stanford artificial intelligence project i · pdf filestanford artificial intelligence project...
TRANSCRIPT
i I i I I I I i
STANFORD ARTIFICIAL INTELLIGENCE PROJECT MEMOAIM-143
COMPUTER SCIENCE DEPARTMENT REPORT NO. CS-209
CD QO
CM
Q
PROJECT TECHNICAL REPORT
BY
John McCarthy, Arthur Samuel, and the Artificial Intelligence Project
Edward Feigenbaum, Joshua Lederberg anJ the Heuristic Programming Project Staff
MARCH 1971
AR PA Order No. 457
! ■
1
COMPUTER SCIENCE DEPARTMENT
STANFORD UNIVERSITY
D D CV 0"
Riproducsd bv , _ . , NATIONAL TECHNICAL INTORMATION SERVICE
Jpnn«(i«:i*. V». 2JU1
'UN Itt IST! I lüDi^EiTtnslM B
^
STANFORD ARTIFICIAL MEMO AIM-143
INTELLIÜENCE PROJECT MARCH 1971
Project Tecnnlcal Report
öy
John McCarthy» Arthur Samueli ono the Artificial Intelligence Project Eovtard Felgenbaum, Joshua Lederberg and the Heuristic Programming
Project Staff,
ABSTRACT! An overview Is presented of current research at Stanford In artificial intelligence and heuristic programming, This ■ "»rt Is largp'y the text of a proposal to the Advancao Research ejects Agenry for fiscal years 1972-3,
The research reported here was supported In part by the Advanced Research Projects Agency of the Office of the Secretary of Defense under Contract SD-183 and In part by the National institutes of Mental Health under Grant PHS MH 066-45-08,
The views and conclusions contained In this document are those of the authors and should not be Interpreted as necessarily representing the official policies» either expressed or Implied» of the Advanced Research Projects Agency or the U.S, Government,
Reproduced In *he USA. Sclertlflc and Techplca Price: full size copy»
Available from the Clearinghouse for Federal I Information» Springfield» Virginia 22151. $3,00J microfiche copy, $,95,
Table of Contents Dl3«
1, Artificial Inte I I jgence Project 1 1.1 Analysis of Algorithms ,,« , ,.,,,,...,. 3 1.2 Machine Translation .., 7 1.3 Interaction with the Physical World •.«•i< H
1.3.1 Hand-Eye ,, 11 1.3.2 Speech Recognition .............. t, 16
1.4 Heuristics 17 1.4.1 Machine Learning • 17 1.4.2 Automatic Deduction , 17
1.5 Mathematical Theory of Computation 20 1.5.1 Rocant Research ..,....,. ,..,.•• •• 20 1.5.2 Proposed Research , 21
1.6 Representation Theory ,.,.,.,...,...., ...,..,, 23 1.7 Computer Simulation of Belief Systems .,..,..,. 24 1.8 Facilities , 29
2, Heuristic Programming Project .,.,..,. 32 2.1 Introduction 32 2.2 Change of Project Name ..........,.,..:. 32 2.3 Proposed Work for New Contract Period ,, 33 2.4 Historical Synopsis ., ,. 38 2.5 Views of Others Concerning This Research .,....,. 39 2.6 Review of work of the Current Period ,. 40 2.7 Heuristic DCNORAL as Application to Chemistry:
Possible NIH Support , 41 2.8 Computer Facilities , •..«•• 42 2.9 Budgetary Note Concerning Computer T|me •ci..t.< 43 2.10 Budgetary Note Concerning Personnel t,.,, 43
3, Budget 47 3.1 Summary of Budgets for Continuation of SO-183 (FY 1972) 47 3.2 Summary of Budgets for Continuation of SO-183 (FY 1973) 48 3.3 Artificial Intel IIgence Budget .., 49 3.4 Heuristic Programming Budget .,. 51
4, Cognizant Personnel .....t ,... 52
Appendices
A, Publications of Project Members A-l
B, Theses B-l
C, niir Reports ., .,. C-l
D, Artificial Intelligence Memos ,. D-l
E, Operating Notes , E-l
I I I
1, Artificial Intelligence Project
the excer linenta I and theoretlca processes using comouters. Its
Artlflcltj Intelligence Is perceptual and Intellectual goal Is to understand these processes well perceive, understand and act In ways now on This understanding Is at present In a Nevertheless» progress In Identifying and mechanisms Is being made and the range of problems that computers can be trade to solve Is Increasing, The unders*andlng so far achieved has Important potential practical applications, The development of these applications Is worth undertaking,
study of u111 mate
enough to make a computer y possible for Humans, very preliminary state.
duplicating Intellectual
The Stanforo Artificial Intelligence Project Is concerned with both thft centra' problems of artificial Intelligence and some related subfleld» of computer science, The proposed structure of the Project Is given In Figure 1, The scopes of some continuing activities have beer modified and two new research areas have been added: Analysis of Algorithms and Machine Translation,
n Figure l. Structure of the Stanfo InteI|Igence Project
rd Artificial
-, ^ I
| Interaction I I with the j iPhyslcal Worldl I _l B |nford#reIdman Kay» Samuel
[Analysis j I of | IA Igor ithmsl
Heur1stIcs
Luckham,Samue
Mathematical I Theory of I
Computation I ::._..I Floyd, Knuth Manna, McCarthy
iMechanlcal I I TranslatIonl
I Knuth Scnank,Wi
McCarthy
I
Iks
Representation Theory
McCarthy
Models of I CognltIve I Processes I
Colby
[1
"Analysis of Algorithms" is headed by Professor Doneld Knuth and directed to an understanding of the quantitative behavior of oartlcular algorithms, The oro^artles of many algorithms that are of central Importance to comouter science are known only In a Qualitative or crudely quantitative way, Knuth and his group are emninylng analytical techniques to deepen our knowledge of this area.
The problem of machine translation will be approached anew from two directions: artificial Intelligence and linguistics. This small project witj Involve representatives of both disciplines who propose to test their Ideas Initially on a restricted formal language.
"Interaction with the Physical World" Includes continuing projects on computer vision and control, as well as speech recognition research. During Prof, Feldman's sabbatical leave, (academic year 1970-71) responsibility for Hand-Eye research has passed to Drs, Thomas Binford and Alan Kay, Work on speech recognition was curtailed with the departure of Professor Reddy but Is continuing in the area of syntax-directed recognition.
Work on Heuristics continues in the areas of machine learning and autoTatic deduction, Board games such as Checkers and Go are the prlitary test vehicles for Ideas In machine learning. Theorem- proving is the current objective of our research on automatic deduction,
John McCarthy's Representation Theory work will continue on esc I sterro loglca I problems (I.e. choosing a suitable representation for situations and the rules that describe how situations change),
Research |n Mathematical Theory of Computation Is expanding somewhat, oartly through other sources of support, A practical goal of this work Is to replace certain time-consuming and uncertain program debugging processes with formal proofs of the correctness of programs,
The work on Models of Cognitive Processes shown in Figure 1 Is an affiliated project not Included |n th|s proposal. It will be supported by the National Institutes of Mental Health under Grant MH06645-10,
Subsequent deta I i i
sections cover the proposed research In somewhat more
l! [
!j
L
'..
y y L1
i i
I I I
1,1 ANALYSIS OF ALGORITHMS (Donald Knuth)
oroblems .re usually Investigated? algorithms. Two kinds of
A. Quantitative analysis of an a is usualjy to determine the
rerrents of a given algorithm, The n an essentially
goal requ can be done expressing the algor necessari|y a forma step Is executed. analysis" (the max
aorlthm, In this case the running time and/or memory space
determination of running time 4. . '"achlno-independent manner bv thm in some machine- independent i.nguaSe language) and counting the number of times
counts t imas
Usua mum
ly these number of
performed, taken over some specified set of a 'best case analysis" (the minimum number case of
analysis» nputs),
(the average number of times u ► xu '* ,s |n fact desirable to
about the distribution of ibution of the dlstr
the
(not
< . . - each
Include a "worst case that the step can be inputs to the algorithm); of times), and a "typical for a given distribution have
h n^.r of ullT'll'TTSZ input,
out, A Cl» PP.
11
s uiu.n! I ^t,0[!.0f MoPt,m*lM algorithms, In s usua |y to fine the "best possible» algorithm of algorithms. Me set up some diflnj? Jn of reflects, as realistically as
cs of the hardware which typical example of this sort of
goal 0 lass which character ist algorithm, A the problem of
this case the In a given
"best possible" possible, the pertinent s to be associated with the
»3 nxeo, is d.^n^;^'^ n: ^iiS*^x'-
^Tniif z\ ^r^^i^oT'^Se^icT^r'"9 two1
d,f * particular computer for sol. ^AUlu^.Tf lolV^^sT' there Is usually more than one way to solve a orohiam * nCe
IHHIT. T b9 Very he,Dfu, '" deciding"hljh 0f ' "" ?«?« JhS .?508??: O^asionaily type A analyses .?o2,iJ! "?*',thmS them9•lv•8, f^ «xample,
«i ?tJ jLthe data has been ^-Wormed enough of Iteration complete the • " Job In a reasonable
It iray seem that type B we will have found the performing type A ana Is only true to a definition of "best algorIthm
analyses of several algorithms
are also incorporated the "spectral test" of Iteration untli it to let another type
amount of time.
analyses "best" al
are far superior to type gorltnm once and for a|1.
yses of all algorithms in the Iimited extenti since si 1 possible" can significantly affect
analyses; nstead of
class. But this ght changes In the
which g e
• Ilff.inatlon method for matrix Inversion uses the tn|n|mupi number of arithmetic operations, provided that whole rows are always operated on at a time» bft Stressen C4] nas recently discovered that substantially fewer operations are needed If the row restriction Is dropped,
Another problem with type B analyses Is that, even whan a simple definition of "best possible" Is postulated, the determination of an optimal alqorlthm |s exceedingly difficult. For example, the following basic problems are among those not yet completely reso;Ved:
(a) The m|n|mum number of multiplications to compute xtn g|ven x w|th n f I xep,
(b) The minimum number of arltnmetlc operations to compute a general polynomial a<n)xtn + ..i + a(l) ♦ a(0J, given x, for fixed values of the coefficients,
(c) The minimum number of steps needed to multiply two given binary n-blt numbers,
(d) The minimum number of steps needed to recognize whether a given string belongs to a given context-free language,
.e) The minimum number of steps needed to multiply two given n x n matrices, when n Is known.
(f) The m|n|mum number of comparison steps needed to sort n 1 I
e len-ents.
Asyrrp so lut po lyn bound so lut a I low answe disco the s suff i sor te bit I
totlc ion omlal s f 0 Ion t ed, r as vered Imp I | c lent d by nspec
so IutI on to (b) I s; but |n r the d o problem The evl
a functlo w|thout
fled defl ly rea I I means of t Ion or b
s ar s kn cas
esl r (f)
dene n of exh
nitl stic como y us
e kn own es ( ed is
e I n h
aust ons
si ar !s Ing
own with c), quan know n ca as n I ve of " tuat ons iden
for p in 1 (d), title n whe se (a o s| trla
best lonsi at el 1111 a
roo|ems or 2 opa and (e) s are n s Imu ) sugges mpie fo l-and-er possible
for e l, they s I |ke
(a) ratio the K far a taneo t str rm w ror s " oft xampl can b
and <f ns, for nown Jpp part, N us comp ong|y th hlch w| earch, en fa>I e, Items e sorte
), and ••almost er and 0 asymp ar isons at the 1 I eve Further to repr need n
d by
the all- lower totlc
are exact r be more, esent ot be using
m|n{a,b) s (a + b - ib-al)/2
considered, If only the number of comparisons Is considered, other Important characteristics of the sorting problem (e.g., the logical complexity of the program and of the data structures) are Ignored, Therefore although type 0 analyses are extremely Interesting, type A analyses more often pay off |n practice,
H
11
I I r i
superf Ida exchang« sorting« |s oartain form of "tr the iraxlmum of a set of the number of cVc of this algorithm | by the "reservoir coming to be a cohe
time ed Is
By any definition of of algorithmic
n.*. CCre of th8 subject» It I« It -"«y be "argued Jha?0nSr^:u?5 f!Udy has been devoted
i Comouter Sclencei th« fiai^ • nalvsl. probably He, In th. cant , " aome»nat surorlslna that so to It. It ,av -• arsuao tnat .a anoulon-t ap.ndt«
"P toT'ilnt'"-" n""r 5e, anyth analyzing Is quite person to are many
a e
are .no.rsto.oo i, coB
ybUd;:;'s:;.t
nc^?i-'-.dT;:n';..ii?»j|tjj.nju.t >,«.
RE.FERE.WCES
CU 0°T.Ü:iU,&^' 4rt 0f COmDÜ,'r Pf<'-"'"«. VOM U^laon- t2] ''X'.h*^^' Art " C~' Progr.TOl„s, V0|i2 Udd|son.
133 '^/'irVuSbV;}1 :;^:;reoh;ä"j;o"on,tn' ""^"'»t,on Unoar algabralc lyillm. o. 'I^*!?"' '°r th« "I""»" of
Daoartmant, Juni 14, 19«5>, «Stonford: Computar Sclanea
C4: "'Jr.Ä'p";."^.8^? •I'"-«'«" 's no» opt,«.,., Num, Hath
C5] Nlklaus Wtrthi "Closing wopd", Tenth Anniversary ALGOL Colloqulm» Zurich, May 31, 1968 CALCiOL Bulletin 20, p, 16],
Ü
L
D D 0
Li
11
I.
L L
■Mb-: 'il,irH^'TriTiifTrii«ri-iriii-i'T''"T'-
I
j
I
IS
a
1,2 M«chan|cal Translation (J, McCarthy, R, SchanK, Y, Wl|ks)
Me plan to undertake a small effort In mechanical translation, The first effort w||l be to create programs to translate from a festrjoted formal language RFL Into both tngllsh and French, The Idea Is that RFL will be used to express the semantic content of the sentences Independent of grammar and without the syntactic and semantic ambiguities, There are two views of semantic content of natural language he|d In whe project and elsewhere* and both will be explored» probably to the extent of making two translators,
The first view (the linguists In the project are betting on this one) Is that semantic content (at least to the extent necessary for translation) can adeauatety be expressed by some form of non-|oglca| representation» such as a network structure,
The second view (held by AI people like McCarthy and Sandewal i) Is that the semantics of natural language w|l| have to be developed along lines similar to those taken In mathematical logic, lie. the notion of denotation for phrases end sentences of natural language will have to be formalized. From this point of view, the first cut at RFL shoula be based on the predicate calculus, and a major effort should go Into dev|s|ng predicates that will enable the content of a wide class of sentences of natural languages to be expressed.
Froir. the linguistic end, Yorlck wjlks and Roger Schänk wl|| lead the work aided 1/4 time by Dr, A,F. Harker-Rho'des and with Or. Margaret Masterman as a consultant, From the Al direction, McCarthy and perhaps Patrick Hayes w||i take part. Several research assistants w| II also be Invo h'ed,
One rrajor reason for reopening th» mechanical translation problem Is that considerable advances have been m-ide towards the satisfactory semantic representation of natural language material, Another Is that there has been an Increase in the general level of sophistication about the delicacy with which real natural language forms must be treated.
HT went through a certa was the word-for-word tra When that fallad, Iingu Their emphasis on syntact fortral methods In MT, When that approach fal something called semant sure what that was. Work those who thought that those who felt that the I understanding and generat MT could be expanded) and approaches that they fa that the goal was Just ov
n number nslatlon ists were Ic struct most of led, it ics ough
in MT great of an must
those wh tled to s er the ne
ars a
ack ion
of c| stage, cat le
ure ea them becam
t to b began dea adeou
be dea o ware ee tro xt mou
early de at th
d on to used a based on e Incre e added, to falI
more lex ate theo It with so heav
ubles be ntain.
fI nab I e inc remedy shift the w
aslngi but n Into
Icogra ry of before lly I cause
e phases, eptlon o the situ towards
ork of Ch y clear o one was three g
Phy was n human la more eff
nto thel of their
Thera f MT. at Ion,
more omsky,
that Quite
roups: eeded; nguafle ort on r own bei|ef
»■.. — .-
The ALPAC Reeort c«used most governmant flnancino to be withdrawn frotr MT, largely because |t was felt that MT was too expansive as human translators could do the Job more cheaply. Since the ALPAC Report C1966], researchers have made considerable strides In the develoorrent of a theory of natural language understanding (for discussion of this, see J. M»y [1970]), The Impetus for these has been provM.d by the growth of time-shared computer systems that oertrlt on-line dialogues between man and machine, The desire to use nature language as the medium pf c0mmunlcati0n In these Dr0Jeots has necessitated the development of a sufficiently r|ch forma! structure that, can represent the conceptual content of each natural language sentence typed Into the machine, Most Importantly, these formal structures have been built to handle linguistic Input at a higher level than that of sentences, and In conjunction with memory of eerner Input, and the long-term memory of the computer model. One aim of these Interlingual formal models has been that Inferences, logical operations and Implications may work In conjunction with the analyzed content of an utterance so as to establish the intent of the utterance and Its affect on the memory, It was found by a number of researchers that formal Interlingual structures could be made to direct language analysis so as to eliminate previous reliance on syntactic analysis, and replace It with a more heuristic approach to sentence structure.
We propose using some of the approaches that have recently been developed for dealing with computational linguistic and formal IIngulstlc and formal logical problems |n order to enrich the emerging theory of huma.i language understanding and generation, and to aooly these theoretical and practical advances tp the problem of translating languages by computer. For example, we would combine the approaches of the following: (1) John McCarthy's suggestions for the construction of a log|c-based intermediate language between source language (S) and target language m, <2) Schank's system of representing semantic structure based on an Interlingual vocabulary and a system of dependency relations, (3) the EnglIsh^French MT work of Shi | lan and Rüther ord based on a theory of phrase-for«phrase translation, (4) wilks' system of semantic structure representation using a mlxtura of dependency and phrase-structure rules ranging over semantic objects, A number of para'lels a-.d contrasts between these approacl.es could be explored, but what Is most Important is that they all share certain assumptions, namely that conventional grammatical analysis Is not fundamental, whereas some Intermediate and Interlingual representation between S and T Is, so then It is proposed to make a new attack on the MT problem. Using an intarllngua to detect and transform semantic content, thus the approach w||| be, from the start, in principle Interlingual though the Inltla' system set up w|i| be oetween two languages only- English and French, ««BHBS wmy
Int^rlinguas for MT can be of two kinds:
L *
I.
I j!
i
I a
D H
Wllks' work has triad to connection with an jnterllnfl patterns that can be transf work has bean on the semantl has Imposed semantic patte coded text, Its most genera bigger patterns In language» lengtht that are genuinely Interlingual Uy at that la obvious failures of intt'lin where the actual linguist towards either the S or T la
tackle another question that come ua: namely, how large are the erred from one language to another c discourse structure of paragrap rns, or templates, onto the segmen I assumption has been that It applying over the longest avallab InterlInguajJ and, moreover
vel can to some extent compensate fluallty at the bottom or coding Ic codings used are Inevitably nguage,
s up In content , This hs and ted and Is the je text » that for the l«ve|, biased
at the cLhHd« . p0S8,b|! t0 co'nb'n« **• approaches developed ax t^e Cambridge Language Research Unit with those developed at Stanford a! as to Integrate a system that utilizes semantic n.tS"^^ Je^«.'?
I
the Interlingual content of a pjece of dlscours«, Tht yosslblllty of utilizing networKs by conjunction with certain associative criteria and properties of a general memory structure to establish the Tearing ot the speaker as opposed to the meaning of the sentence Is currently being Investigated by Schänk. Presumably the entire procedure can be combined with certain high level logical operations In order to create a Mna! representation that could serve as the starting point of a generation routine for natural language. A major goal of th|s project, thereforei for those mathematical- Interllngua oriented» but who have become practically wary. Is to develop Interlingual systems that are formally and ajgorIthmlcal ly Interesting, yet which can have natural language dictionaries made for tnem In a simple and straightforward manner. The stress should not be on contrast and comparison between the constituent approaches to this proposed project, but or. the degree to which the different approaches complement one another, and supply elements missing In the n others,
REFERLNCES
Cl] ALPAC Report, "Language and Machines", U.S. Government Printing Office, Washington D.C., 1966,
C2] M, Masterman, "Semantic Interiinguas for Message Detection", Final report on ONR Contract, Cambridge Language Research n Unit, 1970. H
Ca"» J, McCarthy and P. Hayes, «Some Philosophical Problems from the Standpoint of Artificial Inte11|gence". In Machine Intelii- gence 4, Edinburgh. 1969, (I
[43 K, Montague, "English as a Formal Language»', Llnguaggl ne|ia Socleta e neila Technica, Milano, 1970,
C5] R, Schänk, L,Tester, and S, Weber»"ConceptuaI Case-Based Natural Language Anaiysls", Stanford Artificial Intel'Igence Project Memo AIM-1SW, 1970, U
C6] K, Simmons, "Natural Language Question Answering Systems", University of Texas, 1969, [J
C7] Y, W|lks,M0n-L|ne Semantic Analysis of English Texts",Journal of Mechanical Translation and Computational Linguistics, December 1968, L-
C8] J, Mey, "Towards a Theory of Computational LIngul8tics"i paper given at 1970 ACL, Ohio State, L.
10
i.
i I I I
SI
1,3 INTERACTION WITH THE PHYSICAL WORLD
Computar vision In the three-dimensional world and manipulation with mechanical arms continues to be a central Interest of the project. Speech recognition reseach continues at a modest level,
1,3,1 Hand-Eye Research (Thomas Blnford, Jerome Feldmani Alan Kay).
In our research we aim to develop the ability to see and manipulate In simple Industrial situations. These simple problems are beyond our abilities now, but we expect modest gains In the next two years which will stimulate use of newly available nardware, Mechanical arms now have a pUpe In Industry; touch sensing and computer control have potential advantages In extending versatility, Me anticipate the time when It wl|l be simpler and cheaper to use a gereral purpose device than to make special purpose machines» Just as computers have replaced many special purpose control devices.
Visual perception solutions have applications 'n advancing the capabilities of computer systems, by making easier communication with computers» and as visualIzat! n tools in problem solvln- and natural language.
Arm and eye modules are sufficiently standard that a handful of people use them routinely. The Hand/Eye submorltor Cl] coordinates about 7 cooperating Jobs (there wjl| be 3 or 4 more soon) which comirunicate by message procedures and a global model In the common upper segment, An ALGOL style language SAIL« implemented by Robert Sproull and Dan Swinehart C2»3] is In general use by the Hand/Eye group and others at the project,
A user package for the arm has made it widely available« and generalized manipulation procedures and Improved solutions have stltrulated extensive us» of the arm, A new arm has been constructed and ooerated [4], a oynamic trajectory servo C5] using a Newtonian mechanics model has operated the new arm, Non-iinear calibration« a user package for the new arm« and obstacle avoidance will be done in the first half of 1971,
The visual system is distinguished by automatic sensor accommodation as an integral part of the recognition process C5], An edge follower has been written whfch uses automatic accommodation to enhance its dynamic range and selectivity [73, Camera calibration £6] converts pan and t|lt of the camera to space angles and positions on the support plane, Color Identification routines have been written by Tenenbaum C6] and B|nford, The SIMPLE body recognizer C9] Identlfle? isolated objects from their outlines, The COMPLEX body rioegnizer C9j Is designed for Incomplete edge information and incorporates prediction and verification teohnioues for missing edges, G, Grape £10] Is developing an improved program for organization cf line drawings from raw edge data« using edge prediction and verification.
11
In order to develop and test the system and to standardize the Indeoendent modules, the task of "Instarv. Insanity", a puzzle Involving colored cubes, was successfully carried out, The strategy Job was the work of Bob Sproull, Jerry Feldman and Alan Kay,
We propose to work toward understanding of mors complex objects In more cotrplex scenes, such as tools on a workbench or outdoor scenes. Many scenes w||| be too complex, Over the course of two years we will use new visual functions, organize the scene In new levels, and use a repesentalon of complex objects; we will Incorporate these modules In a goaI-or I anted system which coordinates them.
We will make color permits a brightness In uni form co|or prltritlve color Perception, What Is needed Is understanding of color perception, color constancy, sight,
a region finder which uses color, The new function new level of organization! regions of homogeneous three colors will be grouped Into supar-reglons» of
, We w||| benefit from the use of color even with a a fundamental We have none In
We will organize color super- regions Into higher super proximity In space and color. Color super-regions are connectivity: the set of points wltn a nearest neighbor same property. In very simple scenes of objects with unlf reglon-orlented or edge-oriented processors are adequate, look out a window, we see sky, trees and grass. In none areas will regions based on cont|ngu|ty enable us to de area as a unit. We see patches of fresh green among brow old grass. Patches of sky show through the trees, Pate In the sky are separated by clouds, If we group together super-regions Into higher super-regions based on or position and color, I.e., we group regions which are nea connected, we describe that outdoor scene In terms of t parts: clouds, sky, grass, earth, trees, some of which o have added another level of organization that In some case sense of what appeared a jumble,
•regions by defined by having the
orm faces, But when we
of these scribe that n clay or hes of blue the oofor
oxlmlty in rby but not
few main verlap. We s e&n make
We will criticize the super-regions according to how they simplify the scene, We w||| find relations within the super-regions and relations among the super-regions. We require many levels of organization. Grouping based on some similarity) sub-grouping based on differences. Group Into super-regions based on proximity In space and some attribute (or vector of attributes).
We will form super-regions based on proxlm'ty In space and shape, size and directionality, We suggest that we can organize simple textures. The natural language description of texture seems a reasonable representation: the set of texture elements and the geonre^ric rejatlons among them, As texture e|ements we can work w|th lines and blobs (whose shape we can describe, perhaps In a primitive way r>ow), we seek to Isolate repeated elements, Relations among repeated elements w|I | be examined. This w|I| be a multi-level
12
I I I
orocess n which the finding of some relations win help !n establ shlng others. Dealing with profiles of linear textures, a orlJTltlve program with that structure was able to describe tha textures CWolfe and Blnford], There are three basic operations Involved: finding spatial features, two-dlmenslinal shape description, and organization of features, Me can do each operation well enough on simple cases to make progress by combining them,
If we ware to orga the cost would be distances from an factor the problem (search for a mate I Ike eo tors for prohibitive, A te In n-spaee, This c and storage by the the example of f our-dlircns|ona I dimensions), In h|gh-dlirens|ona| reg Ion-orlented st Is adequately desc for then there are proximity among organization costs handled before,
We
n|ze by proxlml prohibitive, element to a|I Into a two-d
h among regions regions nearb
chnlcal aid to an be Implement technique or m
color super-re soace (two using other
spaces. This rücture In that rlbed by region few regions a the regions
something are those
ty usln These other Imenslo nearby
y In organlz ed at r ultl- e glons, color attr|bu Is a the co
-srowln nd | It , Tne
9 usual "clustering" methods, methods rely on computing
elements, Even If we were to nal problem Plus a search In space, or search among
space), the cost would be atlon is the use of proximity easonable cost In computation ntry coding CBInford3, In we Imp! led proxlrjty In a
dimensions and U;o spatial tes we work with Jlmllar reasonable extension of the
st is sllgnt when the scene g or edge-finding techniques, t|e effort Is involved In
cases where higher level cases which could not be
warn against the Illusion that the visual problem wl|| be solved by one technique. For each new facility we will have "counter- examples , problems It cannot overcome. But we greatly Increase the set of problems which are routine for the system which includes that fee MIty,
We will use stereo correlation to obtain motion parallax and
Irf!^wnd/th?l:k?r0und !KP!raJ,0n ,n1stereo "«ages. As emphasized by
McCarthy, this Is a method of overall characterization of the scene, separating the soene jnto potentially significant areas. There ere some simple cases:
(a) with camera motion or small angle stereo, a|| disparities are smaI I»
(b) the distant parts of the scene have smell dlsparltyi (c) disparity of ground or floor can be approximately
predicted, traced by continuity,
14
I
-^Ji^-jM-^f-^—. ■~- :-„:1,'^,^, .. -.-.■jr,|-„ ,■- , ■■ -■■-;- ^-"
Color and other properties (for examoje, the "eneray". a mjasure of contrast) narrow the range of search, A good starting place for many objects Is the assumption that t'iey rest on the groundi
We Mill use depth Information In the shape representation. Stereo correlation provides a depth mapping, we can use the shape reore- sentatlon to build up a mode) of the object. We w||| also us* the representation to organize data from a direct-ranging experiment by tr langulat ion,
We HIM use visual feedback In a variety of ways to control the arm, Immeaiate|y, we w|i| control stacking o|ocKs and putting a square pin through a square hole,
As our visual facilities become strongeri we will use visual feedback tn tracking the handi screwing a bo|t Into a nut* and picking up blob- type objects,
RLFERE.NCES
Cl] R. Sproull and K,K, Plngle, |n preparation,
C2] 0, Swlnehart and R. Sproull, SAIL» Operating Note No, 57,1* Stanford Artificial Intelligence Project. (Soon to be superceded by an updated version),
C3] J, Feldman and P,D. Rovner* "An Algol-based Associative Language"* Cofrm, ACM, August 1969,
[4] V, Schelnman, "Design of a Computer Controlled Manlou|atorMi AIM-92, Stanford Artificial Intelligence Project.
C5] R. Paul, in preparation,
[63 J.M. Tenenbaum, "Accommodation In Computer Vision", Thesis Electrical Engineering, Stanford University* November 1970,
[7] J,M, Tenenbaum and K.K, Ping|e, |n preparation,
y C8D I. Sobel, "Camera Models and Machine Perception", AIM-I2l| Stanford Artificial Intelligence Project,
C9] G, Falk, ''Computer Interpretation of Imperfect Line Data as a Three Dimensional Scene", Thesis* Computer Science Department, Stanford University, July 1970, ,.
C10] G. Grape, unpublished, *-
Cll] Wolfe, P.O., Binford, T.u,, "Some Methods for Analysis pnd I* Description of [Jnear Visual Texture", (unpublished), y
1121 sÄolnt if\r^Vt9\ ?•; *S?7? Ph|losophlcal Problems from th« (0 SlJh?! 1/ * JJ?'!' '"r*1 ,fline',,' M«hlne Intel llgenca 4. iu. Mlchl«, ed,), Edinburgh University Press, 1969,
C133 Binford, T.O., "Proximity In N-0Imenslons". (in Preparation).
I I
15
1.3,2 Speech Recognition (Arthur Samuel)
by Ken Sleberz and by some new ideas wnlch this worK has prompted,
Gary Goodman Is continuing his thesis research on a syntax-controlled speech recognition system, it Is anticipated that several first year graouate students will undertake studies related to speech recognltIon,
REFERLNCtS
Cl] Astrahan, M,M., "Speech Analysys by Clustering, or the Hyperohoneme Method. AIM-124, Stanford A,!, Project, May 1970,
C2] i^hlta, G,, "Machine Learning Through Signature Trees, Application to Human Speech", AIM-136, Stanford A,I. Project, October 1970,
16 it
I I I I
I
n
U
i
1,4 HEURISTICS
1,4,1 Mtchine Learning (Arthur Samuel)
Work K,0. redue had speed requl rathe store game near I olay modi f of th being schem
Is continuing on Ha.ieon, A major e the core and d| gotten complete! ed up and It now rerrents of SgK, r more than a fac ge techniques h situations which y completed the to be used by the leatjons of the p ese would require
made to evalu es before actual I
the checker program w|th reProgramming effort has sk storage requirements o y out of hand. The pi uses 35K of core as c
The disk requirement tor of 2, At the same ave been altered to per could not previously be s
Insertion and editing learning program, Se
rogrem are now under cons a major programming ef
ate the potential usefu y starting the programmin
continued help from Just been completed to f the program which ay|ng program has been ompared with earlier s have been reduced by t|me the book game
mit the storage of end aved, Mr, Hanson has of 3 books of end game veral rather drastic Ideratlon, Since each fort, an attempt Is Iness of the different g.
The Go Player, capab le the in objecti pattern help th has be the iror local erttlca oonslde large b
program now Work is c
of handlIng herent oomp ves at dlffe -recognition e program in en observed • usual gene lookahead is i to the pos red a spec ranching fac
plays a better game than the beginning human ontlnuing on development c an evaluation function ail stages of the game. Difficulties arise from lexlty of Co and from the vast differences in rent times during the game, A straightforward learning scheme Is expected to be working soon to exact placement of stones, Considerable success from substituting a local lookahead technique for
ration and search of a (global) move tree. The on the board which seem to be so this technique can be
heuristic for trees with very
initiated at points |t|on as a whole» la|-(*urpose pruning tor,
1,4,2 Automatic Deduction (David C, Luckham)
Recently» the Interactive resolution theorem-proving pr extended and Improved. The program and Its applica axiomatic mathematics Is discussed in Cl, 2, 33. impr •xample, in the algorithm for the replacement rul (Paramcduiation C43) have led to furtner successful e which dependencies between axioms in Marshall axlomatlzatlon of Croup Theory [5, p, 63 have been program |s being used to search for proofs of theor research interest In several different axiomatic theo improvements Include (|) the extension oV the inp many-sorted logic, and (11) the facility for using natu this we mean tha sort of relational structures that occ use, e.g. a multiplication table for a finite arrangement of objects in a room) in the model rela strategy. In addition to experiments |n theorem-proving
17
ogram has been tlons to basic ovements» for e for equal Ity xperlments in Hull's third found. The
ems of current riss, Otnar ut language to ral models (by ur In everyday group» or an tlve deduction • the program
Is currently being Incorporated Into some computer-aided Instpuctlon orograms at the Stanford Institute for Mathematical Studies In the Social Sciences, The objective here Is to Increase the flexibility of the instructional programs for high school mathematics by extending the logical inference system to which the student must conform,
One lack trtvi flroun at h deve I domai metho out t a he* satis many- can of tr descr that false Advlc fruit
of the of edit ai ded ds, Su and. b op stra ns, to as for hat the vy cost f lab! I I sorted be ach odeis. lot ion are t
), Pre! e Take fu| are
In addi other | followin
t |on Ines y. (I)
obvl Ing s uct io ch de ut n tegle
do using use in c
ty re node I leved
( of a rue imlna r Pr as in
to of
ous weak tratesie ns wh|c duct|ons ot |n s of a " this s natural
of anyth omputatl lative t s), It by work
For the model Is of the ry «xper eject C which t
nesse s Cl] h ca are
every seman ort mode Ing b on t| o a s now a Ing w pur
a sp mode Iment 63 o o app
s of to
nnot trivl cont
tic" of e is fo ut th "»e» g truct opear ith o pose ecif I I an s sho r co iy th
the theo e|imlna be exc
ai in th ext, It nature
ditlng, r tn|s p e most eneral ure is s that perators of thl
cation o d some w that p rrectnes ese stra
pem-pr te th luded e cont is th for We ha
urpose rlvlal beeau ieiigt
uch e on (p
s dis f some of the rob lern s of tegies
oving e |ar on P
ext of erefor specie ve exp , How struc
se the hy pro ditlng artlai cusslo of th state
s to eomput t
ppogr ge n upe|y
the e nee ized
er ime ever, tures sval
cess app
) des n» a ose s ments do
ep pr
am Is the umber of
logical ppobiam
essary to ppoblem
nted with It turns Involves
uation of (even for Iications cr Iptlons
partial tataments that ape
with the ogams ape
the development of the semantic editing strategies, research currently being pursued include the
Addition of a procedure for question answering, It has recently been shown that the ppocedupe C7] constpucts Interpolation formulas In the sense of [8],
(ID construction of a system fop checking proofs ties of computer ppogpams, where the proofs in a fopm similar to that described in [9],
ill) Experiments with the reduction of second order proof ppovablilty within two-sorted f|pst-order theor le8C103,
of ppopep- are given
mm
0
L
18
I I I
REFERtNCES
CID Allen, j, and Luckham, D,, MAn Interactive Theorem-Proving Proflratn-, Machine Intelligence 5, {Meltzer, b, and Michfe, D,, Eds,), Edinburgh University Press, March 1V70, Also, AIM-103,
[23 Kleburtz, R., LueKham, D,, "Compatibility of Refinements of the Resolution Principle", submitted to JACM|
C3] Kleburtz, Rf, Luckham, 0,, "Compatible Strategies for the Resolution Principle", forthcoming A.I, Memo,
C4] Robinson, G,, and Wos, L,, "Paramodulat|on and Theorem- Proving In First-order The0r|BS with Equality«, Machine Intelligence JV, 0, Michie (ed), Edinburgh University Press, 1969,
C5] Ha)I, M,, "The Theory of Groups", MacM|||an, New York, 1959,
C6D McCarthy, J,, "Programs w|th Common Sense«, |n M, M|nsky (ed). Semantic Information Processing, MIT Press, Cambridge, 1968,
171 Luckham, D,, Ni|sson. N,, "Extracting Information from Resofutlon Proof Trees", submitted to the Artificial Intelligence Journal, and forthcoming A,I. Memo,
C83 Siegle, J,, "Interpolation Theorems for Resolution In Lower Predicate Calculus", JACM, Vol, 17, No, 3, pp. 53=5-542,
C9] Burstall, R,, »Formal Oescrlotlon of Program Structure and Semantics In First-Order Logic"» Machine Intelligence 5, Edinburgh University Press, pp, 79-98,
C103 Skolom, T,, "Reduction of Axiom Systems with Axiom Schemes to Systems with only Simple Axioms", ülalectlca» Vo|, 12» pp. 443-450, 1958,
19
1,5 Mathematical Theory of Computation (Robert Floyd# ^onald Knuth, Zohar Manna, John McCarthy)
1.5,1 Recent Research I Subst metho Const • Igor for s synta async to trans the non-d para I there cnst cor re Furth such
antlal ds of der Ir.g Ithms» Imp le ctlcal hronou more format proper eterml lej p fore p ructlo spondl er res areas
worh Flo the Ash
para I ly s par cornp Ion o ties nlstl rogra resen n o ng n earch as t
has b yd an
Incr croft lei or s|m 'e al lei I Icate f Para of wh
c Prog ms th ted wh f a on-det
s pr me-sha
een don d Mann eased and Man ograms,
the process d pro I le| pr Ich hav rams a ey cor ch» fo s Imp I
ermjnj s oceedl n r Ing an
e In deve a, for pr Interest na (1970)
A Ithoug y do es, The grams, ogams |nt e been fo re In g respond r a given e equjva tic rrog p» emphas d mu11i-p
lop |n ovlng In have
h the exhib forma The
o no rmaj I enera to. para lent ram Izlng roces
g teohni prepertl
the clas extended Programs it mte iIzation method
n-determl zed In Ma I much A simpjlf I lei prog papal|a|
is of r prajtica
sing.
auesi es of s o thes con
ractl can is
nisti nna ( large I cat I r-am, pro
eason I app
usl algo
f P e tec slder on be e based c pr 197Pa r th on me al lo
-■jram, able
cat
ng the rIthms, aral |ei hnlques • d are between xtended
on a ograms, ). The an the thod Is ws the
whose size.
Ions In
Manna (1970b) demonstrates conclusively that ail properties regularly observed In programs (deterministic or non-determlnlstIc) can be fortruiated In terms of a forma I I zat I on of «partial correctness' Ashcroft (1970) 'explains' this by formulating the notion of an Intuitively «adequate' definition (in predicate calculus) of the semantics of a language or a program, He shows the relationship between a forma I I ?at Ion or Partial conwctness of a program and an •adequeto' logical definition of its semantics, These two works give a general theory unifying the various logical approaches inoludlno those of burstail, Coooer, Floyd, and Manna.
Manna and McCarthy (1970) formalize properties of Usp-||ke Programs using Partial function logic, where the partial functions occurring in the formulas are exactly those computed by the programs, They distinguish between two types of computation rules -- seouentlal and paraliei. ^ McCarthy is trying to further develop ax|om«t|c theories to handle 'undefInedness* In a natural way, Among other things, It may avoid paradoxical statements.
Igarashi (1970) has deveiooed axiomatic methods for the semantics of Algol-like languages, mainly based on nis earlier studies, but allowing the r.ethods of Floyd to be carried out within the formalism, A rretathaorem Is Included which can be interpreted as a ppoof of correctness of a conceptual compiler for the progra-ne treated by the forme I Ism,
Manna and Waldlnger (1970) outlined a theorem-proving program approach to automatic program synthesis. In order to construct a program satisfying certain specifications, a theorem induced jy those
<1 B
20
I
specifications Is proved, and the desired program is extracted from the proof, The use of the Induction principle to construct programs with recursion Is explored In some detail,
Other theoretical research In progress Is mainly orientated towards practical applications, For example, Ashcroft, Manna and Fnuejl h>ve oxtended the class of Schemas for which various properties are decldable, (These results give the decidability of the equivalence problem for lanov schemes as a trivial case), Other work, by graduate students, Is directed towards finding more powerful methods of proving equivalence of programs (Mess), detecting parallelism In sequential programs (Cadlou), and proving correctness of translators (Mor rIs),
Currently our main emphasis Is on preliminary studies for the construction of an Interactive verification system. He w(sh to develop a practical system for proving programs correct that will be powerful enough to handle real programs,
1,5.2 Proposed Research
n
H
In the following we outline several research topics that we wish to undertake In the near future. Note that most of these topics are already being actively pursued,
1. To develop further the theory of equivalence, termination and correctness of computer programs,
2. To develop further the theory of semantic definition of programming languages, the formal description of translation algorithms, and the correctness of compilers,
3. To try to develop a theory of parallel processes adequate to prove their correctness and especially their mutual non-interference.
4, To develop a formal system of logic In which proofs of termination, equivalence, correctness, and non-interference can be conveniently express,
5, To pursue whatever new theoretical avenues appear likely to contribute to the goal of making checkout by proving correctness practical,
6, As soon as possible, Stanford graduate students in computer science will pe asked to prove some of their programs correct as part of their course work so as to check out the techniques developed.
21
RCFERtNCES
Cl] E, A. Ashcroft (1970), «Mathemttleal logic Apolled to the Semantics of Comoutor Prograns", Ph.D, Thesis, to be submitted to Iirperlal College, Lonoon,
C2] E, A, Ashcroft and Z. Manna (1970)» "FormalIzatlon of Properties of Parallel Programs", Stanford Artificial IneI I Igence Project, Memo AIM-110, to appear In Machine Intelligence 6, D, Mlchle (ed,)» Edinburgh Unlv, Press,
C3] S, Igarashj (1970), "Semantics of Algol-like Statements", Stanford Artificial Intelligence Project, Memo AIM-129,
C4] i. Manna (1970a), "The Correctness of Non-determn«stic Programs", Artificial Intelligence Journal, Vol, 1, No. 1,
C5] Z. Manna (1970b), "Second Order Mathematical Theory of Computation", Proc, ACM Symposium on Theory o-' Computing, May
C6D 2. Manna and j, McCarthy (1970), "Properties of Programs and Partial Function Logic", in Machine Intelligence 5, Edinburgh University Press.
C73 2. Manna and R, Waldlngar (1970), «Towards Automatic Program Synthesis", to appear In Comm, ACM,
22
I I
1.6 Representation Theory (John McCarthy)
McCarthy will continue his Investigations of ways of formally
J;S^!?l2fl M8,tUt?,on!: !?W9 B^,ng th9 effect8 of «otlons. and laws of trotlon, New ax lomat Izat Ions of Knowledge and "can" ara In th« worWs ■r* *"'
Recent developments Include work performed during a v|8|t to Stanford by Er.k Sandewal I of Uppsala University on expressing natural language Information In predicate calculus CIJ and work by McCarthy
languages In which not all sentences have truth values.
Recent work In mathematical theory of computation by McCarthy, Ashcroft, and Manna on parallel and Indeterminate computations and the correctness of non-halting programs has a direct application to representation theory because it permits proofs of correctness of strategies while processes other than the activity of the machine are
" Th, f ch,cker development under the Mathematical Computation oroject w||| also promote this.
going on. Theory of
In the next period, work |n representation theory w||| be carried out by McCarthy, Patrick Hayes, possibly David Uuckham, and by graduate students.
il
C1J E, Sandewal
REFERENCE
# "Representing Natural-Language Information in Predicate Calculus-, AIM-128, Stanfor A, I, Project, July 1970,
■*
i I
1
? J
I 23
1 1,7 COMPUTER SIMULATION OF BELIEF SYSTEMS CK«nneth Colby» Frank
Hilf, Malcolm Neweyi Roger Schanki Dave Smith, Larry _. Tesler» and Sylvia Weber]
Kenneth Mark Colby, M.D,, who Is a Senior Research Associate In the Computer Science Deoartment, terminated his orlvate practice of jj psychiatry to devote full time to Investigations In this area of computer simulation. The National Institute of Mental Health n sponsored two projects under Ur. Colby's direction, One of these Is a Research Career Award and the other Is a research project which continues the Investigations In which Ms group has been engaged for the past seven years.
Introduction and Specific Alms
The clinical problems of psychopathology and psychotherapy require further Investigation since so little Is known about their essential processes. Some of this ignorance stems from a lack at a basic science level of dependable knowledge regarding higher mental processes such as cogrltlon and affect, The research of the project atten-pts to approach both the clinical and basic science problems from the viewpoint of Information-processing models and computer simulation techniques, This viewpoint is exemplified by current work In the fields of cognitive theory» attitude change» belief systems. n computer simulation and artificial InteliIgenee,
The rationale of our approach to these clinical problems lies In a conceptualization of them as information-processing problems involving higher mental functions, Computer concepts and techniques are appropriate to thjs level of conceptualization, Their success In other sciences would lead one to expect they might be of aid In the areas of psychopathology and psyohotheraoy,
situations.
Methods of Procedure
We have now gained considerable experience with methods for writing programs of two types. The first type ot program represents a computer model of an Individual person's belief system. We have constructed two versions of a model of an actual patient In psychotherapy and we are currently writing programs which simulate the belief systems of two normal individuals. We have also constructed a model of a pathological belief system In th^ form of an
24
j
The specific aims of this project relate to a long-term goal of developing more satisfactory explicit theories and models of psychopatho|oglcal processes. The models can then be experimented with In ways which cannot be carried out on actual patients. Knowledge gained |n this manner can then be applied to clinical L
[
II L
L
I I
3
•rtlflcla paranoia. A second type of program represents an ntervlewlng program which attempts to conduct en on-ilne dialogue Intended to collect data regarding an Individual's Interpersonal relations, We have written two such Intevlewlng programs and at present we are collaborating wjth psychiatrists In writing a program which can conduct a diagnostic psychiatric interview,
A corrput proeedur consists represen persons col lecte of the rr is ask» particut we I l-kno data-col on-i ine atteirpt in iraeh this pro
er mode es for of eon
ts an of impo d from odel is d to c ar mode wn ef lection man-ma
to writ ine ut b|em is
I of proc
cepts lnd|
rtanc each also
onf lr I wh| facts and
chins e |nt I I iza redu
a bei essin and vidua e to indi carr
m or ch re
of ver |f
dla ervle tlon ced w
ief sys g the I bei|ef I's co him
vldual led out diseonf present human Icat |on logues wing pr of na
e must
tern eons I nformat|o s organl nceptualI In his informant In Inter
Irm the o s h|s oel Intervle
s should and thi
ograms, turai ian use human
sts o n It zed zatlo life by I
views utcom Ief s wer Ideal s Is Howe
guage inte
f a ia cental in a n of space
ntervl in wh
e of e ystem, bias, ly be a maj
ver, remai
rvlewe
rge da ns, Th struc
himsel . Thi ev 3, v ich th xper im
Bee the carr I
or rea the d n grea rs.
ta-bas e data ture f and s dat erif Ic e Info ents o ause o proces ed ou son fo Iff leu t and
e and -base which other a Is at Ion rmant n the
the of by
our ittes untl I
We have written one type of therapeutic Interactive program which Is designed to aid language devejopmsnt in nonspeaklng autistic CK -r!n, ,!! haVeI
uSed J* for **• past tw0 *•"* on eighteen children with considerable success (80X linguistic imorovements). We ntend to continue using this program and to Instruct professionals in psychiatry and speech therapy In how to write, operate and Improve such therapy programs for specific conditions.
This research has s computer sciences.
Significance of this Research
Igniflcance for the psychiatric, behavioral and
Psychiatry lacks satisfactory classifications and exp|a psychopathology. we feel these problems should be concent terms of pathological belief systems. Data coll psychiatry Is performed by humans whose Interactive »f believed to account for a large percentage of the unrell psychiatric diagnosis, Diagnostic interviewing should | conducted by computer programs. Finally, the process and of psyohotherapy are not well understood. Since experlmen computer models Is more feaslele and control |a experimentation on patients, th|s approach may contrlbut understanding of psychotherapy as an informatlon-prooessing
nations of uaijzed In eotlon In facts are abi jity In dealty be mechanisms tatlon on ble than e to our problem.
It is estimated that 90X of the data collected In the behavioral sciences Is collected through Interviews. Again, a great deal of the variance should be reduced by having consistent programs conduct
25
^-■-^
Interviews, Also* this research has significance for cognitive theory« attitude change and social psychology.
Conouter science |s concerned with Problems of man-machine dialogue In natural language, with optimal memory organization and with the search problem |n large data-structures. This research bears on these problems as well as on a crucial problem In artificial InteIMsence, I.e., Inductive Inference by Intelligent machines,
Co I laboratIon
we are collaborating with two psychiatric centers for disturbed children and a local VA hospital. We are also collaborating wjth residents In the Department of Psychiatry and with graduate students In corrputer science, psychology, education and electrical eng IneerIng,
References 1. Colby, K.M,, "Experimental Treatment of Neurotic Computer
Programs", Archives of General Psychiatry, 10, 220-227 (1964),
2. Colby» K.M, and Gilbert, J.P,, "Programming a Computer Modal of Neurosis," Journal of Mathematical Psycho logy,1,405-417 (1964),
3. Colby, K.M. "Computer Simulation of Neurotic Processes", In Computers In Biomedlcal Researcn, Vol. 1., Stacey, R.W, and Waxnran.B,. Eds., Academic Press, New York (1965),
4. Colby, K.M., Watt. J, and GllbertiJ.P, "A Computer Method of Psychotherapy," Journal of Nervous and Mental Disease, 142, 148-152 (1966),
5. Colby, K.M, and Enea, H, "Heuristic Methods for Computer Understanding |n Context-restricted On-line Dialogues", Mathematical Biosdences, Vol. I* 1-25 (1967),
6. Colby, K.M. "Computer Simulation of Change In Personal belief Systems," Rehav|oral Science, 12» 248-253 (1967),
7. Colby, K,Mf "A Programmable Theory of Cognition and Affect in Individual Personal Belief Systems," In Theories of Cognitive Consistency, Ubelson,R., Aranson, E,. McGuire,W(l Kjwcomb,T,, Tannebaum, P. Eos.,) Rand-McNal|y, New York» N.Y, (1968;,
8. Colby, K.M, and Enea» H, "Inductive Inference by Intelligent Machines." Sclentla. 103, 1-10 (1968).
26
I I I I
T I
I o il
9, Tesler.L.» Enea, H, and Colby» K.M, "A Directed Graph ReDresentatIon for Computer Simulation of Belief Systems»" Mathematical Blosclences, 2, 19-40 (1968),
10, Colby» K.M, "Comouter-alded Language Development In NonspeaMng Autistic Children." Technical Report, No, Cs 85 (1967)» Stanford Department of computer Science, Archives of General Psychiatry» 19, 641-651 (1968),
11, Enea, H, "MLISP". Technical Report» CS 92 (1968), Stanford Department of Computer Science,
12, Colby, K,M,, Tes'er» L.» tnea,H, "Search Experiments w(th the Data Base of Human Belief Structure", Proc, of the International Joint Conference on Artificial Intelligence» Washington» D.C., (Wamer ana Norton, Ed»,), 1969,
13, Co|by, K.M,, Tes|er, L, and Enea, H, "Experiments with a Search Algorithm on the Data Base of a Human Belief Structure," Stanford Artificial Intelligence Project Memo AIM-94, August 1969. Proceedings of the International Joint Conference on Artificial Intelligence, Walker and Norton (Eds,), 1969,
14, Colby» K.M. and Smith» D.C,, "Dialogues Between Humans and An Artificial Belief System," Stanford Artificial Intelligence Project Memo AlM-97. Proceedings of the International Joint Conference on Artificial Intelligence» Walker and Norton (Eds,)»1969,
113, Colby» K.M, Critical Evaluation of "Some Empirical and Conceptual Bases for Coordinated Research in Psychiatry" by Strupp and BerSln, International Journal of Psychiatry» 7, 116-117 (1969),
16. Colby» K.M., weber»S. and HI|f,F. "Artificial Paranoia," Stanford Artificial Intelligence Memo AIM-125, July 1970.
17. Colby» K.M., Hilf. F, and Hal|»W. "A Mute Patient's Experience with Machine-Mediated Interviews", Stanford Artificial Intelligence Project Memo AIM-113» March 1970,
18. Hl|f, FM Co|by» K.M,, W|ttner» W. "Machine-Mediated Interviewing", Stanford Artificial Intelligence Project Memo AIM-112# March 1970.
19. Colby, K.M. "Mind and Bra|n, Again." Stanford Artificial Intelligence Project Memo AIM 116, March 1970,
20. Colby» K.M. and Smith, D.C, "Computer as Catalyst in the Treatment of Non-Soeaklng Autistic Children," Stanford Artificial Intelligence Project Memo AIM-120» April 1970,
27
21. Schänk, R. " »Sstnantlcs' In ConctPtual Analysis," Stanford Artificial Intelligence Project Memo AIM-122, May 1970,
22. Smith, D.C, "MS.ISP," Stanford Artificial Intalllgenee Projact Hemo AIM-135I October 1970,
23. Schänk, R.C. "Intention, Memory, and Computer Understanding," " Stanford Artificial InteI IIgence Project Memo AIM-140. January
1971.
i: n 1!
26
t I I
1.8 FtclI|t|es
By the time work on this proposal Is to be Initiated d July 1971) the computer facility will Include the following.
Central Processors: Digital Equipment Corporation POP-10 and POP-6
Pr lirtry Store:
Swapping store:
Fl le Store:
PerIpherals:
Terminals;
65K Words of DEC Core (2us) 65K Words of Amoe* Core (lus) 131K words of new core '(2us)
Llbrascope disk <5 million words, 22 million bits/second transfer rate)
IBM 2314 disc fI|e ( leased)
4 DEC tape drives» 2 nug tape drives, line printer, Caleomp Plotter
15 Teletypes, 6 III displays, 1 ARDS display, 30 Data Disc displays, 3 inuAC remote displays
Special Equipment: Audio Input and output systems, hand-eye equipment (2 TV cameras, 3 arms), remote- control led cart
-? it
if
II
l.B.l Processors
The system operates about 23,5 hours per day every day and Is heavily loaded about 70% of this time. Lengthy compute-bound Jobs such as computer vision, theorem-proving, and machine learning programs often bog down during high-load conditions to the point where It Is difficult to complete segments of useful size in as much as an hour. With this problem |n mind, we have explored alternative ways of Increasing performance and concluded that the most productive approach will be to design and construct a special processor that is optiTized for the class of problem we are dealing with,
Before the initiation of the proposal period, we expect to have completed the design of the new processor, submitted It to ARPA for review, received approval, and initiated fabrication. Funds already available under the existing contract should be sufficient to eomplete this project.
si It
In the event that the design does not materialize In a timely way or that this project is disapproved, some other solution to the processing bottleneck w||| be needed. All known alternatives are more expensive and would exceed available funds, The budget of this proposal assumes that the new processor will be completed successfully.
29
I
We understand that on« of the conditions for final approval of the new orocessor project will be that 5\ä% of Its capacity be made available to other participants In the ARPA net, This will be acceptable» however if the external usage makes substantial demands on other system elements (e.g. primary or secondary storage) It may be necessary to supplement these facilities. We have budgeted no fund« for this,
f;
1,8.2 PrImary Store
We expect to have Procured and Installed core memory by the beginning of the proposal period» using funds made available under a supplement to this contract for Mathematical Theory of Computation, It may be desirable to augment this memory or replace some of the less reliable oortlons subsequent to the advent of the new processor,
l.S.S Swapping Store
The Llbraseope disc file crashed some time ago, destroying half Its capacity. Since the system Is totally dependent on this unit for efficient operation, it Is v<|nerab|e to another crash. On behalf of the U.S. Government, Stanford is pressing a claim against Llbraseope (a division of Singer) regarding shortcomings in this eaulpment, It Is hoped that there will be sufficient recovery under this suit to replace the disc with a more reliable swapping store.
1.8.4 F| ie Store
I
u
Cost/performance considerations and the need to store somewhat more picture Information for our computer vision research call for replacement of the e-d|sc IBM 231<$ file (leased) with a 4-d|sc version of the IBM 3330 (also leased).
1.8,5 Peripherals
Ü
i *
\1 Existing peripherals appear to be adequate with few exceptions. We have a need for a high speed printing device with general graphics capability (l.e, the capabilities of a plotter and the speed of a line prlnte^), Existing devices In this class all seem to use special paper, which makes their operating costs too high. Devices that work with ordinary Paper are under development by more than one manufacturer and we will likely want to procure one with available funos when they become available.
i U
1,8.6 Terminals
Existing terminals appear adequate for the foreseeable future.
I
30 B
I
0
1.8,7 Special Equipment
Ongoing hand-eye research and othgr work In computer vision w||I require additional cameras with color vision capability* manipulators» and other Instrumentation, Funds have been budgeted for these Items,
il
Ü
11 U
if
SJ 31
1
1 .
rv
, . , ——. in — •■*
2, HEURISTIC PROGRAMMING PROJECT (I ncorporatIng Heuristic DENDRAL)
(E.A. Felgenbaum, J, Lederberg, B, G, duchanan, R, S. F.ngelmore)
TABLE OF CONTENTS
Introduction Change of Project Ncme Proposed Work for New Contract Perlodi its Relation to long Term Goals
1. Theory Formation In Science 2. Representation 3. Generality and Problen Solving 4. The Nature of Programs and the Organization of Heuristic Programs 5. New Artificial Intelligence Application
Historical Synopsis Views of Others Others Concerning This Research Review of Work of the Current Period Heuristic ÜENDRAL ^s an application to Chemistry; Possible NIH Sunoort Computer Fad I Itles Budgetary Note Concerning Computer Time Budgetary Note Concerning Personnel Blbllography
2.1 INTRODUCTION
Under previous ARPA contract support, the work of the Heuristic OENDRAL project has Leen focused on understanding the processes of scientific Inference |n problems Involving the Induction of hypotheses from empirical data, and on the Implementation of a heuristic program for solving such problems In a reel scientific setting. The Heuristic DENDRAL program now does a creditable, often extremely good. Job of solving a vaclety of mass spectral analysis problems In organic chemistry,
It was necessary to Invest the effort to construct this complex performance program as a foundation of understanding and a mechanism fror, which to build toward more Interesting and Important programs to do scientific theory formation. We have begun this building.
What we Intend to do in the next two years is the subject of this proposal. The phase of ARPA support of the performance program writing and tuning (the Heuristic DENDRAL phase) will end with the expiration of the present contract period, though funds are being sought from NIH to continue that part of the work,
2.2 CHANGE OF PROJECT NAMt
The focus of our work under ARPA support is expanding, and Its nature Is changing, Our desire to convey this explicitly leads us to want to change the project name to Heuristic Programming Project, This desire was reinforced by our observation that "Heuristic DENDRAL" has become, among computer scientists, « technical term referring to a
32
specific program, rather than a covering term for'« group of otoola working on programs which model scientific thought pJocSLes.
2.3 PROPOSED WORK FOR NtW CONTRACT ^RIOO: ITS RELATI0N TQ LONG TERM GUALS
It Is a paradox that the success of the Heuristic DENDRAL oroaram in
discussions In our r.search «siting „" t', lilt til III,I n? Drobl.ms, Ilk. „ost others, t.nd to h,y, , n.tu?« tlm. o/ J^I» r
;u,?;.ü?!;onth,rt,?:ii'r';u "•'*• "•*" •'« "io ;::so:n/ • e'; frustration. It is toward an understanding of the ovaraii «h-n. «,
our work that we offer the follow.ng comments con^M^'ioa^ 0f
Science and the work of scientists Is our object of study We s.ftu
";«Tn.f",u:n;so,r.db:::v:;;:::;«:x-r?:::ri:icre;-h^,:.ffl,;.ns
l!!-'"!.the t00|! an,l ':on'!,Dt8 '" "-tl'lol.i int.liioono.
Ui
•
1, Theory formation In Science r
FUe years . ago we began our work b- seeking a ansrlMrt-.u
th th. conc.otua. too,. ».'„.J ^"ZÜV'^lltä JSS;0^ Wi _ _ . .__
33
L
L
oroblem area needed to have as Its essence the Inductive ana .-sis of empirical data for the formation of exolanatory hypotheses, This is the type of Inference task that calls for the use of a scientific theory by a Performance program» but not the formation of that theory, we did not have the Insight, understanding! and daring at that tiire to tackle ab Inltio the problem of theory formation (and Indeed It would have been foolhardy to do so then),
Now we feel the time Is rioe for us to turn our attention In a major way to the problem of theory formation. Our understanding and our technical tools have matured along with the Heuristic uENQRAl, program to tne point where we now see clear ways to proceed. The effort, which began in a small way a few montns age Is called Meta-OENüRAj.,
As always» proper choice of task environment Is crucial, but for us the choice was absolutely clear, The theory formation task most accessible to us Is the task of forming mass spectral theory. Hence, the notion of building a level of programs "meta" to the DENDRAL performance program.
DENDRAL already contains an excellent mass spectral theory, Me, therefore, have a clear Idea of what a "correct answer" Is like, DENDRAL's theory Is represented In at least two different forms at present, so that we have a pretty good idea of the Issues Involved In reoresentlng mass spectral theory for a program. The Predictor program Is an interesting kind of artificial experimental environment In which to perform certain kinds of Internal "experlmentsH
systematically, thereby generating a kind of systematic data that may not be available in the natural world, A theory language of notations, data structures, and primitive concepts (with which we are intimately familiar because we developed It) Is available, People who are expert in the discovery of mass spectral theory are members of our research team. Many programs for manipulating mass spectral data have already been developed by us and are ready to be exploited as Meta DENÜRAL tools.
The goal of the Meta-DENDRAL program Is to Infer the theory that the performance program (Heuristic OENDRAL) needs to solve problems of mass spectral analysis.
The following tap Ie attempts to sketch some differences between the programs at the oerformance level and the meta-level,
Heuristic QENDRAL Meta-üENÜRAU
Input Tne mass spectrum of a A large number of recorded nrolecuie whose structure mass spectra and the associated Is not known (except, (known) molecular structures, of course» In our test cases)
6A
Output *„-;;-';; structu,. » ,., of cl..v.9. and r,.rP.no.B.nt
inrarrea from the data rules constituting
a subset of the theory of mess spectrometry
Exarrple uses aloha-carbon Discovers (and valldetes)
rutes^n'nl^nn^r^ a I pha-CLrbon fragmentation ruies in p ann no ann miac i^ - __^-. •# _ « i. ; c« of Dossibl eavage. uses
in !!IM J ann,n9 ana •""•es In a space of possible ,n ^"dation patterns of cleavage, ulll
set of primitive concepts but dees not Invent new primitives.
In our view, the continuity evident |n this table reflects - continuity In the processes of Inductive explanat on \Vllill .
for describing regularities In the data. fleneral rules
35
1. ä
L
I I I
I I
Representation
We have seen In our previous about the (DENDRAj.) wor I problem solving and to Improveo Performance. Be representation Is Important to artlflela because " ■ ■
work that the form In which knowledge a Is represented is crucial to effective
^uT":9 '*8 k"o«l^fl- structure fo? .ause of this, because the problem of
intelligence research, and because we felt ♦•«»+ th« „.«^i-- -.II»-MB« researcn, ano
work, begun In the current period. w||| be pushed. Th. w«Pb !i«. I
more recent (and more satisfactory) proSJJtloJ-JClI fnr! TK*
useo to produce appropriate and '.■-- r"rtUf? c«"B','»tor wll I be Prealctor to produce an artificial to the "i — consIderable
3.
systematic Input to the Mo|d"
'new« representation win b.^.r^.^^fo^i^jo^j'^ inkage to the Meta-OtNDRAL work previ;;;iy0Sn^d
,;as
Generality and Problem Solving
compounds. That I s° "„«til la?*» JSj «,;?cu?°y0"f ««""«•*"<"««l
36
I
I 4. The Nature of Programs and the Organization of Heuristic Programs
I One of tne most Important sources of transfer from our present work to our future work and to the work of others Is likely to result from a aetajjed examination of the OENORAL programs as an organized sequence of manipulations on a symbolic world Internal to the LISP "machine". In our research discussions» we return repeatedly to problems Involved in trying to understand systematically that universe of entitles knovn as computer programs» and In particular the subclass of heuristic programs, Why?
First* as builders of perhaps the largest heuristic program that exists» we are forever reafll ly accomplished undiscovered something next steps to be made previous steps.
frustrated that our next steps are not more than our last steps« that there remains
systematic and understandable that will permit more scientifically and less artisan-like than
Secona. the programming task Itself presents a problem domain worthy of intense application-oriented research by the A,I, community, It is almost certainly true that two or three decades ,ience most computer programs (as we know them today) wilj be written by computer program. The necessary Initial explorations should begin now (as indeed they have begun at a few places)•
! i
Third these the and n cut-o But t eng in heur I seien reorg on), least emp I r
• ou ores
organ ot me ff d here eer In stic ce ( anize The pro
leal
r work ent parti iztion o rely rout ec isions Is so I it g specif Programm Invent, , Invent processes gremmlng sc lent ist
is or I cuiar f our |ne s as opp tie he Ic he Ing I organ I
more of th ass I s
, the
mar I|y co problems programs ymbol ma osed to s ur 1st I c P ur 1st Ic s a scl ze» Itnpl , try a e automat tant) nay problem d
ncern and c we ar nipul ome I rogra oroce ence ement gain» Ic he be s
omaln
ed with he hallanges, e wr iting atlons (f Ist-struct mm|ng scle duresl T at al|, I
» try, o observe a
urIstIc pr mllar to of our pr
urlstl At s
heurIs or ex ure re nee to o the t Is a bserve nd Int ogram the pr Imary
c prog ome I tic pr amp le, organl draw exte
n expe » in erpret writer ocesse Intare
rams and evel In oeedures
search zatlon). upon In nt that rlmentai teroret; ; and so (or at
s of the st.
11
it II
Tor all of these reasons, we plan to invest some of our tjme and resources in a exploratory effort to better understand programs, program construction» and In particular the organization of heuristic programs, whether we pursue our quest for understanding by program writing is not now clear. It will probably depend on what Individuals» particularly graduate students» become Interested In pursuing these questions,
5. New Artificial Intelligence Application
During the current period» we have soent much effort In search of a new application In which the techniques and concepts of artificial Intelligence research can be applied to a problem of Importance to
37
I 1 I I
I I I I
it «•
• •
I
Science. To serve adequately, the Droolem domain must be such tha* ccmplex reasoning processes play a significant role In the discovery of problem solutions (Interestingly, many scientific tasks do Jot
this property). In line with our general Inclinations, problem ry interest to us are those Involving Inductive and hypothesis formation from sets of empirical data
It |S tn h- «f i I1" ch»Pac*»«-lst|os tnat a orooiem area must have If it is to be of Interest and use to here.
have domains of prim generalIzatlon There are many other
us, but we will not d sctss these
During the remainder of the will probably be selected, we work during the period of Initial testing for feasibility,
bMng the application to fruition will pröbab I y^be "sought suoport from NSF
to or
current period, a new appllca* on area Intend that the project sut.ort this problem formulation anfl the oerlod of
If tne Idea Is viable, sustaining
2.4 HISTORICAL SYNOPSIS
molecules (in 11la I Iy mass spectra, later including nuclear magnetic resonance spectra as auxl i lary data), The heur 1st lc p ogam "luin to solve such problems consists broadly of two pnales ^„^esls generation and hypothesis validation. nypotnesis
I«!hi«^Ü!8I$ 9enerat!°n phase ,s « juristic search m which comblnatorla. space of possible candidate molecular structures
This Program Is for the I8H 360
written In LISP, In t*rms of bytes of core memory (,nclud,n9 necessary binary program space and
adequate free storage space) a typical "working" core Image wMi occupy over one million bytes (usually run In three .350K job steps)!
36
Other DENDRAL code used In the past und sometimes In the present constitutes another half million bytes«
The program has solved hundreds of structure determination problems. For the supra-family of organic molecules upon which we have focused most Intensely (saturated* acyclic« monofunct|onaI compounds) the performance Is extremely good» measured In t|me-to-solutIon and in quality of solutions! even compared with the best human performance. In other families and suora-faml|les» the osrformance Is sometimes good sometimes npvIce-I iKe, The basic processes are completely general» however» so that Increments of new chemical knowledge w||| reaally give rise to better performance on a broader range of prob I ems ■
This work has been reported to the computer science community in the following publications: C7], C12I3, [163, C17J» [233, Sines the program |s of considerable Interest as an application in chemistry» we have produced a series of papers for the chemical literature entitled "Applications of Artificial Intelligence for Chemical Inference"» of which six papers have appeared C14]» 1152, 1192, £203» [213, [223» and two more are |n progress. The work has been discussed In terms of philosophy of science in C183,
2.5 VIEWS OF OTHERS CONCERNING THIS RESEARCH
The publication of this work has engendered considerable discussion and comirent among computer scientists and chemists. Professor H, Gelernter (SONY, Stony Brook)» at an SJCC 1970 panel of the use of computers in science gave an extended discussion of the program» in which he said that It was the first Important scientific application of artificial Intelligence, Or. W, H. Mare (RAND Corporation) In a recent article entitled "The Computer in Your Future" In the collection Science and Technology In the World of the Future (A. B. Bronwell. ed.. Wiley Press. 1970) said:
"Thus» much of engineering will bo routinized at a high level of sophistication» but what about science? An Indication of what Is coming at a higher level of Intellectual performance is a computer program called Heuristic DENDRAL, which does a task ihet a physical chemist or biologist concerned w|th organic chemistry does repeatedly."
Professor J. Welzenbaum of MIT, in an undergraduate computer science curriculum proposal fo'- MIT entitled "A Fjrst Draft of a Proposal for a New Introductory Subject In Computer Science (September 1970)" Included Heuristic DENDRAL in h|s "group 4" series (three lectures) entitled Great Programs} and he said recently (personal comirunlcation)» commenting on recent work and plans»
39
SBB^ammuLM.
I I I
31 sib
11
Ü u
o
"I see the work you are now beginning as a steo In the direction of composing explicit models of just such programs (tnat build expertise In an area). The Implications of a success In such an effort are staggering, I therefore believe your effort to be worthy of very considerable Investment of human and financial resources."
In his paper presented at the Intelligence Workshop, Professor Saul used Heuristic UENDRAL to Illustrate a theoretical knowledge, rie wrote!
Sixth International Machine Amarel (Rutgers University)
point about programs which use
"The DENDRAL system provides an excellent vehicle for the study of uses of relevant theoretical knowledge |n the context for formation problems," Cfrom "Representations and MooeMng In Problems of Program Formation", Saul Amarel, In Machine Intelligence 6 («, Maltzer and D, Michie« eds,) Edinburgh University Press {|n press)].
Dr. T ACM s FJCC progr obser and I the Br iti the adapt Brltl Unl ve spect of th from Gerira
. G. E IGART,
r'.eet I am |n vatlon ntense artlf I sh com Unlver
and m sh ch rslty rorretr e prog one o
ny) al
vans ( In In
ng of the that
ef f or ctal I puter stty arket em tea I of G 1st. ram th f the so rea
A I r Fore troduc in
SIGART worId" this pr
t than a ntelI ige science of Edin the use
comoan oteberg has aske ere, 0 leading uested I
e Cambri g a talk , cal |e (and fo ogram ha ny other nee fiel consulta burgh r of the
y. A in Swe
d for II n a rece mass spe 1stIngs
dge R on H
d the I I owe d pro s I ng
d). nts h eques
Heu mass den, sting nt v| ctrom of th
esearch eur|stI progra
d this Pab|y r le prog At a pr eaded b ted an rIstlc specto headed
s and h sit to eter ma e progr
Labs c UEN m "or
wit ecel v ram I act lc y Pro d obt 0ENÜ
metry by
e|p I Stanf nuf ac am,
), Pr DRAL obab I h th ed a n th ai le f esso alned RAL
lab an
n beg ord r turer
eslde at
y the e In more e hi vei,
r D, perm
progr orato eml ne Innin eores s (Va
nt of the smar
terea susta story a fir Mlchl I sslo am t ry at nt g the entat r lan-
the 1970 test ting ined
of m of e of n to 0 a the
mass use
1 ves MAT,
2.6 REVIEW OF WORK OF THE CURRENT HERIUD
In the previous proposal, we outlined work to be undertaken during the current period. There has been substantial progress on much of this work, though we are only a year Into the current period. Some of the work has not been attempted because Its time did not seem to be ripe,
We have already mentioned that dramatic progress was made In the Improveirent of th© performance of the program as an application to chetristry. Our first paper on ringed structures was published and more coirplex work (on steroids) Is now being done, Other functional groups were added to the Planner, N.M.R. analysis was brought to bear !n a meaningful way at the level of planning, As we Indicated earlier, the program Is moving to the stage at which it can be exported to the scientific community,
4B
Our work on reoresentat Ion of knowledge, so dominant a theme |n our previous prooosal, has been multi-faceted, The knowledge we are dealing with Is the theory of the mass spectrometrlc processes of fragmentation and recombination, This theory» the basis of our ** Predictor, has been rerepresented In a particular Production rule ("situation-action" rule) form as a preliminary to writing programs IT that will perform this representational transformation automatically, fj
Li
This new representation Is also the Internal representation for the knowledge acquired by the Interactive Dialog program for eliciting knowledge about chemical structures and processes from practitioners, A rather soeciaMzed chemistry language In which to conduct this dialog has been created, as well as the Interpreter for that language. This work will be the subject of an A,I, Memo later In the ll contract period.
There are at least two aspects to the prooiem of the effective use of general knowledge about a world by a problem solver: at what point In the problem solving process the knowledge |s deployed, and the representation chosen for Its deployment at that point, We have done substantial work on both of these proulems with respect to these problems, the most dramatic event of th|s period was the construction of the Planning Rule Generator (described In A,I, Memo 131 C233). This Is the program that deduces, from some general first-order mass spectral the0ry t^at Is basic t0 the Predlct0r's activity, the specific family-related pattern recognition heuristics used In the Planning process. It deduces "experts" for specific chemical families for deployment early In the hypothesis generation phase. Of the other experiments done with respest to point of deployment of knowledge, some have had spectacular effect In search reduction, as for example, the Introduction of the N.M.R, analysis during Planning rather than as a terminal evaluaion step, (The former Is discussed In A,I, Memo 131 (23), the latter In C19],)
L
D
Scientific report? on our experiments with representation and design will be forthcoming as A,I, Memos during the Spring and Summer of . 1971, |
In this period, also, we have been able to formulate and begin the groundwork for the next period's work on meta-DENDRAU» discussed ear I ler, u
2,7 HEURISTIC DENDRAL AS APPLICATION TO CHEMISTRYI p POSSIBLE .jIH SUpp0RT
• ■ It should be clear from the earlier sections of this proposal that we have demonstrated the feasibility (at least} of applying techniques of artificial Intelligence research to structure determination problems In organic chemistry In a meaningful and practical way. Feasibility, however, is not realization, A very considerable amount of hard work by chemists and OENORAL programmers remains to be done to rske a comprehensive practical scientific tool, ARPA Is not being
41 i
I
If
i t
(
asked to support th|s part of the research and development,
An MH grant application was submitted last Aprlj and given an extraordinarily comprehensive scientific review by Nlhl, The apoMcation was approved and Is now awaiting funding by the Division of Research Resources, which funds computer facilities, mass spectrorretar facilities and other expensive Instrumental tools of bio-rredlcal research, along with research into methods for their most effective us«.
The N1H proposal calls for?
1. Experimental work with a new mass spectrometer, 2. Organizing and programming existing and newly-developed knowledge
about mass spectrometry to Improve the breadth and qualM-y of the performance of the Heuristic ÜE.NDRAL program,
3. The control of a mass spectrometer with a gas-liquid Chromatograph In real-time by the Heuristic DCNDriAL program, such that the whole system Is solving a problem rather than merely collecting data for later analysis,
4. Meta-DtNDRAL research on theory formation In mass spectrometry (a very small fraction since th|s work Is advanced A,I, research and Is central to the ARPA -sponsored effort),
2.6 COMPUTER FACILITIES
Fortunately, the project Is blessed with excellent computer facilities at the moment, so that the only budgetary proposal that neeos to be made In this regard Is for the purchase of services and not for the development of a resource. Our programming Is done almost entirely In Stanford 360/LISP.
On the 360/67 at the Computation Center's Facility, the following Is aval lable:
Remote Job Entry to Batch Partitions via the WYLBUR Text Editor. with Job output available at the terminal. Partition Sizes for Batch Partitions: I3IK bytes In separate high-speed partition for diagnostic runs 280K bytes, normal partition size 411K bytes, large partition size 800K bytes, "giant" partition size, available on overnight runs Interactive t|me-shared LISP Interpreter »-'i compiler, available under ORVYL time-sharing submon|tor
On the 360/50 at the Medical School's ACME Computer Facility, the following Is available under th^ ACME tIme-snarIn» monitor (non-swapping)»
360/LISP (interpreter and compiler)
42
LJ
i Airount of memory: un davtlme operation, request, bo to memory aval lab ie at n
to a few hundred thousand variable» deoendlng upon 1,6 ml ! I Ion byte?» of s oht and at off hours,
bytss In the our Immediate ow-speed core
2.9 BUDGETARY NOTE CONCERNING COMPUTER TIME
The need to hold the overall annual budget constant as we move to the next contract period, coup|ep w|th the need to absorb expenditure Increases In a variety of budgetary categories, necessitated the budgeting of reduced computer time expenditures from the budgeted $6002/ month of the present contract tc $4000/ month. The posslbjs averse Impact of this reduction can (hopefully) be mitigated byi
a. \.he use of spme of the MH gpant funds (|f our proposal Is funded) for certain parts of the work»
b. use of the Artificial Intelligence Project's facilities for part of the work,
c. use of ARPA network facilities, where feasible and appropriate,
Thus, the fallback positions appear at present' to be adequate,
2.10 BUDGETARY NOTE CONCERMING PERSONNLL
In addition to the p -p|e mentioned In the ARPA-supported budget fTe Igenbaum, Lederbt»rg, Buchanj»"-, En^lmore, and graduate students)» other project scientists are provided by the Stanford Mer» Spttctroiretry Laboratory, and the Instrumentation Research Laboratory
M
(i
i ! f i ■
0
o' the Genetics Dspartmer.t, Addlti the MH grant application to carry
onal positions are requested out tasks called for there.
BIBLIOGRAPHY
C1J Ji Lederberg, ,,DENQRAL-e4-A System for Computer ConstructIcn, Enurreratlon and Notation of Organic Molecules as Trre Structures and Cycllu Graphs", (technical reports to NASA, also available from the author and summarised In Cl23)s
Cia] Part I, Notatlonal algorithm for tree structures. (1965) CR.57029
CJb3 Part II, Topology of cyclic graphs (1965) CR!68398 Llcl Part III, Cc p|ete chemical g«"-.phs;
rings In trees (1969) C2] J. Lederberg, "computation of Molecular Formulas for M^ss Spectroiretry", Holden-Oay, Inc, (1964),
embead'n8 i_
C3] Nat.
J, Lederberg, "Topo I og I ca | Mapping of Organic Molecules'*! Proc, Acad. Scl,, 53:1, January 1965, pp, 134-139,
[4] j, Lederberg, "Systematlcs of organl molecules, graph topology and Han-llton Circuit"?. A general outline of the DLNORAL system," NASA CR-48899 (196
5).
43
r
1 ah-
1 1
I f I I
ll
if
! i
i i
Po lyhedra C5.1 J, Leaerberg, "Hamilton circuits of Convex Trlvalent Po|yhedr (UD to 18 vertices), Am. Math, Monthly, May 1967,
C6.1 G.L, Sutherland, "DENDRAL - A Computer Program for Generating an Filtering Chemical Structures", Stanford Artificial Intelllgenc Project Memo No, 49, February 1967,
C7] J, Uederberg and t,A, f-e I get.baum; "Mechanization of Inductive Inference In Organic Chemistry", In B. Klelnmuntz (ed) Korma! Representations for Human Judgment, (Ml|ey, 1968) (ajso Stanford Artificial Intelligence Project Memo No, 54, AuSust 1967).
[8] J. Lederberg, "Online computation of molecular number." NASA CR-94977 (1968),
formulas from mass
C9] E.A. Felgenbaum and ß,C. Buchanan, "Heuristic DENDRAL: A Program for Generating Explanatory Hypotheses |n Organic Chemistry"» In Proceedings, Hawaii International Corference on system sciences, B.K. Kinarlwala and F.F, Kuo (eds), January 1968,
[103 B.G. Buchanan, G.L. Sutherland, and E.A. Felgenbaum, "Heuristic DENDRAL: A Program for Generating Explanatory Hypotheses In Organic Cherrlstry", in Machine Intelligence 4 (B, Meltzer and D. M|ch|e, eds) Edinburgh University Press (ly69), (also Stanford Artificial Intelligence Project Memo No, 62, July I'ftS).
Cll] E.A, Felgenbaum, "Artificial Intelligence: Themes In the Second Decade", In Final Supplement to Proceedings of the IFIP 68 International Congress, Edinburgh, August 1968 (also Stanford Artificial Intelllgenc« Project Memo No, 67, August 1968),
C12J J. Sc lenses Research Sc i ences 37-51,
Lederberg, "Topology of Molecules", in The Mathematical - A Collection of Essays» (ed,) Committee on Support of in the Mathematical Sciences (COSRIMS). National Academy of
- National Research Council» M.i.T. Press, (1969), pp.
C133 G. Sutherland, "Heuristic DtNDRAL: A Family of LISP Programs", to appear In 0, Bobrow (ed)» LISP Applications (also Stanford Artificial Intelligence Project Memo No, 80» Maren 1969),
[14] J. Lederberg, G.L. Sutherland» d,G, Buchanan, E,A. Feigenbaum. A.V, Robertson, A,M, Duff laid, and C, TJerassI, "Applications o, Artificial Intelligence for Chemical Inference I. The Number of Pc sible Organic Compounds: Acyclic Structures Containing C, H, o and N", Journal of the American Cnenjcal Society, 91ili (May 21, 1969).
[15] A.M. üuffield» A,v. Robertson, C, üjerassl, B.G. Buchanan, G.L. Sutherland.« E.A. Feigenbaum, and J, Ledfrberg, "Application of Artificial Intelligence for Chemical Inference II, Interpretation of Low Resolution Mass Spectra of Ketones", Jour al of the American
4 A
Hherrlcal Society, 9IJII (May 21, 1969»,
RECENT PUBLICATIONS
C16] B.C. Buchanan» G.L. Sutherland, E ', Felgenbaum, "Toward an Understanding of Information Froceses o Scientific Inference In the Context of Organic Chemistry", In Machine Intelligence 5, (B, Maltzer and D. Mlchle, eds> Edinburgh University Press (1969), (also Stanford Artificial Intelligence Project Memo No, 99, September 1969),
C17] J, Lederberg, G.L, Sutherland» B.fi. Öuchanan, and E.A. Feigenbaum, "A Heuristic Program for Sojvlng a Scientific Inference Problem: summary of Motivation and Imp IementatIon", In Theoretical Aporeaches to Non-NumorIca| Problem Solving (R, BanerJI and M.O, Mesarovic, ads.) SorInger-Verlag. New rork (1970), (Also Stanford Artificial Intelligence Project Memo No, 104, November 1969),
[18] C.w, Churchman and B.C. Buchanan, "On the Design of Inductive Systems: Some Philosophical Problems", British Journal fcr the Philosophy of sclence, 20 (1969), pp, 311-323,
C193 G. Schroll, A.M, Duffleld, C, Djerassl, B.C. Buchanan, G.L. Sutherland, E,A. Felgenbaum, and J. Lederberg, "Application of Artificial Intelligence for Chemical Inference III. Aliphatic Ethers Diagnosed by Their Low Resolution Mass Spectrs and MMR Data". Journal of the American Chemical Society, 91:26 (December 17, 1969),
I
[20] A, Buchs, A.M, Duffleld, G. Schroll, Ü. Djerassl, ...d, Delflno, B.G, Buchanan, G.L. Sutherland* E,A. Feigenbaum, and J, Lederberu, "Applications of Artificial Intelligence for Chemical Inference ly. Saturated Amines Diagnosed by their Low Resolution Mass soectra and Nuciear Magnetic Resonance Spectra", Journal of the American Chemical Society, 92:23 (November 18, 1970),
C213 Y.M. Sheikh, A, Buchs, A,B. Delflno, B,G. Buchanan, G.L. Sutherland, E.A. Felgenbaum, and J, Lederberg, "Applications of Artificial Intelligence for Chemical Inference V, An Approach to the Conputer Generation of Cyclic Structures, Differentiation Between All the Possible Isomeric Ketonas of Composition C6H100,»I Organic Mass Soectrometry (in press),
[22] A, Buchs, A.B, Delflno, A.M. Duffleld. C. Djerassl» R.G. Buchanan, E.A, Feigenbaum, and J, Lederbert, "Applications of Artificial Intelligence for Chemical Inference yl, Approach to a General Method of Interpreting Low Resolution Mass sPectra with a Computer". Helvetica Chemlca Acta, 5316 (1970),
[23] L,A, Feigenbaum, B.G, Buchanan, and J, Lederberg, ,{0n Generality and Problem Solving: A Case Study Using the DENDRAL Program", In Machine Inteiliganca 6 (B. Meltzer and D, Mlchle, eds) Edinburgh University Press (In press). (Also, Stanford Artificial
45
L
iT
.
I I
^—-i-^— -
Intelligence Project Memo No. 131),
C24] B.G, duchanan and T.E, Haadrlck, "Some Speculation about Artificial Intelligence and i.egal Reasoning". Stanford Law Review, November» 1970, (Also, Stanford Artificial Intel Igence Project Memo No. 123),
1251 B.G, Buchanan» A. . Duff|e|d and A.V, Robertson, "An Application of Artificial Intelligence to the Interpretation of Mass Spectra". In Mass Spectrometry, 1970 (G.W, Ml|ne. ed,) Wiley (In press).
1
46
immmmmsr*?!"''
1 ^1 if
r
ss j
p ,
.1
I I
II
i i
f
11
U
3.BUDGET
3.1 SUMMARY OF BUDGETS FOR CQNTINUATIÜN OF SQ-183 (Flsctl Year 1972)
BUDGET ITEM 1 JUL 71 TO 30 JUN 72
A.I.
Salarles
Staff Benefits
Trave I
Capital Equipment
Equipment Rental
Computer Time
tqulpment Maintenance
Comrrunlcat Ions
PublIcatlons
40,000
14,400
14,000
Other Operating Expenses 32,942
Indirect Costs
Tota I s
322,262
H.P.
$467,526 S 79,628
70,051 11,931
30,500 3,700
108,000 - . .
50.319 5,400
... 46,081
1.500
1,600
1,800
48,360
t 1,150,000 i 200,000
47
TOTAL
$ 547,154
81,982
34,200
108,000
55,719
46,081
40,000
15,900
15,600
34,742
370,622
S 1,350,000
3.2 SUMMARY OF BUDGETS FOR CONTINUATION OF SD-183 (Fiscal Yetr 1973)
BUDGET ITEM 1 JUL 72 TO 30 JUN 73
A,I. H.P.
TOTAL
76,900
30,502
79,855
56,700
Salarles
Staff Benefits
Trave I
Capital Equipment
Eaulpment Rental
Computer Time
Equipment Maintenance
Comirun I cat Ions
Pub IIcatIons
Other Operating Expenses 32,942
Indirect Costs
$467,526 S 79,628
40,000
13,084
3,700
5,400
44,200
14,400 1,500
14,000 1,600
32,942 1,800
337,169 49.088
$ 547,154
89,992
34,200
79,855
62,100
44,200
40,000
15,900
15,600
34,742
386,257
Tota Is S 1,150,000 S 200,000 S 1,350,000
48 1 1
iLUMaaäJMaJiMMgatM^^"^^^' .-■-;hi.. [-|..M-r-^.-.-^ -. .. ü, - ^ - "'-niri — ---^— ~ ■ ^uirm'Qn\'tmnäimm&ia*MiaLdMm^imifb** ■¥Tnr-li ■
1
I
i n
3.3 ARTIFICIAL INTELLIGENCE BUDGET
l-JUL-71 TÜ
30-JUN-72
I. TOTAL ARTIFICIAL INTELLIGENCE SALARIES - 467,5?6
II. STAFF BENEFITS-
13.95i tc 8-31-71 10,831 15,2% to 8-31-72 59,220 11,344 16,7X to 8-31-73 65,064
TOTAL STAFF BENEFITS 70,051
III. TRAVEL
6 Foreign trlüs,1200/ea. 7.2^0 20 Trips east,450/ea, 9,0^0 5 Professional staff moves
to Stanford,1900/ea, 9,5*90 Local travel 4i800
TOTAL TRAVEL 30,500
IV. CAPITAL EQUIPMENT
5-IBM 3336 Disk Packs 5,000 Test Equ|pment<Ann and
camera Instrumentation) 70,000 Color Equipment (Camera,
mount» f I | ters) 33,000 Computer peripherals
and Test Equipment 79,855
TOTAL CAPITAL EQUIPMENT 108,0^0
V. EQUIPMENT RENTAL
IBM Disk Fi|e and Packs (2314, 3330) 50,319
IBM 3330 Disk F|Is 56,700
TOTAL EQUIPMENT RENTAL 50,319
VI. EQUIPMENT MAINTENANCE - — 40,000 (based on past experience)
l-JUL-72 TO
30-JUN-73
467,526
76,908
30,500
79,855
56,700
40,000
I I
49
j
VII. COMMUNICATIONS 14,400 14,400 *' (Teiiphon«»» dataphon«s» teistyo«») -.
VIII, PUBLICATIONS COST (Past Experience) 14,003 14,000 »*
IX, OTHER OPERATING EXPENSES 32,942 32,942 T (e.g, office Supplies, postage» Freight, li Consulting, Utittles)
X. INDIRECT COSTS |
59% of salaries to 9-1-71 45,974 46X of modified direct costs
thereafter (direct costs less L capital equipment) 276»208
TOTAL INDIRECT COSTS 322,262 337,169
TOTAL ARTIFICIAL INTELLIGENCE BUDGET S 1,150,000 1,150,000
50
i
1 1 1 1
I I
3,4 HEURISTIC PROGRAMMING BUDGET
l-JUL-71 l-JUL-72 1 TO TO
40-JUN-72 30PJUN-73
XI. TOTAL HEURISTIC PRÜGRAMKING SALARIES —- S 79,628 79#628
XIJ. STAFF BENEFITS-
n 13,9X to 8-31-71 1,845 15,2X to 8-31-72 10,086 2,002 6,7X to 8-31-73 11,082
TOTAL STAFF BENEFITS n o-.- «, - J XIII. TRAVEL 11»931 13,084 2 foreign trips. $1200, ea. 2,400 2 Trips East, $450. ea, 900
J Local Tpavei 400
n TOTAL TRAVEL i Tan 3,70P 3,700
-
j
it
J
XIV. EQUIPMENT RENTAL (Wyibur Term I na IS-36Z/67)- 5,400
XX. TOTAL SD-183 BUDGET-
5,400
«81 44,200
00 1,500
XV, COMPUTER TIME (IBM 360/67 t|me) 46,
XVI, COMMUNICATIONsCTelephones, Dataphones) 1,5
XVII, PUBLICATIONS 1,602 1,60«5
XVIII.OTHER OPERATING EXPENSES 1,803 1,800
XIX, INDIRECT COSTS
59X of salaries to 8-31-71 7,850 46X of modified direct costs
thereafter (direct costs less computer time) 40,530
TOTAL INDIRECT COSTS----- 48,360
TOTAL HEURISTIC PROGCAMMING BUDGET — $200,03P 49,088
200,000
•$1,350,000 1,350,000
51
4, Cognizant Personnel
For contractual matters:
Office of the Research Administrator Stanford University Stanford» California 94305
Teleohone: (415) 321-2300. axt. 2883
For technical and scientific matters regarding the Artificial Intejijgenco Project:
Prof, John McCarthy Dri Arthur Samuel Mri Lester Earnest Computer Science Department Telephone: (415) 321-2300, ext. 4971
For technical and scientific matters regarding the Heuristic programming project:
Prof, Edward Feigenbaum Computer Science Department Telephone: (415) 321-2300, ext, 4878
Prof, Joshua Lederberg Genetics Department Telephone: (415) 321-2300, ext, 5052
For administrative matters, including questions relating to the budget or property acquisition:
Mr, Lester 0, Earnest Mr, Norman 8rIggs Computer Science Department Stanford University Stanford. California 94305
Telephone: (415) 321-2300, ext. 4971
52
Äi^ÄJl-r - —
APPLNDIX A
PUBLICATIONS
Articles and books by members of the Stanford Artificial Intelligence Project are listed here by year, Only publications following the Individual's affiliation with the Project are given.
1.
2.
3.
4.
1.
1.
2.
3,
4.
5.
1963
J. McCarthy, "A Basis for a Mathematical Theory of Computation." In P, Blaffort and 0, Hershberg (eds,), Computer Programming and Formal Systems. North-Ho I land, Amsterdam, l'öj,
J, McCarthy, "Towards a Mathematical Theory of Computation»" In Proc, IFIP Congress 62, North-Holland, Amsterdam, 1963,
J, McCarthy (with S, Bollen, E, Fredkln, and J.C.R, Lickjlder), "A Time-Sharing Debugging System for a Snail Computer," In Proc, AFIPS Conf, (SJCC), Vol. 2i, 1963,
J, McCarthy (with F, Corbato and M, üaggett), "The Linking Segment Subprogram Language and Linking Loader Programming Languages," Comm, ACM, Ju|y, 1963,
1965
J,. McCarthy, "Problems In the Theory of Computation," In Proc, IFIP Congress 65, Spartan, Washington, D,C,, 1965,
1966
A, Hearn, "Computation of Algebraic Properties of Elementary Particle Reactions Using a Digital Computer," Comm, ACM, 9, pp. 573-577, August, 1966,
J. McCarthy, "A Formal Description of a Subset of Algol," In T, Steele (ed.), Formal Language Description Language?, North-Holland, Amsterdam, 1966,
J, McCarthy, "Information," Scientific American, September, 1966.
J, McCarthy, "Time-Sharing Computer Systems," in w, Orr (ed,). Conversational Computers, Wiley, 1966,
U, Reddy, "Segmentation of Speech Sounds," J, Amer,» August 19*6,
Acoust, Soc.
A-l
uräcr.-^^..■I-. -■- .. ■■ . ■ - ,.-„■ ml ■ ■■-
1967
!• S, Brodsky and J, Sullivan, "W-^oson Contribution to the Anomalous Magnetic Moment of the rtLon," Phys Rev 156, 1644, 1967,
2, J, Campbell, "Algebraic Computation of Radiative Corrections for Electron-Proton Scattering," Nuciear Physics, Vol. 81, DD. 236-300, 1967, ' ' ^
3, t, Feigenbaum. "Information Processing and Memory,« |n Proc, Fifth Berkeley Symposium on Mathematical Statistics and Probability. Vo|, 4, u.C. Press, Berkeley, 1967.
4t J. Goodman, "Digital Image Formation from Electronically Detected Holograms»" In Proc, SPIE Seminar on Digital Imaging Techniques, Soc. Photo-Opt Ica| Instrumentation Engineering, Redondo Beach, California, 1967,
5t J. Goodman, "Digital Image Formation from Electronically Detected Holograms," Applied Physics Letters, August 1967,
6. A, Hearn, "REDUCE, A üser-Orlented Interactive System for Algebraic Simplification, Proc, ACM Symposium on Interactive Systems for Experimental Applied Mathematics, August 1967,
7. J, Lederberg, "Hamilton Circuits of Convex Trlvalent Polyhedra»" Arrerican Mathematical Monthly 74, 522, Ma.* 1967.
8« J« McCarthy, D. 8r|an, G, Feldman, and J, A||en, "THOR—A Display Based T|me-Shar|ng syste.n," AFlPs Conf, Froc.i Vol. 30, (FJCC), Thompson, Washington» D.C., 1967,
9» J. McCarthy, "Computer Control of a Hand and Eye," In Proc. Third Ail-Union Conference on Automatic Control (Technical Cybernetics), Nauka, Moscow, 1967 (Russian).
10. Jt McCarthy and J, Painter, "Correctness of a Compiler for Arithmetic Expressions," Amer, Math, Sop., Proc. Symposia In Applied Math,, Math, Aspects jf Computer Science, New York, 1967,
11. D, Reddy, "Phpneme Grouping for Speech Recognition," J, Acoust, Soc. Amer,, May, 1967,
12. D, Reddy, "Pjtch Period Determination of Speech Sounds," Comm. ACM, June, 1967,
13. D, Reddy, Computer Recognition of Connected Speech," J. Acoust. Soc, Amer,» August, 1967,
A-2
II
II
I [j
I I I
14, A, Samuel. "Studies In Machine Learning Using the Game of Checkers. II-Recent Progress," IdM Journali Novemfceri 1967,
15, G, Sutherland (with G,W. tvans and '».F, Wallace), Simulation Using Digital Computers:. Prentice-Hal 11 Engelwood C|lffs, N,J., 1967.
1968
1. E, Felgenbaum, j, Lederberg and B, Buchanan, "Heuristic Dendrai", Proc, International Conference on System Sciences, University of Hawaii and IEEE, University of HawaM Press, 1968,
2. E, Felgenbaum, "Artificial Intelligence: Themes In the Second Decade", Proc. IFIP Congress, 1968.
3. J, Fe|dman (w|th D, Grles), "Translator Writing Systems"i Comm. ACM, February 1968.
A, J, Feldman (w|th P, Rovner), "The Leap Language Data Structure", Proc, IFIP Congress, 1968,
5, R, Gruen and M. Welher, "Rapid Program Generation", Proc, DECUS Symposium, Fa|| 1968,
6, A, Hearn, "The Problem of Substitution", Proc, IBM Summer Institute on Symbolic Mathematics by Computer, Ju|y 1968,
7, D. Kaplan, "Some Completeness Results In the Mathematical Theory of Computation", ACM Journal, January 1968,
8, J, Lederberg and E, Feigenbaum, "Mechanization of Inductjve Infer-nce |n Organic Chemistry", In y, Klejnmuntz <ed,). Formal Representation of Human Judgment, John Wiley, New York, 1968,
9, J. McCarthy, "Programs w|th Common Sense" In M, M|nsky (ed.). Semantic Information Processing» MIT Press. Cambridge, 1966,
10, J, McCarthy, L, Earnest» U. Reddy, and P. Vlcens, "A Computer with Hands, Eyes, and Ears", Proc. AFiPS Conf. (FJCC), 1968,
11, K, Pjngle, J. Singer, and W. Wlchman. "Computer Control of a Mechanical Arm through Visual Inout", Proc. IFIP Congress 1961', 1968,
12, D, Reddy, and Ann Robinson, "Phoname-to-Grapheme Translation of English", IEEE Trans. Audio and L | ectroacoust|cs, June 1968,
13, D. Reddy, "Computer Transcription of Pnonemlc Symbois", J, Acoust. Soc, Amer,, August 1968,
A-3
14. D, Reddy, and P, Vlcensi "Procedure for Segmentation of Connected Speech", J. Audio Eng, Soc.i October 1968,
15. D, Reddy, "Consonantal Clustering and Connected Soeech Recogn| t Ion", Ppoc. Sixth International Congress on Acoustics, Vol, 2, CJ, C-57 to C-60, Tokyo, 1968,
16. A, Silvestr! and J. Goodman, M0|g|tal Reconstruction of Holographic Images", 1968, NERE.M Record, IEEE, Vo|, 10, pp. 118-119. 1968,
17. L. Tesler, H, Enea, and K, Colby, "A Directed Graph Representation for Computer Simulation of Belief Systems", Math, Bio, 2, 1968.
1969
1, J, faeaucnamp (with H. Von Fperster) (eds,), "Music by Computers'S John W|ley, New York, 1969,
2, J. Seeker, "The Modeling of Simple Analogic and Inductive Processes In a Semantic Memory System", Proc, International Conf, on Artificial Intelligence, Washington, D.C,, 1969,
3, B, Buchanan and G, Sutherland, "Heuristic Dandral: A Program for Generating Hypotheses In Organic Chemistry", In 0, Mlchle (ed,), Machine Intelligence 4, American Elsevier, New York, 1969,
A. B, Buchanan (with C, Churchman), "On the Design of Inductive Systems: Some Philosophical Problems", British Journal for the Philosophy of Science, 2Z, 1969, ppf 311-323,
5, K, Co|by, L, Tesler, and H, Enea, "Experiments with a Search Algorithm for the Data Base of a Human Belief System"i Pr*c. International Conference on Artificial Intelligence, Washington, D.C,, 1969,
6, K, Colby and D. Smith, "ü|a|ogues between Humans and Artificial Belief Systems", Proc, International Conference on Artificial Intelligence, Washington, D.C,, 1969,
7, A, Ouffleld, A. Robertson, C, UJerassi, B, Buchanan, G. Sutherland, E. Felgenbaum, and J. Lederberg, "Application of Artificial Intelligence for Chemical Interference II, Interpretation of Low Resolution Mass Spectra of Ketones", J. Mirer, Chem, Soc,, 91:11, May 1969,
8, J. Feloman, G, Teidman, G, Falk, G, Grape, j, Paarlman, I, Sobal, and J, Tenenbaum, "The Stanfo-d Hand-Eye Project", Proc. International Conf, on ArtifK a| Intelligence, Washington, O.C., 1969,
i .
A-4
n 11
9. J, Fe|dman (with H, Rovner), "An A|goi-based Associative Language", Comm, ACM, August 1969,
10. T. Ito, "Note on a Class of Statistical Recognition Functions", ItEt Tpans, Computers, January 1969,
11. D, Kaplan,"Regular Expressions and the Completeness of Programs", J. Comp. & System Scl,. Vol. 3, iMo. 4, 1969,
J 12. J. Leder^erg, "Topology of Organic Molecules", National Academy of Sclerc.., The Mathematical Sciences: a Coljectlon of Essays
n MIT Press, Cambridge, 1969,
13. J. Lederberg, G, Sutherland, B, Buchanan, E. Felgenbaum, A. Robertson, A. Duffleld, and C, Djerassl, "Applications of Artificial Intelligence for Chemical Inference I, The number
u of Possible Organic Compounds: Acyclic structures Containing C, H, o, and N", J, Amer. Chem, Soc,, 91:11, May 1969,
14, Z, Manna, "Properties of Programs and the First Order Predicate Calculus", J, ACM, April 1969,
15, ?.. Manna» "The Correctness of Programs", j,System and Computer Sciences, May 1969,
16. 2, Manna and A, PnueM, "FormaI IzatIon of Properties of U Recursively Defined Functions«, proc. ACM Symposium on
Computing Theory, May 1969,
j 17. J, McCarthy and P. Hayes. "Some Philosophical Prublems from the Standpoint of Artificial Intelligence», |n 0, Mlchle (ed.) Machine Intelligence 4« American tisevler, New York, 1969, ' '
" 18, U, Montanari, "Continuous Skeletons from DigitIzed Images", JACM, October 1969,
I u l9- R. Paul, G, Fa|k, J, Feldman, "The Computer Representation of Simply Described Scenes", Proc, 2ND Illinois Graphics
i Conference, L'nlv, Illinois, April I969. I
20. R, Schänk and L, Tester, "A Conceptual Parser for Natural Language", Proc, International jo|nt Conference on Artificial Intelligence. Washington, D.C,, 1969,
21. G, SchroM, A, Duf»le|d, C. D.jerassi, B, Buchanan, G. Sutherland, E, Felgenbaum, and J, Lederberg, "Applications of
j Artificial Intelligence for Chemical Inference III. Aliphatic Ethers Diagnosed by Their Low Resolution Ma-s üpectra and NMR Data", J, Amep, C- .m, Soc,, 9i:26, Docembar 1969.
A-5
! !
1970
1. J, Allen and D, Luckham, -An Interactive Theorem-Proving Program" FHI K' ^*!,U,er and "• ^'^l8 <edsf), Machine Intelligence 5, Edinburgh University Press, 1970, ""!■■« w J,
2. B. Buchanan, G, Sutherland, and E. Feigenbaum, "Rediscovering some Problems of Artificial Intelligence in the Contex? o Organic
intelligence 5, Edinburgh Unlvers.ty Press, 1970,
4, b,
5.
6,
7. K, Colby,"Mind and Brain Again", A, Math, Bio,, 1970, McCullough Memorial Vol, of
8. E. . Feigenbaum, chapter In Readiness to Remember, 0, P. Klmbeil ied,), Gordon and Breach, 1V70, * ' "■"
9' *• *I[!?,9' "vlsual Perception by a Computer", in Automatic I" !rPi;^t,0n anri Classification of Images, Academic Press, New TOrK, 1970 ,
10, Hclr^*\Tlil*?, G; Sf,:e:,and' f« Buchanan, E. Feigenbaum, «A Heuristic Program for So|v|ng a Scientific Inference Problem: suirnary of Motivat.on and Implementation", in M- Mesarjvic (er,,), Theoretical Aporoaches to Non-numerical Problem Solving. Soringer-verlag, New York, 1970, ^«iving,
11. D, Luckham, "Refinement Theorems |n Resolution Theory", Proc. 1968 IRIA Symooslum In Automatic aeduction, Ver zilles. Franre* Sor Inger-Ver lag, 1970, ' B *
A-6
I I
i 4f
mm
11
i ^
13.
14.
15.
16.
17.
18.
2. Artlflcl
D. Luckham (w|th D, Papk «nd M, PB .arson), "On Form»i\*** Cojjyf r Props'«. J. Co.p. . Syst«; Sc ,./Jo"! ^.^r June
H. Manna and J. McCarthy. "Prooertles of Programs and Partial Function Logic- In B. Meltzer and 0. Mlchle (ed, V M-IK !' intelligence 5, Edinburgh Un| vers I ty ,Pressri570. ''' " '^
Manna. "The Correctness of Non-Oetermlnlst lc Programs", «I Intelligence Journal, Vol. 1, No, 1, 197?,
Pri;rHT;%?eC0?d"0rder Mathen'atlcal Theory of Computation". Hroc. ACM Symposium on Theory of Computing, May 197p,
2, Manna and A, Pnue||, "Formal Izat|on of Properties of Recursively Defined Functions", J. ACM, July 1970, r0Deri,es of
U, Montanarl, "A Note on Minimal Length Polygonal Annmy im«t 1«« to a Digitized Contour", Comm. ACM, January 19?S A0DrOX,mat,0n
JACM.' ApMrj^; "^ Um,t Properties in »IS'tlzatlon Sch
19. M semalvlco (with G. Bracchl), "An Interactive Software System llL C?^ut!r-alded Design: An Application to CIrcul? Project« Lotrtn, ACM, September 1970, ^rojact ,
20. 0, Waterman, "Generalization Learning Techniques for Automating U HO^U^9 0f H9ur,st,cs,,' J' Artlflcla| Inte i i igencTvo l!
1971
if
«
!U
1.
2.
3.
t. Ashcroft,"Formal Ization of Properties of Paraliii Ppr,n^m » Machine Intelligence 6, Edinburgh Snlv, PrJss, Im!' Pr09r&msH'
P^hi!!9*^?"?' *• Buchanan' J' Lederberg, "On Generality and MI S I 1 S'r n9: A Case Study "«'"S t^ DENDRAL P ogram" Machine Intelligence 6, Edinburgh Univ. press, 1971
L '
A-7
»KV J ^° ■
-
aü—■■Mjjfci MBMÜiüilMi
^■^£;---—^
AHPE.NÜIX d
1 THESES l •
Thrses that have been published by the Stanford Artificial Intelligence Project are listed here, Several earned degrees at
il Institutions other than Stanford, as noted, Abstracts of recent A. I, Nemos are given jn Appendix 0.
^ * n AIM-43, D, Raj Reddy, AN APPROACH TO CÜMPUTEK SPEECH RECOGNITION BY
DIRECT ANALYSIS QF THE SpEECH WAVE, ph.D, Thesis In Computer Science» September 1966,
11 AIM-46, S, Persson. SOME SEQUENCE EXTRAPOLATING PROGRAMS: A STUDY OF REPRESENTATION AND MODELING IN INQUIRING SYSTEMS, Ph.D, Thesis In Computer Science, University of California, Berkeley,
ll September 1966,
r» AIM-47, Bruce Buchanan, LOGICS OF SCIENTIFIC DISCOVERY, Ph.D, Thesis In Philosophy, University of California, Berkeley, December 1966,
AIM-44, James Painter, SEMANTIC CORRECTNESS OF A COMPILER FOR AN U ALGOL-LIKE LANGUAGE, Ph,D, Thesis |n Computer Science, March
1967,
l| AIM-56, William Wlchman, USE OF OPTICAL FEEDBACK IN THE COMPUTER CONTROL OF AN ARM, Eng, Thesis In Electrical Engineering» August
n 1967,
** AIM-58, M, Callero, AN ADAPTIVE COMMAND AND CONTROL SYSTEM UTILIZING HEURISTIC LEARNING PROCESSES, Ph^, Thesis In Operations Research, December 1967«
U AIM-60, Donald Kaplan, THE FORMAL THEORETIC ANALYSIS OF STRONG
EQUIVALENCE FOR ELEMENTAL PROPERTIES, Ph,D, Thesis in Computer Science, July 1968. H
5 t
y
AIM-65, Barbara Huberman, A PROGRAM TO PLAY CHESS END GAMES, Ph.D. Thesis In Computer Science, August 1966,
AIM-73, Donald Pleper# THE KINEMATICS OF MANIPULATORS UNDER COMPUTER CONTROL, Ph.D, Thesis In Mechanical Engineering, October 1968,
AIM-74, Donald Waterman, MACHINE LEARNING OF HEURISTICS, Ph,D, Thesis In Computer Science» December 1968,
U AIM-83, Roger Schänk, A CONCEPTUAL DEPENDENCY REPRESENTATION FOR A COMPUTER ORIENTED SEMANTICS, Ph,D, Thesis In LIngulstIcs. University of Texas, March 1969,
I
8-1
^SL~~ «=^=3te-
AIM-85, Plepre Vlcens, ASPECTS OF SPEtCH RECOüNITlON BY COMPUTER, Ph,D, Thesis In Computer Science» M^roh 1969,
AIM-92. Victor D, Schelnman, DESIGN OF COMPUTER CONTROLLED MANIPULATOR, Enq, Thesis In Mechanical Engineering, June 1969,
AIM-96, Claude Cordell Green, THE APPLICATION OF THEOREM PROVING TO QUESTION-ANSWERING SYSTEMS, Ph.D. Thesis In ElectpleM tngineerIng, August 1969,
AIM-98, James J, Horning, A STUDf OF GRAMMATICAL INFERENCE, Ph.D. Thesis In Computer Science, August 1969.
AIM-106, Michael E. Kahn, THE NEAH-MINIMUM-TIMt CONTROL OF OPEN-LOOP ARTICULATED KINEMATIC CHAINS, Ph.D, Thesis In Mechanical Englneerlng, December 1969,
AIH-121, Irwin SobBl, CAMERA MODELS AND MACHINE PERCEPTION, Ph.D. Thesis in Electrical Engineering, May 1970.
AIM-130, Michael D. Kelly, VISUAL IDENTIFICATION OF PEOPLE BY COMPUTER, Ph.D. Thesis In Computer Science, July 1970,
AIM-132, Gilbert Fa|k, COMPUTER INTERPRETATION OF IMPERFECT LINE DATA AS A THREE-DIMENSIONAL SCENE, Ph.D. Thesis In Electrical Engineering, August 1970,
AIM-144, Jay Martin Tenenbaum, ACCOMMODATION IN COMPUTER VISION, Ph.D. Thesis In Electrical Engineering, Septan.^, r 1970.
I I
iJ
U
L
B-2
I r i
1
Prin Inte that
1 i 1
1 Alte f rotr.
I 1 I
Appendix C
FILM RLPORTS
ts of the following fllins are available for short-term loan to rested group? without charge, They may be shown on|y to groups have paid no admission fee. To make a reservation, write to:
Artificial Intelligence Project Secretary Computer Science Department Stanford University Stanford, CalIfornla 94305
rnatlvely, prints may be purchased at cost (typically $30 to $50)
C|ne-Chrome Laboratories 4075 Transport St, Palo Alto. Ca| Ifornla
i, Ar* Elsenson and Gary Zeldman, "E|IIs D, Kroptechev and Zeus, his Marvelous Time-Sharing System", 16mm black and white with sound, 15 minutes, March 1967,
The advantages of t|me-shar|ng ovbr standard batch processing are revealed through the good offices of the Zeus time-sharing system on a PDP-1 computer, Our hero, Ellis, Is saved from a fate worse tnan death, Recommended for mature audiences only.
2, Gary Feldman, "Butterfinger", 16mm color with sound, March 1968,
6 minutes.
Describes the state of the hand-eye system at the Artificial Intelligence Pr0Ject In the fall 0f 1967, The PDP-6 computer getting visual Information from a television camera and co trolling an electrical-mechanical arm solves s|mp|e tasks Involving stacking blocks, The techniques of recognizing the blocks and their positions as well as controlling the arm are briefly presented. Rated "G".
3, Raj Reddy, Dave Esoar and Art Elsenson, "Hear Here", 16mm with sound, 15 minutes, March 1969,
co l or
Describes the state of the speech recognition project as of Spring, 1969, A discussion of the problems of speech recognition Is followed by two real time demonstrations of the current system. The first shows the computer learning to recognize phrases and second shows how the hand-eye system may be controlled by voice commands, Commands as complicated aa "Pick up the small block In the lower lefthand corner", are recognized and the tasks are carried out by the computer control led arm.
4, Gary Feldman and Donald Peiper, "Avoid", 16mm minutes, March 1969,
C-i
si lent, color, 5
Reports on a computer program written by D, Pelper tor his Ph.D. Thesis, The problem Is to move the computer controlled electrical-mechanical arm through a spaed fMled with one or more kno*n obstacles. The program uses heuristics for finding a safe oath: the film demonstrates the arm as It moves through various cluttered environments wlti- fairly good success.
C-2
I
il
n
il
il
i li
i
These memos report 1970 and later are see AIM-117,
Appendix 0
ARTIFICIAL INTELLIÜLNCE MEMUS
research results. Abstracts of rramos published In listed here, For an earlier list going back to 1963,
Interested researchers may obtain available copies upon request to: Artificial Intelligence Project Secretary Computer Sclent« Department Stanford University Stanford, California 94305
Alternatively, they are available from! Clearinghouse for Federal Scientific
and Technical Information Springfield, Virginia 22151
The Clearinghouse charges $3,00 per full size microfiche copy,
copy and $.95 for a
1970
AIM-108, Michael D, Kelly, EDGE DETECTION IN PICTURES USING PLANNING, January 1970, 26 pages
BY COMPUTER
s paper describes a program for extracting an accurate outline of iran's head from a digital picture, The program accepts as inout iltai, grey scale pictures containing people standing In front of
Thl a irt dlgl
various backgrounds. The output of the program Is an ordered'I 1st of the points which form the outline of the head. The edges of bacnground objects and the inter lor detal's of the head have been suppressed,
has less detail, Edges of objects are located in the reduced picture The edges found in the reduced picture are used as a plan for finding edge» in the original picture,
AIM-109, Roger C, Schänk, Lawrence Tesler, and Sylvia Weber, SPIfJZA HL "NCEPTUAL CASE-BASED NATURAL LANGUAGE ANALYS!S, January 1970, 107 pages.
This paper presents the theoretlrai changes that have developed In Conceptual Oepender.5y Theory ai.o their ramifications In compute" analysis of natural language. The major items of concern a.-e« the el :ir.lnation of reliance on "grammar rules" for parsing with the emphasis given to conceptual rule based parsin.., the develooment of a conceptual case system to account for the power of
D-l
conceptual Izattons: the categorization of ACT's based on oermlssible conceptual cases and other criteria, These Items are developed end discussed In the context of a more powerful conceptual parser and a theory of language understanding,
AIM-U0, Edward Ashcroft and 2ohar Manna, F0RMALI2ATI0N OF PROPERTIES QF PARALLEL PRQGRAMS, February I97i, 58 pages.
In this paper we describe a class of parallel programs and give a f orira I Izat Ion of certain properties of such programs in predicate ca leu tus.
Although our programs are syntactically simple, they do exhibit Interaction between asynchronous Parallel processes, which is the essential feature we wish to consider, The formalIzatIon can aastiy be extended to more complicated programs.
Also presented Is a method of simplifying parallel programs, i,e., constructing simpler equivalent programs, based on the M|ndependenc•,,
of statements In them, with these sjmp||oations our formaitzatfon gives a practical method for proving properties of such programs.
AIM-Ul, Zohar Manna, SECONü-OROER MATHEMATICAL THEORY OF COMPUTATION, March 1970, 25 pages.
In this work we show tnat it is possible to formalize a|| properties regularly observed In (deterministic and non- deterministic) algorithms in second-order predicate calculus.
Moreover, we show that for any given algorithm it suffices to Know how to formalize its "partial correctness'' by a second-order formula in order to formalize all other properties by second-order formulas*
This result Is of special Interest since "partial correctness" has already been formalized In second-order predicate calculus for many classes of algorithms,
This paper will be presented at the ACM Symposium on Thtory of Computing (May 1970).
AIM~U2, Franklin 0, H|if, Kenneth Mark Colby, David C, Smith, and William K. wittner, MACHINE-MEDIATED INTERVIEWING, March 1970, 27 pages.
D-2 1 I
I I
II
»
Tpr
I
AIM-U3, Kenneth M, Colby, Franklin PATIENT'S EXPERIENCE WITH 1972, 19 pages,
. H||f# Ml I I lam A, Ha| | , A K'UTE MACHINE-MEUUTLO INTERVIEWING, Mapch
lnt«pS Ü! dw Üt /^T* "Pt,C,pat0d ,n sev8n -"achlne-medlated nterv ews, excerpts pf whlcn are presented, After the fifth Interview he began to use spoken language for communication. This nove teohnlque |s suggested for patients who participate In the usual vls-a-vis Interview,
are unable to
AIM"li4rTW?T?,<;T*i?r?nr? TaSS J,A' Feldman' ÜN THE SYNTHESIS OF FINITE-STATE ACCEpTgRS, April 1V70, 31 pages.
Two algorithms are presented for solving the foil ig problem: Given a f|n te-set S of strings of symbols, f|nd .nlte-stjte machine which w|M accept the strings of imweaiaie
The paper gives a method for Identifying a finite-state language from a randomly chosen finite subset pf the language If the subset Is large enough and If a bound Is known on the number of states required to recognize the language, Finally, we discuss some of the uses of fnJ.Ji^I1 hW8 and the,r r8,»tl0"sh«P *<> the problem of grammatical inie r snce,
AIM-115, Ugo Mont^narl, ON THE OPTIMAL DETECTION OF CURVES IN NOISY PICTURES, March 1970, 35 pages,
N of the curves (pr to |og2(N) If an approximate method Is used).
AIM-116, Kenneth Mark Colby, M,D,, MNO AND BRAIN, AGAIN, March 1970. 10 pages,
D-3
Classical mlnd-braln questions appear deviant through the analogy comparing mental processes with computational Problems of reducibiMty and personal consciousness considered In the light of this analogy,
ens of an processes, are also
AIM-117, John McCarthy and the Artificial Intelligence Project Staff, E, Feigenbaum, J. Uederb«rg and the Heuristic OENDRAL Project Staff, PROJECT TECHNICAL REPORT, April 1970, 75 pages,
Current research Is reviewed in artificial Intelligence and related areas, Including representation theory, mathematical theory of computation, models of cognitive processes, speech recognition, and computer vision,
, HEURISTICALUY GUIDED SEARCH AND CHROMOSOME MATCHING, April 1970, 29 pages.
AlM-li8, Ugo Montaner
Heur1st leally guided search is a technique which takes systematically Into account Information from the problem domain for directing the search, The problem Is to find the shortest path in a weighted graph froir a start vertex Va to a goal vertex vz: for every intermediate vertex, an estimate Is available of the distance to Vz, If this estjrrate satisfies a consistency assumption, an algorithm by Hart, Nilsson and Raphael Is guaranteed to find the optlmu'i, looking at the a priori minimum number of vertices, above algorithm Is presented, which I fflfnlrrum amount of storage. An appl chromosome matching problem Is then stage of automatic ambiguities in the
In this paper, a version of the s guaranteed to succeed with the ication of this technique to the shown, Matching is the last
chromosome analysis procedures, and can also solve classification stage. Some pecullaritias of this
kind of data suggest the use of an heur1st leally guided search algorithm instead of the standard Edmonds' algorithm, The method that we obtain In this way is proved to exploit the clustering of chroTOSome data! a linear-quadratic dependence from the number of chromosomes Is obtained for perfectly clustered data. Finally* some experimental results are given.
AIM-U9, j, Becken AN LEVEL COGNITIQN,
INFORMATION-PROCESSING MOHEL OF May 1970, 123 pages.
INTERMEDIATE«
There Is a large class of cognitive operations in which an organism adapts Its previous experience in order to respond properly to a new situation - for example: the perceptual recognition of objects and events, the prediction of the Immediate future (e.g. In tracking a moving object), and the employment of sensory-motor "skills", Taken all together, these highly efficient processes form a cognitive subsystem which Is Intermediate between the low-|eve| sensory-mo-or operations and the more deliberate processes of high-level "thought".
0-4
I
11
ii
The present peport describes a formal information-processing model of this "Intermediate-Lever' cognitive system, The model Includes memory structures for the storage of experience, and processes for responding to new events on the basis of previous exoerlence, In addition« the proposed system contains a large number of mechanisms for making the respcnse-seIectI on process highly efficient, In spite of the vast amount of stored Information that the system must cope with. These devices Include procedures for heurIstIca I |y evaluating alternative subprocessas, for guiding t.ie search through memory« and for reorganizing the information In memory Into more efficient reprssentatIons,
AIM-120, K. M, TREATMENT pages.
Co|by, Ü,C. Smith, COMPUTLR AS CATALYST JN THE OF NONSP£AKING AUTISTIC CHILDREN, April 1970, 32
Continued experience with a computer-aItiod treatment method for nonspeaking autistic children has demonstrated improvement effects on thirteen out of a series of seventeen cases, Justification for this conclusion Is discussed In detail, Adoption of this method by other research groups Is needed for the future development of computer-aided treatment.
AIN-121, Irwin Sobe|, CAMERA MODELS ANÜ MACHINE PERCEPTION, May 1970« 89 pages«
it
■ *•
Me h mövea trans eontr for more the mot io some syste I imlt smal I prese
ave d ble c form oi var measur than a major n and of it mat ic at ion, syste
nt |y a
eve iop amera relatt iables ing th dequat source have d and el Image
We matle ccount
ed a on a
ns ob of th
e mode e for of er
eve lop iminat d|sto have
errors ed for
paramet pan-ti
Ject spa e camera I parame our pres ror In m ed means ing othe rtIons shown due to
r Ic mo It hea ce to j , we c ters wh ent nee ode I me of me
r parts If the how to aspects
del for d. The mage spec onstructe ich has a ds. We asurement asur ing of It.
y become general I of pan-t
a computer-c model expre
e as a functl d a cai ibreti demonstrated
have also I to be undesi
and compensa The system ca
the major ze the model I |t head geom
ontroI led sses the on of the on system accuracy dent if led red Image ting for n measure accuracy
to handle etry not.
We have demonstrated the ,ode|'s application In stereo vision and have shown how it can be applied as a predictive device |n locating objects of Interest and centering them in an Image,
AIM-122, Roger C. Schänk« "SEMANTICS" IN CONCEPTUAL ANALYSIS, May 1970« 56 pages,
Li
This paper examines the question of what a semantic theory should account for. Some aspects of the work of Katiz« Filimore« Lakoff and Chomsky are discussed. "Semantics" is concluded to be the representation problem with respect to conceptual analysis. The beginnings of a solution to this problem are presented in the light
Ü-5
of developments In conceptual dependsncy theory,
AIM-123, Bruce G, Buchanan» Thomas E, Headrlck, SOME SPECULATION ABOUT ARTIFICIAL INTELLIGENCE AND LEGAL REASONING, May 1970, 54 pages.
Legal reasoning Is viewed here as a complex problem-solving task to which the techniques of artificial Intelligence programming may be applied. Some existing programs are discussed which successfully attack various aspects of the problem, In thjs and other task donf.lns, It remains an open question, to be answered by Intensive research, whether computers can be programmed to do creative legal reasoning. Regardless of the answer, It Is argued that much will be gained by the research,
AIM-124, M.M. Astrahan. SPEECH ANALYSIS BY CLUSTERING, OR THE HyPERPHÜNEME METHOD, june 1970, 22 pages.
In this work, measured speech waveform data was used as a basis for partitioning an utterance Into segments and for classifying those segtrents, Mathematical classifications were used Instead of the traaitlonal phonemes or linguistic categories, This Involved clustering methods applied to hypersoace points representing periodic samples of speech waveforms, The cluster centers, or hyperphonemes (HPs), were used to classify the sample points by the nearest-neighbor technique. Speech segments were formed by grouping adjacent points with the same classification. A dictionary of 54 different words from a single speaker was processed by this method. 216 utterances, representing four more repetitions by the same speaker each of the original 54 words, were similarly analyzed Into strings of hyperphonemes and matched against the dictionary by heurIstlea Ily developed formulas. 87* were correctly recognized, although almost no attempt was made to modify and Improve the Initial methods and parameters,
AIM-125, Kenneth M, Colby» Sylvia Weber, and Franklin H||ff ARTIFICIAL PARANOIA, ju|y 1970, 35 pages.
A case of artificial paranoia has been synthesized in the form of a computer model, using the test operations of a teletyped psychiatric Interview, clinicians Judge the Inp1 -output behavior of the model to be paranoid, Formal validation of the model will require experiments Involving IndlstIngulshabI I Ity tests.
AIM-126, Donald E, Knuth. EXAMPLES OF FORMAL SEMANTICS, July 1970, 34 pages.
I A technique of formal definition» based on relations between "attributes" associated with nonterminal symbols in a context-free gramrar, is illustrated by several applications to simple yet typical problems. First we define the basic properties of lambda expressions, involving substitution and renaming of bound variables, Then a simple
D-6
I orogramnrlno language Is defined using several different oojnts of vleiM, The emphasis Is on "declarative" pather than "Imperat I ve" forrrs of definition,
AIM-127, 2ohar Manna and Richard J. Waldlnger, TOWARDS AUTOMATIC PROGRAM SYNTHESIS, July 1970, 54 pages.
II
n
An elementary outline of the theorem-proving approach to automatic program synthesis Is given, without dwelling on technical details. The method Is Illustrated by the automatic construction of both recursive and Iterative programs operating on natural numbers, lists, and trees.
In order to construct a program satisfying certain specifications, a theorem induced by those specifications Is proved, and the desired program |s extracted from the proof. The same technique Is applied to transform recursively defined functions Into Iterative programs, frequently with a major gain in efficiency,
i y
Ü
'0
It Is emphasized that In order to construct a program with loops or with recursion, the principle of mathematical Induction must be applied, The relation between the version of the Induction rule used and the form of the program constructed Is explored in some detail.
AIM-126, Er|k J,Sandewal I, REPRESENTING NATURAL-LANGUAGE INFORMATIQN IN PREDICATE CALCULUS, July 1970, 27 pages,
A set of general conventions are proposed for representing natural language Information |n many-sorted first order predicate cajculus. The purpose Is to provide a tasting-ground for existing theorem- proving programs,
AIM-129, Shlgeru Igarashl, SEMANTICS OF ALGOL-LIKE STATEMENTS, June 1970, 95 oages,
y
u
if u
if
D-7
Finally, a process of decomposition of Algol-like statementsi which can be r«gardBd as a conceotual compiler, or a constructive description of semantics based on primitive actlonsi Is defined and Its correctness Is orovad formally, by the help of the Induced Induction prInclole,
AIM-130, Michael D. Kelly, VISUAL lOeNTIFICATION OF PEOPLE 8V COMPOTF«, Ju|y 1970, 238 pages.
This thesis doscMoes a computer program whloh performs a complex picture processing task. The ♦"ask is to choose, from u collection of pictures of people taken by a TV camera, those pictures that depict the same person, The primary purpose of this research has been dlrecte« toward the deveiooment of new techniques for picture orocessIng,
In brief, the program works by finding the location of features such as eyes, nose, or shoulders In the pictures, Individuals are classified by measurements between such features. The Inter«8t|ng and difficult part of the work reported in this thesis Is the detection of those features In digital pictures, The nearest neighbor method is used for Identification of Individuals once a set of ireasurements has been obtained.
The success of the program Is due to and Illustrates the heuristic use of context and structure, A new, widely useful, technique called planning has been applied to picture processing. Planning !s a term which Is drawn from artificial Intelligence research In problem so I vIng,
The principal positive result of this research Is the use of goal- directed techniques to successfully locate features In cluttered digital pictures. This success has been verlflec* by displaying the results of the feature finding algorithms and comparing these locations with the locations obtained by hand from digital printouts of the pictures, Successful performance In the task of Identification of oeople provides further verification for the feature finding algorithms,
AIM-131, Edward A, Falqenbaum, Bruce G, Buchanan, Joshua Lederberg, ON CENERALITY AND PROBLEM SOLyiNGS A CASE STUDY USING THE DENORAL PROGRAM, August 1970, 48 pages.
Heuristic OENORAL Is a computer program written to solve problems of Inductive Inference In organic chemistry, This paper will use the design of Heuristic OENORAL and Its performance on different problems for a discussion of the following tooles:
1, the design for generality» 2, the performance problems attendent upon too
much generali ty 3, the coupling of expertise to the general problem solving
D-8
I
11 mi
processes, 4, the symbiotic relationship between generality and
txpertnness of problem solving systems,
«l^r^hf; TS DaD8; W!th.a VleW 0f tne desifln for a 9"^«l P^blem solver that Is a variant of the "big switch« theory of generality.
AIM-132, Gilbert Fajk, COMPUTER INTEf^HtTATlüN ÜF IMPERFECT LINE DATA AS A THREE-DIMENSIONAL SCENE, August 1970, 187 pages,
The major portion cf thjs paper describes a heuristic scene description program. This program accepts as Input a lUnl tll'lrloSt *;/ ''"/rawing. Based on a set of known objec? mo^ls the program attempts to determine the Identity and location of each ob.ect viewed. The most significant feature of the p SJr.m Is "5 abTlltv to deal with |mperfect Input data,
AIM-133, Anthony C, Hearn, REDUCE 2, Octooer 1970, pages.
manual provides the user w|th a description of the algebraic mming system REDUCE 2. The capabilities of this smem
This manual orogra Ine lud
1) Expansion and ordering of rational functions of
AIM-134, Jay Martin Tenenbaum, ACCOMMOOATION IN COMPUTER VISION. September 1970, 452 pages,
We an evolving * . *. - computer vision system In which the Srj^V! ^ Ca,nePa are "ntroll.d by the computer It iJ J^t'^i8 8Jn«:0m c0nve"t'?na, p,ctur# Processing sysUms by the fact that sensor accommodation is automatic Integral part of the recognition process.
and treated as an
Inf^m.i?;« tilnVa DePSOn' c0mes ,n contact with far more visual Information than it can process. Furthermore, no physical sensor can simultaneously provide Information about the full range of the • nvlromrent. Consequently, both man and machine must ^«ommoda?;
U-9
their sensors env I rontrent,
to emphasize selected characteristics of the
Accofrmodat Ion improves the reliability and efficiency of machine perception by matching the Information provided by the sensor w|th that reoulred by specific perceptual functions. The advantagee of accommooatIon are demonstrated In the context of five Key functions In computer vision: aciulsltlon» contour following, verifying the oreserce of an expected edge, range-f|ndIng, and color racognltlon.
We have modeled the Interaction of camera parameters with scene characteristics to determine the composition of an Image, Using a prior! knowledge of the environment, the camera Is tuned to satisfy the Information reoulrements of a particular task,
Task performance depends Implicitly on the appropriateness of available Information, If a function falls to perform as expected, and If this failure Is attributable to a specific Image deficiency, then the relevant accommodation parameters can be refined,
This schema for automating sensor accommodation can be applied In a variety of perceptual doma'ns,
AIM-135, Oavld Canf|e|d Smith, MUSP, üctober 1970, 99 pages.
AIM-136, George M, White« MACHINE LCARiMxNG THROUGH SIGNATURE TREES, APPLICATION TO HUMAN SPEECH, October 1970, 40 pages.
Signa I nvas human to e extra made When and featu f eatu the h this
ture tlgat spee
stabi ctors by
the d recog res, res eur Is expos
tree ed f ch, Ish must
compa ata b nltlo Lear
to a tics ition
"mach or t When trend be e ring ase I n ca nlng ch|ev for t
Ine lea he spe the dat s wl th mp loyed Its fe
s rep |e n be results e recog his typ
rnlng" elf lc a base
conf and "
ature te, a ach lev from
nl tlon e of I
i pat p-ob of g Idenc recog value Ms|gn ed by se|ec . Pr earn!
tern re lern of Iven ut e, a nltlon" s w|th ature" the ev
ting an opert|e ng are
cognl t lo compute terances targe n of an
those of trse can aluatlon optlm«
s of slg of prime
n heur I r recog is Ins
umber o unknown known be eo of a s mlnlnm
nature ry Int
stfes are nltlon of ufflclent f feature pattern
patterns, nstructed elect few I set of trees and crest In
AIM-137, Donald E, Knuth, AN EMPIRICAL STUDY OF FORTRAN IN USE, November 1970, 44 pages,
0-10
I
•k
If
!
if I
I L
u
—
i -
i
I
A sairple of programs» written !n Fortran by a wide varl for a wide variety of aoplleatjons, was chosen "at atteirpt to discover quantitatively "what programmers Statlstleal results of this survey are presented here, some of their apparent Implications for future work design, The principle conclusion which may be Importance of a program "profile", namely a table of which record how often each statement |s performed there are standard
strong
ety of people random" In an really do",
together wjth In compljer
drawn Is the frequency counts n a typical run:
ndlcatlons that profile-Keeping should become a practice In all computer systems, for casual
as system programmers, Some new are also suggested. This paper undertaken by the author and repräsentatives of the software
users as we I I approaches to compiler optimization Is the report of a three month study about a dozen students and
ndustry durl.ig the summer of 1970,
AIM-138, E, Ashcroft and Z, Manna, THE TRANSLATION OF »GO-TO' PROGRAMS TO »WHILE* PROGRAMS, November 1970, 28 pages.
In this paper we show that every flowcnart program can be written without »go-to» statements by usjng »while» statements. The main Idea Is to Introduce new variables to preserve the values of certain variables at particular oolnts In the program; or alternatively, to Introduce special boolean variables to keep Information about the course of the computation, The new programs preserve the «topology» of the original program, and are of the same order of also show that this cannot be done in varlables,
general efficiency, without add
We ng
AIM-139, Hohar Manna, MATHEMATICAL THEORY OF PARTIAL CORRECTNESS, December 1970, 24 pages,
In this work we snow that it is possible to express most properties regularly observed In algorithms in terms of «partial correctness» (I.e., the property that the final results of the algorithm. If any, satisfy some given Input-output relation), This result |s of special interest since »Partial correctness» has already been formulated In predicate calculus and In partial function logic for many classes algorIthms,
of
1971
AIM-140, Roger C, Schänk, INTENTION, MEMORY, UNDERSTANDING, January 1971, 59 pages,
AND COMPUTER
Procedures are described for discovering the Intention of a speaker by relating the Conceptual Dependence representation of the speaker's utterance to the computer's world model such that simple Implications can be made, These procedures function at levels higher than that of the sentence by allowing for predictions based on context and the structure of the memory. Computer understanding of natural language Is shown to consist of the following parts: assigning a conceptual representation to an Input; relating that representation to the
D-ll
memory such as to extract the Intention of the sooake-s and selecting the correct response type triggered by such an utterance according to the situation,
AIM-141, Bruce G, Buchanan» Edward A, Felgenbaum« and Joshua Uederberg, THE HEURISTIC DENDRAL PROGRAM FOR EXPLAINING EMPIRICAL DATA, February 1971, 20 pages.
T>a Heuristic DENORAL program uses an Information processing modal of scientific reasoning to explain experimental data In organic eherrlstry, This report summarizes the organization and results of the program for computer scientists, The program Is divided Into three main parts: planning, structure generation, and evaluation,
The planning phase Infers constraints on the search space from the empirical data Input to the system, Tne structure generation phase searches a tree whose termini are models of chemical models using pruning heuristics of various Kinds, The evaluation phase tests the candidate structures against the original data. Results of the program's analyses of some test data are discussed,
AIM-142, Robin MMner, AN ALGEBRAIC DEFINITION OF SIMULATION BETWEEN PROGRAMS, February 1971, 21 pages,
A simulation relation between programs Is defined which Is auasl-orderIng, Mutual simulation Is then an eoulvalence relation, and by dividing out by It we abstract from a program such details as how the sequencing |s eonrrolled and how data Is represented. The equivalence classes are approximations to the algorithms which are realized, or expressed, by their member programs,
A technique Is given and Illustrated for proving simulation and equivalence of programs; th-re Is an analogy with Floyd's technique for proving correctness of programs, Finally* necessary and sufficient conditions fc* simulation are given.
fl
1;
D 0 r
u
D
- !
i
D-12 1 I
Apptndlx £
OPEHATINC NOTES
Stanford Artificial Intel iigsnes Laboratory Operating Notes (SAILONs) describe the operation of computer programs and equipment and are Intended for project use, This annotated list omits obsolete notes.
The laboratory has a dual-processor (DEC POP-10/PDP-6) time-shared computer with 131 thousand words of core memory backed by a swapping disk (20 million bits per second transfer rate) and an IBM 2314 disk file, Online terminals Include 40 display consoles and 15 Teletype terir.inals, Other online equipment Includes TV cameras, mechanical arms« audio Input and output,
SAILON-2.1» W, Weiher« "Calcomp Plot Routines«« September 1968.
SAILON-3,1« 8, Baumgart« "How to Uo It and Summaries of Things'*« March 1969, An Introductory summary of system features (obsolescent).
SAIL0N-8, S, Russell« "Recent Additions to FORTRAN Library"« March 1967,
SAILON-9« P, Petit, "Electronic C|ockM« March |967, Electronic clock attached to the system gives time jn mlero-secondSi seconds« minutes« hours« day« month« and year. You have to remember whether it's B,C, or A.O,
SAILON-ii, P, Petit« "A Repent Change to the Stanford PDP-6 Hardware"« March 1967, The POP-6 has been changed so that user programs can do their own I/O to devices numbered 700 and above,
SAILON-21« A, Grayson« "The A-D Converter"» June 1967,
SAILON-21 Addendum l, E, Panofsky« "A/0 Converter Multiplexer Patch Panel and Channel Assignments as of 1-9*69"« January 1969,
SAILON-24, S, Russell« "POP-6 I/O Device Number Summary"« August 1967,
SAILON-25« S, Russell« "The Miscellaneous Outputs"« August 1967. Gives bit assignments for output to hydraulic arm and TV camera positioning,
SAILON-26.2« P, Petit« "FAIL"« April 1970, OfeScMbes one-pass assembler that Is about f|ve times as fast aa MACRO and has a more powerful macro processor,
E-l
L
SAILON-28,3, L, Guam, "Stanford LISP 1,6 Hanual". SeDtamber 1969, Describes the LISP Interoretar and coitiDller, the editor ALVlNt, and other aspects of this venerated list processing system.
SA1L0N-29, W, Welher, "Preliminary Description of the Display Processor", August 1967, III display system from the orogrammer's viewpoint,
SAILON-31, J, Sauter» "Dlso Diagnostic", October 1967, A program to test the Llbrascopa Disk and Its Interface,
SA1L0N-35.2, K, P|ngle. "Hand-tye Library F|le'S April 1970,
SAILOM-36, G, Feldman, "Fourier Transform Subroutine", June 1968. FORTRAN aubrogtlne performs one-dlmenslonaI Fast Fourier Transform,
SAILON-37, S. Russell and L. Earnest, "A,I. Laboratory Users Guide", June 1968, Orientation and administrative ppocedupes,
SAILON-37, Supplement 1, J. McCarthy, "A,I, Laboratory Users Guide", June 1968, Hard-line administration,
SAILON-38, P, Vicens. "New Speech Hardware", August 1968. Preprocessor for Input to speech recognition systems,
SAILON-39, J. Sauter and D, Swlnahart, "SAVE", August 1968, Program for saving and restoring a slngje user's disk files on magnetic
tape,
SAILON-41, L, Guam, "SMILE at LISP", September 1968, A package of useful LläP functions.
SAILON-4?, G, Falk, "Vldlcon Noise Measurements", September 1968, Measurements of spatial and temporal noise on Cohu vldlcon camera connected to the computer,
SAIL0N-43, A, Moorer, "DAEMON - Disk Dump and Restore", September 1968. Puts all or selected files on magnetic tape, New version described In SAILON-54,
SAIL0N-44, A. Moorer, "FCROX - MACRQX to Fall Converter". September 1968, Converts MACRO programs to FAIL format, with a few annotated exceptions,
SAIL0N-45, A, Hearn, "REDUCE Implementation Guide", October 1968, Describes the orocedure for assembling REDUCE (a symbolic computation system) In any LISP system.
SAIL0N-46, W, Welher, "Loader Input Format", October 1968.
B D n
• &
E-2
i I
i n
1 n I U
n
D
SAILON-47, and 47 Suppl.ment 1, J. Siuter and J. Slngtr, -Known Prooramrlng Olffaraneaa Betwean tht PDP-6 and POP-10.. Novambar
SAILON-49, A. H.arn, "Servlca Routlnaa for Standard LISP Usars". raoruary 1969,
SAlLON-50,2, S. Savltzky. "Son of Stopgap", Apr|| 1970. A
i!h;;?^?;"0r,ente2 t9r ed,tor w,th »^mg iiaroh .nd substitution commands and hypnanlass text Justification,
SAILON-52,1, A, Moorer, "Systam Bootatrappar's Manual", Fabruary 1969, How to bring back the systam from various states of
SAILON-53, R, Neely and J, Beauohamp, "Some FORTRAN I/O Human|zot|on Techniques", March 1969, How to live with FORTRAN crockery,
SAILON-54.2, A. Hecrer. "Stanford A-l Project Monitor Manual: Chapter I - Console Commanas", September 1970, How to talk to the timesharing system.
SAIL0NCh!n?: Ni M00rM^, ^tan'ord A-I Project nonltor Manual: chapter II - Use,- Programming", September 1970. Machine language commands to the timesharing system,
SAILON-56, T. Panofsky, -Stanford A-l Facility Manual", Computer equipment features (In preparation), ^ompuxer
SAILON-57. 0. Swinehart and R. Sproull, -SAIL", November 1V69. ALGOL-60 compiler w|th LtAP constructs and string processing.
SAIL0N;;:;!!;;-n!/. Pet,t' ,,RA!0H' September 1969. Olspl.y-orlanted machine language debugging package.
SAILON-59, A, Moorer, "MONMON«, October 1969. Lets you peer into the is monitOj*.
SAILON-60, L, Earnest, "Oocumentation Services", February 1970. Text F«Mmir#«y lom1ut*r I» often cheaper then typewriters. Faei.ltles for text preparation and reproduction are discussed.
SAILON-61, R, Helliwell, "COPY-, January 1971. A program for moving
I;il!*:r0,n one D,acg t0 ■noth«r' often with Interesting side
I ! E-3