AI Magazine Volume 6 Number 2 (1985) (© AAAI) Selection of an Appropriate Domain for an Expert System David S. Prerau ComputerScience Laboratory, GTE Laboratories,Incorporated, 40 Sylvan Road, Waltham,Massachusetts 02254 the development of an expert, system, and thus a signifi- Abstract cant amount of eflort should go into the selection process. This article discusses t,he selection of the domain for a knowledge-based expert system for a corporate application Background The selection of the domain is a critical task in an expert sys- tem development At the st,art of a project looking into the Interest in artificial intelligence by the corporate business development of an expert, syst,em, the knowledge engineering communit,y has been growing dramatically in t,he last, few project team must investigate one or several possible expert years, and many corporations have set, up AI groups or system domains They must decide whether the selected appli- are in the process of doing so. One of the prime arcas of cation(s) are best suited to solution by present expert system corporate interest is expert systems. Though the nunlber technology, or if there might he a hettel way (or, possibly, no of expert systems actually functioning in a corporate en- way) to attack the problems. If there arc several possibilities, the team must also rank the potential applications and select vironment, is still relatively small, the number of project,s the best availahlc To evaluate the potential of possible ap- looking into expert system development is growing rapidly. plication domains, it has proved very useful to have a set of The knowledge engineering project team working on desired at,trihutes for a good expert system domain. This ar- an expert, syst-em development must investigate possible t,iclc presents such a set of attrihut,es The at,trihute set was application domains. In some cases there is a very specific developed as part of a major expert system development project application, chosen by management,, for which an expert at GTE Lahorat.ories. It, was used recurrently (and modified system is to be developed. In this situation, it is likely that, and expanded continually) throughout an extensive application those who selected the application area had little techni- domain evaluation and selection process cal knowledge of artificial intelligence or expert systems. Thus, the project team must decide whether the selected application is one that is best suited to solution by present This article discusses the srlection of the domain for a expert system technology, or if there might be a better way knowledge-based expert system. In particular, it focuses (or, possibly, no way) to attack the problem. on selecting an expert system domain for a corporate ap- In other cases, the project team is asked to select one plication. The choosing of the domain is a critical task in of several corporate problems or to survey corporate con- cerns to find a good application of expert system technol- ogy. Here, the pro.jcct tcarn must not only decide if an I would like to thank the management and members of the Knowl- edge Based Systems Department at GTE Laboratories fol their con- application is suited to present expert, system technology, tinuing .I suvuortI. in this work EsDeciallv, I would like to thank R,alph but must also rank potential domains and select the best Worrest of the Department for his walk with me on the application available application. evaluation and selection phase of our expert system dcvclopment project and for his suggestions tlnoughout the course of the work To evaluate the potential of a possible application, it that led to this paper. Also, I’d like to thank Shri Goyal, Alan Gun- has proven very useful to have a set of the attributes de- derson, Alan Lemmon, Robert Reinke, and Roland Van der Meer of the Department, and Charles Rich of MIT for their suggestions on sired in a good expert syst,em domain This article pro- t,hr contents of this paper vides such a set of attributes. The set includes technical 26 THE AI MAGAZINE Summer, 1985 attributes as well as ai-tributes related to non-technical likely to find some unique properties of the domain that corporate issucs may require new approaches. There may be a degree of commonality among some of An Application Domain Evaluation Process the attributes listed in this section. However, to encourage consideration of the diffcrcnt, aspects of domain selection, The set of dcsired expert system domain att,ribut,es was these comnonalities were not, eliminated. developed as part of a niajor expert, syst,cnl development, Very few of these desired attributes are absolutr, and prqjcct at GTE Laborat,orics. It, was used recurrently (and it is unlikely that any domain will meet, all of them cou- ruodifird and expanded continually) throughout au cxt,cn- plctely. Fmtherrnore, in each different sit,uation the wright- sive application domain rvaluation and sclcction process. ing of the factors will be different, and additional fact,ors Over 50 corporate managers and “experts” were inter- may apply. This set does provide, howcvcr, a fairly cxt~cn- viowcd, and over 30 extremely diverse possible expert, sys- sive list of aspects to consider in doniain selection. tciri applications areas were considered, at least. briefly. This list was narrowed to eight nla,jor possibilities, and Basic Requirements these were further analyzed and ranked Two primary l The domain as charncterazed by the use of expert Icno~~L candidate areas wcrc studied in great, dct,ail. Finally, one edye, judgment, and experoence. The goal of the pro.jcct is application alca was chosen, and our syst,cm devclopnlent to ext#ract a portion of an expert’s knowledge, .judgnlcnt, was \qpl. and experience, and put it in a prograni At each stage of the sclectiou process, the set, of at,- l Conventional programmrng (algorzthmrc) approaches to tributes proved very useful. In initial iiit,crviews, a dis- the task are not satasfactory. If a conventional approach cussion of the attributes was an excellent. way t,o give our will work well, there is usually less t,cchnical risk to using interviewees; who usually knew nothing about artificial in- it, rather than an expert, systcnl approach. Note, however, telligcucr or expert systems, some quick idea of the sort that expert systcln rnet,hodology lnay offer sonle additioual of application area for which WC were looking As each advantages over conventional tccliniqurs, such as the ex- potential application surfacrd, a brief check through the pected ease of updating and nlaint,aining a knowledge base desired at,tribute list enabled us to identify possible proh- and the ability t,o explain results. lems rclat,ed t,o the candidate area, and then t,o focus our l There are recognized experts that solve the problem today. fiirtlicr questions When the sot of majol possibilities was If an area is t,oo new or too quickly changing, there inay determined, we were easily able to highlightS the good and be no real experts. However, these are often the arcas that, bad points of each potential application Finally, when are siiggest,ed for expert. syst,eiii developinents. the actual application area was decided upon, we used the l The experts are probably better than amateurs m per- att,ribut,e list to justify the decision. One further point: at forming the taslc. Thus, the task does require expertise. each step, the list proved very useful t,o justify the drop- l Expertase as not or wall not be nvaalable on a relaable ping of politically favored candidate areas. and contanuang bnsas, i c , there as a need to “capture” the expertzse. Thus, there is a need for the expert systeni Desired Properties of the Domain For example: (1) expertise is scarce, (2) expert,ise is CX- This s&ion prosmts a set of desired attributes for the prnsive, (3) there is a strong dependence on ovelworkcd domain of au expert system for a corporate application. experts, and/or (4) expertise is available today, but, will Though many of these attributes arc applicable to all PX- be unavailable, or less available, in the future. pert systems, there are some that arc specific to the dcvcl- l The completed system as expected to have a saynzJicant oprnent of an expert system in a corporate eiivironincnt. payof for the corporataon. These involve, for cxanlplc, the likelihood of corporat,e ac- l Among posszble application domaans, the domain selected ceptance of a system, the support, for the system dcvelop- as that one that best meets overall prolect souls regardang rnent by corporate management, etc. There arc probably project payoR versus rask of faalure For cxarnplc, a conscl- analogous pointas that, apply to an academic or other envi- vativc approach would be to attrrnpt to dcvclop a systcrn ronmmt, but these are uot. addressed here. that would lneet some criterion for nlininnun payoff if suc- The attribute set was developed from t,hc perspective cessful. and that seenis to offer the best chance of success. of providing a real working expert, system to solve a cor- porat,e problem, using state of the art expert system tech- Type of Problem niques The discovery of new or better methods for expert, l The task pramarily requares symbolac reasonrng. For a syst,clu tlevelopulcnt was not an ob,jcctivcPmin fact, a do- task t,hat prilnarily involves nunierical coulputlation, con- main that requires a major brcakthough in expert system sideration should also be given to other progranuning ap- methodology is probably not, a good domain to choose if proaches the goal is to maximize the likelihood of success Yet, any l The task requires the use or heurastzcs, e.g., rules of project that is the first, t,o attack a particular domain is thumb, strategaes, etc.