FEATURES

INTRODUCTION . 100 CIARCIA:S CIRCUIT CELLAR: BUILD THE HOME RUN CONTROL SYSTEM, . PART 1: INTRODUCTION by Steve Ciarcia ... . 102 Steve returns to the field of home control in this first part of a three-part-series.

COPROCESSING IN MODULA-2 by Colleen Roe Wilson . .... I I 3 ... This method lets you cooperatively process information by interleaved execution on a single computer. A MILLION-POINT GRAPHICS TABLET by James Hawley ...... 120 Build a graphics pad for less than $200 using the KoalaPad for input.

THEMES

INTRODUCTION ...... 124 ...... ' •...•'

COMMUNICATION WITH ALIEN INTELLIGENCE by Marvin Minsky ...... 126 ...... It may not be as difficult as you would think. THE OUEST TO UNDERSTAND THINKING

by Roger Schank and Larry Hunter...... 143 . . . It begins not with complex issues but with the most trivial of processes. THE LISP TUTOR by John R. Anderson and Brian J. Reiser...... 159 . . The system described offers many of the advantages of a human tutor 100 in teaching LISP programming.

PROUST by Lewis Johnson and Elliot Soloway ...... 179 W. . . . . This LISP program automatically debugs the efforts of novice Pascal programmers.

ARCHITECTURES FOR AI by Michael Deering ...... 193 F. .. The right combination of hardware and software is necessary for efficient processing. THE LISP REVOWTION by Patrick H. Winston ...... 209 The language is no longer limited to a lucky few. THE CHALLENGE OF OPEN SYSTEMS by Carl Hewitt ... 223 Current logic programming methods may be insufficient for developing the intelligent systems of the future. VISION by Dana H. Ballard and Christopher M. Brown ..... 245 Te chnology is st ill being challenged to create reliable real-time vision systems. LEARNING IN PARALLEL NETWORKS by Geoffrey E. Hinton...... 265 . The author presents two theories of how learning could occur in brain-like networks. CONNECTIONS by Jerome A. Feldman...... 277 ...... Massively parallel computational models may simulate intelligent behavior more closely than models based on sequential machines. REVERSE ENGINEERING THE BRAIN by John K. Stevens ... 286 ..• ' The brain's circuitry can serve as a model for silicon-based designs. THE TECHNOLOGY OF EXPERT SYSTEMS by Robert H. Michaelsen. Donald Michie. and Albert Boulanger ...... ' .....303 There's more than one way to transplant expert knowledge to machines. INSIDE AN EXPERT SYSTEM by Beverly A. Thompson and William A. Thompson .315 The authors trace the development of a rule-based system from index cards to a Pascal program. 124

BYTE is published monthly by McGraw-Hill Inc. Founder: James H. McGraw !1860-1948[. Executive. editorial. circulation. and advertis· ing officeI[SSNs: 70 0360-52801 Main St . Peterborough. NH 034 58. phone [603[924·9281. Office hours. Mon-Thur 8:30 AM -4:30 PM. Friday 8:30 AM- 1:00PM.Eastern Time. Address subscriptions to BYTE Subscriptions. POB 590. Martinsville. 08836. Postmaster: send address changes. USPS Form 3 579. undeliverable copies. and fulfillment questions to BYTE Subscriptions. POB 596. Martinsville.Nl 08816. Second-class postage paid at Peterborough. NH 014 58 and additional mailing offices. Postage paid at Winnipeg. Manitoba. Registration numberN! 9121. Subscriptions are 521 for one year. 518 for two years. and 555 for three years in the USA and its possessions. In Canada and Mexico. 521 for one year. 5<12 for two years. 561 for three years. 569 for one year air delivery to Europe. 17.100 yen for one year surface delivery to Japan. 517 surface delivery elsewhere. Air delivery to selected areas at additional rates upon reques1. Single copy price is 51,50 in the USA and its possessions. 53.95 in Canada and Mexico. 54.50 in Europe. and 55 elsewhere. foreign subscriptions and sales should be remitted in United States funds drawn on a U.S. bank. Please allow six to eight weeks for delivery of first issue. Printed in the United States of America.

2 BYTE APRIL COVER ILLUSTRATI ON BY ROBE RT TI NNEY • 1985 VOWME 10, NUMBER 4, 1985

REVIEWS

INTRODUCTION ...... 334 . REVIEWER'S NOTEBOOK by Glenn Hartwiq .. 337 THE ITT XTRA by John D. Unqer ...... 338 ...... An IBM PC-compatible system with telecommunications software. INSIGHT-A KNOWLEDGE SYSTEM by Bruce D'Ambrosio ..... 345 Software to help·you build an expert system and learn about artificial intelligence. REVIEW FEEDBACK . .... 348 Readers respond to pr,evious reviews.

KERNEL

INTRODUCTION 353 ...... COMPUTING AT CHAOS MANOR: OVER THE MOAT by Jerry Pournelfe . . . 355 ...... As construction workers descend on Chaos Manor. Jerry battles the flu to look at more new items. CHAOS MANOR MAIL conducted by Jerry Pournelle . . . . 373 ...... Jerry's readers write. and he replies. WEST COAST. LASERS, OFFICE PUBLISHING, AND MORE by john Markoff and BYTE Phillip Robinson ...... 379 334 Our West Coast editors report on lnterleaf's OP5-2000 and TP5-2000 and on FastFinder for the Macintosh. · U.K.: NEW DATABASE IDEAS by Dick Pountain . . 389 BYTE ...... I.D.E.A. S. is a commercial database-generator package in which all data items are related by a system of coordinates abstracted from the real world. )APAN: THE FIFTH GENERATION IN JAPAN by William M. Raike . . 401 BYTE ...... Our Japan correspondent takes note of the International Conference Generation Computer Systems. the new Hitachi supercomputer. of Fifth and software development in the country. CIRCUIT CELLAR FEEDBACK conducted by Steve Ciarcia . , ...... 408 ...... Steve answers project-related queries from readers.

ED11DRIAL: GOLFERS AND HACKERS .... 6 EVENT 0UEUE ..83

MICROBYTES ...... 9 WHAT'S NOT ... 96 LETIERS. 14 BOOKS RECEIVED ..... 414 ...... •...•...•....

FIXES AND UPDATES 33 PROGRAMMING INSIGHT ...... •...429 .....•...... WHAT'S NEW. . .. 39, 440 UNCLASSIFIED ADS ...... 493 ASK BYTE ...... 48 BYTE's ONGOING MONI1DR Box. . . . • . BOMB RESULTS ... 494 CLUBS NEWSLETIERS . 58 . & READER SERVICE .... 495 BOOK REVIEWS ...65

353

Address all editorial correspondence to the Editor. BYTE. POB 372. Hancock. NH 03449. Unacceptable manuscripts wil! be returned if accompanied by sufficient first-class postage. Not responsible for lost manuscripts or photos. Opinions expressed by the authors are not necessarily those of BYTE. Copyright© \985 by McGraw-Hill Jnc. All rights reserved. Thademark registered in the United States Patent and Trademark Office. Where necessary. permission is granted by the copyright owner for libraries and others registered with the Copyright Clearance Center !CCC) to photocopy any article herein for the flat fee of per copy of the article or any part thereof. Correspondence and payment should be sent directly to the CCC. 29 Congress St.. Salem. MA 01970. SpecifySl.SO tSSN 0360-5280/83. 5!.50. Copying done for other than personal or internal reference use without the permis- sion of McGraw-Hill Inc. is prohibited. Requests for special permission or bulk orders should be addressed to the publisher. BYTE is available in microform from University Microfilms International. 300 North Zeeb Rd.. Dept. PR. Ann Arbor. Ml 48\06 or 18 Bedford Row. Dept. PR. london WCJR 4El England. Subscription questions or problems should be addressed to: BYTE Subscriber Service. POB 328. Hancock. NH 03449 e

SECfiON ART BY DOUGLAS SMITH APRIL 1985 BY T E 3 • I·N·T·E·L·L·I·G·E·N·C·E

THE TECHNOLOGY OF EXPERT SYSTEMS

BY ROBERT H. MICHAELSEN, DONALD MICHIE, AND ALBERT BOULANGER

Transplanting expert knowledge to machines

THE PURPOSE OF this article is to in­ analysis is successful. the computer ferentiate. pursue and test hypothesis. troduce expert systems. Initially. we'll program based on this analysis will be explore and refine. ask general ques­ define these systems. Next. we'll able to solve the narrowly defined tions. and make a decision (see ref­ discuss methods for building them. in­ problems as well as an expert. (For a erence II). cluding the advantages and disadvan­ discussion of successful expert sys­ Experts are capable of tages of each method. Finally, we'll tems. see reference 2 .) review the computer resources Experts typically solve problems Applying their expertise to the solu­ needed to build and run expert that are unstructured and ill-defined. •tion of problems in an efficient man­ systems. usually in a setting that involves ner. They are able to employ plausi­ diagnosis or planning. They cope with ble inference and reasoning from in­ DEFINITION this lack of structure by employing complete or uncertain data. Expert systems are a class of com­ heuristics. which are the rules of Explaining and justifying what they puter programs that can advise. thumb that people use to solve prob­ do.• analyze. categorize. communicate. lems when a lack of time or under­ Communicating well with other ex- • consult. design. diagnose. explain ex­ standing prevents an analysis of all the (continuedi plore. forecast. form concepts. iden­: parameters involved. Likewise. expert Robert H. Michaelsen is an assistant professor tify. interpret. justify. learn. manage. systems employ programmed heuris- of accounting at the University of Nebraska monitor. plan. present. retrieve. . tics to solve problems. Figure I is an (Lincoln. NE 68588-0488). He received his schedule. test. and tutor. They ad­ example of a complex heuristic used Ph.D. in accountancy from the University of dress problems normally thought to by TAXADVISOR. an expert system Illinois. Donald Michie is Director of Research require human specialists for their that gives estate-planning advice (see at the Thring Institute (36 North Hanover solution. Some of these programs reference 17). St.. Glasgow G I 2 AD. Scotland). Formerly have achieved expert levels of perfor­ Experts engage in several different a professor at the University of Edinburgh. mance on the problems for which problem-solving activities. For in­ he is the author of numerous books and ar­ they were designed (see reference 6). stance. the following problem-solving ticles on artificial intelligence. Albert Expert systems are usually devel­ activities have been identified in Boulanger is a scientist for Bolt Beranek and oped with the help of human experts MYCIN (see figure 2): identify the Newman Inc. (10 Moulton St .. Cambridge. who solve specific problems and problem. process data. generate MA 02238). He has a master's degree in reveal their thought processes as they questions. collect information. estab­ computer science from the University of Illinois proceed. If this process of protocol lish hypothesis space. group and dif- at Urbana-Champaign.

APRIL 1985 • BYTE 303 EXPERT TECHNOLOGY

perts and acquiring new knowledge. such cases. we try to represent an distinguishing meaningful hypotheses Restructuring and reorganizing understanding of structure and func­ from coincidences in the data. It is •knowledge. tion. Surface representations are also likely that deep representation Breaking rules. They have almost as often empirical associations but are will enhance the incorporation of the •many exceptions as they have rules. sometimes "compiled" from an un­ last four previously listed expert capa­ They understand both the spirit and derstanding of structure and function. bilities into expert systems. Surface the letter of a rule. In the former case. the association representations have offered little in Determining relevance. They know between premises and conclusions of this regard. •when a problem is outside their ex­ rules is based on empirical observa­ However. surface representations pertise and when to make referrals. tion of past association. Causality is have their advantages if the only con­ Degrading gracefully. At the boun­ implicit in the rule. rather than explicit. cern is problem-solving performance. daries• of their expertise. they become Deep representations enhance the empirical associations. or compiled gradually less proficient at solving explanatory powers of expert sys­ understanding. They should be less problems. rather than suddenly incap­ tems. With surface representations. all costly to formulate than causal able (see reference 4). the system knows is that an empirical models. This lower cost can provide association exists; it is unable to ex­ a reasonable level of explanation Expertsystems have modeled only plain why. beyond repeating the as­ along with a primitive form of knowl­ the first three expert capabilities to sociation. Where more fundamental edge acquisition. If a domain's exper­ any extent. and even explanation and insight is available. deep representa­ tise is based on empirical association. knowledge acquisition . have just tion will enable the system to respond as in many areas of medicine. surface begun. more substantively. If computer induc­ representations are the only kind Expert systems. like human experts. tion is used for knowledge acquisi­ available (see reference 4). can have both deep and surface rep­ tion. a model for understanding The best approach to expert-system resentations of knowledge. Deep rep­ events in the domain (a deep repre­ building is probably to use deep rep­ , resentations are causal models. cate­ sentation) often guides the induction resentations when they are cost-effec­ gories. abstractions. and analogies. In of rules from examples by tive and surface representations for the rest of the system. This approach RULE 216 has already been explicated in a (This rule applies to clients and is tried to find out whether a short-term trust should paper by Hart (reference 12) and im­ be recommended.) plemented in Digitalis Advisor. a sys­ tem that provided advice on digitalis If: 1) The client and/or spouse do wish to shift property income to, another (not dosages for cardiac patients (see ref­ for legal support), etc., for at least 10 years or until the death of the erence 29). beneficiary, 2) The client and/or spouse do desire to eventually reclaim control of this property (for retirement, estate liquidity, etc.), BUILDING EXPERT SYSTEMS 3) The client and/or spouse are in a higher income bracket than the An expert system is able to make beneficiary, decisions on a par with an expert pri­ 4) The client and/or spouse are willing to relinquish control of the beneficial marily because its structure reflects enjoyment of the property, the manner in which human special­ 5) The client and/or spouse are able to provide for their living needs without ists arrange and make inferences from this income, even in the event of disability or unemployment, 6) The client and/or spouse do not plan to have the trust income used to their knowledge of the subject. The pay life-insurance premiums on his/her life without the consent of an system is driven by a database of in­ adverse party, exact and judgmental knowledge that 7) The client and/or spouse do not plan to use the trust for a leaseback of is typically made up of if-then rules assets, and when surface representation is used. 8) A: The client and/or spouse have a person (e.g., a parent) they are or frames and semantic nets when supporting without legal obligation with this property income (will lose a dependent if trust is formed), deep representation is used (see ''A B: The client and/or spouse have a child, not a minor, that they will be Glossary of Artificial Intelligence putting through college with this property income (can set up early Te rms" on page 138). Domain knowl­ and accumulate income without tax problems), or edge is processed in a strict order of C: The client and/or spouse are using some of their after-tax income for deductive inference and is invoked by the benefit of some oth(lr taxpayer (child's marriage and/or home a pattern match with specified fea­ purchase, etc.), Then: It is definite (1 .0) that client should TRANSFER ASSETS A SHORT-TERM tures of the task environment. Figure TO TRUST 3 is an example of pattern matching by TAXADVISOR. Because uncertain­ Figure 1: An example of TA XADVISOR rule. ty is usually involved in expert judg­ a ments. expert systems must allow

304 BYTE • APRIL /985 EXPERT TECHNOLOGY

conclusions to be reached with less CON SULT than certainty. Figure 4 illustrates how TAXADVISOR copes with uncertain­ MAKE DIAGNOI SIS ty during a consultation. (For more in­ _ -- - - IDENTIFY------REVIEW '-...... COL LECT - MAKE formation on uncertainty mechanisms / - in expert systems. see reference 32.) PROBLEM DIFFERENTIAL INFORMATION DECISI ON The type of computer program that PROCESS/ GENERATE \ ESTABLISH GENERAT E PROCE SS is used to develop an expert system DATA QUESTIONS HYPOTHES/IS QUESTIONS \ ------,HARD cannot have its flow of control and data utilization rigidly fixed because such a structure is ill-adapted for simulating a human's responses to a PROCESS. GROUP EXPLORE EXPLORE ASK complex. rapidly changing. and un­ DATUM AND AND AND GENERAL ( /AD .\HE) /'T�"REFINE QUE IONS familiar environment. Instead. such a program must examine the state of / \ \ the world at each step of the decision Q2 Q3 Ql7 TEST TEST '\' ":'\"PURSUE process and react appropriately i'"\ HYPOTHESIS HYPOTHESIS HYPOTHESIS because new stimuli continually arise. (INFECTION) (MENINGITIS) (VIRUS) 1.� The type of program that has been developed to cope with this constant Q4 PROCESS TE STI change is a loosely organized collec­ /IDATA HYPOTHESIS (VIRUS) tion of pattern-directed modules (PDMs) that detect situations and re­ PROCESSI 1.\ DATUM Q9 QIO spond to them (see reference 31). The (FEBRILE ) .� rule in figure is a PDM from TAX­ I ADVISOR. QS!\ Q6 Each PDM examines and modifies data structures that model critical Figure 2: The MYCIN problem-solving hierarchy. Question numbers (02. etc.) aspects of the external environment. correspond to questions asked in the consultation. Solid lines show tasks actually done. In TAXADVISOR. the client's financial­ dashed lines those that might be done. (Figure used with permission; see reference II.) planning situation and objectives con­ stitute the environment. A PDM should be written as a single and Prestored Client's separate unit that is independently Necessary Attributes for Short-Term Trust Value Value Does client wish to shift property income to meaningful within the task domain of 1) another (not for legal support) for at least the program. This aids incremental 10 years or until the death of the beneficiary? yes yes program growth and debugging. since 2) Does client desire to eventually reclaim revision of one PDM does not affect control of the property? yes yes the others. It also provides explana­ 3) Is client in a higher income-tax bracket than tion power: a single PDM can be used the beneficiary? yes yes to explain a recommendation by the 4) Is client willing lo relinquish control of the system. beneficial enjoyment of the property? yes yes 5) Is client able to provide for his living needs Any system composed of several without this income even if disabled or PDMs. one or more data structures unemployed? yes yes that may be examined and modified 6) Does client plan to have trust income pay life- by the PDMs. and an executive pro­ insurance premiums on his life without gram to schedule and run the PDMs consent of an adyerse party? no no is caJled a pattern-directed inference 7) Does client plan to use the trust for a no no POlS leaseback of assets? system (PDIS). In effect. a factors 8) Does client have a person he is supporting complex problems into manageable. without legal obligation? yes yes largely independent subproblems. Figure 3: An example of pattern matching done by TAXADVISOR. performed to SURFACE REPRESENTAT IONS determine if the client should be forming a short-term trust. The Prestored Va lue column Rule-based systems (RBSs) were shows the pattern of attribute values that a client must have before TA XADVISOR originally used in cognitive modeling· will recommend a transfer of assets to a short-term trust. Since the client's pattern of short-term memory. Since expert matches the prestored one. the trust will be recommended. (continued)

APRIL 1985 BYTE 305 • EXPERT TECHNOLDGY

(begin) systems attempt to imitate people. it - -Formation of a Te mporary Trust- was natural that RBSs would also be used in their development. To date. Does client wish to shift property income to another (not for legal support) for at 1) 10 RBSs are by far the most common least years or until the death of the beneficiary? structure for expert systems. Among **YES 2) D.oes client desire to eventually reclaim control of the property? the successful rule-based expert sys­ .. YES tems that have been developed are 3) Is client in a higher income-tax bracket than the beneficiary? the following: .. YES 4) Is client willing to relinquish control of the beneficial enjoyment of the property? MYCIN-diagnoses infections (ref­ • ..WHY erence 26) HEURISTIC DENDRAL-identifies This will aid in determining whether client should form a temporary trust. It has organic• compounds (reference 8) already been established that PROSPECTOR-aids geologists in • Client wishes to shift property income to another (not for legal support) for at evaluating mineral sites (reference 5) 1) least 10 years or until the death of the beneficiary, PUFF-analyzes pulmonary function • 2) desires to eventually reclaim control of the property, and . tests (reference 15) 3) is in a higher income-tax bracket than the beneficiary. INTERNIST- performs diagnosis in internal• medicine (reference 22) Therefore, if XCON (formerly Rl)-configures the • client is willing to relinquish cbntrol of the beneficial enjoyment of the property, VA X-11/780 computer system 1) 2) is able to provide for his living needs without this income, even if disabled or (reference 16) unemployed, SACON-provides engineers with 3) does not plan to have trust income pa'y life-insurance premiums on his life advice• on structural analysis (ref­ without the consent of an adverse party, erence I) 4) does not plan to use the trust for a leaseback of assets,. and A: has a person (e.g., a parent) he is supporting without legal obligation with Because of the popularity of RBSs. S) this property income, several domain-independent systems B: has a child, not a minor, that he will be putting through college with this have been developed to make it much property income, or easier to build rule-based expert C: is using some of his after-tax income for the benefit of some other taxpayer, systems in many fields. The following is a partial list of domain-independent then systems:

client should form a temporary trust EMYCIN (reference 30) • (back to question 4 ...) AGE (reference 21) • OPS5 (reference 16) ..YES • ADVISE (reference 18) • Hearsay-3 (reference 7) Is client able to provide for his living needs without this income, even if disabled • 5) ALIX (reference 23) or unemployed? • (8) i Expert-Ease (Human Edge Software ..YES [Whenever a response s made with less than certainty, the system user • enters a number betwen 1 and 9 indicating his degree of certainty in that response.] Corp .. Palo Alto. California) 6) Does client plan to have trust income pay life-insurance premiums on his life KS 300 (Revamped EMYCIN; from without consent of an adverse party? Te• knowledge Inc .. Palo Alto. Califor- ..NO nia) · 7) Does client plan to use the trust for a leaseback of assets? KES (lntelligenetics Inc .. Palo Alto. ..NO • 8) Does cli.ent have a person he is supporting without legal obligation? California) **YES Personal Consultant (Texas Instru­ • I recommend that the client form a short-term trust. ments Inc.. Dallas. Te xas) [The degree of certainty that the system has in this recommendation is .8. This certaintyfac tor (CF) was calculated as follows. The temporary trust rule's action CF An RBS is composed of PDMs was 1.0 and it had an '1\ND" premise. In such a case, the rule's CF is the minimum called rules. each with a left-hand side CF used in the responses, or .8. Since the system's threshold CF is .2, the (the antecedent. a logical combina­ recommendation was made.] tion of propositions about the data­ (end) base) and a separate right-hand side (the consequent. a collection of ac­ Figure 4: A partial interactive consultation with TAXADVISOR. The user's input is tions). An RBS separates data ex­ in uppercase. amination (done by the left-hand side) from data modification (done by the

306 8 Y T E APRIL 1985 • Inquiry 168

EXPERT TECHNOLOGY Why do major hardware vendors right-hand side of the rule). be found. To satisfy each antecedent. Most RBSs are production systems which represents a subgoal. the sys­ endorse (PSs). in which matching and schedul­ tem collects those rules whose con­ ing are explicitly defined by the oper­ sequents satisfy its value. The process GoLDEN ation of the executive (control) pro­ of working backward through the LISP? gram. The control schema can be rules from consequents to anteced­ COMMON"Wang Laboratories sees the COMMON LISP Standard as an characterized as having four basic ents to consequents in search of a .l'mporta,ntelement of its tirategy parts: causal chain that will satisfy the goal fo r advanced Offtee Automation. is called backward chaining. (For a Wa ng Laboratories and Gold Computers are both com­ I. Selection: select relevant rules and simple backward-chaining program Hill data elements. Selection may be mitted to bringing Artificial written in BASIC. see "Knowledge­ Intelligence technology to Of fice trivial (e.g.. on each cycle all rules and Based Expert Systems Come of Age" Automation and see COMMON all data elements can be considered) by Richard Duda and John G. LISP on the Wa ng productfa mily 0. or quite complex (e.g .. special filters Gaschnig. September 1981 BYTE. as a vital step." can be designed to eliminate from page 238.) Leo Chan consideration many rules that could With antecedent-driven (forward­ Vice President, R&D Wang Laboratories not possibly match the current data). chaining) systems. program execution "Idealfo r entry-level op era­ In TAXADVISOR. rules are organized consists solely of a continuous se­ AI in a hierarchy to narrow the rules quence of cycles terminating when a ,tions, the program package was designed to provide training fo r considered. rule's action dictates a halt. At each programmers as wellas fo r . 2. Matching: compare active rules cycle. the system scans the anteced­ program development. As a against active data elements. looking ents and determines all rules with subset of COMMON LISP, it is for patterns that match. i.e.. rules antecedents that are satisfied by the compatible with Digiral's whose conditions are satisfied. Figure contents of the database. If there is recentlyannoun ced VA X LISP " 3 is an example of pattern matching. more than one such rule. select one Arnold Kraft AI Marketing 3. Scheduling: decide which "satis­ by means of a conflict-resolution Digital Equipment Corporation fied" rule should be "fired:· "Firing" strategy. Perform all actions asso-. consists of accessing and executing ciated with the selected rule and "With the availability of these packages, the DA TA GENERA L/ the procedures associated with the change the database accordingly. For One gives engineering users a pattern elements that matched the example. with Rl (XCON). you enter personal expert system' with current data. If morethan one rule is all the information on the problem both a standalone software satisfied. conflict-resolution heuristics into the database. and the system development workstation and a are used to decide which rule to fire. then applies the rules to reason for­ distributed development tool in a single, portable computer." 4. Execution: fire the rule chosen dur­ ward from the data to the conclusions. Don McDougall, Vice President ing the scheduling process. The result In summary. forward chaining consists and General Manager of execution is a modification of data of putting the rules in a queue and Te chnical PnDduots m'visimi elements or structure. With TAXAD­ then using a recognize-act cycle on Data General VISOR. execution results in an estate­ them. "GOWEN COMMON LISP isthe planning recommendation for a client. Some forward-chaining systems try best AI delivery environment This is illustrated in the test consulta­ to control the search for rules in the available on a PC. Ne tworks tion in Figure 4 (see reference 3I). recognize cycle by grouping rules into of PCs can be connected to Sy mbolics LISP machines to packets. These rule groupings are ap­ provide powerfu l distributed PSs are either consequent-driven pealing conceptual structures. since AI applications at affordable systems or antecedent-driven systems. they group rules according to the sub­ prices." A consequent-driven (backward­ topic that they deal with. Object­ Bruce M. Gras chaining) system. which is the type oriented programming can also be ViCe President, ¥arketing used in TAXADVISOR. uses rule con­ used to organize collections of rules. Symbolics, Inc. sequents (which represent goals) to In object-oriented programming. we guide the search for rules to fire (with give objects behavior. and thus we TAXADVISOR. · estate-planning ac­ can distribute the control of rules into tions to recommend). The system col­ rule. rule-packet. and domain objects. lects those rules that can satisfy the This approach. which has been taken goal in question and tries to satisfy in LOOPS. a domain-independent sys­ the consequents of those rules. which tem (see reference 27). also allows AI Solutionsfo r Personal Computing usually represent the values of vari­ multiple instantiations of the same set 163 Harvard Street, Cambridge ables. In order to find these values. of rules to solve subproblems of the Massachusetts 02139 (617) 492-2071 the values of the rule antecedent must See our ad on page [continued) 129

APRIL 1985 • BYTE 307 Finally, business computer software for hard-nosed.

No one takes a harder look at stands in the way of progress. software than small to mid-sized Thanks to multi-user capability, two businesses. enough for the novice hard-nose or more hard-noses can use the So take a long, hard look at to use within minutes of receiving same application at the same time. The Accounting Solution� a new, the package . Ye t it's also sophisti­ totally integrated software package cated , offering unlimited flexibility Hard-nosed flexibility. from Business To ols, lnc!M and opportunity to the hot-shot With The Accounting Solution, Yo u'll find its breakthrough fea­ hard-nose. And it's designed to it's easy to change your mind tures are designed to pay off where run on CP/M-80, MP/M-80, IBM because the source code is so it counts-on the bottom line. PC and compatibles: simple to modify. Ready to grow? • Great. Yo u can change hardware Hard-nosed economy, $99� Multi hard-nose capability. without spending a dime on new Contrary to popular opinion, The Accounting Solution never software. you don't need a small business loan to buy quality software . Not Ta ke it from hard-nose if you're buying The Accounting Phil Mickelson. Solution. For you get a iiU�I�iiS:� i�iJuL: Phil created The Sensible language/data $99, base manager with TM Solution�· a highly respected soft­ more hard-nose capabilities and ware package. Now he's offering speed than any program available the next step , another break­ at any price; buys the II • through: The Accounting Solution. language plus$249 accounts receivable/ It's simple. Sophisticated. Affordable. payable and general ledger; And backed by Phil's reputation gets you all the above plus inven­$399 [[DUD:LIDg and personal service. If you're tory control, sales order entry, looking for hard-nosed value and purchase order entry and payroll. quality, you'll agree, The Accounting Even more good news for the Salutia ..... Solution is the only solution. budget minded -source code is included with applications. Write or call: Suggested retail price. • •·cP/M-80 and MP/M-80 are registered trademarks of Business To ols, Inc. Digital Research, Inc. ; IBM PC is a registered trade­ Easy for any hard-nose. 4038-B 128th Avenue SE mark of International Business Machines Corporation; The Sensible Solution trademark rights are claimed by The Accounting Solution is easy Bellevue, WA 98006 O'Hanlon Computer Systems.

1-Wa800shington-648 -6State258: (206) 644-2015

Dealer inquiries welcome. 308 BYTE APRIL 1985 Inquiry 53 • Inquiry 169

EXPERT TECHNOLOGY Why do large corporations purchase same type within one session. resented by networks or frames. A The primary difference between semantic network is a graph of the GoLDEN backward and forward chaining is a relations. A frame or script system LISP? COMMON"I'm not a programmer. Having top-down versus bottom-up style of (see references 20 and 24) organize completed the Gold Hill tutorial, linking rules together. Though the the objects and their relations into en­ Ife el comfortable in starting to most common. these are not the only tities (recognizable collections of ob­ write usefu l AI programs. We control structures for rule-based jects). Frame systems also provide a hope to build fr iendlier user interfa ces using expert sys tems. systems. For example. rules are rep­ system to inherit attributes from a tax­ Th is product should help us to resented as an "inference" network in onomy of entities. Thus. a frame develop the in-house expertise PROSPECTOR (see reference 5). system implements the semantics of in a cost-effe ctive manner." some of the relations between ob­ H.M. Seeburg DEEP REPRESENTAT IONS jects. With a semantic-net or frame Program Management Systems Frame- and network-based ap­ system you can represent objects of Hughes Aircraft Company proaches allow the implementation of the domain of expertise as well as the "The AI Group at Arthur Andersen Company is intend­ "deeper-level" reasoning such as process. strategies. etc.. that are also & abstraction and analogy. Reasoning part of the domain. The control of ing to use GOLDEN COMMON LISP as a delivery environment by abstraction and analogy is an im­ frame or semantic-net systems is fo r a major internal application. portant expert activity. Yo u can also usually much more involved than with We are also using it as a vehicle represent the objects (e.g .. "pump" in surface systems and is implemented fo r tra ining a large base offirm figure 5) and processes (e.g .. the "start" in a way that an explanation facility personnel in AI technology." instructions in figure 5) of the domain can't get at. But surface systems are Bruce B. Johnson of expertise at this level. What is im­ "shallow"; a surface system may be Partner in charge of the AIGroup Arthur Andersen & Company portant are the relations between objects. viewed as a projection of deep-level primary concernfo r Litton Deep-representation expert systems knowledge of a domain for a specific 'M Industries has been fi nding perform inference using relations rep- (continued) app ropriate delivery vehicle;m ifo r AI applications. Th e availability of thisproduct op ens up new ISOL ATED TWO -PORT DEVICE avenues of potentia lfo r us." TO ALIGN : •OPEN INLET VA LVE Sy Schoen • OPEN OUTLET VALVE Program Manager for AI Litton Industries "GOWEN COMMON LISP MOTOR -DRI VEN PUMP provides a powerfu l LISP-envi­ TO START : ALIGN PUMP M ronment that allows even a I.2. START MOTOR novice to createmean ingful LISP progmms. With the On-line Help fa cilities ant:{tutorial, the product is an inexpensive entry­ 0 pointfo r companies of all sizes into this critical technology. " Brad Millman Member of AI Core Group Arthur D. Little CEN TRIFUGAL PUMP WHEN START : "GOLDEN COMMON LISP is • OPEN SUCTION VALVE exciting! will do fo 1· AI what It BEFORE START MOTOR Wa ng Laboratories did fo r word • OPEN DISCHARGE VALVE AFTER START MOTOR processing-popularize it. " Dan Corwin Software Architect Wang Laboratories

MAIN CONDENSA TE PUMP TO START: OPEN SUCTION VALVE 2.I. START MOTOR OPEN DISCHARGE VALVE 3. AI Figure 5: Procedure steps are obtained from the subcomponents and abstractions of an Solutions fo r Personal Computing 163 Harvard Street, Cambridge object. here a main condensate pump. This example comes from Steamer. a tutorial Massachusetts system designed to teach operating procedures of shipboard steam plants. 02139 (617) 492-2071 See our ad on page 129

APRIL 1985 BYTE 309 • EXPERT TECHNOLDGY

being told with a single associated file of ex­ Some systems have • analogy amples. Instead. first break it down • example into a main problem and a set of sub­ a built--in capability • observation. discovery, and experi­ problems. even going further (to the mentation• level of sub-subproblems) if the com­ fo r taking a file reasoning from deep structure plexity of the problem domain seems • to call for it. The originators of this The manual acquisition of knowledge style. which is known as "structured of expert decisions from human experts is a very labor­ induction." are Drs. Shapiro and intensive process. There is an ac­ Niblett (reference 2 5). Corporations and generalizing knowledged need to have aids for enjoying the use of powerful inductive knowledge acquisition as part of the generators such as ITL:s FORTRAN­ from this knowledge system. based EXCfRANsystem or Radian Cor­ Methods to speed knowledge ac­ poration's C-coded RuleMaster have quisition are now becoming available applied the approach to the building an executable rule. in the form of machine learning of of complex systems for trouble­ rules from examples. Systems such as shooting large transformers. severe­ expert activity. Expert-Ease have a built-in capabili­ storm warning. circuit-board fault One type of expertise that has been ty for taking a file of expert decisions diagnosis. and user-friendly guidance represented with a deep-level ap­ from you and generalizing from these to set up numerical batch jobs in proach is tutoring (see "The LISP an executable rule. In a sense. you are seismic analysis in the oil industry. 1\.Jtor" by John R. Anderson and Brian able to transplant chunks of decision­ Rates of production of compact in­ ). Reiser on page 159). Here we want making skill from your own brain to stalled code in excess of 100 lines per to convey to the pupil domain knowl­ the personal computer. a possibility worker day are now commonly re­ edge that is best represented at the foreseen as early as 1966 by Earl ported. deep level: concepts. abstractions. Hunt and his colleagues. Any robust expert system takes a analogies. and problem-solving strat­ The machine procedure that allows tremendous amount of resources to egies. this skill transplant was developed develop. Once developed. the knowl­ Steamer is a training aid developed from a Pascal-coded program called edge along with the control structure jointly by Bolt Beranek and Newman ID3 (Iterative Dichotomiser 3) due to can be "compiled out"; that is. the Inc. and the Navy Personnel Research Professor Ross Quinlan of the New system of rules is rewritten into a and Development Center. Its goal is South Wa les Institute of Science and piece of code that performs the same to teach operating procedures of Te chnology. function on a personal computer. For shipboard steam plants. These pro­ A number of conclusions follow example. some expert systems (AD­ cedures consist of a series of steps on from Quinlan's work: VISE. EMYCIN, OPS5-see reference subcomponents of the plant. The 10) can generate code or other I. It is possible. using such a program. components and procedures are rep­ primitive forms of the knowledge for to generate machine-executable solu­ resented as frames in Steamer. as are use on a personal computer. (Systems tions for complex decision problems the abstractions of components and run on a personal computer are usualc ·in a fraction of the time a program­ procedures that experts use in ly referred to as "delivery systems:') mer would need for developing a teaching steam-plant operations. The solution by conventional hand coding. steps of a procedure come from the KNOWLEDGE REPRESENTATION 2. The resulting solutions are super­ abstractions and subcomponents of As AI researchers point out. a robust efficient as compared with those ob­ the device the procedure applies to. expert system that can explain. justify. tainable by the old hand methods. The ordering of the steps comes from acquire new knowledge. adapt. break 3. It is important to make up your a third represented entity: operating rules. determine relevance. and mind in advance whether super­ principles. These principles are culled degrade gracefully will have to use a efficiency is all you demand of a from experienced operators and multitude of knowledge representa­ machine-executable solution. or represent "compiled" knowledge of tions that lie in a space whose dimen­ whether you also want the resulting steam-plant operation (although they sions include deep/surface. qualita­ rule base to be understandable on are not represented as rules but tive/quantitative. approximate (uncer­ inspection. frames). tain)/exact (certain). specific/general. If the answer to the third statement and descriptive/prescriptive. Systems KNOWLEDGE ACQUISITION above is that user transparency of in­ that use knowledge represented in The following are ways of acquiring duced rules is desired. then (unless it different forms have been termed knowledge in a form that can be used is a very small one) do not treat your multilevel systems. Steamer is an exam­ by an expert system (reference 19): problem as one big superproblem ple of one such expert system.

310 BYTE • APRIL 1985 Inquiry 170

EXPERT TECHNOLOGY Why do the AI experts recommend Steamer uses the following represen­ plex expert system using an existing tations: domain-independent system or if the GoLDEN system has a rule-compilation facility LISP? I. A graphical (icon) representation of "We are excited about this that allows applications to be run on COMMON the objects of the Steamer domain, product and its potential to personal computers. then a personal such as va lves. pumps. tanks. and sys­ op timize the method by which computer (preferably with 512K bytes) people learn.Included in our tems composed of these. is sufficient. If all you need are fo rthcoming book on Automatic 2. A frame representation of Steamer resources to run an existing expert Deduction and Theorem Prov­ objects. procedures. and operating ing will be software written in system. a large personal computer principles. This is used for describing. GOLDEN COMMON LISP Th is should nearly always be sufficient. will give studentsfir st-hand explaining. categorizing. abstracting. There is no obvious line of demar­ experience with advanced pro­ and referring. cation for a given project. However. gra ms written in the standard 3. An assertional database where LISP dialect on their own PC." certain barriers make personal com­ assertions about Steamer entities can Woody Bledsoe, President puter use less desirable as system size be made and retracted. American Association for and complexity increase. 4. A quantitative numerical simulation Artificial Intelligence of the steam plant that is used in il­ Michael Ballantyne SYSTEM BARRIERS University of Te xas, Austin lustrating cause and effect and rami­ Many high-level languages do not of­ fications of the application (or misap­ "I'm used to working on a fer the right primitives (i.e .. program­ Sy mbolics yet I am quite of procedures. 3600, plication\ ming-language statements) for devel­ comfortable moving to the PC using GOWEN COMMON LISP. Wo rk is just beginning in building oping expert systems. Among the de­ GCLISP is a very respectable such multilevel systems. and they will sirable primitives are subset of the COMMON LISP be a major research topic for this A parser or interpreter that parses dialect.... In summary: this is decade. Wo rk needs to be done in a superb product. It puts state­ state• ments during program run time. studying and representing in a of-the-art LISP programming Without this. you have to write a technology into the hands of general way the different problem­ parser for the rules. anyone who can affo rd a PC." solving activities an expert does (see List and nonnumeric processing David To uretzky reference 3). When you build expert • primitives. Computer Science Department systems. you realize that the power Carnegie Mellon University A language design that allows in­ behind them is that they provide a cremental• compilation and other fast "Gold Hill has an enormous regimen for experts to crystallize and competitive advantage in the prototyping facilities. Incremental codify their knowledge. and in the AI game. It is located next door compilation enables you to recompile knowledge lies the power. to MIT, and has direct access to a function or other portion of a file the students and fa culty of the without recompiling the entire file. MIT AI Lab. Th e people at Gold RESOURCES NEEDED Hill have done some highly orig­ Before resource needs are discussed. The view that many people in the inal thinking about how to dra ­ you must precisely define the type of field are adopting is that high-level matically increase the amount of comp uting power available expert system you want to build. If languages like Pascal. Ada. and C are to personal computer users." you wish to build a large. "custom" acceptable for the delivery system. Howard Austin, President, model expert system (i.e .. it is not but for prototyping. a language like Knowledge Analysis Inc. feasible to use many of the smaller LISP or Prolog is preferred. Program­ "There are a lot ofpeople domain-independent systems that are generation tools are then used to eager to get their hands on this available). you will need substantial write the system in the delivery stuff Ith ink this will bring AI resources: large memory. good lan­ language. to the masses." guage support. and fast execution of The knowledge-intensive approach Patrick H. Winston the code. Yo u may need to develop to expert systems implies that the Director of the AI Lab, MIT President-elect, American such a system in LISP on hardware memory will be highly utilized in all Association for Artificial specialized to processing the lan­ but the most nontrivial applications. Intelligence · guage. or on time-sharing machines ALIX is one example that ran on a with a large address space. Such 64 K-byte machine. but it was a small "custom" systems are usually referred expert-system shell. Since memory to as "prototype" or "development" prices have gone down and many D H I L l systems. They can either be devel­ small machines have broken the 64K­ P u T e R: s AI oped for a specific domain (e.g .. byte barrier. we can expect that more Solutions fo r Personal Computing MYCIN) or be domain-independent expert systems can be developed. at 163 Harvard Street, Cambridge 02139 (617) 492-2071 (e.g., ADVISE). least for the delivery system. on per- Massachusetts See our ad onpage If you are able to build a less com- (continued) I29

APRIL 1985 BYTE 311 • EXPERT TECHNOLOGY

ed. London and New Yo rk: Gordon and T!ie International Journal of Knowledge Engineer­ Some researchers Breach. 1982. ing. October. 1984. page 149. 3. Chandrasekaran. B.. and Sanjay Mittal. 18. Michalski. R. S .. A. B. Baskin. A. predict that memory "Deep Ve rsus Compiled Knowledge Ap­ Boulanger. R. Reinke. L. Rodewald. M. proaches to Diagnostic Problem-Solving:· Seyler. K. Spachman. and C. Uhrik. 'A International Journal of Man-Mac!iine Studies. Te chnical Description of the ADVISE Meta needs of advanced #19. 1983. page 425. Expert System:· Department of Computer 4. Davis. R. "Expert Systems: Where Are Science. University of Illinois at Urbana­ We? and Where Do We Go From Here?" Champaign. 1983. expert systems will AI Magazine. Spring 1982. page 3. 19. Michalski. R. S.. ). Carbonell. and T. 5. Duda. R .. ). Gaschnig. and Hart. Mitchell. eds. Mac!iineLearn ing: An Artificial P. drive development of "Model Design in PROSPECTDR Consul­ Intelligence Approac!i. Los Altos. CA: Tioga tant System for Mineral Exploration." Publishing Company. 1983. £SMA . 1979. page 153. 20. Minsky, M. "A Framework for encyclopedic memories. 6. Duda. R. and E. H. Shortliffe. "Expert Representing Knowledge." T!ie Psyc!iology of 0.. Systems Research." Science. April 198 3. Computer Vision. P. Winston. ed. New Yo rk: · page 261. McGraw-Hill. 197 5. 7. Erman, L. D. . E. London. and S. 21. Nii. H. and N. Aiello. "AG E (Attempt sonal computers. Some researchers P. F. P.. Fickas. "The Design and Example Use of to Generalize): A Knowledge-Based Pro­ predict that the memory needs of ad­ .. vanced expert systems will drive the Hearsay 3 . Proceedings of /fCA no. 7. 1981. gram for Building Knowledge-Based Pro­ development of encyclopedic memo­ page 409. grams." 1/CA . 179. 1979. page 645. ries for personal computers. 8. Feigenbaum. E. A.. B. G. Buchanan. and 22. Pople. H. E. J. D. Myers. and R. A. ). Lederberg. "On Generality and Problem Miller. "Dialog: A Model of Diagnostic Solving: A Case Study Using the Logic for Internal Medicine." IJCA . 175. CONCWSION DENDRAL Program." Mac!iine Intelligence 1975. page 848. 6. Expert systems can be built in many B. Meltzer and D. Michie. eds. New Yo rk: 23. Reiter. ). "ALIX: An Expert System ways, involving rules. networks. Edinburgh University Press and Halsted Using Plausible Inference." Intelligent frames. and combinations thereof. Press (Wiley). 1971. page 165. Te rminals Ltd .. University of Edinburgh. with all sorts of variations within these 9. Forbus. Ke nneth D. "Qualitative Process 1980. Theory:· MIT Te chnical Report 789. MIT AI 24. Schank. R. C.. and R. Abelson. Scripts. categories with respect to knowledge P. representation and control. We could Laboratory. May 1984. Plans. Goals. and Understanding. Hillsdale. N): not begin to cover all possible ap­ 10. Forgey. C. L. "Rete: A Fast Algorithm Larrence Erlbaum Associates. 1977. for the Many Pattern/Many Object Match 2 5. Shapiro. A.. and Niblett. 'Automatic proaches to building expert systems. T. Artificial Intelligence. since new ones are being developed Problem." September Induction of Classification Rules for a 1982. Chess Endgame:· Advances in Computer C!iess almost daily. II. Hasling. Diane Wa rner. William ). 3. M. R. B. Clarke. ed. Oxford: Pergamon. Even if the most efficient approach Clancey, and Glenn Rennels. "Strategic Ex­ 1982. has been ascertained for the domain planations for a Diagnostic Consultation 26. Shortliffe. E. H. Computer-Based Medical in question. the most cost-effective System:· International Journal of Man-Mac!iine Consultations: MYCIN. New Yo rk: American computer resource must still be deter­ Studies. January 1984. page 3. Elsevier/North-Holland. 197 6. .. mined. In most cases. approach selec­ 12. Hart. P. "Direction for AI in the 80's. 27. Stefik. Mark. Daniel G. Bobrow. San­ tion at least narrows the choice for 5/GART Newsletter. November 1981. page II. jay Mittal. and Lynn Conway. "Knowledge resources; in some cases. approach I 3. Hollan. )ames. Edwin Hutchins. and L. We itz­ Programming in LDOPS: Report on an Ex­ and resources can be selected to­ man. "Steamer: An Interactive lnspectable perimental Course:· AI Magazine. Fall 1983. Simulation-Based Training System:· AI Magazine. gether. However. this hardly reduces 28. Stevens. Albert. and Bruce Roberts. Summer 1984. page 15. "Quantitative and Qualitative Simulation the complexity of the choice. To make 14. Hutchins. Edwin. Te rry Roe. and James in Computer Base Training:· Journalof Com­ matters worse. computer resources Hollan. "Project STEAMER: VII. A puter Based Instruction. volume 10. numbers are changing as rapidly as the new Computer-Based System for Monitoring I and 2. Summer 1983. page 16. system-building approaches are being the Boiler Light-Off Procedure for a 29. Swartout. R. 'A Digitalis Therapy .. W. developed. The best we can.hope to 1078-Ciass Frigate . NPRDC Te c!inical Note Advisor with Explanations." Technical Report 82-85. August 1982. MIT Lab for Computer Science. convey is an awareness of the oppor­ 176. tunities and complexities involved in 15. Kunz. ) ; C.. et al. 'A Physiological Rule­ February 1977. Based System for Interpreting Pulmonary 30. Van Melle. Domain-Independent the development of expert systems. .. W. 'A • Function Te sts. Heuristic Programming Production Rule System for Consultation Project. Memo HPP-7 8-19. Stanford Programs:· CA . 179. 1979. page 92 3. REFERENCES I/ Bennett. J. S.. and R. S. Englemore. University. 1978. 31. Wa terman. D. A.. and Hayes-Roth. J. F. "SACON: A Knowledge-Based Consultant 16. McDermott. J. "R I: A Rule-Based Con­ eds. Pattern-Directed Infe re nce Systems. New for Structural Analysis:· /fCA . 179. 1979. figurer of Computer Systems." Computer Yo rk: Academic Press. 1978. page 47. Science Department. Carnegie-Mellon 32. Whalen. Thomas. and Brian Schott. 2 . Bramer. M. A. 'A Survey and Critical University. 1980. "Issues in Fuzzy Production Systems:· In­ Review of Expert Systems Research." In­ 17. Michaelsen. R. H. 'An Expert System ternationalJournal of Man-Mac!iineStudies. # 19. troductory Readings in Expert Systems. D. Michie. for Federal lax Planning." Expert Systems: 1983. page 57.

312 BYTE • APRIL 1985