IN 3N 3S P ID ID ID ID IN ID 3N 3N 2D P P P ID ID ID IN 3N 3N 3N P P IC IC IC IC IN 3C 3N 3N P IC iC IC IC IC IN 3N 3N 3N IC IC IC IC IN IC iC 3H 3N ? P ID ID ID ID IN 3N 3N 3N 2N 28 ? ? ? 3C ? ? ? 3N ? ? ? ? 2S ? ? ? 3H Table 7: Some of the responses when partner has started with pass, showing a weak hand without extreme shape. Question-mark indicate unknowncase (no example). Epoch 849 951 1074 1221 1413 1683 Training-level 84.7 83.4 79.1 67.3 45.7 35.6 Lr 6/1024 5/1024 4/1024 3/1024 2/1024 1/1024 Difference 170.7 204.8 256 341.3 512 1024 Thetest-set evaluationsare quite close to the training evaluations, indicating that no over-fitting is achievedyet. An initial qo value set to a reasonableloweracceptablethreshold seemto steer the learninginto realistic areas. Goodruns with decreasinglearningrate wasnot possibleto achievedue to the integer nature of this implementation. Theresulting biddingsystem,studied in detail is also interesting. The responderhave learnt, "understood",that an openershowshearts or spades or a strong handwith the two diamondopening bid and thereby never passes and do not bid twohearts with heart support. The no-trumpbidding is also interesting; two-bidresponsesseemsto be sign off, two NTis neverused and three in a suit seemsto be highly conventional and never passed by the NT-opener.Three NTis often the response with mediumstrength hands and not an extremeshape and four in suit is a transfer bid. Maybean interesting NT-biddingpart of the systemcan be developed if the learning wereconcentratedon this? Anyhow; the learnt systemis not practically useful until the defensive bidders are put into consideration. Moreinformationwill be available in myLicentiate thesis (ongoing work)including a Genetic approach. Table 8: Integer round-offerror 150 i 100 Figure 7: The scores are actually decreasingstep-wisewhen the learningrate reduces. Reducing the Learning Rates The single learner Q-learningtheory says that convergence is assuredif the learning rate is reducedslowlyenough.An attempt to improvethe average score is done by reducing the learningrate witha decayfactor. In our cases, the learning seems to drop offat about epoch 150. By setting the learning rate to 20/1024at this point a experimentswasrun with a slow decay (Figure 7). There are jumpsin the performanceand if viewedin detail those jumpsare positioned according to table 8. Whenthe learning rate steps down to small values, the required difference betweenscore and q-value has to be at least one whenscaled with the learning rate accordingto EquationI. The first noticeable effects appear whenthe learning rate is 6/1024~-, 0.006 and the required difference is then 1024/6~ 170.7. In bridge termsthis meansthat the learner is not able to see the difference betweenpart-scores and concentrateson the large contracts like gamesand slams. Whenlearning rate gets below 3/1024,the game-contract abilities are also lost. This problem with learning could be avoided by using floating-point representation but this consumesmore memory (4 times in this case; 384,4 = 1536MB)and requires more computing time. Conclusions and Further Work The runs showsit is possible to find a reasonable bidding systemin a limited time. A relatively highlearning rate combinedwith highexplorationresulted in the best performance. 76 References Andrews, 1". 1999. Double dummybridge evaluations. http://www.best.com/thomaso/bridge/valuations.html. Bj6rn Gambllck, M. R., and Pell., B. 1993. Pragmatic reasoning in bridge. Technical Report 299, University of Cambridge,ComputerLaboratory, Cambridge,England. Carley, G. 1962. A programto play contract bridge. Master’s thesis, Dept.of Electrical Engineering,Massachussets Institute of Technology,Cambridge,Massachussets. Frank, I. 1998. Search and Planning UnderIncomplete Information. London:Springer-Verlag. Ginsberg, M. L. 1999. GIB:Steps towardan expert-level bridge-playingprogram.ProceedingsfromSixteenth International Joint Conferenceon Artificial Intelligence 584589. Hillyard, R. 2001. Extending the law of total tricks. http : //www. bridge,hotmail,ru/fil e/Total Tricks.html. Lindvlof, E. 1983. The ComputerDesignedBidding System - COBRA.Number ISBN0-575-02987-0. London: Victor Gollancz. MacLeod,J. 1991. Microbridge - a computer developed approach to bidding. Heuristic Programming in AI- The First ComputerOlympiad81--87. Prorok, K. 1999. Gubbeni I§dan (in swedish). Svensk Bridge3 l(no 4, october):26-27. Stanier, A. 1977. Decision-Makingwith Imperfect Information.Ph.D.Dissertation, EssexUniversity. Sterling, L., and Nygate, Y. 1990. P3~thon: An expert squeezer. Journal of Logic Programming 8:21--40. Watkins, C. J. C. H., and Dayan, P. 1992. Q-learning. MachineLearning 8:279-292. Winston, P.H. 1977. Artificial Intelligence. AddisonWesley.