Stanford Artificial Intelligence Laboratory Memo AIM-252 July ...

Stanford Artificial Intelligence LaboratoryMemo AIM-252

July

. -

Computer Science DepartmentReport No. STAN-B-74-466

RECENT RESEARCH IN ARTIFICIAL INTELLIGENCE,HEURISTIC PROGRAMMING, AND NETWORK PROTOCOLS

Edited byLester Earnest

ARTIFICIAL INTELLIGENCE PROJECTJohn McCarthy, Principal Investigator

HEURISTIC PROGRAMMING PROJECTEdward Feigenbaum and Joshua Lederberg,

Co-principal Investigators

NETWORK PROTOCOL DEVELOPMENT PROJECTVinton Cerf, Principal Investigator .

Sponsored byADVANCED RESEARCH PROJECTS AGENCY

ARPA Order No. 2494

COMPUTER SCIENCE DEPARTMENTStanford University

1 974

Stanford Artificial Intelligence LaboratoryMemo AIM-252

July 1974

Computer Science DepartmentReport No. STAN-(X-74-466

RECENT RESEARCH IN ARTIFICIAL INTELLIGENCE,HEURISTIC PROGRAMMING, AND NETWORK PROTOCOLS

Edited byLester Earnest

ARTIFICIAL INTELLIGENCE PROJECTJohn McCarthy, Principal Investigator

HEURISTIC PROGRAMMING PROJECTEdward Feigenbaum and Joshua Lederberg,

Co-principal Investigators

NETWORK PROTOCOL DEVELOPMENT PROJECTVinton Cerf, Principal Investigator

ABSTRACT

This . IS a progress report for ARPA-sponsored research projects in computer science for thepertod July 1973 to July 19’73. Accomplishments are reported in artificial intellig-ence (especiallyheuristic pro~ramm1ng, robotics, theorem proving, automatic programming, and natural languageunderstandIng), mathematical theory of computation, and protocol development for computercommun~rar~on networks. References to recent publications are provided for each topic.

This research was supported by the Advanced Research Projects Agency of the Department o fDefense u.nder Contract LIAHC 15-73-C-0435. T/le views and conclusions contained in thisdocument are those of the authors and should not be interpreted as necessarily representing theofficial policies, either expressed or implted, of the Advanced Research Projects Agency or the U. S.Government.

TABLE OF CONTENTS i

Sect ion Page

1. INTRODUCTION 1 3.2.4 Rule Modification 27

2. ARTIFICIAL INTELLIGENCEPROJECT

2.1 Robotics 42. I. 1 Manrpulation 42.1.2 Assembly Strategies 6

2.2 Computer VIslon 62.2. I Description 62.2.2 Visual Guidance of a Vehicle 92.2.3 Mars Picture Analysis 10

2.3 Mathematical Theo?y ofComputation

2.3.1 [email protected] LCF

2.4 Heurlstlc Programming 122.4.1 Theorem Proving 122.4.2 Automatic Programming 14

2.5 Natural Language2.5. I Speech Recognition2.5.2 Natural Language

Understanding2.5.3 Higher Mental Functions

- 2.6 Programming Languages

2.7 Computer Facilities2.7.1 Hardware2.7.2 Software

3. HEURISTIC PROGRAMMINGPROJECT

3.1 J<nowledge-based Systems Design

3.2 The Meta-DENDRAL Program3.2. I Data Selectron3.2.2 Data Interpretation and

Summary: The INTSUMProgram

3.2.3 Rule Generation: TheRULECEN Program

3

111111

1616

1720

20

212122

23

23

2425

25

26

Sect ioll Page

3.3 Systems for Semi-AutomaticKnowledge Acquisition byExperts 27

3.4 Knowledge Acquisition andDeploy men t

3.4.1 Background3.4.2 Objectives3.4.3 Methods of Procedure3.4.4 Overall design of the program

2829303031

3.5 Knowledge Deployment Research:Inexact Reasoning 35

3.6 Knowledge Deployment Researchfor Real-World Applications

3.6.1 WHY Questions -- Looking atGoals

3.6.2 HOW questions: Looking atPreconditions

37

38

38

3.7 Knowledge Representation:Production Systems

3.8 Application of AI Techniques to aProgrammer’s Task:Automatic Debugging

3.9 Tools and Techniques

40

41

3.10, Technology Transfer: Chemistryand Mass Spectrometry 42

3.11 Technology Transfer: to Biologyand Medicine

3.12 Technology Transfer: to MilitaryProblems

3.13 Publications of the Project,1972/1974

43

44

44

i i

Sect iori

T A B L E O F C O N T E N T S

Page

4. NETWORK PROTOCOLDEVELOPMENTPROJECT 47

4.1 Internetwork Protocol Design 47

4.2 PDP- 11 Expansion 48

4.3 Software Systems 48

Appendices

A. ACCESS TO DOCUMENTATION 51

B. THESES-_

53

C. FILM REPORTS 57

D. EXTERNAL PUBLICATIONS 59

E. A. I. MEMO ABSTRACTS 65

1

1 . I N T R O D U C T I O N

.

This is a report of accomplishments by threeongoing projects that have been supported bythe Advanced Research Projects Agency(ARPA) In the period July 1973 to July 1974.Some related research supported by othera g e n c i e s ( m a i n l y N S F , N A S A , N I H , a n dN I M H ) i s a lso d iscussed. W h e r e n o totherwise stated, the work reported below wassupported by ARPA.

The Artificial Intelligence Project is the oldestand largest of the activities treated here. Itwas organized by John McCarthy, Professorof Computer Science, in 1963 and hasreceived ARPA support _contmuously sincethen. It has Included work in computervision, robotics, mathematical theory ofcomputation, theorem proving, speechrecognition, natural language understanding,programming language development, and anumber of other activities. ARPA budgeted$1.25 million in support of this work for theyear of this report.

The Heuristic Programming Project wasformed in 1965 by Edward Feigenbaum,Professor of Computer Science, and JoshuaLederberg, Professor of Genetics, and wasinitially an e l e m e n t o f the ArtificialIntelligence Project. It became a separateo_rganizational entity with its own budget inJanuary 1970. The central interest of thisproject has been artificial intelligence appliedto scientific endeavor and the problems ofknowledge acquisition, representation, and usethat arise in constructing high-performanceapplications of AI. ARPA support for theyear amounted to 8200K.

The Network Protocol Development Projectwas formed in July 1973 by Vinton Cerf,Assistant Professor of Computer Science andElectrical Engineering, and has beenconcerned with communication protocols forcomputer networks, especially the ARPAnetwork. ARPA support to this activity was#50K for the year.

This report updates and builds upon our tenyear report [ 11. Like most progress reports, itis mainly a concatenation of segments writtenby. the indiv iduals w h o d id t he work .Consequently, there are substantial variationsin style and depth of treatment.

The fo l lowing sect ions summar ize recentaccomplishments and provide bibliographiesin each area. Appendlces list theses, films,books, articles, and reports produced by ourstaff.

Bibliography

[I] Lester Earnest (ed.), FINAL REPORT:The First Tell Years of ArtificialIntelligerlce Research at Stanford,Stanford A. I. Memo AIM-228, July 1973.

I N T R O D U C T I O N

Stanford Arm Assembling Water Pump

2 . A R T I F I C I A L I N T E L L I G E N C EP R O J E C T

Automat ic Programming

The work of the Art~ficlal Intelhgence Projecthas been basic and applied research inartificial intelligence and related fields, such asmathematical theory of computation. Here isa short list of what we consider to have beenour maln accomplishments during the pastyear.

Robotics

We have devrlopcd a two-arm synchronizedmanlpulatlon capabtllty a n d t e s t e d It o nseveral mechanrcat assembly tasks that arebeyond the capahlllty of a-‘single arm. A newhigh-level “hand language” called HAL hasbeen developed for speclfylng advancedmanlputatlon tasks.

A successful new automatic programmingsystem accepts descriptions of library routines,programmmg methods, and programspecifications in a h igh l eve l s eman t i cdefimtion language. It returns programswritten in a subset of ALGOL that satisfy thegiven specifications. Experimentalapplications include computing arithmeticalfunctions and planning robot strategies.

Another system works with algorithmsexpressed in a higher-level language andautomatically chooses an efficlen trepresentation for the data structure. It thenproduces a program that uses thisrepresentation. Representations consideredinclude certain kinds of linked-lists, binarytrees, and hash tables.

Natural Language UnderstandingConi pu ter Visioii

We have used near and far field stereo visronand motion parallax to locate objects spatiallyand to automatlcatly generate contour maps.Another program can recognize thmgs of thecomplexity of a doll or a hammer in variousposltions, using a laser trlangulatlon system.

A system called MARGIE was completed thatlinks natural language understanding,inference, and generation of natural languageoutput. This is the first such system with astrong theoretical basis, namely conceptualdependency.

TraithlgMathemat ica l Theory of Computatioil

-Using o u t LCF proof-checker, we havep r o d u c e d a n axlomatlzation o f theprogramming language PASCAL. Thisrepresents a ma JOI’ step toward using LCF asa practlcat program verification system.

Theorem Proving

During the y e a r , s i x m e m b e r s o f o u r s t a f fpublished Ph.D. dissertations and another 32graduate students received direct support.

The following sections review principalactivities, with references to published articlesand books.

An InteractIve system has been developed forstructured top-down programming inPASCAL. It guides the user in constructing aprogram in successive refinements and inproving its correctness.

4 A R T I F I C I A L I N T E L L I G E N C E P R O J E C T

2.1 Robotics

A group led by Jerry Feldman, Tom Binford,and Lou Paul has been developing automaticassembly techniques using general purposemanipulators, visual representation anddescrrptive techniques using television cameraand other sensory data. The vision work iscovered III Section 2.2.

2.1.1 Manipulatioll

The robotics group has established leadershIpin nian~pirlatlon and notes particularlya d v a n c e s m two a rm synchronizedmanipulation, and III design of a new handlansirage for manipulation.-_

T h e completion l a s t y e a r o f a u t o m a t e dassembly of a water pump by Paul and Belles[Belles] m a r k e d a c h a n g e 111 dlrectlon o fmanipulator research. In the previous phaseof system bulldlng, Paul [Paul] had developedsoftware for control of the Scheinman arm[Schemman] us~n$ t o u c h (one swrtch p e rfincyr wrth 10 g r a m sensrtrvlty) a n d f o r c e(measured from Joint motor currents, about350 gram sensltlvrty). The pump assemblytask showed the use of touch, force, tools, andvision 117 a complete system task. The newemphasis has been on application of thesystem to programming of repetltlve assembly

e tasks, and executing tasks chosen to developnew manlpulatlon abllitles.

Our conceptlun of the assembly task as aplanning task cat.1 red out once on a largesystem, and a small repetltlve esecutlon taskseems suited to Industrial assembly. The plancan be lntelllgently tarlored to the lndivldualtask; the small system repeatedly executes aplan and modifies the plan at runtime to takeinto account part to part variations.

In order to provide this runtime modificationit was necessary to move the arm solutionroutines from the planning stage to theruntlme system and to communicate posItIonsin terms of rectangular table coardmates

Instead of in terms of arm joint angles. Bycommunicatmg part positions in terms ofrectangular coordinates it was possible to‘translate and rotate sets of positions asnecessary to adapt the mampulator to eachactual part and i t s re la t ive mampulatorplacement.

W h e n this work was comple t ed , Bellesprogrammed automatic assembly of the plston-crankshaft subassembly of a two-strokegasoline engine, allowing considerablevariation in work piece positioning. Boltesprogrammed tool changing. T h e a r mautomatically removes one tool and mountsanother from a set of socket tools for anautomatic tool driver. A second arm wasinterfaced to our computer. Paul developed asystem m which two arms could be run insynchronization. He programmed assembly ofa hinge, using two arms. (These examples arereco rded In a film [31). To explore armdynamics, F i n k e l programmed throwingobjects Into bms sorted by size.

These tasks were carried out m the handlanguage WAVE [Paul]. Programming in ahand language gave a generality which mightbe described as: given that we had carried outone task, programming assembly of a similarobject would be simple. For example,programmlng assembly of a generator wouldrequire about 10 hours work, following thepump assembly. A set of macros weredeveloped which were applicable to a varietyof tasks: put a pm in a hole, insert a screw.About 8 hours are required to program amacro of the complexity of Inserting a screw.The hand language has made it simple toteach students how to program the arm. In arobotics course, students programmed “swordin stone”, inserting a shaft into a hole. Thelanguage WAVE is on the level of assemblercode.

In order to take advantage of more than onem a n i p u l a t o r , It IS necessary that themanipulators can be run simultaneously, elthelperforming independent subtasks or acting

2.1 Robot its

together. The cx~stlng language, Wave, wasnot designed to incorporate parallel operation,and was Inadequate to take advantage of theruntinie mocllficatlon feature alreadydcscrlbed. A new language was needed tospecify structures and attachments of parts,and to provide a sultable syntax in which toexpress parallel operation. It was alsonecessary to incorporate general expressionevaluntlon, Including matrlces and vectors. Anew language design was undertaken toincorporate these features. Superficrally, i tresembles Algal, as It provides for structuredprogramming and block structure.

H A L

Paul, Flnkel, i?olles and Taylor have begun anew hand language, HAL. The effort beganas a higher level language version of WAVE,to Include coordinated motion of two arms.The design was broadened to Include somestrategy generatlon. The system is madepartially model-based (versus procedure-basedas WAVE, ALGOL, etc). Some degree ofautoniatrc generatlon of sequences fdr non-independent operations whose order isimportant, has been included rn the design.In order to carry on the next stage ofgenerallzatlon beyond WAVE, the systemmust malntaln models of its world.

C.onsider modifying the pump assemblyprogram to assml)le a generator. An expertprogrammer IS needctl to modify the program,while for a model-hnsecl system, the englnrelcould input a new model (presumably from aprevious clcsrgn) and allow the system to dothe low level InterfacIng. The system could notperform that Interfacing if given only lowlevel traJect0t.y commands. We regard thissystem as a first level of model-based system.A major technlcal advance will be coordlnatcdtwo arm motion, as opposed to independenttwo arm motion. An Important part of thedesign process has been to express a numberof tasks (gedanken experiments) rn the newlanguage. The design of HAL was three-fourths completed during the period of thisprogress report.

New work has gone into .touch s e n s o rdevelopment. A new sensor with adequatesensltivlty and small size was built and testedby-. Perkins. The sensor seems adequate foruse In task execution, but requires moredevelopment in packaging and mounting onusable fingers.

A new collision avoidance package has beenprogrammed by Widdoes. The package findsa collision-free path for the first three joints ofthe arm, usmg an interesting strategy. Theprevious collrsion avoidance program [Pieper]used a very local search around objects andwas very time-consuming.

During the period covered In this report, wehave begun conversion of arm hardware to aPDP-1 l/45 and begun converting to a newhand/eye table which allows room for two armmanipulation.

Bibliography

[Belles] Robert Bolles, Richard Paul, The useof Seusory Feedback in a ProgrammableAssenlbly Systeul, Stanford A. I. MemoAIM-220, October 1973.

[Dobrotinl Dobrotln, Boris M., Victor D.Scheinman, Design of a ComputerControlled Manipula tor for RobotResearch, Proc. Third ht. Joint Conf. onArtificial Intelligence, Stanford U., 1973.

[Paul] R. Paul, Modelliug, TrajectoryCalculation and Servoing of a ComputerControlled Arul, Stanford A. I. MemoAIM-177, Ph.D. Thesis in ComputerScience, September 1972.

[Pieper] Donald L. Pieper, The Kinematicsof Manipulators under ConlputerControl, Stanford A. I. Memo AIM-‘72,October 1968.

[Scheinmanl V. D. Scheinman, Design of aCowputer Manipulator, Stanford A. I.Memo AIM-92, June 1969.

A R T I F I C I A L I N T E L L I G E N C E P R O J E C T

Films 2.2 Computer Vision

[I] Richard Paul and Karl Pingle, InstantInsanity, 16mm color, silent 6 mm, August1971.

[2] Richard P Iau and Karl Pingle AutowatedPump Assembly, 16mm color, Silent, 7min, April 1973.

[3] Pingle, Paul and Belles, AutolnatedAssembly! Three Short Examples, 1974(fot~thcol7llng).

2.1.2 Assembly Strategies

O u r w o r k a t programmIng manlpulatlonsequences applies to programming the class ofprogrammable assembly devices. The goal isto program at th,p level of assembly instructionmanuals: Insert shaft B into hole C. That isto go from a high level program to a programin terms of device motions, includmg f o r c eand con t 1-01 information. There IS a nenormous scope for such applications; the easeof programming assembly devices IS crucial totheir wide application, f r o m h i g h v o l u m especial automation to low volume generalpurpose devices. The effort has producedoutline programs for assembly of the waterpump (without manipulation programmmg)by Taylor, and by Luckham [Luckham]. A

- typical sequencing task IS to choose a sequencewhich does not involve putting down the toola n d picking it up again 111 pulling out theguide pins and insertrng screws to fasten thepump cover. As another facet of compatiblesequences, semarltlc constraints such as 2; on yare translatccl into mathematical constramts;Taylor has programmed a lmear constraintsotutlon package to solve the resultingmathematical conditions.

Bibliography

The theme of our work in visual perceptlon ofcomplex objects has been description and notclassification. We have concentrated onbuilding up capabilities for generatingstructured descriptions useful in a richumverse of objects.

2.2.1 Descriptioll

This proJect has been extended by Nevatia[Nevatia], with programs which recognizeobjects of the complexity of a doll, glove, toyhorse, or hammer. The work has includednew, stable hardware for the lasertriangulation system [Agin]. The programsuse depth data from the laser system, findboundaries of continuous surfaces, and makedescriptions of armlike parts according to arepresentation based on getleralized c o n e s[Binford]. Other groups have begun to usespecial cases of such representations. Theprograms then make complete structureddescriptions of objects as a part/whole graphof parts and relations of parts at joints.

Compact summary descriptions are abstractedfrom complete object descriptions and used toindex into a structured visual memory ofmodels of previously seen objects to locate asubclass of mode1 similar to the test object.The index procedure hmlts comparison of thetest object description to relatively few modelsfrom memory, e v e n wltti a l a r g e visualmemory of many objects (although only aboutSIS objects were used). Models in memorywere ciescriptlons made by the program ofpreviously seen objects, sometimes modified byhand. An important feature of the descriptionmatchmg process is that it depends ongeneratmg structured symbohc descriptions ofdifferences between test object description andthe model.

[Luckham] David Luckham, Jack Buchanan,Automatic Generation of ProgramsConta in ing Conditiollal Statements, Proc.

A.I.S.B. Summer Conference, Sussex,England, July 1974.

The descriptions themselves are intuitivelyfamiliar for humans, so that the decisions ofthe program are easy to understand anddebug. Although a great deal more work is

2.2 Computer Vi5iorl 7

necessary for that system, It represents a firstand slgniflcant step in such ciescrlption andrecognitron, particularly since it can toleratemoderate obscuration. The same techniquesare applicable to edge images from TV data;they give good descriptive ability for thatdomain. However, that is only a small part ofthe necessary system for analysis of TVimages and although useful, in no wayrPqPnlbles a solution to that complex problem.

Stereo Vision

It IS dltflcult for humans to performmanlpulatlon tasks from a single TV Image,wlthout s tereo. We intend to makeconsiderable use of stereo 111 applications ofvehicle navigation and v”rsual feedback forassembly. H a n n a h [Hannah] hasdemonstrated some techniques of stereomatching for stereo pairs with moderate stereoangle, IO degrees , without calibrationinformation. By matching a minimum of sixpoints in the two images, It was possible toobtain the relative calibration of the twocameras. Further search was limited bycalibration informatlon. Techniques weredeveloped to match corresponding areas Inoutdoor pictures from features including color,mean and variance of intensity. A programwas able to define regions bounded by depthdiscontlnulties.

Mot ion Parallax

Thomas and Plngle [Thomas] have appliedmotion parallax to simple scenes. They limltattentlqn to a few pomts defined by varianceor edge measures. These pornts are tracked asthe scene is rotated on a turntable, equivalentto moving the camera. The program requiresonly about 1 second per frame, using the SPS-41 computer. Although the research wasperformed on scenes of blocks, It 1s not limitedto such scenes. The mechanlsrn for selectrngpoints o f lriterest would be Inadequate forscenes with texture, however. Canapathy hasdeveloped techniques for wide angle stereomatching in scenes of blocks. These programs

use a variety of conditions of the form thatplanes remam planar under matching.

Bullock [Bullock] has made a systematic studyof available operators for description oftexture, and made a library of standardtextures. His informal conclusions are thatspatial domam descriptions are necessary, andthat known techniques are very weak.

We have continued our study of techniquesfor visual feedback to deal with scenes withrealistic contrast (not just black versus white)and with realistic complexity (curved objects).Perkins [Perkins] has made a program whichfinds corners for visual feedback In blockstacking. Although block stacking itself is oflittle interest, the program is interesting for itsability to function with realistic contrast levels(no special preparation of scene) andinterestmg for its global edge finding strategy.Perkins also made a program which foundelliptic curves among edge points from theH u e c k e l o p e r a t o r [Hueckel 1 9 7 11. T h eprogram was able to identify cylinders.

Vision Language

Binford has made a beginning on a languagefor vision research. Previously, the laboratoryhas built a hand/eye monitor system tosystematize cooperating routines at a job level;a library of simple procedures has beenimplemented [Pinglel. It has been found thatthe job level is too coarse to be useful foraccomplishing our objectives: to allow researchto build on previously built data structuresand modules; to allow a common vocabulary.The new effort is not predominantly a systemsoftware effort, but a scientific effort, aimed atproviding a language in which strategies canbe expressed. Our experience is that it isdifficult for humans to program in LISP orSAIL, and that we cannot reasonably expectstrategy programs to be expressed at that lowlevel of language. The language will beembedded in SAIL. Our previous work inrepresentation of shape has been significant;n o w tve are extending the study of

8

representation to visual program structure,including intermediate internal structures. Ourexperience with the hand language is that thisis a valuable step.

Polyhedral Modeling

Baumgart [Baumgart 19’73, 1974A, 1974B1 hasdeveloped a system for descripelve computerVISION based on polyhedral modeling andimage contourmg. Baumgart’s overall designidea may be characterized as an inversecomputer graphics approach to computelvlsioti. In computer graphics, the world isrepresented 111 suficient detail s o t h a t t h eImage forming process can be numericallysimulated to generate synthetic televisionimages; 111 the invers;, perceived te levis ionplcturcs are analyzed to compute detailedgeometrrc models. To date, polyhedra (such asin t h e f i g u r e ) h a v e b e e n automatlcallygenerated by intersection of silhouette conesfrom four views of a white plastic horse on ablack turntable. The viewing conditions arenecessarily favorably arranged, but then theclaimed results are honest.

A R T I F I C I A L I N T E L L I G E N C E P R O J E C T

Bibliography

[Agm] Agin, Gerald J., Thomas 0. Binford,. Computer Description of Curved

Objects, PYO.C. Third Internatio?zal JointConf. on Artificial Intelligence, StanfordUniversity, August 1973.

[Bajcsyl Bajcsy, Ruzena, ComputerDescriptioil of Textured Scenes, PYOC.

Third ht. joint Co??f. on Artificia/Intelligence, Stanford U., 1973.

EBaumgart 19731 Bruce C. Baumgart, ImageContouring arid Coniparing, Stanford A .I. Memo AIM-199, October 1973.

[Baumgart 1974A] Bruce C. Baumgart,CEOMED - A Geometric Editor, StanfordA. I. Memo AIM-232, May 1974.

[Baumgart 197481 Bruce G. Baurngart,Geonletric Modeling for ComputerVision, Stanford A. I. Memo AIM-249,October 1974.

[Binford] T.O. Binford, Visual Perception byComputer, Invited paper at IEEE SystemsScience and Cybernetics, Miami, December1971.

[Bullock] Bruce Bullock, in preparation.

[Hannah] Marsha Jo Hannah, CoulputerMatching of Areas iu Stereo Inlages,Ph.L). Thesis in Computer Science, StanfordA. I. Memo AIM-239, July 1974.

[Hueckel 19711 M. H. Hueckel, An OperatorWhich Locates Edges in DigitizedPictures, Stanford A. I. Memo AIM-105,December 1969, also in JACM, Vol. 18,No. 1, January 1971.

[Hueckel I9731 Hueckel, Manfred H., A LocalVisual Operator which Recognizes Edgesand Lines, J. ACM, October 1973.

[Nevatia] R. K. Nevatia and T. 0. Binford,

2 . 2 Coliiputer Vi5ion 9

Structured Descriptions of ComplexObjects, Proc. Third International JointConf. on A./., Stanford Univ., February1973.

c[Perkins] Walton A. Perkins, Thomas 0.

Elnford, A Corner Finder for VisualFeedback, Stanford A. I. Memo AIM-214,September 1973.

[Pingle] Karl I<. Plngle, Hand/Eye L i b r a r y ,Stanford Artrficial Intelligence Laboratory@ptrating Note 35. I, January 19’72.

[ S o b e l ] Sobel, Irwin, On CalibrntillgConlputer Controlled Cameras forPerceiving 3-D Scenes, Proc. Third ht.Joint Conf. on Arti@al Intdfigcnce,Stanford U., 1973; also In ArtificialIntelligence J., Vol. 5, No. 2, Summer 1974.

[Thomas] A.J. Thomas and K.K. Plngle, Inpreparation.

[Yakimovsky] Yaklmovsky, Yoram, Jerome A.Feldman, A Semalltics-Based DecisiorlTheoretic Region Arlalyzer, Proceedingsof the Third International Joint Conferenceon Arti$cial Intelligence, StanfordUniversity, August 1973.

2.2.2 Visual Guidance of a Vehicle

Lynn C$am and Hans Moravec are workingOil a “Cart ~JI’OJPCt” that has as one of Its goalst h e devclopmctit of a set of techniques t oenable a vehicle to guide Itself through aunknown environment on the basis of visualinformation. As a first step, a program hasbeen written which takes a motion parallaxpair of pictures of a scene, finds “interesting”feature points scatterecl over one Image, andtries to locate the same features m the otherImage, deducing their locatlon i n t h r e edlmens1ons.

This program has been tried on about 40 To handle this problem, we have developed apairs of pictures of outdoor scenes, and in all technique for area matching under scalecases was able to line up the horizons change using a mode1 for the camera position

properly. In about 60% of the cases one or twonearby obstacles were located accurately. Inthe remaining 40% the “matching” features‘found were typically pieces of the same roadedge farther along the path than the desiredfeature in the first picture. This kind of errorprecludes exact measurement of distances, butstill provides enough information so that theedge can be avoided.

Significant subtasks completed include theoperator which locates “interesting” features bythresholding and locally maximizingdirectional variation in grey level. A minorinnovatlon is a dlstortion of the pictures inthe hor izonta l directjon, tantamount totransforming the original planar images into acylindrical projection, thus making the scale offeatures invariant over camera pan.

Considerable effort has been expended ingetting our existing cart hardware to the pointwhere these techniques can be tried on arunning vehicle. A set of control routines,which calculate an optima.1 path for arequested position and orientation change andtransmit the appropriate commands to thevehicle, were also written this past year.

Near-field Stereo

Near-field stereo has the problem that a highdegree of distortlon and occlusion occurs Inmost scenes when the basel ine d i s t a n c ebetween the camera positions is comparable tothe distance from either camera to objects inthe scene.

For our immediate cart project goals, we areprimarily interested in objects in the directionof motion. Such objects undergopredominantly a scale factor change, butprevious efforts in area matching have notallowed an unknown scale change between theareas in the two images.


.and orientation of one Image relative to thepreceding Image. Whenever a point in thesecond image IS proposed as a match for apoint in the first image, one can use thecamera posItIon model to determme at whatdepth the 3-space point must lie. The ratio ofthe distances between this point and the twocamera posltlons corresponds to the observedscale factor change. This scale factor ratio IS

used to geometrically scale the points in thearea of Interest in one Image prior tocomputation of the area correlation operator.

This technique was applied to a sequence ofroad Images gathered as a “typIcal” roadenvrronment. The results indicated that areaswith scale changes of up to 1.5 or 2.0 to 1could be efticlently and- lx?kIb~y be matched.An extension of this technique which allowsunequal scale changes in the vertical andhorizontal directions is planned.

2.2.3 Mars Picture Analysis

T h e N A S A Vlklng Project supported afeasibility study at to determlne if computetimage processing techniques could be used foraerlal/orbltal photogrammetry. T h e objectwas to take pairs of orbital photographs ofportions of planets (the Moon and Mars In thestudy) and construct contour maps for theterrain. These techtllques are under study fatcheckmg the sultabillty o f t h e p r o p o s e d

‘landing sites for the 1975 Vlhlng mlsslons toMars.

The approach we took was to first match upas much as possible of the two images wrth ap r o g r a m t h a t u s e d correlation, the regongrower, and an approxm~ate camera modelderived from spacecraft position and polntlngdata. The parallaxes were then converted toelevations by a second program and contouredat the clesrred intervals by a third program.

contour maps in a small fraction of the timerequired by traditional photogrammetrytechniques [@am 19741.

Several articles have recently been publishedbased on earlier work supported by NASA oninteractive analysis of photographs of Marstaken by the Mariner satellites [Quam 1973,Sagan, Veverkal.

Bibliography

[Quam 19’731 Quam, Lynn, Robert Tucker,Botond Eross, J. Veverka and Carl Sagan,Mariner 9 Picture Differewing atStallford, Sky and T&scope, August 1973.

[@am 19741 Quam, Lynn and Marsha JoHannah, An Experimental System inAutomatic Stereo Photogrammetry,Stanford A. 1, Memo, 1974 (forthcow ing).

[Sagan] Sagan, Cart, J. Veverka, P. Fox, R.Dubisch, R. French, P. Gierasch, L.(@am, J. Lederberg, E. Levinthat, R.Tucker, B. Eross, J. Pottack, VariableFeatures on Mars II: Mariner 9 GlobalResults, Journal of Geophysical Research,78, 4163-4196, 1973.

[Veverka] Veverka, J., Carl Sagan, LynnQam, R. Tucker, B. Eross, VariableFeatures 011 Mars III: Comparison ofMar iner 1969 and M a r i n e r 1971Photography, Icarus, 2 1, 3 17-368, 1974.

Early results are fairly promising. Givenimages which are of sufficient resolution,reasonably free from noise ancl have sufficientinformation content, the computer can produce

11

2.3 Mathenlatical Theory of Cnmputatiorl

Several articles based on our eat-her work Inmathematical theory of computation werepublIshed during the year [I, 2, 3, 4, 51 andM a n n a pu,bllshed the first textbook in thisfield 161.

Vulllemln’s Ph.D. thesis examines theconnection between the concept of least fixed-pomt of a contmuous functron and recursiveprograms [71.

2.3 .1 FOL

The FOL project was deslgned by RichardWeyhrauch and John McCarthy to create anenvironment II-I which first order logic andrelated traditional forrnal~~ystems can be usedwith ease and flrxrhiltty. F O L I S a p r o o fchecker based on the natural deduction styleof representlng the proofs of first order logic.The ability to use FOL to do substantiveexperiments IS just becoming feasible. Someof these are described below.

Eventually we expect FOL to act as a practicalsystem in w h i c h t h e verification o f t h ecorrectness and equivalence of programs canbe carried out In the language of ordinarymathematics. T h e theoretical discussion o fh o w this c a n b e accompllshed h a s b e e noutllned In the papers of McCarthy, Floyd,Manna and others. The reduction of theseIdeas to practice is still In the experimentalstage.

The above task requires a system that canrepresent the traditional arguments ofmathematics. Thus a maJor part of our effortis devoted to developing a smooth and usefulembedding of traclitlonal set theory Into thisenvironment, and for ways to deal correctlywith the metamathematics necessary tocompletely represent any substantive part ofmathematical practice.

An example of usmg FOL to prove a verysimple theorem follows. Lines beginning withII GS*Z=Z” are input and the others are output.

wcdIECLf7RE INDWR x y;DECLFlRE PREDCONCT F 1;wwwzTQUT F(x)v-F(x)',1 F(x)v-F(x)

,9*::w31 1, xcy occ 2;

2 3y, (F(x)v-F(y))

****:~VI 2, x;

3 VxJy.(F(xh-F(y))

The first large proofs using FOL are reportedby Marlo Aiello and Richard Weyhrauch [9].They describe a n axiomatlzation o f t h emetamathematics of FOL and prove severaltheorems using the proof checker.

Weyhrauch has also expanded McCarthy’sidea of a computation rule using a notion hehas called semantic attachment. This is auniform technique for using the computationto decide sentences like 32 or 3+7=8 orISON(BLACKKINC, B K N I , B O A R D ) o rHAS(monkey, bananas) . Independently ofthis, Arthur Thomas suggested using FOL ina similar way to explore models of perceptionand their interaction with the actual world.Robert Filman is using these ideas extensivelyto axiomatize basic chess knowledge.

S e v e r a l p r e l i m i n a r y u s e r s m a n u a l s w e r eproduced for FOL, and an AI memo [ 101 willappear soon.

2.3.2 L C F

Progress was made in two directions. Marioand Luigia Aieilo and Richard Weyhrauchproduced an axlomatization o f theprogramming language PASCAL using LCF.This project represents a major step towardsusing LCF as a practical program verificationsystem. This work IS reported on in [S] and[ 111. PASCAL was chosen in order tocompare the techniques of D a n a Scot t ’ sextensional approach to program semanticswith that of Robert Floyd and C.A.R. Hoare.The1 latter approach is represented at the AIlab by David Luckham and his work on t h ePASCAL program verification system [Section2.4.11.


Fredrlch v o n Henke rewrote the LCFprogram and expanded it to Include anaxromatlzatlon of the type-free logic originallydevised by Dana Scott, Robin Milner andRichard Weyhrauch. In addition, von Henkeused the type-free logic to study thefunctlonals definable over data structureswhich have recursive definitions.

Bibliography

[ 11 Ashcroft, Edward, Zohar Manna, AmilPnuell, Decidable Properties of MouodicFunctional Schemas, J. ACM, July 1973.

[9] Marlo Alello, Richard Weyhrauch,Checking Proofs iu theMetamathematics of First Order Logic,Stanford A. I. Memo AIM-222, October

1 9 7 4 .

[lo] Richard W. Weyhrauch, Arthur J.Thomas, FOL: A Proof Checker for First-order Logic, Stanford A. I. MemoAIM-235, June 1974.

[ill Luigia Aiello, Richard W. Weyhrauch,LCFsmall: an Ilrlplemerltatio~~ o f L C F ,Stanford A. I. Memo AIM-241, August1974.

[2] Igarashi, S., R. L. London, D. C. Luckham,I n t e r a c t i v e Prograin Verification: AL o g i c a l Systenl and its Illlplelllerltatiorl,Acta Informatica, (to-appear).

2.4 Heuristic Progranlniing[3) Katz, Shmuel, Zohar Manna, A Ifeuristic

Approach to Program Verificatiou,Proceedings of the Third InternationalJoint Conference on Artificial intelligence,Stanford University, August 1973.

[41 Manna, Zohar, Program Schemas, inCurrents in the Theory of Computing (A.V. Aho, Ed.), Prentice-Hall, EnglewoodCliffs, N. J., 1973.

[53 Manna, Zohar, Stephen Ness, JeanVulllemln, I Inductive Methods for

a Proving Properties of Programs, Comm.ACM, August 1973.

[6] Manna, Zohar, Introduction toMathematical Theory of Computation,VcGraw-Hill, New York, 19’74.

[71 Jean Etlenne Vulllemin, Proof Techniquesfor Recursive Progranls, Thesis; Ph.D. inCom,btlter Science, Stanford A. I. MemoAIM -2 18, @ctober 1973.

[S] Lulgla Alello, Mario Aiello, RichardWeyhrauch, The Semantics of PASCALin LCF, Stanford A. I. Memo AIM-221,August 1974.

Heuristic programming techniques are beingapplied to theorem proving and automaticprogramming problems.

2.4.1 Theoreln Proving

A group headed by David Luckham hasdirected their recent research toward theapplication of mathematical theorem provingin the area of program verification. There arenow have two theorem provers:(1) A general proof program that has been

developed for research in different areasof mathematical problems. This is basedon the Resolution principle and rules forequality. It contains a wide selection ofproof search strategies, and incorporatesan interactive user language for guidingproofs and selecting strategies. It can beused either as a theorem prover or as aproof checker. There is a facility for anautomatic selection of search strategiesbased on an analysis of the problem, sothat prior knowledge of theorem provingtechniques on the part of the user isunnecessary. We summarize recentdevelopments with this program below.

(2) A fast special purpose prover (called the

2.4 Hrurirtic Programming 13

Simpl~ficr) cleslgned spcclfically fotprogram verlflcatlon. This programmakes use of documentation submittedwith the program to reduce theCO~I~~ICN Ity of logical verificationcondltlons (the truth of these conditionsimply the correctness of the program).@riglnally, this program was Intended asa preprocessor for the general theoremprover. However, it includes a limitedtheorem-proving capability aimed atellmlnatlng the “easy work” and this hasturned out III experiments to be apowerful component of the verificationsystem (see below).

A user’s manual for the general theoremprover is available [I]; and publications[2,3,4,5] d e a l with InteractIve applications o fthis program to mathematics and rnfotmatlonretrrcval. Recent experiments III using theprove1 t o obtain n e w characterlzatlons o fvarieties of Groups and to check tedrousproofs in Euclidean Geometry are given in [61.A primitive “HUNCH” language has beenprogrammed by J. Allen. This enables theuset to describe complex proof searchprocedures in which the strategies may varyduring the search for a proof. This language1s currently being extended to permit outlinesof proofs to be described in a natural way.We regard this as a necessary development forFore difficult appllcatlons in mathematics.

During the last year the prover has been usedto provide the automatic deduction capabilityfor the PASCAL program verification system[71. 1~ particular, J. Morales has made anextensive study of the verification of sortingalgovlthms from first prlnclples (including, fore x a m p l e , SIFTUP [S] CUCBLE S O R T 191,and INSERTION SORT [lOI> and is workingon mocllficatlons of the HUNCH language toaId In these problems.

T h e slmpllfie~ is a fast theorem proverIncorpora ted into the program verificationsystem for PASCAL programs [: I 11. Theverification system as originally described in

[7], has been extended to permit the user tosubmit axiomatic descriptions o f datastructures and specifications o f (possiblyuriwritten) subroutmes with the program to beverified. The simplifier uses thesedocumentation statements either as algebraicsimplification rules or as logical goal reductionrules in a depth first proof search. Amethodology of using the verification system todebug ad verify real life programs dependingon nonstandard data structuresdeveloped [ 121.

is beingA user ’s m a n u a l for this

system is available 1131, and experiments usingthe Simplifier to verify sorting and patternm a t c h i n g programs on the basis of user-defined concepts are reported in [ 12,141. Aversion of this system for PL/l (including thedata type POINTER) is being programmed.

Further developments and applications o fheuristic theorem proving are described in thesection on Automatic P r o g r a m m i n g (c.f.Luckham and Buchanan), and an ambitiousproof checking system for higher order logichas been developed by R.W. Weyhrauch (seeSection on Mathematical T h e o r y o fComputation).

Bibliography

[II Allen, J.R., Preliminary IJsers Manual forau Interactive Theorem-Prover, StanfordArtificial Intelligence LaboratoryOperating Note SAILON-73.

123 John Allen, David Luckham, AuInteractive Theoremproving P r o g r a m ,Proc. Fifth International Machinelntdligence IYorkshop, EdinburghUniversity Press, Edinburgh, 1970.

[3] Richard B. Kieburtz and DavidLuckham, Compat ib i l i ty and Complexi tyof Refinemel1t.s of the ResolutionPrinciple, S I A M J. Cornput., Vol. 1, No.4, December 1972.

I4 ARTIFICIAL INTELLIGENCE PROJECT

[4] David Luckh am, Refinemellt Theorems inResolut ioll Theory , Syn/osium onAutomatic Dtvnonstrntion, Springer-VerlagLecture Notes in Mathematics No. 125,Berlin, 1970.

[S] David Luckham, NIIS J. Nilsson,E x t r a c t i n g Informatiou from ResolutiorlProof Trees, Artificial Intelligence, Vol. 2,No. 1, Spring 1971.

[6] Jorge J. Morales, Interactive TheoremProving, Pm. ACM Nntionnl Conference,August 1973; also Tileorcul Provirlg inGroup Theory aud Tarski’s Geometry ,forthcomrng A.I. Memo.

[7] Igarashl, S., London, R., Luckham, D.,Auutoiriatic P r o g r a m Verification I,Logical Basis aud Its IIllplelllelltatiorl,Stanford A. 1. Memo AIM-200, May 1973;to appear ~ii Acta lnformatica.

[81 Floyd, R.W., Algorithm 245, TREESORT3, Comm. ACM, June 1970.

[9] Knuth, D.E., The Art of ComputerProgramming, Vol . 3 : Sortiiig arldSearching, Addison-Wesley, Reading,Mass., 1973.

[lOI Pratt, V.R., Shellsort and Sort ingNetworks, Ph.D. Thesis in Camp. Sci.,

- Stanford Univ., February 1972.

[1 11 N. Wirth, The Programming LanguagePASCAL, ACTA Informaticu, I l., 1971.

[12] F. von Henke, D. Luckham, AMethodology for Verifying Programs,A.I. Memo, forthcoming (submltted to the1975 lnternatlonal Conference on ReliableSoftware).

[ 131 N. Suzuki, Short Users Manual for theStanford Pascal Program VerifierA.I.Lab. operating note (forthcoming).

[14] N. Suzuki, Verifying Progranls byAlgebraic aud Logical Reduction, A.I.Memo, forthcoming (submitted to the 1975International Conference on Reliable

. Software).

2,4,2 A u t o m a t i c P r o g r a m m i n g

Research in automatic programming hasprogressed on several fronts, summarizedbelow.

A u t o m a t i c Prograul Generatio

A heuristic theorem proving system for aLogic of Programs due to Hoare [ 11 forms thebasis for a successful automatic programmingsystem that has been developed over the pasttwo years. This is an experimental system forau toma t i ca l l y generatmg simple kinds ofprograms. The programs constructed areexpressed in a subset of ALGOL containingassignments, function calls, conditionalstatements, while loops, and non-recursiveprocedure calls. The Input to the system is aprogramming environment consisting ofprimitive programs, programming methods forwriting loops, and logical facts. The input isin a language similar to the axiomaticlanguage of [ 11. The system has been used togenerate programs for symbolic manipulation,robot control, every day p l a n n i n g , andcomputing arithmetical functions.

Two papers concerning the theory andapplications of the automatic programmingsystem have been written [2, 31. Applrcationsof the system to generatrng assembly andrepai r procedures wlthin the HANDCONTROL language [til for simplemachinery are described In [2]. Report [?,Ipresents a full overview of the system withmany examples. Details of the implementationare In Buchanan’s thes is [41. The l oopconstruction and program optimizationmethods have been extended by John Allenand more ambitious a p p l i c a t i o n s i nprogramming and mechanical assembly arebeing tackled.

15

Arrtonlatic Sclcction of Data Structures

A system has been developed which, given analgortthm expressed in terms of high-levelinformation structures such as sets, orderedsets, and relations, automatlcally choosese+Ytcient representations and produces a newp rogra m that uses these representations.Representations are plcked from a fixedlibrary of low-level data structures includmglinked-lists, binary trees and hash tables. Therepresentations are chosen by attempting tomlnlmrze the predicted space-time integral ofthe user’s program execution.

PredIctions a r e b a s e d u p o n statlstlcs o fInformation structure use provided directly bythe c1ses and collected by monltorlngexecutions CJf the user program using defaultrepmen t a t Ions for the high-level structures.In performance tests, this system has exhIbItedbehavior superior to human programmers,and IS at the stage where it could beImplemented In a very high level language,like SAIL. Thrs work is reported in JimLow’s thesis 161.

Program Ullderstarldirlg Systems

Progress has also been made 111 the design ofsystems which can be said to have some“program understanding” ablllty. In our case,the primary evidence for such ability lies inthe capabllity to synthesize programs, eitherautomatlcally or semi-automatlcally, but such acapablllty alone is lnsuficient forunderstanding: the line of reasoning which thesystem follows ciLm1g the synthesis processmust * also sLlppOl’t the claim of“understarlding”, and we feel that most of oursystems behave well III this regard.

One experimental system used its knowledgebase of “programming facts” to synthesize(among others) programs which interchangeelements, perform a 3-element non-recursivesort, and find the integer square root, basingchoices at declslon points on user responses toquestlons posed by the system. Another

experimental system can synthesize programsfrom example input/output pairs, and haswritten about 20 simple list-processingfunctions.

These experiments have led us to severalpreliminary conclusions and to a view that twoof the major research areas in programunderstanding systems are the exploration ofvarious manners of program specification, andthe codlficatlon of programming knowledge.

Looking at the two experimental systemsmentloned above, we see two different methodsof specifying the desired program: exampleinput/output pairs and user responses toquestlons from the system. There seem to bem a n y other ways in which the desiredprogram could be speclfied, ranging from veryformal to very informal. A unifying paradigmwould seem to be a kind of dialogue betweenthe user and the system. In such a dialogueany of these methods (or even several of them)might be employed, depending on suitabilityfor the program, and preferences of the user.Work is currently progressing on variousmethods of modeling and conducting suchdialogues.

Our experimental systems and numerous handsimulations of program understanding systemsindicate that satisfactory behavior can only beexpected when the system contains a largebody of knowledge. For the understanding ofprograms in a given domain, there isconsiderable domam-specific knowledge.Additionally, there seems to be a large body of“pure” programming knowledge which isrelatively domam-independent. Much of ourwork is aimed at isolating, codifying, andrepresentmg this knowledge.

Our early experimental systems as well asdiscussions of co11~lu~1or1~ and future plans arereported in [71 and in papers by Green andBarstow, and by Shaw, Swartout, and Green,which are in preparation.


Bibliography .

[I] C.A.R. Hoare, An Axiomatic A p p r o a c hto Computer Progran1wing, Cotnm, ACIZ1,October 1969.

[2] David Luckham, Jack Buchanan,.Automatic Generation of ProgranlsContaining Conditional Statenlerlts, Proc.A.I.S.B. Strmmcr Conference, Scissex,England, July 1974.

[3] Jack Buchanan, David Luckham,OnAutomat ing the Construction ofPrograms, Stanford A. 1. Memo AIM-236,May 1974, (submltted to the JACM).

[4] Jack Buchanan, A Study in AutonlaticProgram 111 i ng, Ph.LK-Thesis in ComfmtcrScience, Stanford A. I. Memo AIM-245,September 1974.

[5] R. Bolles, L. Paul, The Use of SensoryFeedback in Prograulmable AsseulblySystems, Stanford A. I. Memo AIM-220,October ! 973.

161 Low, Jim, Autoulatic Coding: Choice ofData St ructure$, P&P. Thesis in ComjwterScience, Stanford A. 1. Memo AIM-242,(forthcoming).

[7] Green, Cordell, R.J. Waldinger, David R.- Barstow, Robert Elschlager, Douglas B.

Lenat, Brian P. McCune, David E. Shaw,LOUIS I. SteInberg, Progress Report onProgralll-~lIlder,stalldillg Systenls,Stanford A. I. Memo AIM-240,(forthcoming).

[$I Manna, Zohar, Autonlatic Progranwling,Proceedings of the Tllird InternationalJoint Conference on Artificial Intelligence,Stanford University, August 1973.

2.5 Natural Language

Research continued on three aspects of naturallanguage:. 1) speech recognition, which typically deals

with acoustic waveforms,2) natural language understanding, which

generally starts with text, and3) hjgher mental functions, which deals with

psychiatric problems manifested throughnatural language.

A lack of funding suppor t for speechrecognition has resulted in a progressivereduction of that activity.

2.5.1 Speech Recognition

During the past year the focus of speechrecognition research at Stanford has changedf rom mach ine teaming b a s e d p h o n e m erecognition HI to linguistically structuredacoustic-phonetic processing [21. Thephilosophy of the research has been to attemptto extract a m a x i m u m o f linguisticinformation from the speech signal. This ledto using waveform type segmentation, pitchsynchronous analysis o f voiced regions,waveform level steady state detectors andsyllable detectors. The maJor effort has goneinto developing algorithms whichautomatically extract the linguistic informationat each level; waveform and short timefrequency spectra.

Neil Miller has developed a semantic pitchdetector which used the expected pattern ofexcitation and exponential decay of theacoustic signal during voicing. The purposeof the earliest version of the pitch detector wasto mark the beginning of each pitch periodduring voicing, making a voicing decisionalong the way. Various versions of thisprogram find the pitch in less than real time[41.

Waveform level acoustic segmentationalgorithms were developed by JimHieronymus 131. On the sub phonemic level,

2.5 Natura l Language

areaS of steady frequency spectra wherecontinuant phonemes most closely approachtheir target frequency values are found bypitch synchronous waveform comparisons. Aprocess for segmentation into continuants andtransitlons was developed based on a model ofthe way a h u ma n visually compareswaveforms. An algorithm for waveform typesegmentation into voiced, nasalized vowel,voiced fricative, unvoiced fricative, and silencewas developed based on amplitudes, integralsunder the acoustic peaks in a pitch period andzero crossings.

Pitch synchronous short time frequencyspectra were found to contain clearly delimItedformants, so that linear predictive modeling ofthe spectrum was not necessary In order toreadily find the formants. In addition, pitchsynchronous anatysis preserves the maximuminformation about formant transitions.Transrtions in and out of stop consonants are

clearly seen. A formant extractron algor’thmwas developed by Arthur Samuel to pick theformant peaks from the pitch sychronous FFTspectra. Visual comparisons with the outputof the M.I.T. Lincoln Labs formant trackerand sonograms have been made. C;cncrallythe formant tracking is as good as or betterthan m ii c h more comyhcated trackingprograms using LPC data. Pitch synchronousanalysis also preserves t h e t r u e formantbandwldths, which may be useful 111 nasal

‘detection.

Moorer has developed a very efficient schemefor performing pitch period analysis [5].

A system for displayrng speech waveforms,their frequency spectra, and for hearing theutterance being examined has been developed.Hard copy plots can be made from the drsplayprogram using the Xerox Graphics Printer.

A f t e r April 1974, the group worked onrefin I ng the pitch detector, sylta ble detectionand rate of speech measures based on syllablecounts. A plan to do some research 111 contestdependent vowel recognition was formulated,

17

s ince this is a significant problem area inpresent speech recognition systems.

This work is continuing at a very low level forlack of funding. Several articles are inpreparation from the research work done in1973-74.

Bibliography

[I] Thosar, Ravindra B., Recogrlitiorl ofCoutinuous Speech: Segnientation arldClassification using Signature TableAdaptation, Stanford A. I. MemoAIM-2 13, September 1973.

[21 Hleronymus, J. L., N. J. Miller, A. L.Samuel, The Atnan ueusis SpeechRecognition System, Proc. IEEESymposium on Speech Recognition, April1974.

131 Hieronymus, J. L., Pitch SynchronousAcoustic Segmentation, PYOC. lEEESymposium on Speech Recognition, April1974.

[4] Miller, N. J., Pitch Detection by DataReductioll, Proc. IEEE Symposium onSpeech Recognition, April 1974.

[5] Moorer, James A., The Optinwm CombMethod of Pitch Period Analysis ofCorlthuous Speech, lEEE Trans.Acoustics, Speech, and Signat Processing,Vol. ASSP-22, No. 5, October 1974.

2.5.2 Natural Language I.Jnderstarlding

This was a transitional year for our programof research in natural language. RogerSchank, who previously directed some of thework, was on leave at the lnstituto per g/iStudi Semantici e Cognitizri, in Switzerland.He continued his research into conceptualdependency theory for natural languageunderstanding at the institute [2]. His work,along with that of Chris Riesbeck, NeilGoldman, and Charles Rieger, led to thecompletion of the MARGIE system [I].


One aspect of this is reported in Rleger’sthes is [l I], wh ich deve lops a memoryf o r m a l i s m a s a basis for cxamlnIng t h elnferentlal processes by which comprehensionoccurs. Then, the notion of inference space IS

prescntcd, a n d sixteen classes o f conccl,ruali n f e r e n c e a n d their ~mplementntion in t h ecomputer model are examined, emphaslzlngthe contrlbutlon of each class to the totalproblem of understandIng. The idea of pointsof contact of Information structures iI3inference space IS explored. A point of coIltactoccurs when an inferred unit of meaIling fromo n e starting point within one utterance’smeaning graph either confirms (matches) orcontradicts an inferred unit of meaning fromanother point within the graph, or fromwithln the graph of ansther utterance.

The work of the other members of the groupwill b e p u b l i s h e d in the coming year,including a book edited by Schank,summarizing research in conceptualdependency.

Yorlck Wilks continued his work on machinetranslation [3, 4, 5, 6, 7, 8, 9, 101. Inparticular, he studied the way I~I which aPreference Semantics system for naturallanguage analysis and generation tackles adifficult class of anaphoric inference problems(finding the correct referent for an Englishpronoun In context). The method employed

- converts all available knowledge to a canonicaltemplate form and endeavors to create chainsof non-deductive Inferences from theunknowns to the possible referents.

Annette Herskovits worked on the problem ofgenerating French from a semanticrepresentation [ 133. She concentrated on thesecond phase of analysis, which bindstemplates together into a higher level semanticblock corresponding to an English paragraph,and which, in operation, Interlocks with theFrench generation procedure. Frenchsentences are generated, by the recursiveevaluation of procedural generation patternscalled stereotypes. The stereotypes are

semantically context sensitive, are attached toeach sense of English words and keywordsand are carried Into the representation by theanalysis procedure.

In addition, members of the translation groupentered Into discussions with others in thelaboratory in a series of conversations dealingwith some of the issues connecti,ng artificialIntelligence and philosophy [ 141. The majortopics included the questlon of what kind oftheory of meaning would be involved in asuccessful natural language understandingprogram, and the nature of models in A Iresearch.

Terry Winograd spent the year at Stanford asa visitor from MIT, and continued his workon natural language understanding and itsrelationship to representation theory. Hepublished a number of papers outlining histheories [ 15, 16, 17, 18, 201 and anintroduction to artificial intelligence and theproblems of natural language [191. He gave anumber of talks, including a lecture series atthe Electrotechnical Laboratory in Tokyo, theTutorial on N a t u r a l L a n g u a g e a t t h eInternational Joint Conference on ArtificialIntelligence ( P a l o A l t o , A u g u s t 1973), a ninvited lecture at the ACM SICPLAN-SIGIRinterface meeting (Washington D.C.,November, 1973), and “A Computer ScientistLooks at Memory”, a part of Sigma XiLecture Series (Palo Alto, February 1974).

Bibliography

[ 11 Schank, R oger C., Neil Goldman, CharlesJ. Rieger III, Chris Riesbeck, MARGIE:Memory, Analysis, Response Generatioalld Inference 011 English, Proceedings ofIhe Thiyd Internutionul Joint Conferenceon Artijiciul Metligencc, StanfordUniversity, August 1973.

[21 Schank, Roger C., Kenneth Colby (eds),Computer Models of Thought andLanguage, W. H. Freeman, San Francisco,1973.

2.5 Natural L a n g u a g e 19

[3] W i l k s , Yorick, T h e Stailford MachirleTranslation aud Ifrlderstarldiilg P r o j e c t ,in Rustin (ed.) Nnttsrnl LunguugeProcessing, New York, 1973.

141 Wilks, Yorlck, Understalldillg WithoutProofs, Proceedings of the Thirdlnternutional Joint Conference on ArtijciulIntelligence, Stanford University, August1973.

[5] Walks, Yorlck, Annette Herskovits, AIIIlltelligent Analyser aud Gerlerator o fNatura l Language, Proc. ht. Conj. onCom/~tr.tutional! Linguistics, P tsa, 1 taly,Proceedings of the Third hternationulJoint Conference on Artificial Intelligence,Stanford University, August 1973.

[6] Wilks, Yorick, The Computer Analysis ofPhilosopl~ical Argwnents, CIRPHO, Vol.1, No. 1, September 1973

[7] Wilks, Yorlck, An Artificial IntelligenceApproach to Machiile Translation, i nSchank and Colby (eds.), Com/Jufcr n/lotlt*tsof Thought anti Language, W. H. Freeman,San Francisco, 1973.

[S] Wilks, Yorrck, One Small Head -- ~iodclsand Tlleories in Linguistics, Foundationsof Lunpsugc, Vol. 10, No. 1, January 1974.

[9] Wilks, Yorlck, Preference Setnalltics, E.Keenan (ed.), Proc. 1973 Colloquium onFormul Sonuntics of Nutud Lunguuge,Cambridge, U.K., 1974.

[lo] Walks, Y . , Machine Trarlslatiorl,inCu,rrent Trentls in the Lungtruge Sciences,T.A. Sebeok, (ed.), forthcommg.

[ 111 Rieger, Charles J., Comeptual Memory:A Theory and Computer Prngraw forProcessing the Meanillg Content ofNatural Language Utterances, StanfordA. I. Memo AIM-233, June 1974.

[I21 Wilks, Yorick, Natural LarrguageInference, Stanford A. I. Memo AIM-2 11,September 1973.

[ 13) Herskovlts, Annette, The Gerleratiorl ofFrench from a Semantic Representation,Stanford A. I. Memo AIM-212, September1973.

[ 143 Anderson, D. B., T. 0. Binford, A. J.Thomas, R. W. Weyhrauch, Y. A. Wilks,After Leibrliz . . . . Discussions onPhi losophy arld Artificial Iiltelligellce,Stanford A. I. Memo AIM-229, April 1974.

[ 151 Winograd, Terry, A Process Model ofLanguage Ullderstauding, in Schank a n dColby (eds.), Computer Models of Thoughtand Language, W. H. Freeman, SanFrancisco, 1973.

[16l Winograd, Terry, The Processes ofLallguage Understanding in Benthal l ,ted.), The Limits of Human Nuture, AllenLane, London, 1973.

[17] Winograd, Terry, Language alld theNature of htelligewe, in C.J. Dalenoort(ed.), Process Models for Psychology,Rotterdam Umv. Press. 1973

[181 Winograd, Terry, Breaking theComplexity Barrier (again), Proc.SIGPLAN-SIC;IR Interface Meeting, 1973.

[19] Winograd, Terry, Artificial Intelligellce-- Wheli Will Computers UllderstandPeople?, Psychology Today, May 1974.

[2Ol Winograd, Terry, Parsing NaturalLanguage via Recursive Trarlsitioll Net,in Yeh (ed.) Applied Computation Theory,Prentice-Hall, 1974.


2.5.3 H i g h e r M e n t a l Fullctions

T h e H i g h e r Menta l Functions Project isdirected by Kenneth Mark Colby and is ’supported by the National lnstltute Of MentalHealth. The overall objective of the projectare to develop computer models for problemsin psychiatry.

One model is a simulation of paranoidthought processes (PARRY) which can beinterviewed using unrestricted naturallanguage input. Another involves a computer-aided treatment for nonspeakmg autlstlcchildren.

Bibliography

[ll Schank, R.C., Colb< KM. (eds.), ComputerModels of Thought and Language, W.H.Freeman and Co., San Francisco,California, 1973.

[2] Colby, K.M. Artificial Paranoia, PergamonPress, N.Y., 1974.

[3] Enea, H. and Colby, K.M. IdiolccticLanguage-Analysis for IJnderstarldingDoctor-Patient Dialogues, PYOC. Thirdlnternntionnt joint Confwvw on ArtifiiinlIntelligence, Stanford University, August1973.

- [4] Colby, K.M. and Parkison, R.C. Pattcrn-matching rules for t h e Rccoguitiou o fNatural Language Dialogue Expressions,American Journal of ComputationalLinguistics, 1, 1974.

[5] Hllf, Franklin, IJse of ComputerAssistance iu Euhmciug Dialog BasedSocial Welfare, Public Health, audEducational Services iI1 DevelopingColt ii t ries, Proc. 2nd Jertrsaltvn Conf. onInfo. Ttdnology, July 1974.

[6] Wilks, Y. S enlantic Procedures andIrifornialioii, in Studifs in theFou.ntlations of Communication, R. Posner(ed.), Sprrngcr, Berlin, forthcomrng.

[7l Colby, KM. and Kraemer, H. ObjectiveMeasurement of Nonspeaking Childrerl’sInteractions with a Computer CorltrolledPrograni for the Stimu)atioii o fLanguage Development, (in press 1974).

2.6 Programming Languages

We continue to find it profitable to invest aportion of our effort in the development ofprogrammmg languages and related facilities.We have already discussed the development of“HAL”, the advanced “hand language” forrobotics research [Section 2.11. We expect thatwork on automatic programming [Section2.4.21 will greatly increase the power ofprogramming, though such systems are notvery practical yet.

The languages LISP [6,7,81, FAIL [ 101, andSAIL [I 13 carry the bulk of the programmingworkload here and at a number of other PDP-10 installations. We have contmued to makemodest improvements In these systems, whichwe originated.

The Higher Mental Functions group, underNIMH sponsorship, has been developing apattern-directed extensible language calledLISP 70 [91.

Professor Hoare spent a sabbatical here,continuing to develop his ideas on structuredprogramming and related concepts [2,3,4].

Swinehart completed a dissertation on anInteractive programming system that controlsmultiple processes [53.

Bibliography

[I] Feldman, Jerome A., James R. Low,Comment 011 Brent’s Scatter StorageAigoritlrm, Comm. ACM, November 1973.

2.6 Progranlnl i Ilg Language.5 21

[2] Hoare, C.A R. Parallel Prngraintuiilg: at1Axionlatic Approach, St~niot~cl A. I.Memo AIM-219, October 1973.

[3] Hoare, C.A.R., Recursive Data Structures,Stanford A. 1. Memo AIM-223, December1973.

[4] Hoare, C.A.R., HilIt. 011 ProgramnhgLanguage Design, Stanford A. I. MemoA 1 M -224, December 1973.

[S] Daniel C. Swinehart, COPILOT: AMultiple Process Approach to InteractivePrograllllllitlg Systems, Ph.D. Thesis inCop7iptcte’r Science, Stanford A. I. MemoAI M-230, August 1974.

[6] O,l~aam, Lynn, Whitfield Difie, StanfordLISP 1.6 Manual, Stanford A. I. Lab.Operating Note SAJLON-28.7, 1973.

[73 Smith, David C., MLISPZ, Stanford A. I.Memo AIM-195, It-1 dlskfile:MLISP2[AJM,DOCl, May 1973.

181 Smith, DC. and Enea, H. Backtrackingi II M LIS P2, Pm. Third InterncltionntJoint ConfcrEncc on Artificial Intelligence,Stanford Unrversrty, August 1973.

191 Tesler, L.C., Enea, H., and Smith, D.C._ The LISP70 Pattern Matching System,

Proc. Third lnternationnl joint Conferenceon Artificial Intelligence, StanfordUniversity, August 1973.

[IO] F.H.C. Wright II, R. E. Corm, FAIL,Stanford A. I. Memo AIM-226, April 1974,in diskfile: FAJL.REG[AJM,DOCl, update

\ i n FAJL.UPD[AIM,DOC].

2.7 Computer Facilities

Our primary computer facility continues to bePDP-IO (KA-10 processor) with 68 displayterminals online. It a rather efficient system inthat it can gracefully carry a load of forty-some sizable Jobs. Even so, it is chronicallyoverloaded by the local demand.

2.7.1 Hardware

In late 1973 to early 1974, we received thecomponents of a new realtime system, namelya PDP-I l/45 processor, an SPS-4 1 processor,and a 192K:::lG bit Intel MOS memory. Thissubsystem is connected to the PDP-10 and isbeing developed mainly for computer visionand manipulator control.

Late in 1973, we installed an audio switch thatconnects any of 16 audio sources to any of 64speakers on display terminals. This permitsgeneral audio responses from the computerand also supplies sound to go with televisionimages that are available on the video switch.The cost of the audio switch was kept low byusing digital gates for switching. The audiosignal is encoded as a pulse-width modulatedsquare wave at a frequency of about 100KHz.

In December 1973 we received a BBN Pagerthat had become surplus at NASA-Ames andconnected it to our KA- 10 processor. Systemchanges to exploit the pager are underdevelopment.

We replaced our six IBM 3330 disk driveswith four double density drives in June 1974.This increases our file system capacity to 136million words but reduces the monthly leasecosts slightly.

[ 1 I] VanLehn, Kurt, (ea.), SAlL UserManual, Stanford A. 1. Memo AIM-204,July 1979; in dlskfile:SAJL.J<VL[AJM,DOC], update inSAJL.UPD[AlM,DOC].


2.7.2 Software

Generally, recent changes in the timesharingmonitor w e r e m a d e o n l y t o accomodatehardware chatlges. Documentation was greatlyimproved by the new monitor commandmanual 141 and program interface manual [2].

PUB, the document compiler [I], had a fewbells and whistles added (mostly by LarryTesler, who IS now at Xerox PARC) and wasused to produce this report.

The onlrne newswire system called APE hasbeen superceded by NS 131, which has anumber of new capabilities and accesses bothAssociated Press and New York Timesnewswires.

Our display-or-tented text editor “E” had a fewfeatures added, and much improveddocumentation 151. Though It IS not complete,it still appears to be the world’s best texteditor.

Baumgart Improved his geometric editor [6],which facllltates the interactive design ofthree-climenslotIal objects and p rocl ucesvarious graphical and photographicrepresentations of them.

Our interactive drawing program for cllgltall o g i c design [7] continues In u s e a t M I T ,

- DigItal Equipment C o r p o r a t i o n , Carnegie-Mellon University, and here.

Bibliography

[ 1) Tesler, Larry, PlJB -- The DocumentCompiler, Stanford A. I. Lab. OperatingNote SAILON-70, September 1972; indslkfile: PU B.TES[S,DOC], supplement inPU BSUM.LES[tJP,DOC].

[3] Frost, Martm, Readitlg the WireserviceNews, Stanford A. I. Lab. Operating NoteSAILON-72.1, m diskfile: NS.M E[S,DOC J,September 1974.

[4] Harvey, Brian, Monitor ComnlarldManual, Stanford A. I. Lab. OperatingNote SAILON-54.4, September 1974; indiskfile: MONCOM.BH[S,DOC], updatein MONCOM.UPD[S,DOC].

[5] Samuel, Arthur, TEACH, in diskfile:TEACH[UP,DOC], 1974.

[Gl Baumgart, B.C., GEOMED, Stanford A. I.Menlo AIM-232, May 1974.

[7] Helliwell, Dick, S tanford ‘LlniversityDrawhg System, in diskfile:SUDS.RPHCUP,DOC], April 1974.

[2] Frost, Martln, 1J1JO Manual, Stanford A.I. Lab. Operatins Note SAILON-55.3,December 1973; 111 diskfile:UUO.ME[S,DOCl, update InUUO.UPD[S,DOC].

23

3 . HEIJRISTIC P R O G R A M M I N GP R O J E C T

We begin this annual report by mentioningone of the tasks that the ARPA IPT Officeasked one of the co-Principal Investigators,Professor Feigenbaum, to perform during theyear: to write a paper explicating the currentgoal structure of Artificial Intelligence as ascientific and technological endeavor, andsuggesting a set of most fruitful lines ofadvanced research and exploratorydevelopment ovei the next five years. Thistask was completed in November, 1973, and areport prepared for ARPA (available as diskfile AI.RPT[I,EAF] at SU-AI on the ARPAnet). -_

That document is used as the basis oforganizing the material contained in thisannual report, since portions of it provide anexcellent framework for the activities of thisproject. Where quotation marks otherwiseunidentified are used, the quotation is fromthe Feigenbaum report to ARPA (sometimesslightly edited).

The project’s research activities continue to bemotivated by this global view of AI research:“Artificial Intelligence research is that part ofComputer Science that IS concerned with thesymbol-manipulation processes that produce-intelligent action. By ‘intelligent action ismeant an act or decision that is goal-oriented,arrived at by an understandable chain ofsymbolic analysis and reasoning steps, and isone rn which knowledge of the world informsand guides the reasoning.” The project aims atcmltlng computer programs that act as“intelligent agents” to human problem solversin areas of scleri tific problem solving,hypothesis induction, and theory formation;dragnosis and treatment of program failures(automatic debugging) and medical problems.It aims also at a general understanding of theinformatron processes and structures needed tocarry out these types of intelligent agentactivity; and the construction of necessary

programming tools to facilrtate the building ofsuch programs.

3.1 Knowledge-based Systems Design

“The AI field has come Increasingly to view asits main line of endeavor: knowledgerepresentation and use, and an exploration ofunderstanding (how symbols inside acomputer, which are in themselves essentiallyabstract and contentless, come to acquire ameaning).‘*

“In this goal of AI research, there are fociupon the encoding of knowledge about theworld in symbolic expressions so that thisknowledge can be manrpulated by programs;and the retrieval of these symbohcexpressions, as appropriate, in response todemands of various tasks. This work hassometimes been called ‘applied epistemology or‘knowledge engineering’.”

Two of the major subgoals of the work inknowledge-based systems design constitute themost important thematic lines of research inthis project. They are:

“A. How is the knowledge acquired, that isneeded for understanding and problemsolving, and how can it be mosteffectively used?

B. How is knowledge of the world to berepresented symbolically in the memoryof a computer? What symbolic datastructures in memory make the retrievalof this information in response to taskdemands easy?”

Significant advances on these problems havebeen made in the past year. They are detailedbelow.

“The paradigm for this research is, verygenerally sketched, as follows: a situation is tobe described or understood; a signal input isto be interpreted; or a decision in a problem-solution path is to be made.

24 HEURISTIC PROGRAMMING PROJECT

Example: The molecule structure-generator must choose a chemicalfunctional group for the ‘active center’ ofthe molecular structure it IS trying tohypothesize, and the questlon is, ‘Whatdoes the mass spectrum indicate is the‘best guess’?”

Specialized collectlons of facts about thevarious particular t a s k domarns, sultablyrepresented 117 the computer memory (callthese Experts) can recognize situations, analyzesituations, and make decisions or take actlortswith in t h e domm of their specializedknowledge.

Example: In Heuristic DENDRAL, theExperts are those._ that know aboutstability of organic molecules in general,mass spectrometer fragmentationprocesses in particular, nuclear magneticresonance phenomena, etc.”

“ W i t h i n this paradigm lie a number ofimporta tit problems to which we haveadd ressed ourselves:

a. Since It IS now widely recognized thatdetalled specific knowledge of taskdomains IS necessary for power IIIproblem solving programs, how 1s thisknowledge to be Imparted to, or acquiredby, the programs?a 1. By rnteractlon between human expert

and program, made ever moresmooth by careful design ofinteraction techniques, languages‘tuned’ to the task domain, flexibleinternal representations. Our workon sltuatlon-actlon tableaus(productlon systems) for flexiblytransmlttlng from expert to machinedetails of a body of knowledge IS anexample.

a2. ‘Custom-crafting’ the knowledge In afield by the painstaking day-after-day process of an AI sclentlstworking together with an expert inanother field, eliciting from thatexpert the theories, facts, rules, and

heuristics applicable to reasoning inhis field. This was the process bywhich Heuristic DENDRAL’s‘Expert’ knowledge was built. Wehave also used it in AI applicationprograms for treatment planning forinfectious disease using antibiotics,and protein structure determinationusing X-ray crystallography.

a3. By mductlve inference done byprograms to extract facts,regularities, and good heuristicsdirectly from naturally-occurrlngdata. This is obviously the path topursue if AI research is not to spendall of its effort well into the 2lstCentury, building knowledge-basesin the various fields of humatendeavor in the custom-craftedmanner referred to above. Themost important effort in this areahas been the Meta-DENDRALprogram described below.”

3.2 The Meta-DENDRAL Progranl:Krlowledge Acquisition by TheoryFormatiotl Processes

The research task mentioned above as (a?.)has been studied in the context of a specificinductive theory formation task, a task whichis ideally matched to the project’s researchhistory: inferring parts of the theory of massspectrometry of organic molecules (i.e., rules offragmentation of molecules) from instrumentdata (i.e., mass spectra). This is an area inwhich we have not only a vast amount ofempirical data, cooperative collaboratingexperts in the area, and a considerableunderstanding of the structure of knowledgeIn the area; but also we have an operatingperformance program capable of using(thereby testing) the knowledge inferred. Thisprogram is the Heuristic DENDRAL program,developed previously wrth ARPA support(and substantial NIH support).

3 . 2 T h e Mcta-DENDRAL Program 25

Compared to grammatical inference, sequenceextrapolation, o r other induction tasks,scientific theory formation as a prototypicaltask for knowledge acquisltlon, has receivedllttle attcntlon f rom workers m Art i f ic ia lIn telllgcnce. This may be partly becausesclen tlfic theory format ion i s a relativelyc o m p l e x t a s k , characterized b y noisy a n dambiguous data whrch m u s t b e orgailizedwlthln Incomplete or controversial models ofthe discipline. However, many task areas towhich AI techniques might be applied havethis character.

Meta-DENDRAL ( M D ) I S a c o m p u t e rprogram that formulates explanatory rules( b u t n o t revolutionary rcformulatlons o fprlnclples) to account fo_r empIrIcal data inmass spectrometry, a branch of analyticalorganic chemistry. The problem IS one ofexplaining the mass spectrometry (ms.) datafrom a collection of chemical compounds -- inother words, of telling why the givencompounds produce the observed data in amass spectrometer. The most recentdescription of the Meta-DENDRAL theoryformatIon work IS given III “Scientific TheoryFormation by Computer”, a paper presented tothe NATO Advanced Study Institute onCompir let Orlen ted Learning Processes(Bonas, France, Aug. 26 - Sept. 6, 1974). Thefollowing summary is taken from that paper.

The rules of mass spectrometry are expressedin texts, and represented in the program, asconditlonal rules (also called “productions”).T h e productlons i n d i c a t e w h a t physlcalchanges (e.g,, bond breakage) we can expect amolec’u le to undergo wrthin the massspectrometer.

A dIscussIon o f o u r w o r k o n productlonsystem representations of knowledge appearslater in thrs report.

The instances presented to the program areparrs of the form <molecular structure> -<mass spectrum>, i.e., pairs of known chemicalstructures and corresponding empirical data

from mass spectrometry. A rule explains whythe mass spectrometer produces some datapoints for a molecule of known structure. Agiven molecule may undergo repeatedfragmentation in several possible ways, and wewant to find a collection of rules to explain thewhole mass spectrum.

The program is organized into four mainsteps:

1) data selectIon2) data interpretation3) rule generation4) rule modification.

3.2.1 Data Select ion

Knowing which data to examme and which toignore is an important part of science. Theworld of experrence is too varied and tooconfusing for an unselective scientist to beginfinding and explaining regularities. Evenwithm a sharply constrained discipline,experiments are chosen carefully so as to limitt h e a m o u n t o f d a t a a n d their va r i e ty .Typically one thinks of a scientist as startmgwith a carefully selected set of data andgenerating requests for new data. from newex perlmen ts, after formulating tentativehypotheses. Both the Initial selectlon and theaddltional requests are in the realm of d a t aselectlon.

The strategy behind data selection is to find asmall number of paradigm molecules - i.e.,molecules that are likely to exhibit regularitiestyplcal of the whole class. Rules formed fromthese can be checked against other data in theinstance space.

3.2 .2 Data Iuterpretatiou andSuwnary: The INTStJM P r o g r a m

Experimental data In science can beinterpreted under many different models.Finding explanatory rules withm a moclel thusforces one to interpret the data under thatmodel. For example, when one is lookmg forbiologlcal rules wlthln an evolutionary model,

26 H E U R I S T I C P R O G R A M M I N G P R O J E C T

the data (even as they are collected) areinterpreted in terms of Colltllllllty o fp r o p e r t i e s , slmrlarrtles of behavior, e t c . T h erules (If any) suggested by the data are alreadypre-formed in that the predicates and relationsused In the rules -- and often the logrcal formItself -- are exactly those of the model.

The data Interpretation part of the MDprogram (named INTSUM) descrlbcs theinstances selected b y t h e d a t a selectlonprogram in terms of one model of massspectrometry. Thts redescriptlon is necessaryfor the reasons Just noted. Without it, ruleswould be formed at the pattern recognltlonlevel of statistical correlations between datapoints and molecular features. Although rulesat this level are sometimes helpful, our intentis to produce rules that begin to say somethingabout WHY the correlations should beobserved.

Four points are Interesting here because theyseem common to screntific rule formation tasksbut unusual for other induction tasks:

I>

2)

-

T h e amblgulty o f t h e Interpretntlons.T h e mapptng f r o m d a t a potnts t op r o c e s s e s I S a one-mariy m a p p i n g .SometImes a data point actually (or mostprobably) results from multiple processescompounding the same result. At othertimes a data potnt is most probably theresult of only one process, even thoughseveral processes are plaustbleexplanations of i t . And, a t still othertimes a data point IS most probably a“stray”, In the sense that tt was producedby an Impurtty in the sample or noise inthe Instrument, even though severalprocesses may be plauslhle explanationsof It. This ambiguity 111 the instancesmakes the Induction task harder.

Theneed

fact that the presented Instancesrernterpretlng at all.

3) The strength of evldcnce associated wrthprocesses. The data are not merely

bmary readings on masses of fragments.(Most concept formation or grammaticalInference programs use only a binaryseparation of instances - “hit or miss”,although Winston’s program uses theclassifica tlon o f “near miss” toadvantage.) The strength of evidencereadings on n1.s. data points can be usedto focus attention on just a few of themany processes the program c a nconsider.

4) There is more than one rule to beformed to explain the data. In thepresentation of instances, there is noseparation according to the rules to beformed: Instances of many rules arethrown together. T h e p r o g r a m m u s tseparate them.

3.2 .3 Rule Geueratiou: T h e R U L E G E NProgram

The collected observations of INTSUM, aswith Darwin’s carefully recorded observations,constitute only the weakest sort of explanation.After clescrlblng features and behavior, andsurnmarizmg the descrlpttons, such a recordcan only give a weak answer to the question,“Why is this S an A?” The answer is roughlyof the form, “Because all X’s seem to be A’s.”

The rule generation program, RULEGEN,provides an addItional level of explanation bydescrlbmg w h a t attrlbutes o f the inputmolecular graphs seem to “cause” the moleculesto behave in the observed way.

R U L E G E N normally begins with the mostgeneral class of rules: The bond (or bonds)break regardless of their environment(situation). The program successivelyhypothesrzes more specific classes of rules andchecks each class against the data. Likelyclasses are expanded into specific rules forwhich addltlonal data checks are made. Theprocess terminates whenever (a) the morespecific classes fall the data checks, or (b)individual rules fail the data checks. T h i sprocedure is outltned below:

3.2 The Meta-DENDRAL PrograIn 27

START WITH INITIAL CLASS OFRULES

GENERATE NEXT MORE SPECIFICCLASS

painstakingly custom-crafting theknowledgelexpertrse necessary for high levelsof performance by the programs.”

SEE IF THE CLASS PASSES THEFILTER

4.3A. IF NOT, STOP

ESPAND CLASS INTO INDIVIDUALRULES

3 . 3 Systelrls for Selrli-AutonlaticKrlowledge Acquisitioll b y E x p e r t s

5C.6.

EVALUATE RULESPRUNE WITH REGARD TOEVALUATION

7.6A. IF NO RULE REMAINS, STOPF@R EACH REMAINING RULE,WRITE OUT RULE AND DO 2 - 7,

Previously we mentioned that one of themodes of knowledge acquisltlon (a.2) In wideuse is Expert-Computer Scientist interaction.Currently this mode is slow, palnstaking, andsometimes ineffective. Therefore, we h a v ebeen conducting research aimed at introducingintelligent interaction into the process ofextracting and “debugging” knowledge fromExperts.

3.2.4 R u l e Modificatiorl

While the programs described so far arepresently operatlonal, the addition of a rulemodification phase is still under way. Theprogram for rule modification will duplicatefor its own purposes the steps alreadydescribed: data selectjon, data interpretationand rule generation. D a t a selectIon in thiscase will be for the purpose of finding datathat can resolve questions about rules. Thosedata, then, ~111 b e tnterpreted a n d r u l e sformed from them, as described above. Theresults of rule generation on the new data willthen be used to modify the previous set ofrules. The data gathered In response to thequestions asked about the old rules ~111determine, for example, whether those rulesshould be made more specific or more general.Rule modlficatlon opens interestingpossibilities for feedback in the system thatremain to be exploited.

T h e Meta-DENDRAL p r o g r a m i s t h ekeystone of our study of automatic knowledgeacquisrtion. “Though the main thrust of AIresearch is in the direction of knowledge-basedprograms, the fundamental research supportfor this thrust is currently thin. This is acritical ‘bottleneck’ area of the sclencc, since (asw a s p o i n t e d o u t earlisr) It is inconceivablethat the AI field will proceed from oneknowledge-based program to the next

Knowledge acquisition is an Importantcomponent of an intelligent system because acomplex knowledge base will almost certainlyhave to change as the system is applied to newproblems. Gaps and errors in the knowledgebase will have to be considered. We h a v erecently designed a system that will provide anexpert with high level access to the knowledgebase of the system. (Work is in progress onthe implementation of these ideas.)

The specific task that is the base for this studyis a “diagnosis and therapy” advice systemdeveloped by researchers and students of ourproject, in collaboration with clinicalpharmacologists, under an NIH grant -- theMYCIN system.

The knowledge modification andaugmentation system ~111 have two entry-points: (i) the expert supplying the knowledgecan enter the system during the course of aconsultation if something seems wrong, or (Ii)at the end of a session, the post-consultationreview mechanism automatically enters thesystem to validate the program’s performance.

From the questlons that the expert asks i nattempting to find the error (or perhaps as a


result of what he decides the error is) theproblem IS classified according to one of theerror models. We may view this clnssrficationscheme as stirplclity, ignorance, incompetence,and system errors. T h u s there IS either someIncorrect rule in the current knowledge base,some rule is missing, a ‘concept (predicatefunction) is missing, or there is an error in thecontrol structure.

In the first case ‘diagnosis’ and ‘therapyroutines in the appropriate error modelattempt to discover and remedy the problem.Heavy use IS made of contextual information(what subgoal was being traced, whichquestion the user found odd, etc.).

In the second case, the-ttherapy’ is to invoke arule acquisition system. This consists of a ruledeciphering module and an incorporationmodule. The former relies on domain andcon tex t specific knowledge to aid ininterpreting the expert’s newly offered rule.The latter uses the current knowledge base ofrules and an understanding of the system’struth model to insure that the new rule isconsistent with those currently in the system.

While there appears to be little this systemwrll be able to do beyond recognizing the lasttwo types of errors, this can at least providethe hooks by which future, more sophisticatedroutines can be applied intelligently. In

- addition, the incompetence case IS clearlyrelated to ignorance -- in the latter the systemlacks a rule it is capable of expressing, whilein the former it lacks the concept necessary foreven expressing the rule. Thus failure of theignorance model to handle the problem shouldresult in the suggestion to the incompetencemodel that it may be able to recognize what’sreally wrong.

The error models are characterizations of thetypes of errors 111 the knowledge base that weexpect the system might encounter. Therelevance function for each would take a lookat what was known about the current state ofthe world to decide if rt was relevant. The

model which found itself most relevant wouldtemporarily take control, using its diagnosticroutines to attempt to uncover the source of

the problem, and its therapeutic routines toattempt to fix the problem. Thus it effectivelyoilers Its advice on how to proceed by actuallytaking control for a time.

The rule models used by the rule acquisitionsystem are slightly different in two ways. Thetask here is to decipher the new rule whichthe expert has typed In, and in this case themodels offer advice on what the content of thenew rule IS likely to be, relying on contextualinformation to hypothesize the type of theincoming rule. Hence they do not directly takecontrol, but more passively offer advice aboutwhat to expect. In addition, the large numberof rules currently in the system makesconceivable the automatic generation of rulemodels. By using a similarity metric to formanalogy sets, and then describing theproperties of the rules in the set in generalenough terms, we obtain a set of models onwhich to base acquisition. The error models,on the other hand, are both few enough, andspecialized enough to require creation byhand.

3 . 4 Kllowledge Acquisitiorl audDeploynlent: A New AI Appl icat ionto Scientific Hypothesis Formation

“We have felt for some time that it isnecessary to produce addrtional case-studyprograms of hypothesis discovery and theoryformation (i.e., induction programs) indomains of knowledge that are reasonably richand complex. It is essential for the science tosee some more examples that discoverregularities in empirical data, and generalizeover these to form sets of rules that canexplain the data and predict future states. It islikely that only after more case-studies areavailable will AI researchers be able toorganize, unify and refine their ideas

3 . 4 K n o w l e d g e Acquisitiorl and Deploylnent 29

‘concerning computer-assisted inductlon o fknowledge.”

I n s e a r c h i n g f o r n e w case-studies, t h eHeurlstlc Programming Pro~cct has developedcriteria for a successful application, c7xploredseveral task domains, and has brought onenew applrcation, dIscussed below, far enoughalong to submit grant applications (to NSFand NIH) for further research. Meanwhile,other laboratories h a v e m a d e slgmficantprogress in the design and implementation ofAI programs 111 this general area -- notably theSOPHIE system for electronic troubleshootlng(BBN) and the HASP-Sonar work (see Sectlon3.7).

3 . 4 . 1 B a c k g r o u n d -_

Our choice of the protein crystallographyproblem as a task domain in which to studyinformation processes of sclentlfic problemsolving followed several informal discussionswith P r o f e s s o r J o s e p h K r a u t a n d h i scolleagues a t IJCSD, w h o w e r e cager t oe x p l o r e n e w computational techniques f o rprotein structure eluctdation. They explamedto 11s how 3-clmicnslonal structures ofcrystallized protein molecules are Inferredfrom x-ray and amino acrd sequencing data.It was clear from these discussions that, Inaddition to the necessary but more or lessstraightforward tasks of data reduction,Fourier analysis and model refinement, therewas also a conslderable amount of heuristicdata interpretation involved in structuredetermlnatlon. T h e m o d e l builder, f o rexample, AS often faced with a number ofpossible substructures which are consistentwith an electron density map, and must basehis choice on “rules of thumb” as well asprlnclples of chemistry and geometry. Thetask domain seemed wel l suited for theapplication of heuristic programs which couldgenerate )I1 auslble hypotheses about a protein’sstructural elements and test these hypothesesagainst the crystallographic data.

Professor Kraut suggested that our efforts

would yield a high payoff If we couldsomehow eliminate any of the mainbottlenecks In determmmg protein structures.A maJor bottleneck is obtaining phase data,which is required to construct an electrondensity map of the molecule. These data areusually obtained by the process of multipleisomorphous r ep l acemen t (MIR), in whichheavy atoms are implanted in the crystallizedmolecule. The determination of many proteinstructures has been delayed for months toyears because of the difficulty in obtainingMIR data.

Kraut suggested that a way to eliminate thisbottleneck is to use the parent protein dataonly, in con junction with all t h e “non-crystallographic” knowledge which the expertwould bring to bear on each specific problem.For example, the “unknown” protein IS oftenone member of a family of similar proteinshaving known characteristic structuralfeatures. We assume that one or more ofthese features is present and test t h a tassumpt ion agamst the data. Thus is donereadily by first reducing the data to a functionof the three crystallographrc space variables.This function, the Patterson function, has asimple physical interpretation, which facilitatesthe verification process.

Havrng delineated the task domain, wecontinued to work closely with our UCSDcolleagues, and developed a program, PSRCH,whose main purpose is to test the feasibility ofinferring structure without phase informatlon.We have thus far applied the program to dataobtained from two proteins whose structuresare already known. In one case we searchedfor a characteristic tetrahedral structure ofiron and sulfur in the protein called HIPIP,by starting with its known relative coordinatesand looking f o r t h e orientation(s) a n dposittons in the unit cell which give the bestconfirmation of the experimental data. Thesearch was successful; however, the task wasan easy one and we could only conclude thatthe procedure might work on more subtlecases. We then moved on to a slightly more


drfficult case, searchrng for the posltlon of theheme strirctu re in the protein cytochrome C?.(IncIdentally, HIPIP a n d cytochrome C2 a r etwo proteins whose structures were firstdlscovered and reported on by the LJCSDgroup. There are currently only about threedozen proteins whose complete, i.e., tertta ry,structures are known today.) Here, too, tt wasposstble to find the orlentatlon of the hemegroup properly. T h e translation s e a r c hyielded several postttons which were highlyconfirmed by the data, including the correctone.

At this potnt in our research we entered tntodlscusslons with a member of the ComputerApplications to Research section of theNational Science Foun_datlon, which led to ap r o p o s a l submlsslon o n M a y 31, 1974. Anearly Identical proposal was also sent to theNational Institutes of Health in September.Extracts of the research plan, namely ourobjectives and methods of proccdurcs, aregiven below.

C o t n p u t c r networkrng has been and wrllcontinue to be an important component of ourcollaborations with the UCSD group. Untilrecently, we were using our program on theI B M 370/158 at the RAND Corporation inSanta Monica vra the ARPA network. TheUCSD group also has access to the RANDComputation Center through their ARPA

- network connection. We were thus able toexercise the program Jotntly, peruse the storedoutput on other files, or simply use thenetwork as a communlcatton faclllty for rap~cl

interchange of ideas and results. Computatronshave now been shifted to the new SUMEXfacility (a PDP-IOI), located at the StanfordMedical School. S U M E X IS a nationalresource sponsored by NIH for use Inapplying Artlficlal Intelltgence to problems o fblomecllcal lnterrst. SUMEX IS also access~l~le

t o t h e IJCST~ group, as well as CJtherS, by anetwork conrlpctlon. SUM ES ~111 provldc uswith computation unly. Our abtllty to proceedwith the work ~111, of course, depend on thecontinuation of support for the scientists whoare deslgnlng and Implementing the programs.

3.4.2 Objectives

The objective of the research proposed here IS

to apply problem solving technques, whichhave emerged from artificial Intelligence (AI)research, to the well known “phase problem”of x-ray crystallography, 111 order to determinethe three-dlmenstonal structures of proteins.The work IS intended to be of practical as wellas theoretical value to both computer science(particularly AI research) and proteincrystallography. Viewed as an AI problemour objectives will be:

1. To discover from expert proteincrystallographers the knowledge andheuristics which could be used to inferthe tertiary structure of proteins from x-ray crystallographic data, and toformalize this knowledge as computerdata structures and heuristic procedures.

2. To discover a program organization anda set of representations which will allowthe knowledge and the heuristics tocooperate in making the search efficient,i.e., generating plaustble candidates in areasonable time. (This is a central themeof current artificial intelligence research.)

3. To Implement the above In a system ofcomputer programs, the competence ofwhich will have a noticeable impact uponthe discipline of protein crystallography.

3.4.3 Methods of Procedure

Our task can be defined at two levels. At theapplication level the goal is to assist proteincrystallographers in overcoming the phaseproblem in x-ray crystallography. We proposeto do this by developing a computer programwhich Infers plausible “first guess” structures,III the absence of phase lnformat1on, b yapplying as much general and task-specificknowledge as possible to constrain the searchfor a structure consistent with the x-rayintensity data.

3 . 4 Knowledge Acquisitioll and Drploylnrnt 31

At the computer science level, our task is todiscover a pl~oy*am organl2atlon rrnd (7 s e t o frepresentatrons which can effectively utilize alarge and disparate body of knowledge incon junction with experimental data. Thep r o g r a m m u s t a l l o w t h e various fac t s ,procedures and heurlstrcs, which the expertsthemselves routlncly employ, to cooperate inmaking the search efficient, i.e., generatrngplausible candidates In a reasonable time.

The problem of organlting and utilrzlng anon-homogeneous body of knowledge, aten tral p r o b l e m i n current ArtlficlalIntelligence research is particularly acute inprotein crystallography. Generally speaking,we can divide our overall knowledge intothree categories: 1) rules and relatlonshlps, I.e.,knowledge and heurlstlcs for which there areno well-defined algorrthmic procedures; 2)rules and relatlonships expressed asalgorlthmlc proceclures; and 3) data. The.accompanylng table shows some members ofeach category:

KNOWLEDGE- Amino Acid Sequence-structure correlation- Symmetry Information

a) crystallographicb) non-crystallographic

- Stereochemical constraints- Mathematical relationships among structure

factors-- When to use which procedures

PROCEDURES- Patterson search- RQtation Function- Superposition- Trial and Error- Anomalous dispersion Patterson- Direct methods

DATA- X-ray intensities- Amino Acid Sequence- Other chemlcat data- Coordinates of related molecules If

available- Existence of prosthetic groups or cofdctors

- Space group and cell dlmenslons

These varied sources of information areex pressed 117 an equally varied set ofrepresentations. Knowledge about sequence-structure correlatrons IS expressed In terms ofamino acid sequences and nlacl.o-desct-l1)tlonsof structures (alpha-hchx, beta-sheet).Symmetry retatlonshlps are usually defined byrotation and translation operators applied tocoordinate vectors. Stereochemical constraintsare usually expressed in terms of standardbond lengths and angles, allowed values forthe (phi, psi) dihedral angles, minlmumacceptable distances for non-bonded contacrs.Among the various procedures used to extractinformation from the data we find that thePatterson search techmque works in vectorspace, the rotation function in reciprocal space,superposition methods In vector space and itsown superposition space, electron densrt)interpre[atlon in direct space, and so forth.The data as welt are comprised of tables andlists defined 111 different clornalns.

We wish to bring as much knowledge to bearon the problems as we have available, just asa practicing protem cryscaltographer would do.Therefore, we must have the ability to takeinformation obtained by operating on the database with a particrllar p r o c e d u r e a n dcommunicate it to another procedure even ifthese two procedures work with differentrepresentations olF the world. We believe thisproblem can be solved by careful systemdesign, as is discussed in the followrng section.

3.4.4 Overall desigtl of the program

One approach to protein structuredetermlnatlon would be to write a battery ofprograms, each of which had a specificcapabllity -- a Patterson Interpretationprogram, a superposltlon prosram, etc. Theinvestigator would es amine the results fromone of these ~J1~OgM-tlS, decide what to do nest,make appropriate Iransformatlons of the datato conform with the Input requirements of thenext program, and submit the next job. This

32

pt~ocess, althou~,ti cor~cc~piually stl aIgtltt0l LC r~rl,has several drawbacks:

1) There 1s a p,reat deal of mar111n1 shuttlltlgback and forth between programs, withthe concomitant task wf r’e-l’elll’e~entntlo~lo f Inpllt.

2) It IS dlticult to assess the value of eachprogram 111 advancing toward a solirtlon.

3) The ability of the lndlvtdual ~~rc~gt’arns tocoof)erate in ,711 iterative fashion Islimited by the investigaIor’s stamina andwillingness to keep the iteratiuns r,olng.

4) The manual control of the Seq~t~~ce ofprograms u.sect Increases the probabllltyof errors 111 data transference.

5) Unless very careful recorcls are kept, it willbe difficult to trace the reasoning processwhich led to the salutron.

6) As new heurlstlcs are ellclted from experts,it may be necessary to incorporate themIn several dlffcrrnt programs.

Another approach IS to adopt the programorganization rlsrd III the HEARSAY system[I] where COCJ~,ft’ntit-l~ “t?xl~~ts” Work 11-1 a

quasi-parallel fashion to build the hy~~otheslstoward a complete solutlolr. FlEgure 2 showshow this program structure might look f‘or 0111applrcation; it IS essentially isomorphic tofigure I sI10w11 for the HEARSAYorganization. Instead of “recognizers” we havesubstituted “analysts”, experts 11’1 afJfJ\\jlrIg 3

spcclnllzcd techrrlque to the data at hand 111- order to propose and verify a partrnl structure.

At the left of the figure are well-estnbllshedpre-proccsslng routines which can reduce thedata and make the transformatrons Into formsthat can be used by the analyst programs. FotHE’Ah!%Y’S kXkOI1, W e haVe Sl~bStl~llb?d 0111’

o w n drctlonary o f superatoms, l.c., apolyatomlc structure which I S consldrted ;IS aunit. Examples are alpha hel~ces, the amitacid residues, hemc group, and INa sheets.The control ler a t the bottom oi the flgirrcplays the same role as RCIVER, therecognrtlon overlord 111 HEARSAY. Thecontrolltr can examine the list of current Ijcsthypotheses and pass control to the a~~~Jq>l’ia~e

analyst for further synthesis of a structure 01verlficatioti of an extant structure. .

HEIJRISTIC: PROGRAMMING PROJECT

Although the representations of knowledgerequired by the various analysts may be~ncompaflble, t hey commutilcate through a

..standardlzed representation of the hypotheseswhich they can generate and verify. Thehypothesis may be thought of as a global database which communicates information betweenthe analysts, as shown schematically in thefigure. The hypothesis is a partial structure,which may be represcntcd by a list of atomiccoordinates plus a descrlprlon of allowed andforbldden regions of occupancy of the unitcell. We have not yet settled on a singlerepresentation; It is currently under study.

The particular analysts shown in the figureare only a representative subset of the fullpanoply that can eventually be added. Theaddltlon of a new expert to the system wouldb e relarively straightforward. T h e n e wprogram c o u l d b e written and testedIndependently, provtding the designer adoptsthe standard hypothesis representation. Tomerge the analyst with the full system wouldrequrre adding n e w ru1es to the controller’sscheduling heuristics. The controller is drivenby a table of sltuatlon-action rules, as in theplannmg phase of Heuristic DENDRAL.

Although we have used the structure of theHEARSAY program to guide the organizationof our protein structure inference program,there will likely be some sigruficant differenceseven at the schematic level shown in the twofigures. For example, not all analysts willcontain both an hypothesizer and a verifier.Some analytical techniques are capable of oneor the other but not both. Also, the generalknowledge box shown at the top of figure 1may contain subcategories of informationwhich a r e no t compa’lble with all of theunderlyIng analysts. These changes shouldnot Interfere with the basic idea of buildingan hypothesis by a set of cooperatingspeclnllzed procedures, under the coordinationof a rule-driven executive.

I

I-f -I !

I - - -5

-i-- -

-2

k

.. I .4-4

1 PARTIF~LLY ,

. I ..t

...I :.\I .

..:

.

..

Fig. 1. Detail of the Recognition Processin the CMU-HEARSAY1 System

3.4 J<rlowledge Acquisition a11d Deployment 35

Bibliography

[ 11 r). R. Reddy, L. D. Erman, R. Neely, AModel and a .Systeln for MachilleRecog 11 it ion of Speech, lEEE Trans.Audio & Electra-acoustics, AU-? I, 1973.

3.5 Knowledge Deploylneut Research:I n e x a c t Reasohg

Often the knowledge acquired from Experts isintrinsically imprecise, i.e., although itaccurately captures t h e Expert’s bestunders tanding of the situation, hisunderstanding is imprecis?. By what processescan a program be made to reason with suchknowledge of a domain?

The tntultive and Inexact aspects of IT~SOIIII~~

are described by Helmet. and Rescher [l] whoassert that the traditional concept of ‘exact’versus ‘Inexact science, with the socral sciencesaccounting for most of the second class, hasrelied upon a false distinction usuallyreflecting the presence or absence ofmathematical notation. They point out thatonly a small portion of natural science can betermed exact -- areas such as puremathematics and subfields of physics in whrchSome of the exactness “has even been put tothe ultimate test of formal axiomatlzatlon”. Inseveral areas of applied natural science, on theo t h e r h a n d , decisions, p r e d i c t i o n s , a n dexplanations are only made after exactprocedures are mingled with unformahzedexpertise. Society’s general awarenessregarding these observations is reflected in thecommon references to the ‘artlsttc’ componentsin science.

In a recent paper (submitted to MathematrcalBiosciences) we examine the nature of suchnonprobabitlstic and unformallzed reasoningprocesses, consider their retatlonship to formalprobability theory, and propose a modelwhereby such Incomplete ‘artistic’ kno\jrledge

might be quantified. We have developed thismodel of inexact reasoning in response to theneeds of AI Knowledge-based systems. Thatis, the goal has been to permit the oplnlon ofexperts to become more generally usable byprograms and thus more avaIlable t ononexperts. The model is, In effect, anapproximation t o conditlonal probablhty.Although conceived with one problem area inm i n d , i t i s potentially applicable t o a n ydomain in which real world knowledge mustbe combined wi th exper t ise before aninformed opinion can be obtained to explainobservations or to suggest a course of action.

Although conditional probability in general,and Bayes’ Theorem in particular, providesuseful results in decision making, vast portionsof technical experience suffer from so littledata and so much imperfect knowledge that arigorous probabilistic analysts, the idealstandard by which to judge the rationality ofdecisions, is not possible. It is neverthelessinstructive to examine models for the lessformal aspects of decision making. Expertsseem to use an ill-defined mechanism forreaching decisions despite a lack of formalknowledge regarding the interrelationships ofall the variables that they are considering.This mechanism is often adequate, in well-trained or experienced lndivlduals, to lead tosound conclusions on the basis of a llmlted setof observations.

A conditional probability statement is, ineffect, a statement of a decision criterion 01rule. For example, the expression P(HIE)=Xcan be read as a statement that there is a100X% chance that hypothesis H is true givenevidence E. The value of X for such rulesmay not be obvious (e.g., “y strongly suggeststhat t is true” is difficult to quantify), but anexpert may be able to offer an estimate of thisnumber based upon experience and generalknowledge, even when such numbers are notreadily available otherwlse.

A large set of such rules obtained fromtextbooks and experts would clearly contain a

36 IJEURISTIC P R O G R A M M I N G P R O J E C T

large amount of task-specific knowledge usefulto an intelligent program. It 1s conccrvablethat a computer program could be designed toconsider all such general rules and to generatea final probability of each H based uponevidence gathered in a specific situation.

Programs that mlmlc the process of analyzmgevidence incrementally often use this versionof Bayes’ Theorem:

Let El be the set of all observations to date,

and e be some new piece of data.Furthermore, let E be the new set ofobservations once e has been added to El.

Then the probability of hypothesis H onthe combined evidence is expressed as:

P(HIE) =P(e 1 H A El) P(H 1 El)

t: P(e 1 Hi A E,) P(Hi 1 El)

The successful programs that use Bayes’Theorem in this form require huge amountsof statistlcal data, not merely P(H 1 ek) for

each of the pieces of data, ek, in E, but also

the Interrelationships of the ek wlthln each

hypothesis H,.

Bayes’ Theorem would only be appropriatefor such a program, however, if values fol

e P(e l 1 Hi) and P(el I Hi A e$ could be

obtained. These requirements becomeunworkable, even If the subjectiveprobablllties of experts are used, in caseswhere a large number of hypotheses must beconsidered. The first would require acquirtngthe inverse of every rule, and the secondrequires obtalnlng expllclt statementsregarding the Interrelationships of all rules 111the system.

In short, we would like to devrse anapproximate method that allows us to computea value for P(Hi I E) solely in terms of

P(H, I ek), where E IS the composite of all the

observed e’s. S u c h a techmque ~111 not b eexact, but since the conditIona probabilities

reflect Jildgmeiital (and thus highly subjective)knowledge, a rigorous application of Bayes’Theorem will not necessarily produce accuratecumulative probabilities either. Instead wehave sought (a) ways to handle decision rulesas discrete packets of knowledge, and (b) aquantification scheme that permitsaccumulation of evidence in a manner thatadequately reflect; the reasoning process of anexpert using the same or similar rules.

We believe that the proposed mode1 is a

plausible representation of the numbers anexpert gives when asked to quantify thestrength of his judgmental rules. W e call

these numbers “certainty factors,” or CF’s. Hegives a pos i t ive n u m b e r (CF>O) i f t h ehypothesis is confirmed by observed evidence,

suggests a negative number (CF<O) if the

evidence lends credence to the negation of the

hypothesis, and says there is no evidence at all(CF=O) If the observation is independent ofthe hypothesrs under consideration. The CFcombmes knowledge of both P(h) and P(h I e).

Since the expert often has trouble stating P(h)

and P(h I e) in quantitative terms, there is

reason to believe that a CF that weights both

the numbers into a single measure is %ctually amore natural intuitive concept (e.g., “I don’tknow what the probability is that all ravensare black, but 1 DO know that every time you

show me an additional black raven my beliefis increased by X that all ravens are black.“)

In accordance with subjective probabilitytheory, we assert that the expert’s personalprobability P(h) reflects his belief in h at anygiven time. Thus l-P(h) can be viewed as anestimate of the expert’s Disbelief regardrng thetruth of h. If P(h 1 e) is greater than P(h), theobservation of e increases the expert’s Beliefin h while decreasing his Disbelief regardingthe truth of h. In fact, the proportionatedecrease in Disbelief is given by the ratio:

3.5 K llowledge Deploymellt Research: Illexact Reasoning 37

Ph I 4 - W-d_.-1 - P(h) -

We call this ratio the measure of increasedBelief 117 h resulting from the observation of e,i.e., M B[h,e’l.

Suppose, on the other hand, that P(hle) wereless than P(h). Then the observation of ewould ciecrease the expert’s Belief in h whileincreasing his Disbelief regarding the truth ofh. The proportionate decrease in Belief is inthis case given by the ratio:

’ P(h 1 e) - P(h)_-1 - P(h)

We call this ratlo the measure of increasedDisbelief in h resulting from the observationof e, i.e., MD[h,el.

In addition, we define a third measure, termeda certainty factor (CF) that combines the MBand MD rn accordance wrth the followrngdefinrtlon:

P(h) - P(h I e>

P(h)

The certamty factor thus is an artifact forcombining degrees of Belief and Dlsbellef intoa single number. Such a number is needed ino r d e r t o facilitate comparisons of theevidential strength of competing hypotheses.The use of this composite numb?r is describedin greater detail in the paper.

Bibliography

[I] Helmer, O., N. Rescher, 011 theEpistemology of the Inexact Sciences,Project RAND Report R-353, RANDCorp., Santa Monica, Cal., February 1960.

3.6 Knowledge Deploylwwt Research forReal-World Applications: TheProblem of Explaining a Program’sReasoning

As AI’s research in knowledge-based systemsmoves toward application to real-worldproblems, it becomes essential for theintelligent agents so constructed to be able toexplain to their users the knowledge a n dinference paths used in solving problems (01suggestmg solutions). Without this ability, itis our belief that AI systems will not receive

widespread acceptance. Nor will it be possibleadequately to “debug” the knowledge in thesystems’ knowledge bases.

To conduct this research, we turned once moreto the specific task domain and program(MYCIN) for dlagnosls and treatment ofinfectious disease (an NIH-sponsored effort).

Recent progress in the development of theMYCIN system has included the development

of a capability for providmg sophisticatedexplanations of the program’s reasoning steps.

Several aspects o f t h e implementation o fMYCIN facilitate the accomplishment of thisgoal -- the modularity of the program’s rulessimplifies the task of malntaining a record ofthe program’s chain of reasoning, while theuse of an interpretive language llke L ISPmakes feasible the examination by theprogram of its own knowledge base, as well asthe translation of the rules Into ‘English fordisplay to the user. This ability of theprogram to keep track of Its reasoning and toexamine its own knowledge and data is thecentral component in Its ablllty t o e x p l a i nitself.

MYCIN normally takes the lnltlative durmg aconsultation session; the system asks q uestlonsand uses the answers to determine theapplicability o f t h e decision rule it hasretrieved. The user who desires anexplanation of the program’s motivation for aparticular question has avallable to him a set

38 H E U R I S T I C PROGRAMMING. P R O J E C T

of commat~ds d e s i g n e d t o m a k e t h eexamination of the program’s reasoning bothsimple and effective.

3.6.1 WHY Questions -- Looking atGOillS

Since any question IS the result of nn attemptto determine the truth of precoi~ditlons of agiven subgoal, the simplest explanatron of themotlvatlon for a questlon is a statement of thecurrent subgoal. Ey typing WHY, the userwrll get a detailed explanation from the systemof the type of conclusion it is trying to draw,and how the current rule is to be applied inthis case to establish that conclusion. Notethat the system first examines its currentreasonjng chain to determine the “purpose” ofthe question, then examines the current rule todetcrmrne h o w It applIcs i n thus p a t tlcularcase, and finally translates all of thisinformatlon from its Internal LISPrepresentation Into uncierstandable English.

The user may understand why any particulatquestlon was asked, but may be unsure as tothe program’s reason for seeking theconclusion mentioned. He can examine thisnext step in the rcasonlng by simply repeating“WHY”. This process can be repeated asoften as desired, until the entire currentreasoning chain has been displayed.

One problem we anticipated In the use of theWHY command, and one that is common withexplanations In general, is the issue ofpresenting an explanation with theappropriate level of sophistication.Depending on the user, we mght want to (a)display explicitly all steps III the chain o freasoning, (b) omit those which aredefinltlonal or trlv~al, or perhaps, cir the mostsophlstrcated user, (c) display only thehlghllghts and allow him to supply the details.

We have provided this capability by allowingthe user to indicate his level of sophlstlcatlonwith an optlonal a r g u m e n t t o t h e W H Ycommand. This parameter indicates how

large a step In the reasoning process must bebefore it IS to be displayed. Once again, thiscan be repeated as often as necessary, allowing

-.the user to follow the reasoning chain in stepsizes of his own choosing.

3.62 HOW questions: Looking atPrecollditious

We have seen that as the user examines the ’current reasoning chain, he is informed of thevarious subgoals the system needs to achievein order to accomplish the main goal. At somepoint he may wish to examine all the waysany subgoal may be achieved. For thisexamination of additional reasoning chains,he can use the HOW command.

The query allows the user to find out: (a) howa rule WAS used in the reasoning, i.e., whatwas known at the time the rule was Invokedand what conclusions were drawn as a result;(b) how a rule WILL BE used in thereasoning, i.e., what will have to be known forthe rule to be invoked and what conclusionwill be drawn, and (c) how a fact w a sdetermined that allowed some inference to bemade.

Two points should be noted about the designof the program which generates theseexplanations. First, consistent with the generalphilosophy of MYCIN, the approach is quitedomain-Independent. A l though we h a v ewritten programs with explicit knowledge ofwhat is required for acceptable explanations,all task-specific knowledge is obtained byreferring to the information stored in theseparate knowledge base of rules.

Second, in attempting to supply information tothe user, the system examines Its own actionsand knowledge base in order to discover whatin fact it is “trying to do”. The explanationprogram thus “keeps watch” over the actionsof the consultation program by keeping arecord of all of its past actions and mimickingits normal control structure when examiningpossible future actions.

39

3 . 7 I< nowlcdgc RepreserlhtiowProductim S y s t e m

“The problem of representation of knowledgefor AI systems is this: if the user has a factabout the world, or a problem to be stated, III

what form does this become representedsymbolically in the computer for immediate orlater use?”

The formal structures and techniques forrepresenting knowledge berng explored by ourproject are productlon rules and productionsystems, a topic also being pursued vlgoroustyby the ARPA projec t a t CarnPglc-MellonUniversity. Productlot s y s t e m s o f f e r a“natural”, highly flexible, and modular way ofrepresenting knowledge.__ The approach IS

highly promising but much more work by usand others needs to be done.

Judgmental rules of the form ‘If A then E’ arecommonly found In text and handbooks ofsclentlfic and technlcal disciplines. These rulesnot only describe formal relatronshlps of adiscipline but rules of accepted practice andhints of suggestive relations as well. For thesereasons, production systems are importantvehicles for encoding an expert’s mferentralknowledge in Intelligent programs. They havebeen studied and used in such programs asNewell’s PSC, Waterman’s poker-IearnIngprogram, S hortlift’e’s M KIN program andthe Heuristic DENDRAL h ypot heslsformation program.

In the Heuristic DENDRAL program, a tableof productlon rules holds the knowledgeneeded to explam emprrlcal data from asubf ie ld of analytical chemrstry (massspectrometry In partrcular). Part of thesophlstlcation of this program comes fromseparating the program’s knowledge base fromthe routines that use the knowledge forp r o b l e m solvmg. Also, the productionsthemselves are more general than simpleconditional sentences, ‘If A then B’. Theantecedents may be Boolean combinations ofcomplex predicates and the conseq uen ts may

be either (1) one or more processes to executewhen the antecedent 1s true, (2) defaultprocesses to execute w h e n n o special-caseantecedents are true (“else” conditions), or (3)another production. Making the productlonschema recursive (I.e., allowrng any consequentto be another production decreases the runtime and increases the storage eficlency of theprogram because the more general predicatesseparating large classes of special cases need tobe stated and checked only once. F o rexample, the schema might be written as:

If A thenIf Al then (If A 11 then B 11)

(If Al2 then B 12)else B 1

If A2 then B2else B.

Or, less efficiently, it could be equivalentlyrepresented as a set of smlple productlons 111the form:

If A A Al A All then Bll;If A A Al A Al2 then 612;IfAAAl then Bl; - [All A A l 2

are implicit in the ordering]If A A A2 then B2;If A then B.

The knowledge representation ideas developedin the context of Heuristic DENDRAL havebeen successfully mapped Into another A.1p r o g r a m , u s i n g a d i f f e r e n t domam o fknowledge. E. Shortliffe, in consultation withmen1 bers of the Heuristic ProgrammingProject (but under other financial support),developed the MYCIN program for reasoningabout antimicrobial therapy.

The knowledge base of the program IS a tableo f productIons suppl ied by an exper t ,containlng definitions, “common sense” piecesof knowledge, general sclentlfic knowledge andhighly task-specific inference rules. MYCIN isanother successful application of AI to ascientific discipline whose sophistication isd e r i v e d p a r t l y f r o m t h e flexibility of aproduction rule representation of knowledge.


Bibliograf)fly

[l] E. H. Shortllffe, S. C. Axllne, B. C.Buchanan, T. C. Merigan, and S. N.Cohen, An Artificial IntclligctlccPrograul to Advise Physiciails RegardirlgAnt ini icrobiai Tflerapy, Compt~ttrr tzn(iBiomedical Research, 6, 544-560, 1973.

[2] E. H. Shorthffe, S. G. Axhne, B. G.Buchanan, S. N. Cohen, DesignConsidcratiol1.s for a Program to ProvideCorlsuftations irl C l i n i c a l Tflerapeutics,PYOC. Biomdiccll Sym,bosium, San Diego,February 1974.

l[3] E. H. Shortllffe, 13. G. Buchanan, A Model

o f Inesact Reasoni&g it) Mcdicirle,(subm~ttecl to r2fathrmn~icai Biosticnct*s),

[4] E. H. Shortllffe, R. Davis, S. G. Asllne, B.G. Buchanan, C. C. Green, S. N. Cohen,Coinf>utcr-Rascd Coilsuftations iIiCfinicaf Tfler,?fxwtics: Exf>inuatiou alldRule Acquisitiorl Capabilities of t h eMYCIN System, (submrtted to Compu.tzrsand Biomedical Research).

3.8 .4pplicatioll of AI Techniques to aProgrammer’s Task: Automatic

e Debugging

We regard the tasks confronting computelprogra nimers to be especially Interestrng aspotential appllcatrons of our AI techlrlques.In such tasks, the “Expert” and the “CornputelScien t 1st” usually merge into one person,facrlitatrng the development o f Col’lllJks

knowledge bases.

T h e w o r k o n Automatic Programmrng h a sbeen done III the context of a Ph.D. Thtws onAutomatic Debugging of Logical ProgramErrors. The long-term goal of this research IS

to provide a sys tem which ~111 d e t e c t ,diagnose, and treat logical programming

errors. The Detection phase of debuggingmvolves the process of deciding that aproblem Indeed exists In the program. The

Diagnosis phase involves the process ofisolating the problem. The Treatment phaseinvolves the process of determming whatcorrection is to be made in the program andmaking it. We make a distinction betweenthree classes of errors: (A) Syntactic errors, (B)Semantic errors, and (C) Logical errors. Asyntactic error occurs when the text of theprogram does not conform to the syntax rules ’of the programming language. A semanticerror occurs when a syntactically correctprogram attempts during its executionoperations whose results are not defined by thesemantics of the language. A loglcal erroroccurs when a program which is syntacticallyand semantically correct gives results whichare “wrong”. A prototype system is now upwhich IS capable of detecting a small butinteresting class of bugs, and diagnosmg andtreating a subset of the detected bugs. Theprototype has correctly repaired a ‘real’ bug ina ‘real’ program. (See Brown, ‘InternalProgress Memo’, available from HeuristicProgramming Project).

The prototype system operates by requestinginformation from the user (programmer)describing features of the execution of hisprogram. This description is then used todetect errors. More precisely, during theexecution of the program, checks are made onthe consistency of the actual programexecution with the programmer’s expressedintentions. If a discrepancy is detected, thediagnosis phase of the system is invoked. Anattempt is then made to recognize what sort ofe r r o r h a s occured, a n d t o determule i t sprobable source. If this is successful, thetreatmsnt phase uses this information torepair both the current executing environmentand the source program. The execution of theprogram is then resumed In the new, hopefullycorrect, environment. The goals to beachieved In the short term are:

1. to dlscover a language for describmgprograms which:a. is easy and reasonably error-free.

3.8 Application of AI Techtliqucs to a Programmer’s Task: Automatic Debugging 41

b. is able to describe those features ofprograms most useful in debuggingprograms.

2. to provlcle a system which interacts withthe user (programmer) to obtain adescription of a particular program.

3. to provide a system which, given aprogram, its description, and sample data,can debug (or at least significantly help todebug) the program.

The effort in part 1 in designing a languagefor describing programs is closely related toother Automatic Programming research. Thislanguage (and its recognizer) need not be asrich or complete as a full-blown AutomaticProgramming language, since a dialog need--only descrrbe features of an exIstlng program,instead of clescrlblng a program to be created.But the primitives of this descriptive languageshould be useful in a more complete language.,The effort In part 2 is related to the “dragnosisand therapy” paradigm of the MYCIN system(mentioned earlier). In both systems, a dialogbetween the user and the system generatesinformation about an Individual(patient/program). This information IS thenused tn diagnose problems in the individual.The 4 ?I lrt in part 3 involves creation of aknowledge base of common programpathologies. This work has a traditional AIflavor. The knowledge base must bestructured in such a way so as to be easilymodified (changes and additions), but yet beeffective IH accomplishing its given task. ADENDR AL-like productlon system is bemgconsidered for knowledge representation In thesystem bemg built.

3.9 Tools arid Techuiques

The major work during this period falls intothe categories of maintenance anddevelopment of basic utilities. Variousprojects that were completed are:

T R A N S O R

A TRANSOR package in INTERLISP fortranslating LISP/ 360 programs toINTERLISP was completed. Emphasis wason making the package complete (almost allprograms would translate without userintervention) and in increaslng the efficiencyof the resultant code (often a straight-forwardtranslation is not the most efficient).

C O M P A R E

A L I S P p r o g r a m f o r c o m p a r i n g t w oINTERLISP files and discovering thedifferences. Similar to SRCCOM but talloredto LISP expressions. The algorithm involvesheuristics about the most common types oftransformations made in editing a program(extraction, switching expressions, insertions,changing function calls, etc) in a tree searchfor the set of “minimal” transformations fromone file to the next.

S Y S T E M D E B U G G I N G

An incompatibility between KA-l0 and KI- 10TENET was discovered which resulted inobscure problems in INTERLISP on the KI-10. The cause was determlned after muchinvestigation. Other similar problems withthe interface of LISP and TENET have alsobeen investigated and fixed.

T E N E X u t i l i t i e s

The following utilities were produced: READ,a simple minded READMAIL; SYSIN, aprogram for starting up INTERLISP sysoutfiles without going through LISP first (latermodified at BBN); OPSAMP, a program formomtoring the different instructions aprogram is executing.


HPRINT, a program for printing and readingback in circular structures, includmg userdatatypes, arrays, hash tables, along with theinterface to the INTERLISP file package andread table facility.

Systei~i iiiairitcliarice

Solving the problems of transfering filesbetween one TENES site and another viatape where each site has a dlfferentconvention for formatting information on tapeinvolved a significant amount of time.Utllrties f rom other TENEX si tes wereeventually brought over and put up on theSUMES-TENET facility.

3,lO Techllology Transfer : Chemistry audMass Spectronletry

The past year has seen heavy involvement byother granting agencies in research which wasinitiated by ARPA fundmg, or supported III

part by thrs funding. This demonstrates theprograms’ high levels of performance III thetask areas of chemistry and mass spectrometry.These programs are, or soon will be, availableto members of the Stanford ChemistryDepartment and to a nationwide communityof collaborators via the SUMEX computer

- facility and TYMNET and ARPANET.

Applications of and further r e s e a r c h Illtoprograms arising f r o m actlvlties o f r h eDENDRAL project have been funded by theBiotechnology Resources Branch, NationalInstitutes of Health for a period of three yearsbeginning May 1, 1974. In addition, twosmaller grants have been awarded whichsupport ancillary areas of research, againbegun with ARPA support. There are (1) anNSF grant to Dr. Harold Brown to supportfurther research into graph theory as appliedto chemical problems, and (2) an NIH awardto Prof. C. Djerassi and Dr. R. Carhart tosupport applications of DENDR AL programsto carbon- 13 nuclear magnetic resonatice.

Three major AI programs have been planned,developed and transferred to workingchemists. These are outlmed below.

PLANNER - our efforts at modelling theprocesses of scientific inference resulted in aprogram for analysis of mass spectral data.This program has been successfullydemonstrated in applications to importantchemical problems in the steroid field. WithNIH support we are extending the generalityof this program and adding an interactiveuser interface.

STRUCTURE GENERATOR - As in everyheuristic search program, an ex haustlve legalmove generator is central to our applicatio!isprogram. We have finished a comp!ete andirredundant generator of chemicai structuresand recently added lnechanisms forconstraining the generation process.

The generator alone is a useful working toolfor chemists with structure elucidationproblems because it can build moleculargraphs from in fe r r ed units m u c h m o r ethoroughly than a human chemist can. It hasrecently been successfully demonstrated toseveral groups of chemists, and is currently inuse by members of the Stanford ChemistryDepartment. Work will continue under NIHsuppor t to improve the program’sperformance, as it represents a framework formore general applications of AI techniques tochemical structure problems.

DATA INTERPRETATION ANDSUMMARY. The INTSUM program (forINTerpretation and SUMmary) wasdeveloped as the first stage of a program formodelling scientific theory formation. Thisprogram is applied to a large collection ofmass spectral data in an attempt to uncoverregular patterns of fragmentation across aseries of related chemical compounds. It hasproven, by itself, to be a powerful assistant tochemists in their interpretation of largequantities of data. As a result, an interactiveversion of this program is now available and

3.10 Technology Transfer: Chenlistry arId Mass Spectrom.etry 43

is being applied to problems in the massspectrometry laboratory.

3.11 Technology Transfer: to Biology arldMedicine

For many years, ARPA contract support tothis project for basic research in IntelligentSystems has been the seed from which hasgrown grant support from other federalagencies with different missions. For example,the research on Heuristic DENDRAL, lnlttallysupported by ARPA, was later supported byNIH (and recently renewed by NIH). TheAR PA-supported work 6n knowledge-basedsystems led to NIH support for thedevelopment f o r a p r o g r a m t o diagnosebacterlnl mfectlons and advlse treatment (theMYCIN program). NSF has Indicatedconsiderable interest in fundmg the hypothesisformation program in protein crystallography,begun under ARPA support.

The most significant event of this type,involving the transfer of concepts andtechniques developed under ARPA support toanother area of science and another source ofsupport, occurred during this period. The Co-principal Investigators of this project were.successful in obtaining grant funds from theBiotechnology Resources Branch of NIH toestablish a computing facility to satisfy notonly the espandlng needs of this protect, theNIH-sponsored DENDRAL project, and theother - NIH-sponsored activity; but also theneeds of an cmbryonlc but rapidly growingnational community of scientific work in theapplication of artlficlal intelligence techniquesto Biology and Med ic ine . A computmgfacility (with staff) called SUMES has beenestabllshed at the Stanford Medical School.Its complement of equipment is similar to thatat the ARPA-sponsored AI laboratories -- themaln frame is a PDPIOI, operating under theTENEX operating system. It is currentlyconnected to the TYMnet, and in the near

future will be connected to the ARPAnet by aVDH connection.

SOMEX, as mentioned, will serve not onlyStanford interests but also the interests ofAIM (Artificial Inteltigence in Medicine), aname given t o a n a t i o n a l community o finvestigators interested in aPPlYiw thetechniques of AI research to Medicine andBiology. Such investigators include, forexample, professors and s tudents a t theComputer Science Department of RutgersUniversity; and Dr. I<. Colby, now at UCLA.AIM IS constituted to have Its own AdvisoryCommittee to allocate its portion of theSUMEX resource, and to advise NIH and theS U M E X Principal Investigator, ProfessorLederberg, on the needs of the nationalcommunity and on how best to satisfy thoseneeds.

In considering the technology transfer aspectsof SUMEX-AIM, it is important to note:

1. that a federal science funding institutionthat has traditionally been veryconservative in its funding of advancedcomputer science research (NIH) --certainly much more conservative thanARPA and other DOD agenciqs -- hasbeen persuaded to take this major step atthe computer science research frontier.The credit for this is in no small measuredue to the massive evidence developedwith ARPA support that such a stepwould have great payoff to the medicalscience community;

2. that the previous NIH computer fundingpolicy -- of funding computer facilitiesfor geographically local communities ofinterest (like “researchers at the BaylorUniversity Medical School”) -- has beenchanged to one that supports facilities folscientific communities of interest notnecessarily geographicatly local. Thecredit for this is due primarily to theARPAnet, and the networking conceptsdeveloped in conjunction with ARPAnetdevelopment.

44 HEURISTIC PROGRAMMING PROJECT

3.12 Technology Transfer: to MilitaryProblems

At ARPA’s request one of the co-principalinvestigators was asked to investigate theapplicability of the concepts and techniquesdeveloped in the DENDRAL project to asurveillance signal interpretation problem ofconsiderable importance to the DefenseDepartment. Since thrs work is of a clnsslfiednature, it IS being performed not at StanfordUniversity but at a local research company.However, the Heuristic Programmrng Project’swork IS of key Importance m shaping thedevelopment of that military appllcatlon ofartificial intelligence. Further cletallsconcerning this application can be obtainedfrom Professor Feigenbaum or from Dr. J.C.R.Licklider of the ARPA IPT Ofice.

3.13 Publications of the Project, 1972/1974

Research Supported by ARPA and by NIH

Bibliography

[l] D.H. Smith, B. G. Buchanan, R.S.Engelmore, A.M. Dutield, A. Yea, EA.Felgrnbaum, J. Lederberg, and C.

e DJel-ass!, Applications of Artificia)Iiitclligeiice f o r Clteulical IiifercrlceVIII, AII approach to the Computerlllterprctation of t h e High RcsnlutiollMass Spectra of Complcs Molccuks.

. Structure Elucidat iou of EstrogeuicSteroids, Jowna/ of the American ChemicalSociety, 94, 5962-597 1 ( 1972).

[2] B.C. Buchanan, E.A. Feigenbaum, andN.S. Srid haran, Heuristic TheoryFormation: Data Interpretatioll and RuleFornrat ion, 111 Machine Intelligence 7,Edinburgh University Press (1972).

[3] Lederberg, J., Rapid Calculation ofMolecular Formulas f row Mass Values,J. Chemical Education, 49, 6 13 ( 1972).

[4] Brown, H., Masinter, L., Hjelmeland, L.,Constructive Graph Labeling UsingDouble Cosets, Discrete Mathematics, 7(1974), I-30; also Computer Science Memo318, (1972).

[5] B.C. Buchanan, Review of HubertDreyfus’ \Vhat Computers Can’t Do: ACritique of Artificial Reason, ComputingRwiew, (January, 1973); also Stanford A.I. Memo AIM-181.

161 D. H. Smith, B. C. Buchanan, R. S.Engelmore, H. Aldercreutz and C. Djerassi,Applicatiorls of Artificial Iutrlligeuce forChe,ulical I n f e r e n c e I X . Aualysis o fMi,xtures Without Prior Separation asIllustrated for Estrogens, J* AmericanChemical Society, 95, 6078 (1973).

[7] D. H. Smith, B. C. Buchanan, W. C.White, E. A. Feigenbaum, C. Djerassi andJ. Lederberg, Applications of ArtificialIiitelligeiice f o r Cheiiiical Iiifereuce X.Irltsuln. A Data In terpretat ion Progranlas Applied to the Collected Mass Spectraof Estrogeuic Steroids, Tetrahedron, 29,3117 (1973).

[81 B. G. Buch anan and N. S. Sridharan, RuleForulatiorl OII Non-I~olllogelleolls C l a s s e sof Objects, Pvoc. T/lit-d lntcrnational JointConjuuxe on Artificial Intelligence,Stanford, California, August (1973); alsoStanford A. I. Memo AIM-215.

[9] D. Mlchie and B. G. Buchanan, CurrentStatus of the Heuristic DENDRALProgranl for Applyiug A r t i f i c i a lIntelligence to the Iuterpretatiou of MassSpectra, to appear in Computers forSpectroscopy, (ed. R.A.G. Carrington),Adam ‘Hilger, London; also: University ofEdinburgh, School of ArtificialIntelligence, Experimental ProgrammingReport No. 32 ( 1973).

3.13 Publications of the Project, 1972/1974 45

[IO] H. Brown and L. Maslnter, AnAlgor i thm for the Constructioil of theGraphs of Organic Molewks, Disrrcte

Mathtvratics, (In press); also StanfordComputer Science Dept. Memo STAN-CS-73-361, May 1973.

[11] D. H. Smith, L. M. Masinter and N. S.Srldharan, IIeuristic DENDRAL:Analysis of MolecuJar Structure, Pm.N ATOICN N A Advanced Study Instituteo n Computer Reprrscntation nntlManipzdntion of Chmictrt Injormntion, (W.T. Wipke, S. Heller, R. Feldmann and E.Hyde, cds.), John Wiley and Sons, inc.,1974.

[ 121 R. Carhart and C. Djerassi, Applicationsof Artificial Intelligciice for ChemicalInference XI: The Analysis of Cl3 N M RData for Structure Elucidatiorl ofAcycl ic Amimes, J. Chem. SW (Perkin II),1753 ( 1973).

1131 L. Masinter, N. S. Sridharan, R. Carhartand D. H. Smith, Applicatiolls ofArtificial Illtelligerlce for ChemicalIllference XI I : Exhaust ive Generatiorl o fCyclic and Acyclic Isomers, submitted toJournal of the American Chemical Society;also Stanford A. I. Memo AIM-216.

[ 141 L. Masinter, N. S. Sridharan, R. Carhart-and D. H. Smrth, Applications ofA r t i f i c i a l Irltelligcl~ce for Chew icalIllferellce XIII: Arl A l g o r i t h m f o rLabelling Chcnl ical Graphs, submitted toJournal of the American Chemical Society

[ 151 N. S. Srldharan, Computer Gerleratiollof Vertex C;raphs, Stanford CS MemoSTAN-U-73-38 1, July 1973.

1161 N. S. Srldharan, et.al., A HeuristicProgram to Discover Syrlthrses forComplex Organic MolecuIcs, Stanford CSMemo STAN-CS-73-370, June 1973; alsoStanford A. I. Memo AIM-205.

[I71 N. S. Sridharan, Search Strategies forthe Task of Organic Chemical Synthesis,Stanford CS Memo STAN-CS-73-39 I,

. October 1973; also Stanford A. I. MemoAIM-217

[18] D. H. Smith, Applicatiom of ArtificialIntelligence for Chemical Inference XIV:The Number of Structural Isomers ofCz:xN~::yO~::Z, x + y + z = 6. AnInvestigation of Some Intuitions ofCllemists.

1191 D. H. Smith, Applications of ArtificialIiitelligrnce for Chemical Iiifererice X V ,in preparation.

I201 D. H. Smith and R. E. Carhart,Applicatiom of Art i f ic ia l In te l l igence forChemical In ference XVI : On StructuraJIsoulerisnl of Tricyclodecarles, to besubmitted to Journnl of the AmericanChemical Society.

[21] R. C. Dromey, B. G. Buchanan, D. H.Smith, J, Lederberg and C. Djerassi,Applications of Artificial Intelligence forC h e m i c a l Infererlce XVI I : A GeneralM e t h o d f o r Predictiilg Molecular 1011s inMass Spectra, submitted to Journat ofOrganic Chemistry.

Other references, relating to the MYCINsystem, under NIH support.

[22] E. H. Shortliffe, S. G. Axlme, B. G.Buchanan, T. C. Merigan, and S. N.Cohen, An Artificial IntelligenceProgram to Advise Physicians RegardingAntimicrobial Therapy, Computers andBiomedical Research, 6 ( 1973), 544-560.

[23] E. H. Shortllffe, S. G. Axline, B. G.Buchanan, S. N. Cohen, DesignCorlsideratiolls for a Program to P r o v i d eConsultatiorls irl Clillical T h e r a p e u t i c s ,Proc. Biomedical Symposium, San Diego,February 1974.

4 6

[24] E. H. Shortliffc and B. C. Buchanan, AModel of Inexact Reasoning in Medicine,submItted to Mathematical Biosciences.

[25] E. H. Shortlifie, R. Davrs, S. G. Axline, B.G. Buchanan, C. C. Green and S. N.Cohen, Computer-Based Collsultations inClinical Therapeutics: Explmatiorl andRule Acquisitioll Capabilities o f t h eMYCIN System, submitted to Computersand Biomedical Research.

H E U R I S T I C P R O G R A M M I N G P R O J E C T

47

4 . N E T W O R K P R O T O C O LD E V E L O P M E N T P R O J E C T

4.1 Internetwork Protocol Design

During this period, a design for anexperimental internetwork protocol wascompleted [ 11 and has been circulated both toIFIP WC 6.1 and to other interested ARPAresea rch ten ters. In addition, an articledescribing the basic concepts was published inMay 1974 123. An updated and more detaileddesign was prepared and circulated only to thesites participating in = ARPA sponsoredinternetworking and is now undergoingfurther revision.

The participants in the internetworkingexperiment include the University CollegeLondon uncler the direction of Prof. PeterK irstein, Bolt Beranek and Newman underthe direction of Dr. Jerry Burchficl, and theStanford Digital Systems Laboratory underthe clirection of Prof. V. Cerf. Plans werelaid to connect a TENEX system at BEN witha PDP-9 at UCLA and with a PDP-11 at SU-DSL, all running the proposed TransmissionControl Program (internetwork protocol).Concurrently an experiment was outlinedbetween the National Physical Laboratory inEngland under the direction of Dr. DonaldDavies and the ARIA research center neatParis under the direction of Mr. Louis Pouzin.In the. latter experiment, a Modula-I computerat NPL is to be connectecl to a CII 100-70 atIRIA running a protocol proposed by H.Zimmerman and M. Elie of IRIA.

In a concurrent effort, plans were made tostudy t h e p r o b l e m o f connecting theTYMNET with the ARPANET using theprotocol proposed in [ 11. During the periodof this report, only modest progress has beenmade in this effort, but enthusiasm for theproject remained high. It is expected thatmore concrete progress will be made duringthe second year.

IFIP Working Group 6.1 met in June 1973and the National Computer Conference inNew York, in September of 1973 in Sussex asthe NATO Confe r ence on ComputerNetworks, and in January 1974 at the SeventhHawaii International Conference on SystemsScience. Plans were made to meet again atIFIP 74 in August 1974. WC 6.1 wasreorganized into four subcommittees to makeworking together easier:

Committee ChairniariExperiments Prof. P. KirsteinProtocols Mr. L. PouzinLegal and Political Issues Prof. F. KuoSocial Issues Dr. C. D. Shepard

In another step to make W.G. 6.1 as selfsupporting as possible, and in the wake of thereduced NIC services offered by ARPA after 1July 1974, all W.G. 6.1 members were to payfor the cost of reproducing and mailing ofcommittee notes and reports. It was expectedthat this move would also shrink the size ofthe group down to those who were seriouslyinterested in the work.

An agreement was reached regarding acommon basic addressing format for bothprotocols [31 and it IS Intended that the resultsof these two experiments ~111 be used to settleon a final protocol which could be used toconnect all 5 sites.

4 8 N E T W O R K P R O T O C O L D E V E L O P M E N T P R O J E C T

4.2 PDP-I1 E x p a n s i o n

During January through March 1974, thePDP- 1 l/20 installation was expanded usingfunds from the Joint Services ElectronicsProgram sponsored jointly by the Army, Navyand Air Force. The PDP-1 1 facility nowincludes:a) PDP-I l/20 CPU with 28 K l&bit words of

memory (maximum allowed).b) 5.6 M word Diablo 44 moving head dual

platter disk. One disk is removable; eachwill hold 2.8 M words.

c) Unibus repeater to expand the number ofUnlbus slots available.

d) Four asynchronous terminal interfaces, twofor hard-wired use and two for dial upmodems. Two AndcTson- Jacobsen modemsand two Direct Access Arrangementtelephone lines also installed.

e) One OMRON microprogrammed CRTterminal with 4K byte buffer memory.

f) One card reader (not new).g) One upper case only printer (not new).h) Two Dectape drives (not new).i) One RS64 64K byte fixed head disk.j) One 1024::: 1024 CRT (not new) with SU-

DSL designed controller and two joysticks(latter two are new).

k) Three Texas Instruments Silent 700portable termrnals.

1) One 16 bit general purpose dlgttal Interfacefor experimental device attachments.

e m) One 50 Kbitlsecond modem withARPANET VDH Interface for use withthe ELF operating system (PDP-11 isconnected by VDH to SRI 1MP).

The ARPA contract pays for the rental of theModems, TI terminals, and maintenance onthe PDP-1 1 during the summer months; theElectrical Engineermg Department of StanfordUniversity pays for maintenance during therest of the academic year.

4.3 Software Systems

E L F

In January, an ELF I system was installed. Itproved to be fairly reliable although it had afew bugs left. It did not support the DiabloDisk or the dial-up facilities. Nor did it havemuch of a File Transfer Protocol (text filesfrom the net could be prmted on the lineprinter). T h e E L F s y s t e m was usedintermittently during this period for access tothe ARPANET, but owing to shared use ofthe equipment for academic projects, the ELFsystem was not up much of the time.

An attempt was made to integrate ELF withthe Disk Operating System (DOS), but thisproved impossible since DOS is configured forsingle user function and simultaneous use ofDOS with ELF caused ELF to lose contrQl ofIt critlcal interrupts. We investigated thepossibrlity of a Virtual Machine system, butthe PDP-1 l/20 does not have adequatehardware to support virtual memory orprivileged mstruction trapping needed forVirtual Machine Monitors. We concludedthat only a P D P - 1 l/40 with hardwaremodifications similar to those on the UCLAsystem would serve for such a VirtualMachine system and gave up that approach astoo costly and time consummg. Consequently,the system still alternated between DOS andELF usage.

File Transfer Protocol

During the summer of 1974, an FTP waswritten which would accept MAIL files fromthe network and print them on the lineprinter. The program was documented [4]and plans were made to extend the system tofull FTP capability.

Simple Minded File System

As an aid to the ELF user community, weproposed to implement a simple minded filesystem which would permit ELF to read or

4 . 3 S o f t w a r e Systems 49

write contiguous files on the disk. Thedetailed specification and implementation ofthis package was seriously delayed owing tolack of documentation of the new ELF IIsystem to which SMFS was to be interfaced.ELF II did not arrive durmg this period, soonly the basic SMFS design specification waswritten using DOS I/O calls as the model foluser level interface.

Bibliography

113 Cerf, V. C. and R. E. Kahn, TowardsProtocols for IlltcrnetworkCommuilicatioll /F/P W.G. 6.1 Document33, September 1973.

[2] Cerf, V. C. and R. E. !_(ahn, A Protocolfor Packet Network IIltercoilllll~lllication,IEEE Transactions on Coltlmu,2i~~ztion,Volume COM-22, No. 5 , May 1974.

[3] Pouzin, L. (Revised V. Cerf), A PacketFormat Proposal , IF/P W.G. 6.1Document 48, January 1974.

141 Haugland, T., Arl hlplelnetltatiorl of theA R P A N E T F i l e T r a n s f e r Pt~otocol forELF, Stanford University Digitnl SystemsLaboratory Technical Note 46, July 1974.

[53 Cerf, V. and C. S unshlne, Protocols atldGateways for IIltercorlllcctiorl of Packet

- Switchirlg Networks, Proceedings oj theSeventh Hawaii lnte?nationnl Conferenceon Systems Science, January 1974.

[6] Cerf, V. Au A ssessiiieiit o f ArpanetPiotocols, Proceedings of the SecondJerusalem Conference on InformationTechnology, July 1974.

51

Appendix A

A C C E S S T O D O C U M E N T A T I O N

This is a description of how to get copies ofpublications referenced III this report.

External Publications

For books, journal articles, or conferencepapers, first try a technical library. If youhave difficulty, you might try writing theauthor directly, requesting a rep-in t.Appendix D lists recent publicationsalphabetically by lead author.

Artificial Intelligence Memos

Artificial Intelligence Memos, which carry an“AIM” prefix on their number, are used toreport on research or development results ofgeneral Interest, including all dissertationspublished by the Laboratory. Appendix Blists the titles of dissertations; Appendix Egives the abstracts of recent A. I. Memos andlnstructlons f o r h o w t o obtarn copies. T h etexts of some of these reports are kept III ourdisk file and may be accessed via the ARPANetwork (see below).

Computer Sciellce Reports

Computer Science Reports carry a “STAN-CS”prefix and report research results of theComputer Science Department. (All A. I.Memos published since July 1970 also carryComputer Science numbers.) To request acopy 6f a CS report, write to:

Documentation ServicesComputer Science DepartmentStanford UniversityStanford, California 94306

The Computer Science Department publishesa monthly abstract of forthcomlng reports thatcan be requested from the above address.

Film Reports

Several films have been made on researchprojects. See Appendix C for a list of filmsand procedures for borrowing prints.

Operating Notes

Reports that carry a SAILON prefix (astrained acronym for Slanfor~ A. I. LaboratoryOperating Note) are semi-formal descriptionsof programs or equipment in our laboratorythat are thought to be primarily of internalinterest. The texts of most SAILONS areaccessible via the ARPA Network (see below),Printed copies may be requested from:

Documentation ServicesArtificial Intelligence LaboratoryStanford UniversityStanford, California 94306

Workhg Notes

Much of our working documentation is notstocked in hard copy form, but is maintainedin the computer disk file. Such texts that arein public areas may be accessed from theARPA Network (see below). Otherwise, copiesmay be requested from the address given justabove.

Public File Areas

People who have access to the ARPA Networkare welcome to access our public files. T h eareas of principal mterest and their contentsare a follows:[BIB,DOC]

[AlM,DOC][S,DOCl[UP,DOC]kLDOC1[ 11,DOCI

[P,DOCl

bibliographies of variouskinds,texts of some A. I. Memos,texts of most SAILONs,user program documentation,system hardware descriptions,PDP- 11 subsystemdescriptions,“people-oriented” files,including the lab phonedirectory.

52 A p p e n d i x A

Network Access

On the ARPA Network, our system is site 11(decimal), alias SU-AI. We use a heavrlymodified DEC monitor. It types “.” whenevelit is ready to accept a conInland. Allcommands end with <carriage return>. Tohalt a program that is running, type<Control>C twice.

It IS possible to examine the contents of publicfiles wlthout logging rn. For example, If youwish to know the names of al\ files In an areacalled [P,DOC]. Just type:

DIR [P,DOCI

The system ~111 then type the names of allsuch files, their sizes, etc.

To type out the contents of a text file, say“TYPE <file namez-“. For example, to type thecontents of our telephone directory, say

TYPE PHONE.LST[P,OOCl

and be prepared for 18 pages of output.

There may be difticulty 111 prlntlng files thatuse the full Stanford character set, whichemploys some of the ASCII control codes (1 to37 octal) to represent special charactc33.

If your tcrmlnal has both upper and lowetcase characters, let. the nionltor know by saying"TTY FULL". If you are at a typewrlter terminal,you may also wish to type "TTY FILL", w h i c h

a causes extra carriage returns to be inserted sothat the carriage has time to return to the leftmargin before the next line begins.

To get informatlon on other services that areavailable, say "HELP RRPn" or just plain "HELP".

File Transfer

Files can also be transferred to another siteusmg the File Transfer Protocol.Documentation o n o u r F T P p r o g r a m i sl o c a t e d In diskfile FTP.DCS[UP,DOC]. N opasswords or account numbers are needed toaccess our FTP from the outside.

53

Appeudis B

T H E S E S

T h e s e s t h a t have b e e n puhllshed by t h e

Stanford Artlficlal lntell~gcncc Laboratory arellsted h e r e . Several earned degrees atinstitutions other than Stanford, as noted.T h i s list is kept in our system In dlskfileTHESES[BIB,DOC].

D. RaJ. Reddy, AIM-43AII Approach to Computer SpeechRecognition by Direct Analysis of theSpeech Wave,P/L Ll. in Compster Science,September 1966. --

S. Persson, AIM-46Some Sequeucc Estrapolatiug Programs: aS t u d y o f Represeutatiou arid Modclirlg iuJnquiring Systems,P/1.1). in Com~trte7~ Science, University ofCalifornia, Berkeley,Septem her 1966.

Bruce Buchanan, AIM-47Logics of Scientific Discovery,Ph.D. in Philosophy, Unlverslty of California,Berkeley,December 1966.

lames Palnter, AIM-44Seniautic Correctuess of a Compiler for auA Igol-Ii kc La Ilguagc,Ph.D. in Computer Science,March 1967.

Wlltiam WIchman, AIM-56 Claude Cordell Green, AIM-96Use of Optical Feedback ill the Computer The Applicatioll of Theorem Proving toC o n t r o l o f au Arm, ~lestion-answering Systems,Eng. in Electrical Engineering, Ph.D. in Ekctrical Engineering,August 1967. August 1969.

Monte Callero, AIM-58An Adaptive Conlmaud aud Control SystemUtiliziiig Heuristic Learning Proccssc.5,Pk.l). in Ofwrntions Research,December 1967.

Donald I< aplan, A I M - 6 0The Fornlal Theoretic Analysis of StrongEquivalence for Elemental Properties,Ph.D. in Computer Science,July 1968.

Barbara Huberman, AIM-65A Program to Play Chess End Canles,Ph.D. in Computer Science,August 1968.

Donald Pieper, A I M - 7 2T h e Kinelnatics of Manipulators u n d e rCoinputer Coiitrol,Ph.D. in Mechanical Engineering,October 1968.

Donald Waterman,Machine Learnirtg of Heuristics,Ph.D. in Computer Science,December 1968.

A I M - 7 4

Roger Schank, AIM-83A Collceptual Dependency Representat ionfor a Computer Oriented Semantics,Ph.D. in Linguistics, University of Texas,March 1969.

Pierre Vlcens, A I M - 8 5Aspects of Speech Recognition byComputer,Ph.D. in Computer Science,March 1969.

Victor D. Scheinman, A I M - 9 2Desigll of Computer Controlled Manipulator,Eng. in Mechanical Engineering,June 1969.

James J. Horning, A I M - 9 8A Study of Granirnatical Inference,Ph.D. in Computer Science,August 1969.

54 Appelldix B

Michael E. Kahn, AIM- 106T h e Near-111iiliillIrI11-ti111e Corltrol of Open-loop Articulated Ki~enlatic Chains,Ph.Ll. in Mcihanical Engincrring,December 1969.

Joseph Becker, AIM- 119AII 111forlllatioll-~>roces.5iilg Model o fIriterniediate-Level Cognition,Ph. Ll. in Computer Science,May 1972.

Irwin Sobel, AIM-121Camera Models and Machine Perceptioll,Ph. Ll. in Electrical! Engineering,May 1970.

Michael D. Kelly, ._ AIM-130Visual identification of People by Computer,Ph.L>. in Computer Science,July 1970.

GIlbert Falk, AIM-132Coinputer lnterpretatiorl o f Inipcrfcct LineData as a Three-dilneusioual Scene,Ph. I>. in E/t-ctrictl/ Engineering,Augrlst 1970.

Jay Martln Tenenbaum, AIM-134Accolnlnodatiotl in Computer Visioll,P/~.LI. in Electrical Engineering,September 1970.

- Lynn H. Quatn, AIM- 144Computer Coniparisorl of Pictures,Ph.L>. in Computer Science,May 1971.

Robert E. Kling, AIM- 147Reasorlillg by Analogy with Applications toHeuristic Problem Solving: a Case Study,Ph.D. in Computer Science,August 197 I.

Rodney Albert Schmidt Jr., AIM- 149A Study of the Real-time Control of aCotnputer-drivel1 Vehicle,Ph.Ll. in Electriccrl Engineering,August 1971.

Jonathan Leonard Ryder, A I M - 1 5 5Heuristic Analysis of Large Trees asGellerated it1 the Game of Go,

Ph.D. in Computer Science,December 197 I.

Jean M. Cadiou, A I M - 1 6 3Recursive Definitions of Partial Furlctiorlsand their Computatiom,Ph.D. in Computer Science,April 1972.

Gerald Jacob Agin, A I M - 1 7 3Representation and Description of C,urvedObjects,Ph.D. in Computer Science,October 1972.

Francis Lockwood Morris, A I M - 1 7 4Correctness of Translations ofProgralmning Languages -- an AlgebraicApproach,Ph.D. in Computer Science,August 1972.

Richard Paul, A I M - 1 7 7Modcllitlg, Trajectory Calculation andServoirlg of a Computer Controlled Arm,Ph.D. in Computer Science,November 1972.

Aharon Gill, AIM- 178Visual Feedback aud Related Problems inComputer Controlled Harld EyeCoordinatiorl,Ph.D. in Electrical Engineering,October 1972.

Rutena Bajcsy, A I M - 1 8 0C o m p u t e r Identification of Textured VisiualScenes,Ph.D. in Computer Science,October 1972.

Ashok Chandra, A I M - 1 8 8011 the Properties and Applications ofPrograininirig Schenias,Ph.D. in Computer Science,March 1973.

T H E S E S 55

Gunnar Rutger Grape, AIM-201 James R. Low, A I M - 2 4 2Model Based (Interlnediate Level) Colnputcr Autolnatic Coding: Choice of DataVis ion, Structures,Ph.D. in Compirtu Science, Ph.D. in Computer Science,May l97?#. August 1974.

Yoram Yaklmovsky, AIM-209Scene Analysis 1Jsing a Selnalltic Base forRrgiotl Growing,Ph.D. in Computer Science,July 1973.

Jean E. Vulliemm, AIM-2 18Proof Techniques for Recursive PrograIns,Ph.P. in C:ompih+r Science,October 1973. *

Daniel C. Swinehart, -- AIM-230COPILOT: A Multiple Process Approach t oIiiteractive Prograinnkig Systeins,Ph.D. in Computer Science,May 19’74.

James Gips, AIM-23 1Shape Gramlnars alld their UsesPh.D. in Computer Science,May 1974.

Charles J. Rqer III, AIM-233Conceptual Memory: A Theory audCoinpufer Prograni for Processing theMeaning Content of Natural LanguageIJtterallces,Ph.L>. in Computer Science,June 1974.

Christopher I<. Rlesbeck, AIM-238Con1 pu tat iollal U tlderstarldillg: Analysis ofSentences alld Contest,PhII. in Computer Science,June 1974.

Marsha Jo Hannah, AIM-239Computer Matching of Areas in StereoIniages,Ph.L>. in Computer Science,July 19’74.

Appelldix C

F I L M R E P O R T S

Prints of the following films are available forshort-term loan to Interested groups withoutcharge. They may be shown only to groupsthat have paid no admlssion fee. To make areservation, write to:

Film ServicesArtrficlal Intelligence Lab.Stanford UniversityStanford, California 94305

Alternatively, prints may be purchased at cost(typically $40 to $60) from:

Cine-Chrome Labormries4075 Transport St.Palo Alto, California(415) 321-5678

This list is kept 111 diskfile FILMS[BIB,lXXl.

1. Art Eisenson and Gary Feldman, Ellis D.Kroptechev and Zeus, his MarvelousT i m e - s h a r i n g Systenl, 16mm B&W withsound, 15 mmutes, March 1967.

The advantages of time-sharing over standardbatch processing are revealed through thegood of?-ices of the Zeus time-sharing system ona PDP- 1 computer. Our hero, Ellis, IS saved-from a fate worse than death. Recommendedfor mature audiences only.

2. Gary Feldman, Butterfinger, 16mm colorwith sound, 8 minutes, March 1968.

Describes the state of the hand-eye system atthe Artificial Intelligence Project in the fall of1967. The PDP-6 computer getting visualInformation from a television camera andcontrolling an electrical-mechanrcal arm solvessimple tasks involving stacking blocks. Thetechniques of recognizing the blocks and theirposmons as well as controllmg the arm arebriefly preset1 ted. Rated “G”.

57

3. Raj Reddy, Dave Espar and Art Eisenson,Hear Here, 16mm color with sound, 15minutes, March 1969.

._Describes the state of the speech recognitionproject as of Spring, 1969. A discussion of theproblems of speech recognition is followed bytwo real time demonstrations of the currentsystem. The first shows the computer learningto recognize phrases and second shows howthe hand-eye system may be controlled byvoice commands. Commands as complicatedas ‘Pick up the small block in the lowerlefthand corner’, are recognized and the tasksare carried out by the computer controlledarm.

4. Gary Feldman and Donald Peiper, Avoid,16mm silent, color, 5 minutes, March 1969.

Reports on a computer program written by D.Peiper for his Ph.D. Thesis. The problem isto move the computer controlled electro-mechanical arm through a space filled withone or more known obstacles. The programuses heuristics for finding a safe path; the filmdemonstrates the arm as it moves throughvarious cluttered environments with fairlygood success.

5. Richard Paul and Karl Pingle, InstalltInsanity, 16mm color, silent, 6 minutes,August, 1971.

Shows the hand/eye system solving the puzzleInstant Insanity. Sequences include findingand recognizing cubes, color recognition andobject manipulation. This film was made toaccompany a paper presented at the 197 1International Joint Conference on ArtificialIntelligence in London and may be hard tounderstand without a narrator.

6. Suzanne Kandra, Motioll and Vision,16mm color, sound, 22 minutes, November1972.

A technical presentation of three researchprojects completed in 1972: advanced arm

58 A p p e n d i x C

control by R. P. Paul [AlM-1771, v isualfeedback control by A. Gill [AIM-17S], andrepresentation and descr1pt 1cJn of curvedobjects by C. Agin [AIM-1731.

‘7. Larry Ward, Coillputer IrlteractivePicture Processillg, (M AR S Project),16mm color, sound, 8 min., Fall 1972.

This film describes an automated picturedifferenclng technique f o r a n a l y z i n g t h evariable surface features on Mars using datareturned by the Mariner 9 spacecraft. Thesystem uses a time-shared, terminal orientedPDP- 10 computer. The film proceeds at abreath less pace. Don’t blink, or you will missan entire scene.

8. Richard Paul and ICtr1 Pin@, AutomatedPun~p Assembly, 16mm color, silent (runsat sound speed!), 7 mmutes, Apr11, 1973.

Shows the hand-eye system assembling asimple pump, using vision to locate the pumpbody and to check for errors. The parts area s s e m b l e d ancl screws inserred, USII-~g s o m e

special tools deslgned for the arm. Some titlesare Included to help explain the film.

9. Terry Winograd, Dialog with a robot,16mm black and white, silent, 20 minutes,(made at MIT), 1971.

a Presents a natural language dialog with asimulated robot block-manipulation system.The dialog is substantially ehe same as that inUnderstanding Natural L a n g u a g e ( T .W inograd, Academic Press, 1972). Noexplanatory or narrative material IS on thefilm.

10. Karl Pingle, Lou Paul and Bob Boiles,Automated Assembly, Three ShortExamples, 1974 (forthcoming).

59

Appendix D

E X T E R N A L PUB.LICATIONS

Articles and books by Project members thathave appeared in the last year are listed herealphabetically by lead author. Earlierpublications are given in our ten-year report[Memo AIM-2281 and in diskfileP U BS.OLD[ BIB,DCX]. The list below iskept in PUBS[BIB,DCX].

1. AgIn, Gerald J., Thomas 0. Binford,Computer Description of CurvedObjects, Proceetfings of the Thirdlntcrnationaf Joint Conftrencs on Artificiallntdligence, Stanford U_nlversity, August1973.

2. Ashcroft, Edward, Zohar Manna, AtnitPnueli, Decidable Propert its of MonadicFuncfional Schemas, J. ACM, July 1973.

3. Ba jcsy, Ruzena, Computrr Dcscriptiotl ofT e x t u r e d Scelles, Proc. Third ht. JointConf on Artificicrl Intelligence, Stanford U.,1973.

4. Brown, H., Masinter, L., H jelmeland, L.,Constructive Graph Labeling UsingDouble Cosets, Discrete Mathematics, 7,1974.

-

5. Buchanan, Bruce, N. S. Sr~clharan,Analysis of Behavior of C&cm icalMolecules: Rule Formation on NOII-

IIolnogeneous Classes of Objects ,Proceedings of the Third InternationalJoint Conference on Artificial Intelligence,Stanford University, August 1973.

6. Carhart, R., C. Djerassi, Applicntiot~s ofA r t i f i c i a l Intclligencc for ChemicnlIlrference XI: The Arlalysis of Cl3 N M RData for Structure Elucidation ofAcycl ic Amiues, J. Chem. $0~. (Perkin II),1753, 1973.

7. Cerf, Vinton G., R. E. Kahn, A Protocolf o r Irlter-network Colll~llullicatiolls,IEEE Trans. Communications, May 1974.

8. Cerf, V. C., D. D. Cowan, R. C. Mullin, R.C. Stanton, Networks alld GeneralizedMoore Graphs, PYOC. Manitoba Conf. onNumerical Math., 1973, (to appear).

9. Cerf, V. C., D. Cowan, R. C. Mullin, R. C.Stanton, Topological DesigllCollsiderations irl Computer-Conmunicatiorl Networks, in R. L .Grimsdale, F. F. Kuo (eds.), ComputerCommunication Networks, Academic BookServices Holland, Netherlands, 1974.

10. Cerf, V., C. Sunshine, Protocols andGateways for Intercoimectioil of PacketSwitcbillg Networks, Proc. 7th HawaiiInternational Conf. on System Sciences,Western Periodicals Co., Hawaii, January1974.

11. Cerf, V. G., R. E. Kahn, A Protocol forPacket Network Irltercomt~unicatio~l,IEEE Trans. Communication, Vol. COM-22, No. 5, May 1974.

12. Cerf, V. C., D. D. Cowan, R. C. Mullin,R. C. Stanton, A Partial Cerlsus ofGeneralized Moore Graphs, PYOC.

Australian National CombinitoricsConference, May 1974.

13. Cerf, V. C., An Assessment ofARPANET Protocols, Proc. JerusalemConf. on Information Technology, July 1974.

14. Chowning, John M., The Synthesis ofComplex Audio Spectra by means ofFrequency Modulatiorl, J. AudioEngineering Society, September 1973.

15. Colby, Kenneth M., Artificial Paranoia: AComputer Simul;ttion of the ParanoidMode, Pergamon Press, N.Y., 1974,

60 A p p e n d i x D

16. Colby, K.M. and Parklson, R.C. Patter-m-luatching rules for the Recoguitiou ofNatural Language Dialogue Espressious,Amcl ican Journnl of ComputationalLinguistics, 1, 1974.

17. Dobrot~n, Roris M., Victor D. Schernman,Design of a Computer ControlledManipulator for Robot Research, hoc.Third lnt. Joint Conf. on ArtificiczlIntelligence, Stanford U., 1973.

18. Enea, Horace, Kenneth Mark Colby,Idiolcct ic Lailgllage-Allalysis forIluderstaudiug Doctor-Patient Dialogues,Proceedings oj the Third InternationalJoint Conftwnce on Artificinl Intelligence,Stanford University, August 1973.-_

19. Feldman, Jerome A., James R. Low,Conllllent 011 Brent’5 Scatter StorageA l g o r i t h m , Conzm. ACM, NoVf’mtXY 197%

20. Hieronymus, J. L., N. J. M~llt~, A. L.Samuel, The Amauueusis Speechlircogrlit inn System, Pros. lEEESymposium on Speech Recognition, Aprd1974.

21 . Hleronymus, J. L., Pitch SyncllrouousAcoust ic Segmeutatioll, Proc. IEEESymposium on Speech Recognition, April1974.

22. Hilf, Ft-anklin, IJse o f C o m p u t e rAssistance iu Euhallcirlg Dialog BasedSocial Welfare, Public Health, andEducational Services iu Devrlopillg

+Couutries, Proc. 2nd Jerusalem Conj. onInfo. Tcchwx’ogy, July 1974.

23. Hueckel, Manfrcd l-l., A Local VisualO p e r a t o r w h i c h Rccoguizcs Edges audLines, J. ACM, October 1979.

24. Igarnshl, S., R. L. London, D. C.Luckham, Interactive P r o g r a mVerification: A Logical System aud itsl~llplclllerltatioll, Acta Informatlca, (toappear).

25. Katz, Shmuel, Zohar Manna, A HeuristicApproach to Program Verification,Proceedings of the Third Intet’nationalJoint Conference on Arti$ciat Intelligence,Stanford Umverslty, August 1973.

26. Luckham, David C., Automatic ProblemSolving, Proceedings of the Thirdlnttwational Joint Conference on ArtificialIntelligence, Stanford University, August1973.

27. Luckham, David C., Jack R. Buchanan,Automat ic Ceneratiou of ProgramsContaining Conditional Statements, PYOC.

AISB Summer Conference, U. Sussex, July1974.

28. Manna, Zohar, Program Sctlemas, inCurrents in the Theory of Computing (A.V. Aho, Ed.), Prentice-Hail, EnglewoodCliffs, N. J., 1973.

29. Manna, Zohar, Stephen Ness, JeanVulllemin, Inductive Methods forProvhg Properties of Programs, Comm.ACM, August 1973.

30. Manna, Zohar, Automatic Progranmhg,Proceedings of the Third InternationalJoint Conference on Arti.cial intelligence,Stanford University, August 1973.

31. Manna, Zohar, Intt’otluction toMathematicat Theory of Computation,McGraw-Hill, New York, 1974.

32. Masinter, L., N. S, Sridharan, R. Carhart,D. H. Smith, Applications of ArtificialIiitelligeiice for Cbeuiical I n f e r e n c e X I I :Exhaust ive Generation of Cyclic andAcyclic Isomers, J. Arn~. Chem. Sot., (toappear).

33. Masmter, L., N. S. Sridharan, R. Carhart,D. H. Smith, Applicatiolls of ArtificialIntelligence for Chemical IiiferericeXII I : AI) A l g o r i t h m f o r LabellirlgChemical Graphs, J. Amer. Chem. Sot., (toappear).

E X T E R N A L PIJBLICATIONS 61

34. M~ch~e, D., Bruce G. Buchanat~, CurrelltStatus of the liruristic DENJ)RftLProgram for Applying ArtificialIutelligeuce to the Interpretation of hlassSpectra, 111 R. A. C. Carrjngton (ed.),Complci~vs fog Spt~trosco~~y, Adam Hilger,London, (to appear).

35. Miller, N. J., Pitch Detcctioll by D a t aReduction, Proc. IEEE Symposium onSpeech Recognition, April 1974.

36. Moorer, James A., The Optiuluul CombMethod of Pitch Period Aualysis ofCoutiiiuous Speech, IEEE Trms.Acoustics, Speech, and Signal Processing,Vol. ASSP-22, No. 5, October 19’74.

37. Jorge J. Morales, Interactive TheoreulP r o v i n g , PYOC. ACM Nntiond C:onft*rencc,August 1973.

38. Nevatla, Ramakant, Thomas 0. Bmford,Structured Descriptions of ComplexObjects, Proceedings of the Thirdlnternntional Joint Conference on ArtijkinlIntelligence, Stanford University, August1973.

39. Quaa, Lynn, Robert Tucker, BotondEross, J- Veverka and Carl Sagan,M a r i n e r 9 Picture Differcuciug atStauford, Sky and Telescope, August 1973.

-

40. Sagan, Carl, J. Veverka, P. Fox, R.Dubrsch, R. French, P. Gierasch, L. @am,J. Lederberg, E. Levinthal, R. Tucker, B.Eross, J. Pollack, Variable Features 011

Mars II: Mariner 9 Global Results, J .C;eophys. Rts., 78, 4 163-4 196, 1973.

41. Veverka, J., Carl Sagan, Lynn Qam, R.Tucker, 6. Eross, Variable Fcaturcs 011

Mars I I I : Coulparisou o f M a r i n e r 1069arrd Marirler 19’71 Photography, Icnms, 21,3 17-368, 1974.

42. Schank, Roger C., Neil Goldman, CharlesJ. Rieger III, Chris Riesbeck, MARGIE:Memory, Analysis, Resporlse Geueratiouand Iufereuce 011 English, Proceedings ofthe Third International Joint Conferenceon Artificial Intelligence, StanfordUniversity, August 1973.

43. Schank, Roger C., Kenneth Colby (eds),Computer Models of Thought andLanguage, W. H. Freeman, San Francisco,1973.

44. Shortliffe, E. H., S. C. Axline, B. G.Buchanan, T. C. Merigan, S. N. Cohen,An Artificial Iutelligeuce P r o g r a m t oAdvise Physicians RegardingAtltimicrobial Therapy, Computers andBiomedicd Research 6, 544-560, 1973.

45. Shortliffe, E. H., S. C. Axline, B. G.Buchanan, S. N. Cohen, DesignCousultations iu Clinical T h e r a p u d i c s ,Proc. Biomedical Symposium, San Diego,February 1974.

46. Smith, D. H., B. C. Buchanan, R. S.Engelmore, H. Aldercruetz, C. D jerassi,Applications of Art i f ic ia l Intelligence f o rCheulical Iufereuce I X . Analysis o fMixtures without Prior Separation asIllustrated for Estrogens, J. AmericanChem. Sot., Vol. 95, No. 18, page 6078,1973.

4’7. Smith, D. H., B. C. Buchanan, W. C.White, E. A. Feigenbaum, J. Lederberg, C.Djerassi, Applications of ArtificiaIIiitelligeiice f o r Cheinical Iiifererice X .Intsuur. A Data Interpretation P r o g r a mas Applied to the Collected Mass Spectraof Estrogeuic Steroids, Tetrahedron, Vol.29, page 3117, 1973.

48. Smith, D. H., L. M. Maslnter, N. S.Sridharan, Heuristic DENDRAL:Analysis of Molecular Structure, Proc.NATOICNNA Advanced Study Instituteon Computer Representation and

62 A p p e n d i x D

Mnnif~tllation o f Chmicc71 Irlfclmcrtion,

John Wiley and Sons, 1974.

49. Smith, Davlcl Canfield, Horace J. Enea,Backtracking in MLISK?, Proceedings ofthe Third Intcvnntionnl joint Conft*~enceon Artificial Intelligence, StanfordUnrverslty, August 1973.

5 0 . Smith, Lelantl, Editing a n d P r i n t i n gMusic by Computer, J. hdusic Theory, Fall1973.

5 1 . Sobel, Irwin, On Calibrating C o m p u t e rControlled Cameras for Perccivirlg 3-DScelles, Proc. Third lnt. Joint Conf. onArtificial Intt*lligencc, Stanford U., 1973;also In Artlficlal Intelligence J., Vol. 5, No.2, Summer 1974.

52. Srldharan, N., Search Strnlqics for theTask of Organic Chcnlicnl Synthctis,Proctqetiings of the Third lnttv nntionnfJoint C.‘onft*jt+nce on Artificicll Intdligence,Stanforcl Unlverslty, August 1973.

53. Tesler, Lawrence G., Horace J. Enea,David C. Smith, The LISP70 PatternMatching System, Proceedings of theThird International Joint Conference onArtificial Intelligence, Stanford Unlverslty,August 1973.

- 54. Walks, Yorlck, The Stallford MachineTranslatioll aud IJllderstaudiug P r o j e c t ,In Rustin (ea.) Natural LanguageProcessing, New York, 1973.

55. . Walks, Yorlck, Ilnderstauding WithoutProofs, Proctdings of the ThirdInternational Joint Confe) enie on ArtificialIntelligence, Stanford University, August197%

56. Walks, Yorlck, Annette Herskovlts, An

I n t e l l i g e n t Analyser and Ceuerntor o fNatura l Lallguage, Proc. lnt. Conf. onComputationa/ Linguistics, Piss, Italy,Proceedings of the Third Internation Joint

Conference on Artificia/ Intelligence,Stanford LJniversity, August 1973.

57. Walks, Yorlck, The Computer Analysisof Philosophical Arguments, CIRPHO,Vol. 1, No. 1, September 1973

58. Walks, Yorlck, An Artificial IntelligenceApproach to Machine Translation, inSchank and Colby (eds.), Computer Modelsof Thoug-ht and Language, W. H. Freeman,San Francisco, 1973.

59. Wilks, Yorick, One Small Head -- Modelsand Theories in Linguistics, Foundationsof Language, Vol. 10, No. 1, January 1974.

60. Wilks, Yorick, Preference Semantics, E.Keenan (ed.), Pm. 1973 Colloquium onFormnl Semantics of Natural Language,Cambrldge, U.K., 1974.

61. Walks, Y. Semantic Procedures andIilforniatioii, in Studies in theFoundations of Communication, R. Posner(ed.), Springer, Berlin, forthcoming.

62. Winograd, Terry, A Process Model ofLanguage Understandillg, in &hank andColby (eds.), Computer Models of Thoughtand Language, W. H. Freeman, SanFrancisco, 1973.

63. Winograd, Terry, The Processes ofLanguage Ullderstandillg in Benthall,(ed.), The Limits of Human Nature, AllenLane, London, 1973.

64. Wmograd, Terry, Lallguage and theNature of Intell igence, In C.J. Da lenoor t(ed.), Process Mode/s for Psychology,Rotterdam Univ. Press, 1973

65. Wmograd, Terry, Breaking theComplexity Barrier (again), Proc.SIGPLAN-SIG/R Interface Meeting, 1973.

E X T E R N A L PLJ~LICATIONS

66. Wlnograd, Terry, Artificial Intelligence-- When Will Computers IJr~dr~*standPeople?, Psychology Today, May 1974.

67. Winograd, Terry, Parsing NaturalLanguage via Recursive Trarlsitiorl Net,in Yeh (ed.) Applied Computation Theory,Prentice-Hall, 1974.

68. Yakirnovsky, Yom-n, Jerome A. Feldman,A Sewantics-Based Decisioll TheoreticRegion Analyzer, Proccdin~s of the TlrirtfInternational joint Conference on ArtificialIntelligence, Stanford Umversity, August1973.

63

65

Appeljdix E

A. I. MEMO ABSTRACTS

A b s t r a c t s a r e g i v e n h e r e f o r Art&la1Intelligence Memos that we have published Inthe last year. For earlier years, see our ten-year report [Memo AIM-2281 or diskfileAI MS.OLD[ BI B,DOC]. The abstracts beloware kept in diskfile AIMS[BIB,DOC] and thetitles of both earlier and more recent A. I.Memos are in AIM LST[ BIB,DOC].

In the listing below, there are up to threenumbers given for each report: an “AIM”number on the left, a “CS” (Computer Science)number in the mlddle, .-and a NTIS stocknumber (often beginning “AD...“) on the right.Special symbols preceding the “AIM” numbe;indicate availabrlrty at this writing, as follows:

+ hard copy or microfiche,03 microfiche only,::: out-of-stock.

If there IS no special symbol, then it isavailable In hard copy only. Reports that areout-of-stock are likely to stay that way becauseof peculiat governmental contractualrequirements. Reports that are in stock maybe requested from:

Documentation ServicesArtificial Intelligence LaboratoryStanford University

- Stanford, Callfornla 94305

Alternatively, reports may be ordered (for anominal fee) in either hard copy or microfichefrom:

Natlonal TechnIcal Information ServiceP. 0. Box 1553SprIngfield, Virginia 22 15 1

If there is no NTIS number given, then theymay o r may not have the report. Inrequesting copies 111 this case, give them boththe “AIM-” ancl “CS-nnn” numbers, with thelatter enlarged into the form “STAN-CS-yy-nnn”, where “yy” IS the last two digits of theyear of publication.

Memos that are also Ph.D. theses are somarked below and may be ordered from:

University Microfilm_ P. 0. Box 1346

Ann Arbor, Michigan 48106

For people with access to the ARPA Network,the texts of some A. I. Memos are storedonline in the Stanford A. I. Laboratory diskfile. These are designated below by “Diskfile:<file name>” appearing in the header.

r:c AIM-21 1 CS-383Yorick Wilks,Natural Language Illfereuce,24 pages, September 1973.

AD769673

The paper describes the way in which aPreference Semantics system for naturallanguage analysis and generation tackles adifficult class of anaphoric inference problems(finding th correct referent for an Englishpronoun in context): those requiring eitheranalytic (conceptual) knowledge of a complexsort, or requiring weak inductive knowledge ofthe course of events in the real world. Themethod employed converts all availableknowledge to a canonical template form andendeavors to create chains of non-deductiveinferences from the unknowns to the possiblereferents. Its method of selecting amongpossible chains of inferences is consistent withthe overall principle of ‘semantic preference’used to set up the or ig inal meaningrepresentation, o f which t h e s e a n a p h o r i cinference procedures are a manipulation.

AIM-Z I2 cs-3s4 AD769379Annette Herskovlts,T h e Generat ion of French from a SemanticRepreserltation,20 pages, September 1973.

The report contains first a brief description ofPreference Semant ics , a system ofrepresentation and analysis of the meaningstructure of natural language. The analysisalgorithm which transforms phrases intosemantic items called tenplates has been

66 Appelldix E

consIdered 11~ rletnll e l s ewhere , so this reportconcentrates on the second phase of analysis,which hinds templates togethct- ltito a hIghet*l e v e l semantic block corresponding t o a nEnglrsh pat.agrnph, and which, in operation,llltel~lucI,s Wlth the French genera tlonprocedure. Durrng this p h a s e , t h e semanticrelntions between templates are extracted,pronouns are referred and those worddisamt~i~uations are done that requrre t h econtext CJf a whole paragraph. These tasks

reqult-e Items c a l l e d parnplatt-s which a r eattached to keywords such as preposltlons,Sll b JUllCt IOIlS a ncl relative prctnouns. Thesystem chooses t h e represcntr4tlon w h i c hm a x i m i z e s a ca re fu l ly de f ined ‘semanticdensity’.

A system f o r t h e @neratlon o f F r e n c hsentences IS clescrtbed, based on the f~cnerntlollo f F rench sentt:nce.s IS descrhecl, b a s e d on the

recurslvc evaluation of procedural generatronpatterns called sr~r~ot~pcs. The stereotypes aresemantically context sensitive, are attached toeach sense of English words and keywordsand are carried into the rcprescntatlon by theanalysis procedure. The reprcsentatlon of themeaning of words, and the versatlllty of thestereotype format, allow for fine meaningdlstlnctlons to appear in the French, and forthe constructlon of French dlffet-rng radicallyfrom the English orlgln.

_ AIM-213 cs-385Ravlncira B. Thosnr,Rewgnit iotl of Colltilluous Speech:Scgnicntation arid Cla.5sificatinn usingS i g n a t u r e Table Adaptntion,37 -pages, September 19711.

Thus tT[lOt~t C!x’~~~~J~‘f?s the }~osskJtilty O f LlSlll~ 3

set of features for segmentation andrecognition o f COlltlllllOllS speech. Thefeatures are not necessarily tlistinclive 01’mlnlmal, in the sense that they do not dividethe phonemes Into mutually exclusive subsets,and can have high redundancy. Thus conceptof feature c a n thus avoid aprtorl b indingbetween the phoneme categories to be

recogmzed and theparticular system.

set of features defined in a

An adaptive technique is used to find theprobabllicy of the presence of a feature. Eachfeature JS treated independently of otherfeatures. An unknown utterance IS t h u srepresented by a feature graph with associatedprobabilities. It is hoped that such arepresentation would be valuable for ahypofhesize-test paradigm as opposed to a onewhich operates on a linear symbolic Input.

AIM-2 I4 CS-3S6Walter A. Perkins, Thomas 0. Binford,A Corner Firlder for Visual Feedback,59 pages, September 1973.

In visual-feedback work often a model of anobject and its approxmiate location are knowna n d It IS on ly neces sa ry t o determme itslocation and orientation more accurately. Thepurpose of the program described herein is toprovide such Information for the case inwhich the model is an edge or corner. Givena model of a line or a corner with two or threeedges, the program searches a TV window ofarbitrary size looking for one or all cornerswhich match the model. A model-drivenprogram directs the search. It calls on anotherprogram to find all lines inside the window.Then it looks at these lines and eliminateslines which cannot match any of the modellines. It next calls on a program to formvertices and then checks for a match ingvertex. If this simple procedure fails, themodel-driver has two backup procedures.First it works with the lmes that it has andtries to form a matching vertex (corner). Ifthis fails, it matches parts of the model withvertices and lines that are present and thentakes a careful look in a small region in whichit expects to find a missrng line. The programoften finds weak contrast edges in this manner.Lines are found by a global method after theentire window has been scanned with theHueckel edge operator.

A . I . M E M O AIlSTRACTS 67

::: AIM-215 CS-387 AD 769380Bruce G. Buchanan, N. S. Srldharan,A nalysis of I3ehavior of Chew ical Molecules:R u l e F o r m a t ion 011 Non-homogeneousClasses of Objects,15 pages, September 1973.

An information processing model of someimportant aspects of Inductive reasoning ispresented withln the context of one scientificdisclpllne. Clven a collection of experimental(mass spectrometry) data from several chemicalmolecules the computer program describedhere separates the molecules into zl~ll-behaz&subclasses and selects from the space of allexplanatory processes the chilrnitt-ristic

processes for each subclass. The clefinrtlons ofwell-b~Anvt~l and characlcristic embody severalheuristics which are discussed. Some resultsof the program are discussed which have beenuseful to chemists and which lend credlbllityto this approach.

::: AIM-2 16 CS- 389Larry Masmter, N.S. Sridharan, J. Lederberg,S. H. Smith,Appl icat ions of Art i f ic ia l lntelligencc forC h e m i c a l Infercllcc: XII. EshaustiveGeneratio of Cyclic and Acyclic Isomers,60 pages, September 1973.

A systematic method of identlficatlon of allpossible graph isomers consistent with a given‘empirical formula is described. The method,embodied in a computer program, generates acomplete list of Isomers. Duplrcate structuresare avolded prospectively.

::: AIM-2 17 cs-39 1 AD770610N. S. Srldharan,Search Strategies for the Task of OrgallicChein ical Synthesis,32 pages, August 1973.

A computer program has been written thatsuccessfully discovers syntheses for complexorganic chemical molecules. The definition ofthe search space and strategies for heuristicsearch are described in this paper.

AIM-2 18 cs-393Jean Etienne Vuillemin,Proof Techniques for Recursive Programs,Thesis: Ph.D. in Computer Science,97 pages, October 1973.

The concept of least fixed-point of acontinuous function can be considered as theunifying thread of this dissertation. T h econnections between fixed-points and recursiveprograms are detailed in Chapter 2, providingsome insights on practical implementations ofrecursion. There are two usualcharacterizations of the least fixed-point of acontinuous f u n c t i o n . T o the firstcharacterization, due to Knaster and Tarski,corresonds a class of proof techniques forprograms, as described in Chapter 3. Theother characterization of least fixed points,better known as Kleene’s first recursiontheorem, is discussed in Chapter IV. It hasthe advantage of being effective and it leadsto a wider class of prrof techniques.

::: AIM-219 cs-394 AD769674C. A. R. Hoare,Parallel Progranmillg: an AxiomaticApproach,33 pages, October 1973.

This paper develops some ideas expounded in[l]. It distinguishes a number of ways ofusing parallelism, including disjoint processes,competition, cooperation, communication and“colluding”. In each case an axiomatic proofrule IS given. Some light is thrown on trapsor ON conditions. Warning: the programstructuring methods described here are notsuitable for the construction of operatingsystems.

AIM-220 CS-396Robert Belles, Richard Paul,The use of Sensory Feedback in aProgrammable Assembly Systems,26 pages, October 1973.

This article describes an experimental,automated assembly system which uses sensory

68 A p p e n d i x E

feedback to control an electro-mechanical armand TV camera. Visual, tactile, and forcefeedback are used to improve posrtionalinformation, guide manipulations, andperform inspectlons. The system has twophases: a p lann ing p h a s e i n which t h ecomputer IS programmed to assemble someobject, and a worliing phase in which thecomputer controls the arm and TV camera inactually performing the assembly. Theworking phase IS designed to be run on amini-computer.

The system has been used to assemble a wntelpump, conslstlng of a base, gasket, top, and SIXscrews. This example IS used to rsplaln howt h e s e n s o r y d a t a i s incorpotatt?d into t h econtrol system. A movie showing the pumpassembly is available from the StanfordArtlficlal Tntelhgence Laboratory.

AIM-221 cs-447Lulgra Alello, Mario Aiello, RichardWeyhrauch,The Seulantics o f P A S C A L in L C F ,78 pages, October 1974.

We define a semantics for the arithmetic partof PASCAL by giving It an rnterpretatlon InLCF, a language based on the typed X-calculus. Programs are represented 111 terms oftheir abstract syntax. We show sample proofs,

_ using LCF, of some general propertIes ofP A S C A L and the correctness of someparticular programs. A programimplenicntlng the McCarthy Airlinereservation system IS proved correct.

AIM-222 CS-467Mario A~cllo, Richard Weyhrauch,Checking Proofs in the Metamathematics ofFirst Order Logic,55 pages, (forthcoming).

This IS a report on some of the firstexperrments of any size carried out using then e w first order proof checker FOL. Wepresent two different first orderaxiomatlzatlons of the metamathematics of the

logic which FOL itself checks and showseveral proofs using each one. The differencebetween the axiomatizations is that one defines

‘the metamathematics in a many sorted logicthe other does not.

+ AIM-223 cs-400C. A. R. Hoare,Recursive Data Structures,32 pages, December 1973.

A D 7 7 2 5 0 9

The power and convenience of aprogramming language may be enhanced forcertain applications by permitting- datastructures to be defined by recursion. Thispaper suggests a pleasing notation by whichsuch structures can be declared and processed;it gives the axioms which specify theirproperties, and suggests an efficientimplementation method, It shows how arecursive data structure may be used torepresent another data type, for example, a set.It then discusses two ways in which significantgains m efficiency can be made by selectiveupdating of structures, and gives the relevantproof rules and hints for implementation. It isshown by examples that a certain range ofapplications can be efficiently programmed,ithout introducing the low-level concept of areference into a high-level programminglanguage.

8 AIM-224C. A. R. Hoare,

cs-403

Hints 011 Progralnnling Language Des ign ,29 pages, December 1973.

This paper (based on a keynote addresspresented at the SKACT/SIC;PLANS y m p o s i u m o n Priniipks of ProgrntnmingLanguages , Boston, October 1-3, 1973)presents the view t h a t a programmlnglanguage IS a tool which should assrst theprogrammer in the most dificult aspects of hisart, namely program design, documentation,and debugging. It discusses the objectivecriteria for evaluating a language design, andillustrates them by application to languagefeatures of both high level languages and

A . I . M E M O A B S T R A C T S 69

machIne code programming. It concludes withan annotated reading list, recommended for allintending language desrgners.

(9 AIM-225 CS-406W. A. Perkins,Memory Model For a Robot,118 pages, January 1974.

A memory model for a robot has beendesigned and tested in a siml,~le toy-blockworld for which it has shown clarity,efficiency, and generality. I n a constralnedpsuedo-English one can ask the progran~ tomanipulate objects a n d query it about thepresent, past, and possible future states of I&world. The progra ni has a goodunderstandIng o f its = world a n d givesIntelllgcnt answers in reasonably good Englrsh.Past and hypothetlcal states of the world nrehandled by changing the state the world In animaginary contest. Procedures interrogate andmodify two globnbl databases, one whichcontains the present representation of theworld and another which contains the pasthistory of events, conversations, etc. Theprogram has the ability to create, destroy, andeven resurrect objects in its world.

+ AIM-226 cs-407F.H.G. Wright II, R. E. Corin,F A I L ,$1 pages, April 1974.

This IS a reference manual for FAIL, a fast,one-pass assembler for PDP-IO and PDP-6machine language. FAIL statements, [JSelldO-

opera+tlons, macros, and conditional assemblyfeatures are described. Although FAIL usessubstantially rnOl’f? main memory thanMACRO- 10, It assembles typical programsabout five times faster. FAIL assembles theentire Stanford time-sharing operating system(two millron characters) in less than fourmrnutes of CPU time on a KA-IO processor.FAIL permits an ALGOL-style block structurewhich provides a way of localizing the usageof some symbols to certain parts of theprogram, such that the same symbol name can

be used to mean different things in differentblocks.

AIM-227 c s - 4 0 8A. J. Thomas, T. 0. Binford,hformatioil Processing Analysis of VisualPerception: A Review,? pages, forthcoming.

We suggest that recent advances in theconstruction of artificial vision systems providethe beginnings of a framework for aninformation processing analysis of humanvisual perception. We review some pertinentInvestigations which have appeared in thepsychological literature, and discuss what wethink t be some of the salient and potentiallyuseful theoretical concepts which have resultedfrom the attempts to build computer visionsystems. Finally we try to integrate these twosources of ideas to suggest some desireablestructural and behavioural concepts whichapply to both the natural and artificialsystems.

8 AIM-228 cs-409Lester Earnest (ed.),

A D 7 7 6 2 3 3

FINAL REPORT: The F i rst Teu Years ofArtificial htelligence Research at Stanford,118 pages, July 1973,

The first ten years of research in artificialintelligence and related fields at StanfordUniversity have yielded significant results incomputer vision and control of manipulators,speech recognition, heuristic programming,representation theory, mathematical theory ofcomputation, and modeling of organicchemical processes. This report summarizesthe accomplishments and providesbibliographies in each research area.

@ AIM-229 cs-4 11D.B. Anderson, T.O. Binford, A.J. Thomas,R.W. Weyhrauch, Y.A. Wilks,AFTER LEIBNIZ...: Discussions 011Phi losophy aud Artificial Intelligence,43 pages, April 1974.

70 A p p e n d i x E

This is an edIted transcript of informalconversatlolls which we have had over recentmonths, III which we looked at some of theissues which seem to arise when artificialintelligence and philosophy meet. Our aimwas to see what might be some of thefundamental principles of attempts to buildintelligent machines. T h e major t o p i c sc o v e r e d a r e t h e relatlonship o f A l a n dphilosophy and what help they might be toeach other; t h e machanlsms of naturalinference and deduction; the question of whatkind of theory of meanrng would be InvolvedIn a successful natural language understandIngprogram, and the nature of moclek III AIresea rch.

e AIM-230 cs-442Danrcl C. Swrnehart,COPILOT: A Multiple Process Approach toI n t e r a c t i v e Prograil~iniilg Systms,Thes i s : Ph.D. 171 Cornput~r Science,2 13 pages, August 1974.

The addition of multiple processing facilitiesto a language used in an interactivecomputing environment req u 1 res newtechniques. This dlssertatlon presents o n eapproach, emphasizing the characteristics ofthe Interface between the user and the system.

We have designed an experlmentnl interactive_ programmlng sys tem, COPILOT, as theconcrete vehicle for testilig and tlescrlbing ourmethods. COPILOT allows the user to create,modify, investigate, and control programswritten In an Algal-like language, which hasbeeu augmented with facllitles f o r multipleprocessrng. Alt bough COPJ LOT is compiler-based, many of our solutions could also beapplied to an Interpretive system.

Central to the design IS the use of CRTdisplays to present programs, program data,and system status. This CO~I~II~ILIO~S display ofinformation III context allows the user toretain comprehension of complex programenvironments, and to Indicate theenvironments to be affected by his commands.

COPILOT uses the multiple processingfacilities to its advantage to achieve a kind ofInteractIve control which we have termed non-preemptive. The user’s t e r m i n a l i scontmuously available for commands of anykind: program editing, variable inqui ry ,program control, etc., independent of theexecution state of the processes he iscontrolling. No process may unilaterally gainpossession of the user’s input; the user retainscontrol at all times.

Commands in COPILOT are expressed asstatements in the programming language.This single language policy adds consistency tothe system, and permits the user to constructprocedures for the execution of repetitive orcomplex command sequences. Anabbreviation facility is provided for the mostcommon terminal operations, for convenienceand speed.

We have attempted in this thesis to extend 5:t!facilities of interactive programming systemsin response to developments ;n languagedesign and in% matton display technology.The resuicant system provides an interfacewhil:h, we think, is better matched to t h einteractive needs of its user than are itspredecessors.

cs AIM-231James Cips,

cs-4 13

Shape Crmmars and their Uses,Thesis: Ph.D. in Computer Science,243 pages, August 1974.

Shape grammars are defined and their usesare investigated. Shape grammars provide ameans for the recursive specification of shapes.A shape grammar is presented that generatesa new class of reversible figures. Shapegrammars are given for some well knownmathematical curves. A simple method forconstructing shape grammars that simulateTuring machines is presented. A program hasbeen developed that uses a shape grammar tosolve a perceptual task mvolvmg the analysisand comparison of line drawings that portray

A . I . M E M O A B S T R A C T S 71

three-clImensiotlal objects of a restricted type.A formalism that uses shape grammas togenerate paintings i s defined, ItsImplementation on the computer is described,and examples of generated paintings areshown. Th-e use of shape

e AIM-232 cs-4 14Bruce C. Baumgart,CEOMED - A Gcomtric E d i t o r ,45 pages, May 1974.

CEOM ED is a system for doing 3-Dgeometric model’mg; used from a keyboard, rtIS an Interactive drawing program; used as apackage of SAIL OI LlSP accessiblesubroutlnes, it IS a graphics language. WithGEOMED, a r b i t r a r y p-olyhedra c a n b econstructed; moved about and viewed inperspective with hidden lines eliminated. Inaddltron to polyhedra; camera and imagemodels are provided so that simulatorsre.levant to computer vision, problem solving,and animation may be constructed.

e AIM-233 cs-4 I9Charles J. Rieger, III,Conceptual Memory: A Theory arIdComputer Program for Processing theMeaning Content of Natural LarlguageIJtterances,Thesk Ph.D. in Computer Science,393 pages, June 1974.

e

H u m a n s perform vast quantities ofspontaneous, subconscious computation inorder to understand even the slmplcst naturallanguage utterances. T h e computation i sprincipally meaning-based, with syntax andtradltional semantics playing insrgnlficantroles. This thesis supports this conjecture bysynthesis of a theory and computer programwhich account for many aspects of languagebehavior in humans. Tt is a theory of languageand memory.

Since the theory and program deal withlanguage in the domain of conceptualmeanmg, they are independent of language

form and of any specific language. Input tothe memory has the form of analyzedconceptual dependency graphs which representthe underlying meaning o f languageutterances. Output from the memory is also N-Ithe form of meaning graphs which have beenproduced by the active (inferential) memoryprocesses which dissect, transform, extend andrecombine the input graphs in ways which aredependent upon the meaning context in whichthey were perceived.

A memory formalism for the computer modelis first developed as a basis for examining themferential processes by which comprehensionoccurs. Then, the notion of inference space IS

presented, and sixteen classes of conceptualinference and their implementation in thecomputer model are examined, emphasizingthe contribution of each class to the totalproblem of understanding. Among the sixteeninference cl asses are: causativelresultativeinferences (those which explain and predictcause and effect relationships relative to thememory’s mode1 of the world), motivationalinferences (those which infer the probableintentions of actors), enabling inferences (thosewhich predictively fill out the circumstanceswhich were likely to have obtained at the timeof an action), action prediction inferences(those which make guesses about what aperson might be expected to do in somesituation), knowledge propagation inferences(those which predict what knowledge isavaIlable to a person, based on what thememory already knows or can infer he knows),normative Inferences (those which assess the“normality” of a given piece of information),and state duration inferences (those whichpredict the probable duration of specific statesin the world). All inferences are probabilistic,and “backup” is d e e m p h a s i z e d a s aprogrammlng tool.

The idea of points of contact of informationstructures in inference space is explored. Apoint of contact occurs when an inferred unitof meaning from one starting point within oneutterance’s meaning graph either confirms

72 Apperldix E

(matches) or contradicts an Inferred unit ofmeaning from another point wIthIt the graph,or from within the graph of another utterance.The quantity and quality of points of contactserve as the primary definition ofunderstandIng, since such points provide aneffective measure of the memory’s ability torelate and fill in information.

Interactlolls between the Inference processesa n d (I) 7uor(i sense /Tro?nolion ( h o w meaningcontest influencrs the language analyzer’schoice of lexical senses of words during theparse), and (2) the processes of reference (howmemory pointers t o t o k e n s o f real wo~~lcientltles are established) are examined. I nparticular, an important inference-referencerelaxation cycle IS Identified and solved.

The theory forms a basis for acomputatlonally effective and comprehensivet h e o r y o f language understandmg b yconceptual inference. Numerous computerexamples are included to Illustrate key points.Most issues are approached from bothpsychological and computational points ofview, and the thesis is intended to becomprehensible to people with a limitedbackground In computers and symboliccomputation.

AIM-294 cs431Kenneth Mark Colby, Roger C. Parkison, BitI

‘Faught,Pattern-Matching Rules for the Rccogrlitionof Natural Language Dialogue Expressiorls,23 pages, June 1974.

Mat-i-machine dialogues using everydayconversational English present dlfiicultproblems for computer processing of naturallanguage. Grammar-based parsers whichp e r f o r m a word-by-word, parts-of-speechanalysis are too fragile to operate satlsfnctorilyin real time interviews allowing unrestrrctedEnglish. In constructing a s i m u l a t i o n o fparanoid thought processes, we desie;ned analgorithm capable of hanclling the linguisticexpresslons used by interviewers III teletyped

diagnostic psychiatric interviews. Thealgorithm uses pattern-matching rules whichattempt to characterize the input expressionsby progressively transforming them intopatterns which match, completely or fuzzily,abstract stored patterns. The power of thisapproach lies in its ability to ignorerecognized and unrecognized words and stillgrasp the meaning of the message. Themethods utilized are general and could serveany “host” system which takes naturallanguage input.

AIM-235 cs-432Richard W. Weyhrauch, Arthur J. Thomas,FOL: A Proof Checker for First-order Logic,57 pages, forthcoming.

This manual d e s c r i b e s a machineImplementation of an extended version of thesystem of natural deduction described byPrawitz, This language, called FOL, extendsPrawitz’s formulation to a many-sorted logicallowing a partial order over sorts. FOL alsoallows deductions to be made in someintuitionistic, modal and strict-implicationloglcs. It is intended to be a vehicle for theinvestigation of the metamathamatics of first-order systems, of problems in the theory ofcomputation and of issues in representationtheory.

AIM-236 cs-433Jack R. Buchanan and David C. Luckham,On Automating the Construction ofPrograliis,65 pages, May 1974.

An experimental system for automaticallygenerating certam simple kinds of programs isdescribed. The programs constructed areexpressed in a subset of ALGOL containingassignments, function calls, conditionalstatements, while loops, and non-recursiveprocedure calls. The input is an environmentof primitive programs and programmingmethods specified in a language currently usedto define the semantics of the outputprogramming language. The system has been

A. 1. MEMO ABSTRACTS 73

used t o g e n e r a t e p r o g r a m s f o r s y m b o l i cmanipulation, robot contlul, everydayplanning, and computing arithmeticalfunctions.

A I M - 2 3 7 G-4 36Y o r i c k Walks,N a t u r a l L a n g u a g e Illlderstalldillg SystelllsWithill the AI Paradigm -- A Survey aridS o m e Coinparisoiis,26 pages, forthcommg.

The paper surveys the major pt*oJects on theunderstancllng of natural language that fallwithln what may now be ca l led the artlficlalin te l l igence paradigm for natura l languagesystems. Some space is devoted to arguutgthat the paradigm is now a reality anddi f ferent in s igni f icant respec ts f r o m t h egenerative paradigm of present day linguistics.The compar isons between systems centera r o u n d questlons of the relative perspicuity o fp r o c e d u r a l a n d static representat ions; theadvantages and disadvantages of developmgsystems over a per-rod to test their limits; andthe degree of agreement that now exists onwhat are the sorts of information that must beavailable to a system that is to understandeveryday language.

8 A I M - 2 3 8 cs-437Christopher I<. Riesbeck,Computat ional IJnderstarlding: Analysis of!$erltences aild C o n t e x t ,Thesis: Ph.D. in Computer Science,245 pages, forthcoming.

The goal o f th is thes is was to develop asystem for the computer analysis of writtennatural language texts that could also serve aa theory of human comprehensron of naturallanguage. Therefore the construction of thissystem was guided by four basic assumptionsabout natural language comprehcns~on. First,the primary goal of comprehension IS alwaysto f ind meanings as soon as possible. Othertasks, such a s discovering syntacticrelatlonshlps, a r e p e r f o r m e d o n l y w h e nessential to decisions about meaning. Second,

an attempt IS made to understand each wordas soon as it is read, to decide what it meansand how i t re la tes to the rest o f the text .Th i rd , comprehension means not onlyunderstanding what has been seen but alsopredic t ing what is l ike ly to be seen next .Fourth, the words of a text provide the cuesf o r f i n d i n g t h e i n f o r m a t i o n n e c e s s a r y f o rcomprehending that text.

+ AIM-239 cs-43sMarsha Jo Hannah,Computer Matching of Areas in StereoImages,Thesis: Ph.D. in Computer Science,99 pages, July 1974.

This dissertation describes techniques f o re f f i c i e n t l y matchmg corresponding arias of astereo pair of images. Measures of matchw h i c h a r e s u i t a b l e f o r t h i s p u r p o s e a r ed i s c u s s e d , a s are methods for pruning thesearch for a match. T h e m a t h e m a t i c snecessary to convert a set of matchings into aworkable camera model are given, along withcalculations which use this mode1 and a pairof image points to locate the correspondingscene point. Methods are included to detectsome types of unmatchable target areas in theo r i g i n a l d a t a a n d f o r d e t e c t i n g w h e n asupposed match IS inval id . Region growingtechniques are discussed for extend matchingareas Into regions of constant parallax and fordeltmlting uniform regions In an image. A l s o ,two algorithms are presented to show some ofthe ways in which these techniques can bec o m b i n e d t o p e r f o r m u s e f u l t a s k s i n t h eprocessing of stereo images.

+ AIM-240 cs-444C. Cordell Green, Richard J. Watdinger,David R. Barstow, Robert Elschlager, DouglasB. Lenat, Brian P. McCune, David E. Shaw,and Louis I. Steinberg,Progress Report 011 Program-understandingSystems,47 pages, August 1974.

This progress report covers the first year and

74 Apperldix E

one- half of work by our automaticprogrammtng research group at the StanfordArtificial Intelligence Laboratory. Majoemphasis has been placed on methods ofprogram specification, codification ofprogramming knowledge, and implementationof pIlot systems for program writing andunderstancltng. List processtng has been medas the general problem domam for this work.

using default representations for the high-levelstructures. A demonstration system has beenconstructed. Resul ts us ing thae system arepresented.

+ AIM-241 CS-446Lulgla Aiello, R~charcl W. Weyhrauch,LCFswall: ati iiiipletiietitatiott o f L C F ,45 pages, August 19’74.

This IS a report on a c o m p u t e r p r o g r a mimplementing a simphfied verston of LCF. Itis written (wtth mtnor exceptlow) entirely inpure LISP and has none of the user orientedfeatures of the implementation described byM Ilner. We attempt to represent directly incode the metamathematical notions necessaryto describe LCF. We hope that the code issimple enough and the metamathematics isclea 1 enough so that properties of thisparticular program (e.g. its correctness) caneventual ly be proved. The program I S

reproduced in full.

ti AIM-242 cs-4 52James R. Low,Auto~tta~ic Coding: Choice of DataStt~ttclrtr*e~,

- Thesis: Ph.D. in Computer Scitwsc,1 10 pages, August 1973.

A system IS descr ibed which automaticallyc hooscs representations fat high-levelinfdrmatlon structures, such as sets, sc)qitcnces,and relations for a given computer program.Rcpr~sSntations a r e plcked from a f i x e dlibrary o f low-lcvcl d a t a structurc?s tncludlngIlnkeci-lists, binary trees and hash tabh. Therepresentations are chosen by attempting tomlnlmize the predicted spacc+tlmc Intcpral o fthe user’s program execution. PredIcttons arebased upon statlstlcs of information structureuse provided directly by the user and collectedby monitoring executions of the user program

Stanford Artificial Intelligence Laboratory Memo AIM-252 July ...

Documents

Transcript of Stanford Artificial Intelligence Laboratory Memo AIM-252 July ...