You are on page 1of 16
System-Building Process sto make our discussion of the expert-system-builling process as conezete as possible, we now presenta hypothetical seen.cio Hlustratiag some ofthe steps that a knowledge engineer might take to develop ai expert system. The problem domain tor this scenario is te culation sonal iniury cli relied tu pradact tability. Keep in mind that the Scena ts someihat idealized. Real experts are fot as cooperative understanding, esplicit, or organized 3s the hypothetical expert de- sorbed heee—nor are real knowledge engineers so untalingly insightfal and accurate, If they were, building an expert s:stem would take mutch less time than it actually does, The Problem The vice president of Acme Insurance has a problem. The volume of claims the company has received has risen dramatically curing the last few years while the number of experienced claims adjuste-s working for the company has decreased. The company’s costs have been unusually high. and many product liability cases have been setied for large amounts of money. There is concern that cases like these are being set led for amounts substantially higher than necessary. The sice president must find a way to help the 20 to 30 claims adjusters in each of the 28 branch oifices of Acme Insurance evaluate cases more accurately The vice president assigns the jab of solving this problem to the senior claims manager, who has just finished an evening course in com puter programming, The claims manager, having heard about expert systems daring the course, devides that perhaps an expert computer system could sssit the cluims adjusters by acting as an expert advisor and possibis even a tutor. Atter 3 long struggle to convince the vice ppresiloat that this approach has ment, the projest is finally approved The claims manager nose searches tor and tinds sn experienced Arias! edge engineee willing to take on the job of building an wxpert system in vs domain, The knowledge engineer, Me Jones ss hired as a fulltime: consultant, aint srk buyin — Characterizing the Problem Kaowledge engineer Jone: Pui wow begins the tirst phase of expert ses the task i810 ley the participants ots sd sosion chamteratics. Asa first step, he Becomes tae lar with the problem oy reading books, studying product liabiliey ca: histories, and talking to chims adjusters and atterneys about how thew cesaluate caves. During this time he characterizes the fypes of reason is the expect system is likely to pertorm and decides thar deter case settlement value is probably a primary task, sshule estimating, feral damages is probably a secondary task During nis bo weeks of informal study, the knowledge ei talks to the senior claims manager about a domain expert for the projec. The claims manager suggests Ms. Smith because she is generally ac knowledged to be one of the best claims adjusters. Knoswledge engineer Jones meets Ms, Smith znd tinds her articulate and knowledgeable about computers and their applications. Jones. atsare that he has found an ideal expert, arrangesto have her collaborate with him in building the expert system. Knovwledse engineer Jones decides to implement a small working system or demonstration prototype as quickly as possible to test ideas about problem scoping and representation, He further decides that the nly stants necessary to Bul this smal inital system are domain ‘expert Sth and himselt During ais two weeks of intormal study, Jones also considers v3 ‘ous options regarding resources needed to build the expert system. He ins a commitment from expert Smith ta devote half of her time oF the first three months of the projec. In addition, Jones evaluaces the availble computing resources and notes that Acme has an in-house UNUCGBSD. He tentatively picks this as the hard 143 Prootype development He ct make a fina desi about haha Luni the expert-s¥stem-builing tools sel, since the fool mest fom the hardware. Tot espert Smith, During these meetings, Jones gives Smith closed cau Ble and intormally discusses them with her. He istens carefully to hee analyses of the cims, tying to understand how she characterecs oe, problems and dewomposes them into important subprobline lores ate, {ocuses om expert Smith's way charatterzing the data and ihe pots Jem solutions. Based on these mevtings, Jones creates the prblets ter Sctiption shown in Table 151 Knowledge engineer Jones eelizes that de problem description is {ar tom complete buvever gives him an inal understands oh Problem, the data, and the types of solutions that cocld ana — What Concepts Are Needed to Produce a Solution? Sith, presenting sample cases and noting how she analrens vont One ofthe irs eesult ofthis series cf interactons is more reise {iron ofthe terms and concepts used by the exper. For sey Jones finds that specal damages refer to spectic econoinclosses wen ‘ng from an injury, such as medical expenses and lost mene es {general damages include othe effets of tm injury, such asthe haus ‘he injury othe fear af losing a faculy, such ssightorheasng Another result ofthese interactions (san understanding of the rela- {during the identification process. Jones notes that the ereer conics {he bins loss, defendants lab. pint’ responsibly ee characteristics, and caso context in the folowing way to saree te case value VALUE ~ LOSS » LIABILITY «(1 - RESPONSIBLY) x ‘naRacTeRIsTics «| CONTEXT, tnhere the loss is the sum of special and general damages. Jones isn't Sure at this point exactly what the characteristics and sonont secant Pass, but he knows they have a value greater than one whee te eee 144 gory Prat gu ain probe: Subproblems: rable sharers lenporant concept. Solio Dewrpion Hep cans adjusters mune assay cvalate past habits cases Determine setement value fer product iby eas Caletating special damages Colca gener! danas Determining Zeerduats hat Incomplete Jat Cine vale depen om bth estive and subjective consiertons The seventy ofthe pants bas (erg. medial reports. photogs) ‘he sreumstincee of he oe (ewes ceports, other eile Product characteristics (e.g. expert testimony witness reper) Fatipanis’ characteristics (erg. cae histones, personal eservaon) “The extent ofthe plant’ os. ‘The degree ofthe delendan's lability ‘The plans responsiblity fo the Loss. Subjective charactesstes (er ltgants, lawyers, judges, urisicions), Case content sete, ning, claim 'yp). ‘The defendant's lability (@ numeral estimate between and 1) The plants tore (a'mamerial estimate in dls) The settlement value (a numeri estimate in dll) ‘charaueristis ae more favorable to the plaintiff than the defendant and value less than one in the opposite situation 'Now Jones decides 1 explore the idea ofthe paint’ loss in more detail, concentrating on the rather dificult concept of general damages. Jones gives expert Smith a case from the dosed claim files of Acme Insurance, including actual medical reports, correspondence between, 145 altomneys, and other eeevant mater, He also provides Smah with the following sunimory ofthe case (n Devembe 1, M89, claimant yas opening a bate a champagne {Sled bythe inured, Oxnard Vintners. Claimant had never opened = ‘himpogte boll bie: He rly drank and the Bote ad been ‘rousnt to him by 2 guest Climant had partly hosened the wine Sige hen the cork oreealy shot out of the Bote, stung ian i the nigh eye Ti blow produced great poin and saints Bide the ese, Chimant wis holding the ork foward his face when he 38 Ison te re age CCsimant was then the nearest ersergeneyfoom and then transfered tothe eve cin atthe lo snveraty hosp: He hal stired 9 Cleened stina,Oastors surely repaid the eye, but or ur Sars {urns id ot pve whether he ould regain won nth ee ‘Ghimantscontion i now stable, He has sight impaired val ‘Shit nthe eye and mut aow weae gloss all Sines. Claimant 2 Sryearold mule nd prior to thie ume id not wear ses, Bocuse of the injury. cman hae a ive 9 ten peecent chance of sutlering hums. CCoiman’s speci damages ttl $4500, Jones now begins to explore the concept of general damages with the exper. KNOWLEDGE ENGINEER: Now let's consider the champagne bottle fase. We know the special damages, which are medial expenses thd lost income, are $4,300, s0 we ean concentrate on what {guess 5 Called general damages, What do you think the general damages, ‘would be ina case ike this? EXPERT: Well, (pause), they would probably be about $0,000—uh, rho, make that $6000. KNOWLEDGE ENGINEER How did you arrive at that gute? sexpeet vehad many yeas of experience analyzing cies like this Tidon't use any fixed rules likes "pain and sutfenng equals three times the medical expenses” Ijustfook at the fet of he ase and estimate the value Jones realizes that most experts have difficulty expressing the rules they use and offen belive that they reason stslly by intuition devel- ‘oped through ther extensive experience. He probes further to uncover the roles and concepts that Smith used to artve atthe snsoee. KNOWLEDGE ENGINEER: What factor seemed the most important in deciding general damages for this case? exer: Well uh, probably the most important factor hee was, uh tfivas an eye injury. This Can have a debilitating ettet on the pine Til He deserves some compensation for sulleting the dignity of being injured this way NOWLEDCE ENGINE: $0,000? xPERT. No, the injury self was only worth about S18 000. KNOWLEDGE ENGINEER: 1 sex, You's saying the plain should be Compensated forthe tac that he sufered an injury, kind ot Hke the lion of being pad for pain and suffering extent, Yes, that's ght EER And the pain and sulfering was worth s.000 " fxrEnT, Well nor exactly. Ithink the tum of the injury isl! as teorth only about S100. RNONLEDCE ENGINEER: But you mentioned a figure of $18,000 EXPERT, Yes, the other $8,400 was because the plant spent four {ys bind in one ove. The doctors did't know winether he would sein sight in hat eye, Something like that would be athe Eight say. NAORLEDGE ENGINEER: So you're saying that $10,000 should o 10 the pain for the trauma ofthe injury itself and another $8,00 for the fea suffered. Then the 38000 kind of presents the inconvene tener the plant went through while the $10,000 represents the cual pain and shook ofthe injury. arent Thats ght, except that I wouldn't cll his fear of being blind an inconvenience. The fact that he had to wear gasses alte the injury although he never wore them before isan example ofan inconvenience caused by the injury NOWLEDGE ENGINEER. Would that add to general damages? TNPERT, Yes, i sy that fas worth about $2,500, KNONLEDGE ENGINEER So the plaintiff should be compensated 52.500 for having 2 wear passes. ExPExr Not exacly. He should get 2,500 forthe inconvenience of having to wear glasses, He should also get anotherS2, because of ths pata focally loss His visual acuity was slighty reduced by the injury. Knowledge engineer Jones now sees pattern emerging expert Smith seems tobe assigning values to various aspects ofthe injury, such srthe trauma, fee, inconvenience, and faculty loss sufered by the 147 ini then ang then ysthor tin Hille Jones {ute imerestatn these concep sn! dey th part was the He fas be sen not oly for sc damon conceal fr base statis the expat used wh lst pero kn wats 9 keow how Sih justiies her ww of tari ad te gi expet stem il net this typeof infesation 40 provide nquate fsplination about its operation I an attempt ty mike her sty hee prion sbout inconvenience sr aculy hn hnes continue the dar fogue a its since the ANOWLEDIE ENGINEER: Five thousand dollars seems like 3 fot of money just for having 0 wear glises [xPExT: Not lly, Itwoull have Been monet the plant's appear nce had been particulary important for hisjob, like 2 mavie star oF TV personaity SNOWLEDGE ENGINEER Justa second —looking through 3 stack of Popes) it sas here thatthe plait sa broadcaster oe 3 lal radio Station, but that prior to his injury he had interviewed for ob as 3 local television sportscaster. He claims he later svt get the job because of is need to wear glasses NPeRT Yes, that’s what I'm talking about His appearance is worth ‘money. That should increase general damages by $3,000 more KNOWLEDGE ENGINEER: If the injury had caused a sar on his face, ‘would that have been the same sort of thing? xreRr, Wel, that would be worth considerably more, butt does {all into the Same category of disfigurement. OWLEOCE ENGIVEER: Let's see now: $10,000 forthe injury trauma, ‘58,000 for fear, $2,500 for inconvenience, $2,300 for culty loss, and $5,000 for distigurement. Ate there any other factors youre consid feng that we haven't hit on yet? EXPERT I recall thatthe medica! report mentioned the possibility of laucoma a5 a result ofthe injury. A 10 percent chance, I believe KNOWLEDGE ENGINEER: Yes, that's what it said, Would this affect {general damages? EXPERT It would hive signiant effect; should be worih about 500,000 because of the possibilty the plain will contact aur coma, This is a serious illness KNOWLEDGE ENGINEER: So a possibilty of a future teauena should always be taken into account? 148 xr, Absolutely, KNOWLEDGE ENCINEER: Adding inthe future trauma factor the stm ‘comes 193,000, That'salmost exactly what you peedicted initially. ExTERT. Yes, $58,000 35a youd approsimation of general damages for this case AU this point fones ends the interview and begins t organize the concepts uncovered dung the dilogue. The resut is a set of invhal ‘Concepts thatthe knwwledge engineer believes is eebvant 19 determin ing injury valve 4+ Task: Determination of value of injuey 1 Atenbutes of injury ‘Type of injury: Eye Value ofinjury: Some dollar amount Extent of injury: One eye Result of injury: Viewal acuity slighuy seduced ‘Temporary blindness in one eye Duration of blindness four days Uncertainty about permanence of blindness 10 percent chance of contracting glaucoma Must wear glasses. Attabutes of plainti ‘Dis he wear glasses originally? Is his appearance important for work? ‘+ Attnbutes of glaucoma: ‘Type: llness| Degree: Serious + Factors contributing to value of injury Tajury trauma factor Fear factor Inconvenience factor Faculty loss factor Disfigurement factor Future trauma factor ‘The knowledge engineer realizes that these concepts represent only « fraction of those used by the expert during problem solving ia the ‘domain. However, he decides that it will be useful to try to formalize them even at this early stage of development. He knows that he can always go back and continue wit the conceptualizaton process at any time during the expert system development, 149 How Can the Knowledge Be Formally Represented?:: Knowiedge engineer Jones aow enters the nest phase o expert system bouiiling rican, Sore te hey cong, subprens mlcante fo tures ate mye inks 4 more rem epnsenat sugested fa expert syotoniulding tot. He decides to test the appropriateness ol the exper Syatem buiking language, ROSIE, by using to ceprewat the concepts ad ides ust extracted from the eapert Figure 1.1 shaws the result of this anaivsis Although this may ik ihe prose its actully ROSIE, code, caput uf being wvecuted by a computer, Tones now goes Lack theough his notes tring ta ind ations bes thewon the varus concepts shown in Figure (31+ One he establishes them he fels read 10 jump ahead tothe implementation phase. Jones lize shat be wil Ee returning tothe tormalaation phase many tines bbut he wants 1 see how esi the relations between the concepts can be represented in ROSIE, FIGUNE 15.1 A formal rpresentston of concepts from the chsmpagnebste Assan the plait race an eye in. Ascon ine plait’ injury does inv ust one eye Assen te plas visual suis sighyrecuced! by he iu. Aster te las injury di cause (a tempera Bienes in ono eye) and It the ewan of at blnanass be 4 days, Asset the plaitis doctors ware net certain abaut he Bindness being Ass ie pari dos have (a chance of contracting geome) and tat chance was eaused bythe paint’ tury and lt the value of hat chance be 10% | Asser ihe pitts injury does requice (he plait 6 wear glasses} ‘Assen te pat didnot wear glasses bear ne iu, ‘sso ta planis appearance is important for work Assert gaucama isa seseus snes, Asser ach of inary wauma, fa, iconvenionee, facay less, ‘Ssigurement and future wauma is factor tat does corrbute © tho value ote ny. 150 —— Tmplementing 2 Prototype System Knowkege engineee Jones note enters the inplemsration phase im fohch ete ccs and relations bse the ino # working com> Jats progam These fev tentative ROSIE rules are the ist tep toward {he implementation ofa prosotype expert sytem. | land the value of that chance = 10% os anya eur “The rules lock good to Jones, and they seem to predict the answer Smith gave during the last interview session, Jones cow decides that he must hove the expert Help him test the rules — Validating the Rules Jones now enters the esting phase, He attempts to aida theres that fave been jontuleted. Ina more fully developed system Jones would run the progeam on a fee cases and ask Smith how well t predicted the 151 sofe'or general dames. However, with this small ptiminay so Tee Was belicves til suffice simply t show theres the expert wane ore opmion of them. At ths stage the howfesge engineer tps the rales to be ovely spastic ad hopes to uncover Same Hane PiBiievet principles the expert using thot are eemplifed by these Towlevel rules ‘ing the nes interview sesion, fone shows the expt one of he rules he generated and asks for her opinion of is aeuraky ANONLEDGEENCINEER: Har eae Toeote to capture the dea tRat SOUESliy ota furce tauma could ad to the value ofthe oe Panes Rane Smith a pec of paper with the fllwing le bait) [a] the plait does nave (a chanes of contactog ga.coma) Sedna chance was eaused bythe paris | Sd e va of nat ehance = 10% | pcaae ne ttre kauma factor by $30000 xreRT: That rule looks fine to me. KxoNLeD&e ENEINEER: How does glaucoma compare with those f+ jes? Trek, That depends om the severity ofthe potential:njury, Acatay Mbp Injury Ee parapegia would be worth more han a modes ie injury like» shortened imo. tomusnce ENCINEER. How does glaucoma compare with thos it~ juries? vert Its somewhere in between, Td rate glascoma a8 a sus CAPER per than catastrophic or moderate, Is in the same cate sory as eplepay and heart disease aeeywtence ENGINEER: Would al ofthese serious illnesses or in ree $3000 if there was a 10 percent chance oftheir occu fence in the Fate? txteRr: Yes, $30,000 is about sight lovontence ENCINEER What if there were a 20 percent chance of arent. Well, then it would be worth more. I wouldbe worth about ‘Sion if there was anywhere from a 5 to 15 percent chance of GNOWLEOCE ENGINEER What would ithave been werth i he injury res ehnitely caused glaucoma rather than just presenting the Pos siblty of causing 1? 152 xrekT: Glaucoma woul he worth about SHW,002, But since there vvas only a5 to (3 poreent chance oF, the pluinblt should only be «compensated for 30 percent of that amount, 30.000, fomus nov revises the rule, making it more general and thus move usetu tu the expert system. To do this he intact new Concepts. 1. dass of ilnesses rated 3s Serious, and 2) the value of contracting an iness. The new concepts nd eevised rule ape shown below | Assan eaen ot gaucoms, aplepsy, and neat diseases 2 serous | Lotine vaue of glaucoma be $100000, [2] ine plant does nave (a chance of contacting. an ross) | 07“ tre'natenance was caused bythe pals hry | ana me value of at chance > Ste | Sr mar aioe == 155 i nd ine ptonal hess o the pain isa sas ines, Incase ie ure trauma lac by 30% of ha vaio he Slanuls potent sino). Ina similar manner the knowledge engineer quetes Smith further about rule 2 fo make i more general He finds that she arrived at the figuce $8000 by multiplying the number of days during which the plane ‘lf as partially blind (our days) by $2000. He adds tis to the rule and Introduces the concept ofa disability ofan important funtion. The new concept and rule are shown below, Asser esc of bindness in ne aye, severe zines, paralysis of one imo, and lss one imo fea dsb of an important hanchon [2] He plats iqury i cause (a temporary sabi ofan ‘poration and he plaints decors wore oot certain about (th teparary rate of hat aby), increase te Tear factr by (2000 % te duration 3 that sabi). "Now knowledge engineer Jones has a set of rules that seem to sat- isfy the expert, at least on an individual basis. He futher test the rules by devising hypothetical problems, presenting them tothe expert, and ‘comparing her answers with the ones that the rules would produce During this refinement process, he finds that some rules are too general ‘and must have additional conditions added to reduce thei generality For example, Smith indicates, among other things, tat the distigure iment factor also depends on the plaintif's age and that the injury ‘trauma factor also depends on whether or not the plant required Sur gery. Jones takes these refinements inte account and produces a new set of rules » 159 | (11 ie psi 3 ecene an eye ir) find natin dd quite Sey ARS ne pla’ sual att sgt reduens oy ta ny 25 tne feeoery rom tat ay is amos compet Spd he pani’ condoned | iectease tne muy wauma factor by $10000, | ex Socauepancaeteytb # eee ee Ewe rao bt os | wieee Se se en gomonet a oy eee tome ctcoengnne | me a i She ny eae of tt chance > 5% She thatlaue © = 15% S25 Re pototat lines of te lan is 2 serous ess occa te fture trauma Incr By 20% of he ae oe ‘nts pot ess). {s)he lint roceve an eye iy Sn the plant's spur does vce jst 008 oF RG he p's vous acu fs sgl ecuce bythe iy. pcroaae th tclty loss factor by $2500. {1 tine pants inry sees rqure (ne pit wear 38395) | ‘dnd the paint Sc not wear glasses batons the inky vranae ne nconveniance factor by $2500. {et te plas iry das que (ne plat 9 wear se) ‘Tad sant appesance is mpotant or wrk 25 te og fl te pa) a te me ote uy) > 25 SPS ne pt id hot woar gaseesbelore Me IL increase the estigurement factor by $5000 | Fa nr at not a a hat fair the ganeral damages Jones continues these interview sessions with Sith to uncover the confepts ules, and problem-solving strategies she uses. In odsiton Jonestths to understand how South jis the ees and strategies, Jeo osacescombeling angument to convince her pees that het 2 ‘eal vaid He deces that one of hs goals so bu a8 much of Fi ustieaton knowledge into the program as pesible ie tiand of wo months, Jones decides that ROSIE san appropt ate ool forthe problem and begins to turn the Feasibility demonstration ito a fulledged research prototype system. A yea and many inter area ater the knowledge base reaches a sigaificant size (any Run 154 reds of rules and data objects), Jones now concentrate on testing and ‘elining the reseasch prototype by I) bringing in addiional experts to help validate the systems accuracy, and 2) using test cass not encoun tered by the system during e previous development. ut now something diferent happens. The research prototype has become so unvisidy, because ofall the patches and changes made, that Jones decides to reinyplement i om the ground up. Reining the do- train knowlege rom the onginal prototype, Jones quicly designs and bills a better organized and more efficent exper sytem. He then Spend a year reining and testing this new system inthe field. Now the Feld prototype i given to a team of Support programmers who feprogiam it 3 lowerlevel, more efficient language. The expert sys- fem is now ready t9 be introduced into the 2ctual work environment at ‘some Insurance ‘So ends our hypothetical excursion int the lie of a knowledge engi eer Please remember that thie scenario is designed t ilustrate the phases of expert system building, It does so in a rather oversimplified fray. making no attempt to convey the actual compleity of the process 155 forts [ irae tors of the XperRule range of sofware for Knowle Hh based development, resoutee optimization, and data ning, [70200.2035 @compuservecom) San sc EER UAE! lecyersyten:oonss velar vatanguioond amen threw [lmlon aes tae | finghnware Gompoiation Home page for Brightwae including demensteations, duct information, ad upeoming events, a —$—$ | [LISP-based language allows development of Case based keasoning applications CECAS® Unversityot wei @eecaseukans edu ‘Kapsas ———————__| 7 Sian SRS | om, , wledge-based expert systems that use expertise to assist AP sree [entte os intani sia (= feendction mi at i SSS STE Info@ gensym.com: service @zensym co [a92156.185.2) _ [A division of American Heursice Corporation whose] mission so facilitate the wanserof new techeologios rough first-rate, hands-on, intensive ining courses. New| Het af [cours are added tothe cusculum ss ached Gordian astute. technologies become sufcienlly mature te waren comneral tention, [Experienced in symbolic processing, elecuonte publishing formation inludes product and service information, pplication, and mere, orks @harlequin com) [ j Tle) eee cen poten peace ae (ato econ Cig LEAL OM © O° Viiome pose fr Iteenc including demensaton, product | ace [foros and upcoming ren then Inference Comm \ tel A/S produces intelligent soitware for creating Fintemet agents, expert systems for HelpDesh, suppor, sates vic, et rio

You might also like