From: AAAI-86 Proceedings. Copyright ©1986, AAAI (www.aaai.org). All rights reserved. A Case Study of Incremental Concept Abstract Applicat,ioli doltlailis of niactiirie promises search t~ff;~ctiverlcss in complex of iricrorrierital, induction domains effective requires lacking. for ctiaracterixirig iriducLior1 wllicti syst,enis ltlarriifig. relate ‘t’he dimensions 1ivtl Irlt:rit,s of 4 iricrerrleritat effert,ive rlific-;Llktly used IlI3. iriductiori det,racting lo compare This coniparisori the quality cremerltal concept intensive (deplh-first which and/or objoct,s makes concept tive techniques rrial hypothesis is obtairled. the of update, of respec- reduce learning keeping without sig- of induct4 knowt- require and more a search isfici?i y. Introduction sporist~ process titrle precludes tid preclude OII t Iit> problem of concept, esuttlples, comeptutrl of concept induction ttltly rquire frorrl for rc~;ilizClt.iori that irlcrt~rrlcrital riurllber 1~6). as new basis pararrlount SctlIlIrllll. world x tIIJIIl~~, 01 ( II rrcarit learriirlg 496 / SCIENCE for rcactirlg t,rlvirt)rlltlr~rll,s lW5) which sys t t>rlis. white on with prorrlisth to in- pTYJl.'t!l-tie.5 duct.iori for st,orcb 111ay be to nt’w sLirnuli; Lhe ticvclol)rrle~lt X: tloocl, ltlc As iliterest, ttlc ~lorl- rllot,ival,iorl Lo pustl wtlicti Simon wliicti possibility places car1 allow are of tligli hypothesis, has or opti- This optirrial such saton re- solutions, hypotheses. rapid control termed comt.raints for optimal for a will generally in search (1960) of (i.e., a stable of obtaining scharcti strategy irlcreasingly Mo- usable is (‘rl(.ollrlt,erc~d, (~1arbor~ctII the a sat,isficiIlg on fOr sal,isfactory Lhe explicit trypottieses t,o clcal of‘ obsc~rval,iotls, a knowledge otily inforrnatiou systems to solutions, solutions. In fact, update rrlerilory and does and quality. t,tiat of tirtle. arc’ rtquired iIISt,ilIICCa iJrol”:rty 5irrl~Jl;ilc4 a span t hc: primary ~1151~1iriirrg ,L corit.iriual is btbcorliirlg iri is Lo occur t)(a c.orri~)~~tatiorlally is 1,hat c;Ktl rnaj0rit.y triIigc& prilnarily divctrsity riot Spt~cific~ally, induct,iorl updat4 the executioll, over systerrls arid rrlay far, induction sy5t.clrrls its learnirlg syslthrrls incrchrrlerllal rai)idly objects but learm’r~y Jrom tlorlincre~rleP1fuI, which 01’ systcrn acccspt ir~crcrrlt~nt.al wil II in great,c:r (,Ilic tlalski, over significantly (e.g., ‘t’hus arc syskrrls t.tie outsel (.rett1etltul systems t,iv;Ltioils irlducf,ion clusterr’rly). all objects t)tl prtw~11 has corlcerltrated seek luxury of the correctness searching a scarctl or opti- the A reduction which erivirori~rierit, At1 riccessil.att3 learning tiypoltiesis. exhaus- a correct systems equivalent to converge instances of hypothe- precludirlg tile guarantee of a filial that Iricrernerital objects 111rty sacrifice or of hy- tfowever, lncrcmentat or search is converged frorltier costly. thus be observed hypothesis guarantee instances to a frontier the incorporation cost past, stable con- Nonirl- breadth-first, previously servca to expand frorit ier of hypottit5es). yields in machine of generally tend maintaining some New rriality n’ork urltit of the update. backtracking, requires a list 1982) rapid systems with which CTlgt~. I induction systenls as a result perf’orrning ses indicates car1 be obtained, frorrl by of illcrernentat emerge OH. lhe of Quirilari’s mandated (Mitchell, incre- cost anti qualily to the straints which we intro- incremental variants program, Irol~~ t>x;llrlples I tlat, cost are paper cost advantages the di5advantagcs potheses discussion of diffcrirrg with come versiorl-space) the developrrlerlt 111 this (iuc(~ :j dimensions of nofi1,earning However, I,he utilily teas beers corriplcx in limits methods. methods. fi>r corriparirig rr~e~~tal melhods l,cx-tlrliques the computational ilitensive cost 01 ciinierisioris inductiori to push incrc~rnerila.1, Along Induction tcor irisLance, vt~rsiiofI space) 1julul107l. t,tiis differeritiatt: 1985; lirriits quality) not. limit atiy c>xhaustive cari bt: irlipI(~r~icrlted iIlc.rc!rrlt,lltally, iLed ulilit,y concept arid but suctl rrlour~ts, explicit we discuss incrcrnentat and WCII its serving rrierital systm~s. as a basis These accepted scarcti tectinique several for evaluating are rliay have dirrlerlsions learners, competing to at a (e.g., instances irlcrerrlental riorlilic:rerrlerltal dimensions one so iib Lo accept, ;tri irIl~~I~jr~t~ritat,ioll ttlis paper iri- characterization are it1 an c~r~vir-o~irIlerrl tit~rnaridirig trl it becorlles thra cu~r~pututiord of increrriental their property ttlat. objects t,hta l~el~uvwrd tirric. learning to Itlake cost, (e.g., Wctijiiqut5, lhUS of in irlcrernerltal irrlportarll lirncufllwhich as incre- The number o/ obseruatiotw l tem to obtain a ‘stable’ required by a learning set of concept sys- occasioiial noisy descriptions. every 0 The u/ u$utifly c o st served objet memory to accorrirtlodate an ob- t. two tree of currlulative expended izing can cost, during which irlto the A last induction quulity ‘I‘ht~ reflects Iearning. iricrernental 0 be corrlbirled a single amount a high derived by a con- of instarlcc3 tal discussing arc these used variants tcrrl of the rriaririer dinrensions. to compare is of the builds decision of t hesrb systems indicates object without rived arid cost implies cau decrease In general, an increase in the entries each distin- number re- quality of tie- rrieritcd approprial of observed objects not a root inforrrialiori II A case study: ID3 1heIi arisen Quirllan’s tree that (1983, 1985) distinguishes 1113 constructs between examples ples of a particular concept. The au tree a collection empty decision riorit~xarriples nurribt~r of those applied The ttlr rrlost to each of the informative value are then for this plied positive or negative. of I tit: coiict3pt ‘l‘tl(b choice root:, II)3 whicll is critical uses an attribute cBac:tl sllt)t,ree. starts and is red). A is used to form a brarich for each thus groups the buildirlg At this point, bc~twecn will the root ot ‘I’tlcl to t hcbir is recursively subtrees. examples urltil eitlker all iristaIrct:s itive or negative) ‘I’able 1 dt3pictS root. ap- pm- tht>n t hc iIlf0rIIli~- ‘I’his attribute as tile root is attl‘il)utt~. ‘I‘he process sul)trrw aw of’ canriot, foI]OWtd l.hC conI iriuc3 orit2 lype relial)ly be irl is tc~rrried gloar~eti over rlur~ib~~r ofsut~stqut~r~l cali credit root (pas- choscri. IliOdifit’d vf:rsi()li I 1)4. previous be rccliost~n. inslarlccs i~i:,t;~rices rrlorc’ stabl<~. ttir> roots iIIlpOrtiirIt previous k.ul,l rW card t r(‘e corn- I,ht> 5111ic titx isiori occurs IIltmur-e <irid ctinllgcs arc) r:il.hcr ( tiosc~I1 root ill rtquires rtnd t)clf’ortl ‘l‘liib prot.c+s of I llc iriforrrlaliorl progrmst3, a poorly 01’ a su bl,t.t~c tliscartls tlxarrlirling irit‘rc’clut’rllly, lleuristic.. (highcbr (stiap irllorrrlatiorl :\s a roots (Iw.~I)(~Tsubtret~ lo the I~~;trIiirig irl t tic t rtY>) bc~corlit~ iri su Gl rt’tf root of’ aI1 ctff>V.t. lri ttlta lollowirig ‘I’hc proce~ arid arc: both if it, is ur1lik~~ly to Iiavta tbrnpty. root, llle dot3 the rrlosl corriputc~ ttlc which instances, it is cllost~rl aI, a is irlc-rcldowll tticre attrilutc5. c.llarIgirlg the the exanlplcs to is Ith Sk[JS allows (:harigirlg in a subtree decision used or a ri(‘w the proceeds arid with positive> was all ~xarIipl<~ is t~ncouutt~red for classilied rriodific~atiorr II)-1 also 3d). examples. of its values. process I.liis irlstarlct tIlerI suril- irlslct11ces tlie ill of and is processed, 1Ike k2 ksl,; this of 1113. This negative ail ‘I’htt ht>arl attribufcls arid at1 ributc, thctri by chance, iIi by a rrieasurc according urilesled uriused usilrg ion tirritb loc.atcd test Otherwise, a value how at tribute into with arid is described color) negative and out to determine positive all of the plet,t~ly discriminates size attributes Jlvided group, until shape, size, rloncxarn- of exarllples and attribute, for each continues with algorilhrn (e.g., between (14acisiori tree instdnces (e.g., attributes t lley discriminate and tt>ac:h object corlcc:pl,. of attribut,es for each is of a arid a discrimination ii consists If’a 511btrtIe previously evali~alcxl al, of tables root. ills1 aric(a rIirsasurc> is tivtt of the to rrlodificaf orie 14:ach tabIt coilrits negative ontk 1 is olm~rveii. lies in a series Classilicatioli have arId increrri6Lri- triaIllit:r. of all (Asu bLrt>e. yet itive 1irritl. ilI1 in is arntlrlable to wlictllr~r or rionexarrlplc. ~iurt~- ;iL. oue for caacti of its attril,iitc~-values accordirig updatcl a largct lo be prcsetitc~d proc~twd of positive count II)3 of’ ilic.rt~rIlcrltal As a new or negative reducing tree. valur3 thra riurrlbclr val lit?. wlic~rc proct3sirlg objects iitl d~~c~isiorl tree for ItIc each aua lysis which in the potent of in- (1985). I I)3 as eat II objet I,(, c.orril,iltal,ioriall~ rriarixm analysis be considerably however, an ‘optimal and which rt’ruli Irid) l.lli:, to insure iI Quirllari for be to allow iIIsl,iirlcc5 is us4 thcb sarriplirig ‘tl algorittlrrl dvailltblc of tticsc> rrioclifications encli tllat cdn tw touuti sirrlply iu a dtaci‘1’0 limit A 11101c de1 ailchd account t hc I I)3 franlework Ltlat l+:ach sys- A formal by an empirical incorporation trees. to find objects so efficier~t, irtc,rerrlerital variety, instances. a significant decision reqllired observed for dirrlerisions 1113 program. examples negative is bolstered that duct4 over from the of several 1085) frorrl trees positive behavior (1983, learning on a cast‘ study Specifically, the of Quinlan’s guish focuses test 111 a betwcerl resultirlg rr~t~ltlod of applyirlg would at a tirrle, paper are instances. large. statistical to cllarlce. force’ llowever, ‘1‘11e rclrnaindcr irlstanco, d e b~rt7’ _ _ 0.f c0nfidcrit.e) is rrot due Orit’ ‘brute system. concept to discrirrlinate t,c> uriricac.msarily 11)X is a rloriillc.rcrrlclrIt bcr uf concept cep t induction for character- is: in the atterript rictgativct rndy II):< arici its rrlcasIIL‘(3 of resources diniension systems mjl’se) a x2 (clli-squared) stances measure arid which growth, factors (or I I)3 will positive siori (with These errors situation, ctloic,c3 tla\‘t‘ loh!~ ilIlillySC’S. II) 1 \Cd5,11)1(1to tiis- I ril)ul.c~ t,lloic (3. ;trlcl t.OJiVt’rge 011 LIXY ds I IJ3. rlon~~xarrlplcs to be acquired. of a rrieasure if good for selectiug decision irlforrnation values 1113 has lhcoretic besl also discriniirialiou trees divide are rri6lasure all beer1 designed trc’r to be ot,lainetl. ot)ject to determine (sub)set at to acconlrlloclate LEARNING / -tc)’ is Inputs: A decision tree, One instance. A tlecisior~ tree. 011tp11t: I. If this instance is positive, increment the total number of positive instances. Otherwise increlnent the number of negalive iristances. 2. If all of the irdanccs tlie decision are positive Compute then return the expected negative If thcrc information scores or Lhe maximal aLtribuLe is riot the a new tree. a test link from the root the root for every then of updating In 1113, to update tree 1 : I’seudo is augmented code (I f?j ;Lrld 16). st!rvt’s to r>rnpiric.al an refine for increrrlerital cart effecting an enipirical incremental ‘I’he introduction the analysis instance ca~illy with additional space indicates quality variants where /II is the and number d is the that, analysis, are during the cost reduced tree number (which of cannot . . . . for a total after every instance, over a single object, then the above over two objects, of and without, signifi- Asyrrlpt,otir:ally, of learning. where the required r~urrlber rooted at node j of’ level the t~unlbc~r of instances of i. objects lcor an i, of t,tie decision tree, r~, ( JL, TL,,) represents of objects required for all nodes at that level a stabIt! discrirninat iI)g altribute. A level cannot until prtavious levels have achieved stability, and all instances seen, the t.hus 7t, ;> 71, 1. Since 1113 retains uurrlt)t:r of objects to c:onsLruct a decision tree of depth d / SCIENCE IA ( is lhe of the of incorporating suflic.ic!rit size, ql, rr~usl bc seer) for It)3 to choose the root at.tribute wllost: values best discrirrlinate objects of the e11viroli~t~crit as a whole. This is true in the creation of all 498 depth rrlcthods techniques, Arr irilportant cotriput,atiorlal measure is the number II):3 of ir~sl~a~lc:es required to c0tisl.ruc.t an optimal tree. choose::, an alt,ribute to form the test for the root based on t,l~ts information that attributtl contains over ttie observed (from the environrric~nt) of A sample of objects instancc5. Lo attain stal)ilk of instances, introduced most iruportant is presuluably term much is ) /I2 since greater than the the of attributes.” In 1111, building ire level, t,ht: 11urrlber frorn j A I). t,he nurriber attributes, cl111 tree a-- (‘4 / riurribcr subtrtlt: roots as well, is 71,, lor the subtree a new PI (1 exceed 1114. of thcxse l&or of incremental be signilicantly the build attri butt:. If a tree is built expense is incurred whic,l~ two is to each node of the tree requires that to determine their values for previThe cost of constructing an entire is attributes, allalysis required to construct choice points, or rnerrrory a tree scratch. CotistructiIig instances be examined ously unused attributes. value of Go to step I with ttre subtree found by following ~lle link for the root attribute’s value in this instance. Table ID4 for all attributes. i. If the maxinlal attribute is x2 tiepentlcnt make it the root of this tree. ii. Make Cost B. is no rout build sample, in order to choose the root attribute in an optimal manner. llowever, because Ii&i does not store all instances encountered, at the next level it must examine mother rll instances because the first n, instances are not available for inspection. Con- for each value present in the ineither the number of positive or the information then object score. instances. Compute a representative same n o instances sequently, the number of instances the tree is the bum of all of the root For each attribute, stance, increment root, or negative tree. Again, assumitlg must examine the a decision of ot?jects times as shown below tree the is pr.oporLional square of the only number to of of objects this to an efficient is substituted into characterization the cost is rod 1. equation When for 1113 we have rank or the sq~arc edge, file, diagonal, tyPc? or otherwise) where or ot,ht>rwise) each (6 attributes), piece (4 attrihutts). resides There and (i.e., corner, art3 a total of six- teen attributes, each with three values. Although there arts fj” i\ (;” h 4” ._ 2, 985, 984 objects possible, ail exhaust,ive enurtreration For 1114, the number of instances tree is larger: 27::: 71,. Substituting sion for object incorporation yields to an optimal decision this into the expres- that these attributes. Four are which stances total required expense to select 71; 1 :- >;f;,’ Tli then is very likely quirc~tl the of the Our 0 bjects empirical ari the subtree number root expensive is probably attribute. than number of inIf 1114. ‘l’his of instances greater re- than the or TL~ r > d. assurned discussion sample of regularity in an environment, we now of and actual are The constructs new only reconstructs been misclassified. instance instance. counts tested tree third variant fourth were randomly is depicted tree 6‘3, instance only has counts of new updates at- is made hoards presented formed in figure perform after for each generated were algoof 1113 scratch updated 164, of version, au in classification positive) decision from is 1114; the variant indi- II)3 version when are an error ‘1‘1rc same The variants. the of the force tree decision pins in terms A smarter instances when ilar to IT,). variants decision The knight objects is a brute is received. the I~‘inally, tribute first rregative and distinct 3,251) actual a new each positive 95,480 orily incremental tested. of the instances a ‘representative’ a rigorous of objects on the tree tree, has lacking distribution since the deecision analysis each 11)s is more case, to construct depth hinges there behaviorally rithrn Comparing of the cates (sim- (KS 69% to these four by ull of the variations 2.’ ’ Idist-bk-knight1 analysis. dist-wk-knight IV (:onsider a hoard ation the task position, and of classifying task versus safety or loss of the black to move. chess attempts Following attainment knight performance a classifier as a win or loss. a corrcept king Empirical Figure (1979), we defjne Quinlan king knight whether and rook or king 1 depicts Given the situ- as determining a white black endgames. to identify results in two a sample a black in the moves board diag/rectl \pther t 0’ ‘0t 0 Figure with 2: IIecision For the three observations the least over more for nurrrber ple picture knight. Iloards were randomly (6 attributes every pair generated (in squares) and bcttween of this type), board of pieces (i.e., whether described each pair relationship they lie on there quickly forms an Figure variant t1ut nurrrhcr ranges of from 3 depicts 2 (averaged as a function a rough should 11ot and tw 01 sim- equated builds more a coniplcte tree, while _instances. 111-l rtquires the on the complete tree after timc each irrstances. is a substantial its ttic in figure gives consistently 20,000 Classification in terms speed, substantially constructs tree l>epth _-- variant aiits of’the distance pin. tree for f-64. by each ID3 rapidly converging Though algorithms, decision built of learning approximately black of the of instances. 1 I)4 requires pinned knight this decision 11% to ttle greatest depth correctrress. largest, of a safe, to forrn 50 execulions) the with efficient required the average I: Example for a safe configura- tion. l<‘igure tree decision etftlctivc perforrnarrctt range tree, classification of tlie three iI1 the each of the for more variants t,htl instances. efficient vari- (averaged instances. over 50 executions) was measured over 1000 _-_ k‘or II):{ and 1114, a 90% effective classification of pieces between the same LEARNING / 499 DEPTH OF DECISION TREE COST PER INSTANCE 30K / Ii% i ID4 / / / / I , ID3 20K / IOK 250 500 750 INSTANCES 1.000 18,500 63 PROCESSED 250 is Ir)r’rrred after ~~t:r Iorrnance as few as 100 instances. of’ 164 rt’d( tit3 75’i’: correct !#‘I c-orrect t I~ough it sp~etf seri first, which is due tests for to the of order al- tIstablisheci learns by JIM. form of the until In figure an four of updating the cant II Iitfw instance. corrll>;lrisorks ~xec’ut ions coricept number decisiori depicted. are of comparisons deeper Icigures per irddfic-e perforrried) it rit’w clccisiori tree desc.riptioris of corriparisons do- in the 4 aritl for each after each is measureci rrracie 5 depict for a lypicd by to irlcorporatt: the cumulative a typical ‘l‘he vert,ic’al (among most IIC~ irislance arid of instarices the clecisiori fied instar1c.e exllibits low, the high, tree from vtxrl ical scale O( ,.,t;‘) _a . of II)3, for bound curve. Ifii which updates I I)3 and becri of 1114 is greater ^_ 1113 (nol~ rrlagnified thaIi the 400 that times). usually the ‘l’tie The low cost it is always ads wtiil(k r t~rriainirig however, i SCIENCE bounded expc~nse per instance important ion (i.e., of the IL)3 of classifying expense of rebuilding an incorrectly this classi- O(jAl’) is clearly expensive attribute The work refiertecI curve counts orily iI1 results if the test Conclusion learning domains, intoIlsive which gorithm wit hiri the rriachine increased nature performs least four. 0( / II’ x let) expense when and of of 11)s is incorrect. corrlpticated sc1arc.h the function ID4 is number version of the reflects irlterrrlittelit processe<i, linear consiclttrably less tharl the latter’s --. pta;ikb. ‘I’he fourth variant, IIM, displays the least expense _-. 01’ tllth ttlree. It asymptotes to a value as srllall as II)3 500 hut IIF1 has to in II&l is nearly step of the as a t’uriction force approach scratch nearly As compartd 1% irislances cumulatl’~e brute curve is encountered. iristarlc32 the intermediate V when is that of each algorithm The The the arid for each by each expensive bound. reveals from iitlt efficierlcy of 1000 refiects accelerating classific~idion coast instance. performance axis rriade the 50 1113 reconstructs ari instmcc of this sequence of’ instances. asyrrlptotic its ‘l‘tlth t’xperisc% of processing price over number curve t tie riurriber execution variant. The 6 the variants displays ‘1’11~cost per slowly. geometrically couritirlg PROCESSED .-. 1113 cost and 1.000 anti quickly. attributes value It5XY 4: II>3 I 750 perfornlarlce, algorithms decisive I;igure it incrcrnerl- relatively these off; ‘l’hub, these classificatiofl important, leaving instances. was achieved with classificatiori t MY’S construction; 750 kvel irisLarlct3, 275 time perfect c.I ;issificatioIl ‘I’tltl apparent effective after corlsiderable hrns to achieve (2 WC) after to 500 INSTANCES classilicatim lorlger classificatiorl take The somewhat c Iassificdtiori may t.al dlgorit good takes I methods interest can have to as they my behave in behavior one tloes in evidcrlt. concept is that observations incremerital applied more of rlorlirlcrerrlerltal, become observatiom thougli, be rriatle process are deficiencies in incremental process point rrlethods the This induction are leas rneth- observed. An noriiIlc~rerrierlta1 an iIicrcrrienta1 at a time). riot insure alfasti- In general, the cornpu- COST PER INSTANCE CUMULATIVE 75 , COST PER INSTANCE ID3 100K 75K ID4 50K II I I-’ _/ 1 iNA __ ” INSTANCES Figure tatiollal 5: ID4 and efficacy for clkal uating of such an inc’rernc~rital t)t:crl outlined. These cost per Figure dimensions have CarbonelI, are: memory corlcept to accommodate necessary of objects PROCESSED Cumulative cost, per J. & flood, Objectives It;). a new and , 1,000 instance. to obtain a stable C;. (1985). The World Modelers Simulator Kutgers II. LeartLing Workshop of the (pp. 14 Ilrliversrty. K nowledge (1985). ftepair t i0n Versus Revolution. Proceedings tio72 ul Machine Workshop description. f’roject: Proceedings Architecture. Muchine lt~ter71Uti07~ul Michalski, 0 ‘l’tr(~ riurriber 800 dirnensions methods Third of updating - _ -_ References Three induction 600 INSTANCES 6: IE _ ..r ----__ --.~--.- 400 instance. algorithm. concept __ _ 200 PROCESSED 64 /’ 1” 25K 250 _/’ ID4 1’. Leurning Mechanlsr~ls: of the (pp. ll:volulntt r71u- Third 1 IG 119). fiutgers IJniversity. ‘[‘he quality l of derived concept descriptions. Mitchull, ‘J’trr>se dimensions have ior of 1 increrriental cascx sludy in the a prorrlising od:, cari domain indication meet c:rlviroIIrrrerrts, been variants the used of chess that meeting the endgames has served induction constraints high t)ehCiv- standards I)iumverirlg Exarlrptes. systerrls and Quinlarl, .I. It. chine Acknowledgements IJisc~ussious with tat irlg to ttie was arid quality in part by tichr grants the [ST-81-20685 trlstitute Naval tfle Ocean National arid Systems raised of learning. the under- grmt initially specifically NO0014-84-K-0391, grants N0001.f-85-K-01154, smrctl KiLlcr paper, cost supported urltlthr Ikrlrlis irk this Altu, Office of Naval (Lge. I ,earrllrlg (19X3) Research SiirIlrrlut, c un- Muchine Re- vcrslty. Voundation I24 19, the ArItly contract (Ed.), khliliburgh fhiirlburgh. ttfficirml icatiori tu ct1ess , . ~tmrIt!lI & ‘I‘ pvl A 71 urtijiclui classific.at.iorr irilrlligerice ‘l’ic )g;i I’ut)lishing prom- uyp Co~ripaiiy. re- Srience under lrltlllctloIl resclarcti and lS’I‘-85- Center leurn~r~g: . C:alllorrila: NOO014-84-K-0345, MI)h903-X5-~X)3~4, by a number the criteria ‘I‘his Rules l’rcss. .. Universit,y of complex lttfrl- ,4 rtific.iul In 11. Michitb ifb the micro (‘OI‘I‘t’CtIleSS. of’ itltlas cbxpressed as Searcll. 226. A meth- of quality 18. 203 as II)3 program. irrcrerrrental computational while to cornpare of Quinlarr’s <:erreralixation ‘I‘. (19832). liyence, and by N66001- h I1ur11e, I>. (1985) plex Kot)o1 World. ~illlOIl, t1. Mass.: Workshop Lcurniny (11)6’3) Procredings ‘h! ‘l’l~e M.l.‘l’. s cirrices 1xarriirlg ~hr~c:~:pts In a Corrl- of the ‘I’hlrd ltltcrnutmnul (pp. 173 176). ftutgcrs Ilni- of the .4rtificd. (Tarrlbritfgr, l’ress t-u-( :-o2s5. LEARNING / 501