101

known connectivities within the protein structure (i.e., the amino acid
sequence, cofactors, disulfide bridges, and coordination bonds to
prosthetic groups). The microstructure knowledge base contains known
facts about protein molecules (e.g., the molecular geometry of the peptide
bond and the amino acid side chains, hydrogen bonding properties, helix
forming propensity, etc.). The macrostructure knowledge base contains
stereotype templates for the plausible major components of the molecule
(e.g., alpha helices and pleated sheets). Our strategy is to isolate an
individual molecule within the map, then determine the path of the main
chain by a skeletonizing procedure (as is done, for example, by J. Greer,
J. Mol. Biol.(1974), vol.82, pp. 279-301). We will then parameterize
the density along the backbone, identify the most obvious regions (heavy
atoms, alpha helices, planar groups) and determine, by region growing and
template matching, the identity and position of side chains.

2, Structure Determination in the Absence of Experimental Phase
Information

X-ray crystallography is the primary experimental technique for
investigating the 3-D structure of molecules. The data so obtained are a
collection of intensity measurements at discrete directions with respect
to the x-ray beam and crystal axes. These intensities are related to the
positions of the atoms in the crystal lattice by a Fourier transformation.
Thus, given the structure of the molecule, its orientation with respect to
the crystal axes and the symmetry properties which determine the molecular
stacking, one can calculate the intensities. However, given the
intensities and the unit cell properties, one cannot go the other way.
That is, the inverse Fourier transform cannot be calculated because the
experiment measures only the amplitudes of the diffracted waves and not
their phases -- the classical "phase problem" of x-ray crystallography.

During our first year we have been investigating and implementing
various computational techniques for inferring trial structures, in the
absence of phase information. Our aim has been to develop a system of
computer programs which apply as much general and task~specific knowledge
as possible to constrain the search for a plausible structure (or partial
structure) consistent with the experimentally measured structure
amplitudes,

A procedure well known to x-ray crystallographers investigating the
structures of small molecules is that of Patterson search. Recently this
technique has been shown to be effective in protein crystallography as
well, when there exists a family resemblance between the molecule under
investigation and a known protein. Patterson search is basically an
image-seeking technique, where one searches for the "Patterson image" of
an hypothesized molecular structure in the Patterson map derived from the

experimental data. The Patterson function does not require any knowledge
of phases.

Patterson search is our primary technique for inferring structures
in the absence of experimental phase information. In order to resolve the
ambiguities which often arise in Patterson search predictions, we are
investigating the use of other knowledge sources, among which are
102

anomalous dispersion Paterson interpretation, Patterson search in
reciprocal space, superposition and Fourier refinement methods. The
integration of these diverse knowledge sources is a primary objective of
the research.

C. Summary of Project Accomplishments

Our activities during the first year fall into three general
categories: augmenting our stockpile of crystallographic computing tools,
applying Patterson search methods to a solved protein structure
(Cytochrome C2) and an unsolved protein structure (Cytochrome F), and
initiating a new research objective.

1. Application of Patterson Search to Cytochrome F

A major goal of our first year of research has been to apply the
method of Patterson search, in conjunction with other analytic techniques,
to solve a real protein structure. Cytochrome F is an excellent
candidate, because (1) phase information is not yet available, and (2) the
protein’s structure is expected to show a resemblance to other members of
the cytochrome family. The current hope is that the family resemblance
will be sufficiently strong that the complete structure can be solved by
standard refinement techniques after one finds the correct orientation and
position of a characteristic substructure. As of this writing a complete
solution has not yet been obtained. A considerable effort, described in
the remainder of this section has been invested in the pursuit of the
correct orientations of the protein in the unit cell. A large number of
Patterson search calculations were performed, exploring the effects of
variations in the search structures, the selection of search vectors, the
choice of measures of fit, and even in the primary data. It now remains
to be seen if some of the candidate orientations proposed by the search
calculations can be verified by other sources of crystallographic
knowledge.

2. Selection of a new research objective.

Shortly after the inception of the project, the two collaborating
groups agreed to the need for an additional scientist with an extensive
knowledge of crystallography and crystallographic computing, and a serious
interest in the application of AI techniques to his field of expertise,

We were fortunate to induce Dr. Carroll Johnson of Oak Ridge National
Laboratories, who well fulfills these qualifications, to join our project
for a one-year period beginning September 1, 1975. His contributions have
been instrumental in defining a new task area for the application of AI
methodology to protein crystallography. After Studying recent work in
visual scene analysis, he noted the similarities of that AI application
with the crystallographic problem of interpreting a 3-dimensional electron
density map, i.e. deriving the coordinates for a trial structure, given
the electron density function, the amino-acid sequence and the
stereochemical principles and constraints known to apply.
103

The task so defined contains most if not all of the ingredients for
the development of a knowledge-based system, in the mainstream of current
AI research. The crystallographer integrates several sources of knowledge
~~ chemical, stereochemical, crystallographic -- as he builds a model of
the protein which is consistent with the given data. He combines this
knowledge with a rich set of heuristics for focussing his attention on
promising regions of the map, for distinguishing characteristic features,
for deciding at what level of detail to stop the interpretation in
different regions, and for evaluating competing hypotheses.

The model builder’s decision-making process is dynamic and flexible,
driven at times by the need to reach specific subgoals, and at other times
by the current state of the model or special features of the data. A
computer program for interpreting the map will require a control structure
which combines both goal-driven and event-driven elements. The design of
a Suitable control structure, and the implementation of a prototype
program for performing the basic interpretive tasks are primary objectives
for our second year of research.

3. Assembly of crystallographic computing tools.

With the assistance of several crystallographers around the country
we have augmented our collection of crystallographic computing programs
and systems (i.e., integrated collections of programs). Those programs we
received and/or implemented on SUMEX include:

a) X-RAY 72. This is a large system of Fortran programs developed by
J.M. Stewart (Univ. of Md.) and others. A version written for the
DEC-10 at the University of Pittsburgh was kindly furnished to us
by Steve Ernst. We have implemented some parts of the system as
separate programs, including the Fourier transform, peak finding,
and bond length and angle calculating programs.

b) Sequence-Structure Correlator. A program which predicts alpha helix
and pleated sheet regions of a protein molecule from its amino-acid
sequence was furnished to us by Ray Salemme (Univ. of Ariz.). The
program was subsequently rewritten in both the SAIL and LISP
languages, The algorithm is based on the rules developed by Chou
and Fasman (Biochemistry, vol. 13, pp. 211-245 (1974)).

c) Oak Ridge Fast Fourier Program (ORFFP). A system of Fortran programs
for generating, analyzing and plotting Fourier maps was obtained
from Henry Levy of Oak Ridge and implemented on SUMEX. The
plotting segment is the ORTEP program written by Carroll Johnson.

d) Greer Skeletonization Program. A Fortran program for reducing an
electron density map of a protein to a set of connected line
segments, following the algorithm proposed by Greer, has been
written and is currently being debugged. This program will play a
pre-processing role in the map interpretation problem, producing a
highly abstracted representation of the map.

e) Huber Rotation and Translation Search Program. This is a Patterson
104

search program which, like our PSRCH, computes correlations of two
vector sets in the vector space representation. The program, in
Fortran, is on file but has not yet been tested at SUMEX.

f) ROTRAN. These programs were written by B.M. Craven (Univ. of
Pittsburgh) and are designed to perform rotational and
translational Patterson searching, employing a method developed by
Crowther. At present we have only a listing of the Fortran program
and instructions for its use.

D. Publications

Feigenbaum, E. A., Engelmore, R. S. and Johnson, C. K., A
Correlation Between Crystallographic Computing and Artificial Intelligence
Research, submitted for publication in Acta Crystallographica.

II. Interaction with the SUMEX~AIM resource

All program development, and most communications between the two
collaborating groups are effected on the SUMEX computer. The UCSD group
has a direct connection to SUMEX via the TYMNET computing network (UCSD
lost its ARPANET connection during the past year). Routine daily
communications now take place using the system’s message facility.
Program files are equally accessible from Stanford and UCSD, so that
either group can construct, edit or exercise the programs. Large data
files are transmitted on magnetic tape.

The greatest benefit of the interaction with the SUMEX-AIM resource
is the opportunity to share ideas, programming experience and utility
programs with other users in the community. The availability of a pool of
INTERLISP programmers, for example, has been of great assistance in our
initial efforts with the electron density map interpretation task.

Members of the SUMEX staff have also been helpful and patient in solving
some of the more mundane problems associated with any computational effort
(e.g., reading magnetic tapes produced at other computer centers).
105

IV.A.2 NATIONAL USERS

IV.A.2.a CHEMICAL SYNTHESIS PROJECT

Simulation and Evaluation of Chemical Syntheses (SECS)

W. Todd Wipke
Department of Chemistry
University of California

Santa Cruz, CA. 95064

I. Summary of Research Program

A. Technical Goals

The long range goal of this project is to develop the logical
principles of molecular construction and to use these in developing
practical computer programs to assist investigators in designing
stereospecific syntheses of complex bio-organic molecules. Previously the
focus of our work had been to represent as accurately as possible the
fundamental chemical transformations and how steric, proximity, and
electronic factors affect these reactions, going into great detail even
involving analysis of three-dimensional models. The goals for this year
focused on developing constraints to help guide the synthesis program in
growing the tree of synthetic precursors, We wanted to utilize high level
information about symmetry, and stereochemistry to set up strategies
defining preferred orderings of making and breaking bonds. We also hoped
to completely separate these strategies from the chemical transforms to
allow experimenting with changing transforms keeping strategies constant
and vice-versa. This separation was also deemed important for ease of
maintenance of large transform libraries. Finally we hoped to use these
strategies for guiding multistep lookahead so the user could see a
sequence developed in the tree at one time.

B. Medical Relevance and Collaboration

The development of new drugs and the study of how drug structure is
related to biological activity depends upon the chemists ability to
synthesize new molecules as well as his ability to modify existing
Structures, e.g., incorporating isotopic labels into biomolecular
Substrates. The Simulation and Evaluation of Chemical Synthesis (SECS)
project aims at assisting the chemist in designing stereospecific
syntheses of biologically important molecules. The advantages of this
computer approach over a manual approach are manyfold: 1) greater speed in
designing a synthesis; 2) freedom from bias of past experience and past
solutions; 3) thorough consideration of all possible syntheses using a
more extensive library of chemical reactions than any individual can
remember; 4) greater capability of the computer to deal with the many
106

structures which result; and 6) capability of computer to see molecules in
graph theoretical sense, free from bias of 2-D projection. SECS was
designed to be able to apply any kind of chemical transformation, and
because of this generality we see it finding application in biogenesis and
metabolism (see section II A below).

C. Progress and Accomplishments

The environment of this project has changed dramatically in the past
year with the move from Princeton to Santa Cruz. SECS was moved from a
hands-on environment consisting of a KA PDP-10 with an LDS-1 graphics
system and standard DEC software to a remote environment with access to a
KI PDP-10 through a GT4O graphics terminal where the host monitor system
is TENEX. The compatibility package of TENEX considerably eased the
problems of conversion. Most problems resulted from differences in file
handling, and differences in the Fortran operating system. Subtle
problems arose from the fact that our files were organized by tapes and
could not simply all be transfered to disk, because of space problems and
naming conflicts. SECS was successfully converted to TENEX and the
graphical interaction was modified for greater efficiency in our new
remote low bandwidth mode of communication.

Progress in developing strategy includes creating a general goal
list structure which allows complex logical combinations of goals to be
expressed, for example, "(break bond 2 or break bond 3) and use group 1".
Thus, instead of one set of "strategic bonds" to be broken, we now can
express strategies involving pairs of bonds, or groups or atoms. We have
succeeded in isolating strategy from the chemical transforms--strategies
can only contain expressions which refer to structural units of the
molecule or changes in those units, and may not refer to any transform by
name. The transforms have been given "character" which describes the type
of structural changes likely to occur if the transform is applied, e.g.,
cleaves ring, removes group, modifies stereochemistry, etc. The SECS
strategy module first sets up standard goal lists based on graph-theoretic
heuristics and then allows the user to view and modify the goal lists. In
this Way the user can place constraints on the syntheses generated, e. Za,
"don’t modify this ring, instead, focus your attention on this part of the
molecule." GOALTST was modified to interpret the new complex goals and
also to test the achievement or violation of a goal as early as possible.
Hence, there is testing before examination of the transform as well as
after interpretation of the transform. The net result of this work on
Strategy is that the user can very closely constrain SECS now to work only
in areas which the user decides is worthwhile, consequently fewer
precursors are generated which the user would delete.

Significant progress was made in the recognition of symmetry and use
of that information in SECS. A general algorithm based on SEMA, our
canonical naming algorithm, was developed and implemented for generating
the entire symmetry group of the molecule, using the stereochemical graph
isomorphism group. We have applied this symmetry knowledge to make SEMA
itself more efficient, and have combined it with symmetry of the
transforms to make application of a transform generate a non-redundant set
of precursors. Thus, if a double bond is introduced into cyclohexane,
107

only one cyclohexene is generated, not six. This algorithm takes into
account all stereocenters in the molecule of both double bonds and
Saturated carbon. Addition of this algorithm reduces execution time of
SECS on certain types of problems by a factor of up to six or more. We
have not yet developed heuristics for creating strategies from this
symmetry group.

Considerable improvement of aromatic chemistry has resulted from the
addition of "character words" to the aromatic transforms and
reorganization of the aromatic module. Electronic perception is only
performed now if SECS is fairly certain that aromatic chemistry will be
used and the user can prevent the MO calculations if he wishes. Work is
still progressing on implementation of strategies to control when to apply
aromatic transforms based on heuristics derived from an extensive
literature study. Many other modifications have been made to improve the
human engineering of SECS. Documentation of modules whose authors have
left the project is still continuing. Now with an expanding users group,
a good user’s manual is required and is under revision.

D. Current list of Project Publications

[1] W.T. Wipke and T. M Dyott, "Use of Ring Assemblies in a Ring
Perception Algorithm," J. Chem. Info. and Computer Sei., 15, 140
(1975).

[2] T.M. Gund, P.v.R. Sechleyer, P.H. Gund and W.T. Wipke, "Computer
Assisted Graph Theoretical Analysis of Complex Mechanistic Problems in
Polycyclic Hydrocarbons. The Mechanism of Diamantane Formation from
Various Pentacyclotetradecanes," J. Amer. Chem. Soc., 97, 743
(1975).

Papers in Preparation:

[1] W.T. Wipke and P. Gund, "Simulation and Evaluation of Chemical
Synthesis. Congestion: A Conformation Dependent Function of Steric
Environment at a Reaction Center. Application with Torsional Terms to
Stereoselectivity of Nucleophilic Additions to Ketones," J. Amer.
Chem. Soc.

[2] W.T. Wipke, G. Birkhead, and T. Brownscombe, "Correlation of
Congestion with Stereoselectivity of Olefin Epoxidation."

[3] W.T Wipke, C. Still, G. Grethe, T.M. Dyott, and P.E. Friedland,
"ALCHEM: A Language for Representing Chemical Transforms. Application
to Heterocyclic Chemistry."

C4] W. Todd Wipke and Hartmut W. Braun, "Graph-theoretical Perception of
Molecular Symmetry.

[5] W.T. Wipke, G. Smith and H. Braun, "SECS-Simulation and Evaluation
of Chemical Syntheses: Strategy and Planning," ACS Symposium

Proceedings.
108

E. Funding Status.

IBM Fellowship supporting S. Krishnan (postdoctoral)
$4000. expires September, 1976

Merck, Sharp and Dohme fellowship supporting Graham Smith
(postdoctoral)
$1000. expires July, 1976

Sandoz Unrestricted grant to support computer synthesis
$2000

Proposal submitted 1 Mar 1976 to Division of Research Resources
"Resource-Related Research: Biomolecular Synthesis"
$391,532 for three years

II. Interactions with the SUMEX-AIM Resource

A. Collaborations and Medical Use of Programs Through Networks.
Since SECS only recently has been operating on the SUMEX-AIM resource,
collaborations are just beginning. However demonstrations of SECS have
been given at the National Cancer Institute and collaboration with
Division of Chemical Carcinogenesis to try to use SECS for metabolism of
compounds to evaluate carcinogenic activity of metabolites is eurrently
under discussion. The National Library of Medicine toxicology program is
also interested in SECS and are planning to access SECS via SUMEX-AIM.
Dr. Steve Heller of the EPA and Dr, G.W.A. Milne of the National Heart
and Lung Institute are currently exploring the possibility of putting SECS
up on the Cyphernetics network. For the past year SECS has been available
over the TELENET from First Data Corporation. Squibb, Merck, FMC,
American Cyanamid, Pfizer and Searle pharmaceutical companies have used an
experimental version of SECS and have provided useful feedback to us about
problems they discovered. We expect increasing numbers of academic users
will be accessing SECS via SUMEX-AIM as they learn of its availability.

The availability of SECS on SUMEX-AIM has also served health-related
research at the University of California, Santa Cruz. For example, model
building using the SECS model builders is being performed for Professor
Edward Dratz (UCSC) to generate conformations of fatty acids isolated from
visual membranes ("Structure and Function of Visual Photoreceptors",
EY00175-05), and for Professor Howard Wang (UCSC) to study how
conformations of steroids may affect the local anesthetic - membrane
interaction ("Role of Membrane Proteins in Local Anesthetic Action,"
GM22242-01).

B. Cross Fertilization with other SUMEX-AIM projects. The SECS
project held joint research group meetings at Stanford with the DENDRAL
and AI groups to discuss common problems and research goals. This has
been very rewarding since the DENDRAL group has useful experience with
symmetry manipulation which SECS was getting into, and the SECS project
has useful experience with representing reactions, which DENDRAL/CONGEN
was getting into. These joint meetings also let the members meet in
person after having met on-line on the network. Last year’s AIM
109

Conference at Rutgers was also a valuable experience, which allowed us to
meet people interested in similar problems in different disciplines, and
it also caused us to think about what we were doing in research with some
new perspectives. We are looking forward to this year’s AIM meeting.

We find the SUMEX-AIM network very well human engineered. The
ability to leave messages on the network, and to LINK to other users on-
line for advice has been extremely useful to us, since we were new to the
TENEX operating system. But more than that, we have been able to utilize
expertise of others which our group lacked, e.g., Trisha Davis (an
undergraduate) has been writing a model builder and display program in
SAIL although there is no SAIL expertise in the SECS group--that would not
have been possible without the network communication features,

C. Critique of Resource Services. The SECS project finds the
SUMEX-AIM staff and community extremely helpful, and anxious to extend
themselves to meet our needs. SUMEX provided a leased line and modems to
us and provided TYMNET access as well. Were it not for SUMEX, this
research effort would have perished since there is no adequate computer
facility on-campus,

We do find we are short of disk space and in our grant proposal we
have requested funds for a disk drive to place at SUMEX to help resolve
this problem. The response time during the day and sometimes even later
is poor for interactive graphics, but hopefully the second processor being
installed will help alleviate that difficulty. We have an additional
problem that it is difficult to transfer files from TENEX to any other
PDP-10 with the files retaining their filenames. This problem may also be
resolved if we are able to write tapes locally from over the network.
Basically we have found that SUMEX=AIM provides a productive and
scientifically stimulating environment and we are thankful that we are
able to access the resource and participate in its activities.
110

IV.A.2.b INTERNIST (DIALOG) PROJECT

INTERNIST - Diagnostic Logie Program
Dr. H. Pople and J. Myers, M.D.
University of Pittsburgh

(Grant HEW MB-00144-01, 3 years, $167,168 last year)

I. SUMMARY OF RESEARCH PROGRAM
A. BACKGROUND AND OBJECTIVES

The principal objective of the MIS laboratory at the University of
Pittsburgh is to develop, test, and implement a computer-based diagnostic
consultation system for internal medicine. Considerable progress towards
this goal had already been made prior to our receipt, in June, 1974, of a
three-year $524,000 grant from the Bureau of Health Resources Development
to establish a "Computer Laboratory Health Care Resource’ at Pitt. At
that time, the medical data base accessed by the internist (formerly
DIALOG) program was estimated to comprise approximately twenty-five
percent of the major diseases of internal medicine, and a number of case
Studies had been run illustrating the power of the INTERNIST heuristic
process in dealing with a variety of complex clinical problems. Our
research plan envisioned a five-year development effort, intended to
yield:

(a) A four-fold expansion of the data base.

(b) Systematic field testing and evaluation of the system in actual
clinical settings.

(c) Eventual implementation, making INTERNIST available for clinical
use on a routine basis.

B. PROGRESS AND ACCOMPLISHMENTS

Shortly after award of the BHRD grant, arrangements were concluded
permitting use of the SUMEX-AIM computer resource for INTERNIST research
and development activities. Although the SUMEX-AIM computer is of the
Same genre as the one used in the original INTERNIST development work,
differences in the LISP language supported necessitated major conversion
efforts.

As mentioned in our last progress report,, the need for rapid access
to large data files motivated the design of an interface between the
INTERLISP host processor and a set of structured files containing the

entire vocabulary and network of associations comprising the INTERNIST
data base.

As of June, 1975, conversion of the data base had been completed and
114

the necessary interfaces had been established to enable INTERNIST
diagnostic programs to work with these revised structures. design and
implementation of an interactive data entry and editing system, was
completed in December, 1975, enabling expansion of the on-line data base
to its present size which is approximately 60% complete.

This newly expanded data base is currently being subjected to
extensive testing in both typical examples of disease and difficult
diagnostic problems. This procedure of systematically checking the entire
clinical data base should be completed by late June, 1976. The planned
field test and evaluation effort will then commence in early fall.

C. PUBLICATIONS

[i] Pople, H.E., "Artificial Intelligence Approaches to Computer-based
Medical Consultation; Proceedings of IEEE Intercon, 1975, New York.

[2] Pople, H.E., Myers, J.D., and Miller, R.A., "DIALOG: A Model of
Diagnostic Logic for Internal Medicine," Proceedings of Fourth
International Joint Conference on Artificial Intelligence, Tbilisi,
Georgia, USSR, 1975.

II. INTERACTIONS WITH SUMER-AIM RESOURCE

Because this year has been largely devoted to system development and
checkout activities, there has been no real opportunity to engage in any
meaningful collaboration via the communication networks associated with
SUMEX-AIM. We fully expect to exploit this attractive feature of the
resource, however, during the evaluation and field test studies planned
for the coming year.

Concerning the service provided by SUMEX-AIM, our only complaint is
the heavy loading during prime hours, which effectively prevents serious
use of the INTERNIST diagnostic programs during certain portions of the
day. We applaud and eagerly await the advent of the SUMEX-AIM dual
processor.
112

IV.A.2.¢ HIGHER MENTAL FUNCTIONS MODELING

HIGHER MENTAL FUNCTIONS MODELING (HMF)
Project Summary - 1976

Kenneth M. Colby, M.D.
Professor of Psychiatry, UCLA

(NIH MH-27132-01, 2 years, $67,000 this year)

Introduction.

One of the oldest and newest applications of computers in artificial
intelligence is the simulation of human cognitive processes. The Higher
Mental Functions project has been modelling belief systems and related
psycho-pathological delusional systems for a number of years. The
specific goal for the past two years has been to construct, test, and
validate a computer simulation of paranoid processes. The development of
such a model has clinical implications for the understanding, treatment,
and prevention of paranoid disorders.

Recently we have been focussing on the origin of beliefs in belief
systems and the criteria by which beliefs are significant to the entity we
are modelling; i.e. the motivation for the beliefs. The motivation for an
entity ’s purposive behavior is based in its affect or emotion system. We
are currently formulating a theory of the motivational influence of affect
on conative (volitional) and cognitive (inferential) processes, with the
intent of implementing this background theory in a simulation model. By
specifying the underlying theory of motivation we hope to make the theory
of paranoia more explicit and the paranoid simulation model more adequate.

The strategy of computer simulation of mental processes can be
characterized roughly by three phases: (1) identification and critical
description of non-random patterns occuring in the phenomena under study,
(2) explanation by postulation of underlying mechanisms which generate,
produce, or are responsible for the non-random patterns, and (3)
validation by repeated attempts to test the reality of the proposed theory
or model. The construction and use of simulation models of mental
processes closely parallels model-building in other sciences. An attempt
is made to reproduce the relevant features of the patterns under study.
This attempt produces simplification and idealization of the phenomena.
Simplification implies that only centrally relevant variables are chosen
for representation in the model. Idealization implies that exact classes
and perfect properties are assumed in the implementation of the model.
Still, the model can provide an explanation of underlying mechanisms which
is useful in understanding and interpreting the observed patterns of
phenomena. Finally, a model can be used in practical situations for
prediction, and for providing suggestions to clinicians for potential
control and change in the phenomena. Such technological purposes are
important for models of mental disorders since the long-range goal of
113

mental health research is to prevent or reduce conditions of human mental
suffering.

Simulations of cognitive processes are difficult for a number of
reasons:

(1) The underlying generative mechanisms of human behavior are
inaccessible to direct observation, and must instead be postulated
as hypothetical constructs that may (possibly) account for the
phenomena.

(2) A simulation must take into account the rich background of
information that a human has available to apply to a contemporary
situation.

(3) Human beings have internal needs which are a function of the
immediate past and present, as well as the long-term past of the
individual. These needs and past experiences color the human’s
response to the contemporary situation.

(4) Human linguistic behavior is the richest source of data for
exploring cognitive processes as well as the most complex and
therefore desirable behavior to simulate. At the same time, it is
difficult due to the variety of behavior possible and the variety of
explanations possible for one specific linguistic action.

(5) Once a simulation model is performing, it is difficult to show the
subtleties of the model’s generative mechanisms. Instead, some
attempt is usually made to reveal the internal workings of the model
and appeal to the observers” intuition and/or introspection.

Our overall purpose, then, is to develop theories of human mental
processes, specifically psycho-pathological processes, and to implement
these theories in computer simulation models. The simulation models help
formalize and explicate their associated theories by forcing them into a
Single notation and requiring the theory-builders to specify the details
of the theory. In addition, the models provide a testing ground for
validating the theory. On the basis of such theories, we hope to explain
the origin of psycho-pathologies and offer principles on which to base
treatment and prevention.

Technical goals.

A. Expand the theory of paranoia. The theory implemented in the
current model (PARRY2), the humiliation theory, postulates that
informational inputs from other people activate a belief in the self’s
inadequacy. The paranoid mode then consists of strategies which forestall
or ward off an impending negative affect experience of humiliation by
negating this belief that the self is wrong and asserting the belief that
the self is being wronged by others. The theory provides generative
mechanisms for explaining the expression of a delusional system by a
paranoid person, the chronic distress felt by paranoid persons, and for
the sudden and extreme displays of fear and anger in interactions with
other people.
114

We plan to extend the theory in two ways:

(1) To cover other paranoid phenomena, such as delusions of grandeur,
the transformation of counter-evidence to evidence supportive of
delusions, and retrospective misinterpretation of input expressions.

(2) To explain the genesis of paranoid patterns of thought; e.g., the
manner in which: (a) normal strategies for dealing with the shame-
humiliation affect are ineffective and paranoid strategies develop,
(b) strategies are selected as being appropriate and are reinforced
when they are successful, and (c) persecutory delusional systems
develop and expand to include much of the paranoid personality’s
cognitive processing.

B. Expand the background theory and model of the motivation of
cognitive processes. The more important characteristics for the model to
have are:

(1) The top-level processes of the model should be purposive intentional
processes guided by the affect system, rather than a question-answer
loop or facsimile.

(2) Every action that the model performs should be motivated by an
intention. These intentions may be explicit in the case of goals,
or implicit in the case of an action appropriate to a situation but
with no explicit end state represented.

(3) Each belief and intention should have an associated measure of its
Significance to the entity. The criteria for measuring the
significance are based in the affect system.

(4) The model should have a number of coping mechanisms for avoiding or
coping with distressing situations. These mechanisms can be
reinforced or discarded as they are proved to be more or less useful
to the model.

(5) The model should be able to change over time to show the development
of psycho-pathologies. The most direct form of change is to the
measures of significance attached to the model’s beliefs and
intentions,

C. Implement the theories in a simulation model. We hope to
construct a model in such a way that the theories of motivation and
paranoia can be represented explicitly, and therefore be open to
inspection and modification. In addition, since we model paranoid
behavior as expressed through linguistic actions, we hope to develop
adequate natural language understanding programs for recognition and
response in dialog situations.

D. Develop further techniques and methods for Simulating cognitive
processes. Specifically, we plan to explore human communication through
natural language in dialog situations and human natural language
interfaces with computers. Also, we will extend our previous attempts at
finding stronger and more sophisticated tests for validation studies. Our
115

results should be applicable to other simulations of human mental
activities.

Medical relevance,

The simulation model of paranoid processes that we are implementing
has implications for the understanding, treatment, and management of
paranoid disorders.

The shame-humiliation theory and its model suggest that paranoid
phenomena be viewed as a consequence of intentionalistic information-
processing strategies which attempt to avoid or minimize the distressing
experience of humiliation. In trying to understand what is going on ina
paranoid patient at a symbol-processing level, this perspective directs
clinicians to look for humiliating and shame-engendering situations in the
patient ’s experience. These may consist of a single, encompassing
humiliating situation such as a demeaning job, or a series of esteem-
damaging defeats such as repeated failures in disappointing love affairs.

Since activation of intense shame is posited to be the core process
in paranoid disorders, implications for treatment involve trying to modify
this central mechanism in some way. One method is to change the patient's
distressing belief in his own inadequacy by exploring topics involving
shame, esteem, and self-censure. Another is to desensitize the paranoid
patient to shame experiences through behavior therapy involving a graded
hierarchy of imagined distress situations and countering procedures.

These treatment procedures may be deduced directly from mechanisms in the
model, and the theory used to predict the outcomes of such procedures.

For management of the disorder, the model predicts that removal from
situational humiliation, as in hospitalization, allows for repair from
breakdowns occurring under repeated activations of shame-humiliation
beliefs. Also, if the patient returns to unchanged situational
humiliation, as in a distressing home life, he risks a relapse.

Current Status.

PARRY2 was completed a year ago and has been available for
interviewing and validation tests on the SUMEX system for the past year.
We are now in the process of writing a new version, PARRY3, incorporating
the theoretical constructs presented in this report. The new version
(which is being completely rewritten) contains mechanisms implementing the
characteristics of models mentioned above with the exception of the
ontogenesis of psycho-pathologies. In addition, it contains a new
language recognizer capable of combining pattern-matching rules with
parsing techniques, and explicit rules for recognizing and interpreting
elliptical expressions in dialogs.

Current publications.

[1] Colby, K.M. Artificial Paranoia, Pergamon Press, New York, 1975.
116

(2] Faught, W.S. Affect as Motivation for Cognitive and Conative
Processes, IJCAI Proceedings, September, 1975.

C3] Colby, K.M., Hilf, F.D., Wittner, W.K., Parkison, R.C. and Faught,
W.S., A Note on Estimating Improvement in a Computer Simulation of
Paranoid Processes. UCLA Department of Psychiatry, Memo ALHMF-2, April
1975.

[4] Parkison, R.C., Colby, K.M., and Faught, W.S., Conversational Language
Comprehension Using Integrated Pattern-Matching and Parsing. UCLA
Department of Psychiatry, Memo ALHMF-5, April 1976.

[5] Colby, K.M. Clinical Implications of a Simulation Model of Paranoid
Processes. Archives of General Psychiatry, 1976.

[6] Faught, W.S., Colby, K.M., and Parkison, R.C., Inferences, Affects,
and Intentions in a Model of Paranoia. Cognitive Psychology, 1976.

Funding Status.

Grant NIMH , 2 years, $67,000 this year.

Interactions with the SUMEX-AIM resource.

The SUMEX-AIM resource and its associated network connections make
possible the merging of artificial intelligence techniques and technology
in psychiatry and the resources of a west-coast center for psychiatric
studies, the Neuro Psychiatric Institute (NPI) at UCLA. Access to SUMEX
from the NPI has brought a new source of questions and viewpoints to
research in mental disorders, as well as an opportunity for the model-
builders to interact with clinicians in elaborating details of paranoid
phenomena. In addition, the current simulation model is being explored
for use as a training device for medical students and residents in the
Department of Psychiatry.

Critique of resource services.

The resource itself has provided excellent facilities and service
for our research needs. We have had almost no trouble developing
Simulations due to the SUMEX system itself, and the cooperation of the
SUMEX staff has been excellent.

In spite of this, problems have arisen with our use of the resource,
due almost entirely to the network aspect of our access. (We connect to
SUMEX through both the ARPA net and TYMNET.) We see the problems of
network use of the SUMEX-AIM facility as falling into four broad
categories:

(a) the keyhole effect. The slow terminal rates typical of a network
connection (due to phone lines or local computer delay) force the
user to peer at his files and communicate with the computer through
117

a small data channel. Additionally, the network user may not have
direct access to a high-speed printing device for listing the day’s
(or even the week’s) work.

(b) the dropped connection, Typically, the network and/or local
connection of the user to the net can drop, leaving the user’s job
in a dormant (non-running) state, forcing the user to reconnect (and
perform the typically elaborate reconnect procedure) before his job
will run again.

(ce) slower interactive computer response due to network and local
hardware, and subsequent greater amount of time necessary to get any
work done.

(d) the difficulty of lobbying for system changes and/or additions from
a distance. We are made acutely aware of this fact whenever notices
are given over the system for classes or seminars explaining new
system features.

Policies and programs that might be useful:

(a) more system programs designed with the network user in mind, such as
status programs to report the status of detached or batch-run jobs,
and cleaner detach and attach programs for reestablishing dropped
connections.

(b) a service-level advantage to non-local users to put them on a more
equal footing with local users.

(c) a special effort to elicit and implement improvements for network

access.

In general, we have found the system reliable, and the staff
courteous and helpful.
118

IV.A.2.d LANGUAGE ACQUISITION MODELING PROJECT

Language Acquisition Modeling (ACT)

Dr. John Anderson
(Univ. of Michigan)

(Grant NIMH $20,000 this year)
I. Summary of Research Program

A. Technical goals:

To develop a production system that will serve as an interpreter of
the active portion of an associative network. To model a range of
cognitive tasks including memory tasks, inferential reasoning, language
processing, and problem solving. To develop an induction system capable
of acquiring cognitive procedures with a Special emphasis on language
acquisition,

B. Medical relevance and collaboration:

1. The ACT model is a general model of cognition. It provides a
useful model of the development of and performance of the sorts of
decision making that occur in medicine.

2. The ACT model also represents basic work in AI. It is in part an
attempt to develop a self-organizing intelligent system.. As such it is
relevant to the goal of development of intelligent artificial aids in
medicine,

C. Progress and Accomplishments:

ACT provides a uniform set of theoretical mechanisms to model such
aspects of human cognition as memory, inferential processes, language
processing, and problem solving. ACT’s knowledge base consists of two
components, a propositional component and a procedural component. The
propositional component is provided by an associative network encoding a
set of facts known about the world. This provides the system’s semantic
memory. The procedural component consists of a set of productions which
operate on the associative network. ACT’s production system is
considerably different than many of the other currently available systems
(e.g., Newell’s PSG). These differences have been introduced in order to
create a system that will operate on an associative network and in order
to accurately model certain aspects of human cognition.

A small portion of the semantic network is active at any point in
time. Productions can only inspect that portion of the network which is
active at the particular time. This restriction to the active portion of
119

the network provides a means to focus the ACT system in a large data base
of facts. Activation can spread down network paths from active nodes to
activate new nodes and links. To prevent activation from growing
continuously there is a dampening process which periodically deactivates
all but a select few nodes. The condition of a production specifies that
certain features be true of the active portion of the network. The action
of a production specifies that certain changes be made to the network.
Each production can be conceived of as an independent "demon". Its
purpose is to see if the network configuration specified in its condition
is satisfied in the active portion. If it is, the production will execute
and cause changes to memory. In so doing it can allow or disallow other
productions which are looking for their conditions to be satisfied. Both
the spread of activation and the selection of productions are parallel
processes whose rates are controlled by "strengths" of network links and
individual productions. An important aspect of this parallelism is that
it is possible for multiple productions to be applied in a cycle through
the set of productions. Much of the early work on the ACT system was
focused on developing computational devices to reflect the operation of
parallel, strength-controlled processes and working out the logic for
creating functioning systems in such a computational medium.

We have successfully implemented a number of small-scale systems
that model various psychological tasks in the domain of memory, language
processing, and inferential reasoning. A larger scale effort is underway
to model the language processing mechanisms of a young child. This
includes implementation of a productions system to analyze linguistic
input, make inferences, ask and answer questions, etc. Also a great deal
of effort is being given to developing learning mechanisms that will
acquire and organize the productions for this language processing. This
learning program attempts to acquire procedures from examples of the
computations desired of the procedures. For instance, the program learns
to comprehend and generate sentences by being given sentences and picture
representations of the meaning of the sentences(actually hand encodings of
the pictures), Although this effort is focused on induction of linguistic
procedures, the hope is to develop a general model of induction of
cognitive procedures and not to place any language-specificity into the
induction procedures.

D. Current list of project publications:

[1] Anderson, J. R. Induction of augmented transition networks.
Cognitive Science, 1976, in press.

[2] Anderson, J. R. Language, Memory, and Thought. Lawrence Erlbaum
Assoc., 1976, in press.

{3] Anderson, J. R. ,Klein, P., and Lewis, C. Language processing by

production systems. To appear in P. Carpenter and M. Just(Eds.)
Cognitive Processes in Comprehension,Lawrence Erlbaum Assoc.

E. Funding Status:
120

The research is currently being funded by a grant from NIMH for
computer simulation of language acquisition. The level of funding for the

year beginning May 1, 1976 has yet to be determined. It was $20,000 for
the past year.

II. Interactions with SUMEX=AIM Resource.

Our period with the project has been too short to develop any
Significant interactions. The ACT program is currently being made a
system which will be available to members of the SUMEX-AIM community,
121

IV.A.2.e MEDICAL INFORMATION SYSTEMS LABORATORY

MISL Project

Dr. B. McCormick and M. Goldberg, M.D.
(Univ. of Illinois at Chicago Circle)

(Grant HEW MB-00114-02, 3 years, $248,793 this year)

I) Summary of research program

A) Technical goals

The major goals of the Medical Information Systems Laboratory fall

into three broad categories, described briefly as follows:

1)

2)

3)

Construction of a database in ophthalmology. This will provide a
trial setting for clinical decision-making research. Four major
activities are involved: implementation of a clinical data network;
design and on-going development of an Eye Outpatient Index; computer
systems development; and installation of a glaucoma clinic satellite.

Network-compatible database design. This will provide cost-effective
distributed data management for clinical records. Current projects
include: design of an intelligent coupler for the database system;
continuing design of various levels of database software (a
relational algebraic language -- RAIN, disk controller, database
skeleton); large database / database software compatibility; design
of a separate database computer; design of database network software.

Clinical decision support

a. Construction of a consultation system for use in the diagnosis and
treatment of retinal/choroidal diseases. The immediate goal is
the development of a system for giving advice about four diseases
prevalent at the University of Illinois Eye and Ear Infirmary:
histoplasmosis, central serous retinopathy, diabetic retinopathy,
and sickle cell disease. Besides actual construction of a system,
the project is concerned with theory of diagnosis and knowledge
acquisition methodology.

b. Interface between a pattern recognition system for Structured
Analysis of the Retina (STARE) and the diagnostic model used by
the retinal/choroidal consultation system.

ce. Initial development of a causal model for motility.

d. Further inter-institutional communications in disease model theory
and development.

B) Medical relevance and collaboration
122

We have chosen to explore inferential relationships between analytic
clinical data and the natural history of glaucoma and selected
retinal/choroidal diseases both in treated and untreated form. These
investigations are intended to provide much more than a simple excursion
of computer technology into Ophthalmology. They address clinical problems
of national interest, as indicated in the Report of the National Eye
Advisory Council Vision Research Program Planning Committee (DHEW
publication No. NIH~75-644).,

Glaucoma, one of the major causes of blindness in the United States
today, is difficult to diagnose in its early stages. Some recent evidence
indicates that enlargement of the optic nerve cups may be the first sign
of glaucoma’s damage to the eye. One of the goals of the present project
is the application of a newly developed technique for quantitative
analysis of the optic nervehead. The technique is sufficiently simple to
permit wide-spread adoption. If this technique is successful in
identifying very early glaucomatous disk changes, it should permit
institution of therapy at a very early stage, and thereby prevent serious
glaucomatous damage from being done to the eye.

Diabetic retinopathy is another principal cause of blindness. Very
little is known about its pathophysiology, and there are many gaps in our
knowledge of its natural course, The present study is designed to elicit
new information about this disease, using a series of new diagnostic tools
which have been developed as part of a system of computerized retinal
image analysis. The need here is great, because at present there is no
proven satisfactory treatment.

Sickle hemoglobinopathy can cause ocular changes that lead to loss
of vision and even total blindness. Little is known about the natural
history of this problem, particularly during its early stages. The
present project is ideally suited to assist in this study; the nation’s
first Sickle Cell Clinic has been established at the Illinois Eye and Ear
Infirmary -- the site of our Ophthalmic Database System,

At present there is great demand in the United States for improved
efficiency in the delivery of medical care. Two ways that this can be
accomplished are: 1) by increased use of paramedical personnel to perform
jobs currently being performed by physicians, and 2) by use of automated
equipment to perform tasks previously performed by the physician. In the
current project we utilize both these methods in the screening of new
ophthalmic patients at the Illinois Eye and Ear Infirmary. If we can show
that these methods are not only feasible, but also improve the efficiency
and reliability of patient care, then a major contribution to ophthalmic
care for patients in large ambulatory care centers around the country will
have been made.

Modeling of clinical decision-making is best carried out in intimate
association with an extensive referral clinic where a sufficient patient
population can be accumulated to provide an adequate biostatistical
Sample. In the prescribed setting, the Illinois Eye and Ear Infirmary,
the Medical Information Systems Laboratory has access to:

- clinical expertise provided by a house staff of 45;
123

- 28 residents, all of whom are required to undertake some
research work as a requirement of their appointment, are
available to explore latent contingencies of the database;

- an indigenous, relatively stable population (25% white,
75% black and other minorities) of a medically underserved
portion of the inner city of Chicago; the clinic provides
ophthalmic services to 50,000 patients per year.

Commonality of the diseases being studied assists construction of an
adequate biostatistical sample. Roughly 2% of the general population
exhibit symptoms of diseases treated at each of three specialty clinics
(Glaucoma, Retina, and Motility), or allowing for multiple presenting of
symptoms, approximately 5% by population.

Besides a strong clinical research orientation, the Medical
Information Systems Laboratory brings to the study of disease a history of
successful engineering-medical collaboration. MISL’s sister project,
"Image Processing in Clinical Ophthalmology," lists the development of the
digital television ophthalmoscope as one of its achievements. This device
will be a major source of clinical data for our Ophthalmic Database.

C) Progress and accomplishments (of the Clinical Decision
Support activity only)

Interaction has continued over SUMEX-AIM with the authors of the
Weiss/Kulikowski glaucoma modeling program. We are now entering cases
into the glaucoma system at the rate of approximately 5 per week.

Work on the consultation system for retinal/choroidal diseases has
progressed along two fronts. While interviewing an expert diagnostician,
in order to build a knowledge base for the four diseases mentioned above,
we have been piecing together a theory of diagnosis for ocular fundus
diseases. We have attempted to incorporate pieces of the expert ‘s
knowledge ina "fuzzy" diagnostic model, based partly on multiple-cue
probability theory, partly on fuzzy set and confirmation theory. The
framework of the model is a hierarchy of disease categories, each with a
significance tempered by functions built into the system. Our efforts
have centered on the acquisition of categories for histoplasmosis, central
serous retinopathy, and diabetic retinopathy, but should soon also include
sickle cell disease.

Our experience in interviewing experts has pointed the way to a
knowledge acquisition methodology that is compatible with our thoughts on
diagnostic reasoning. Specifically we plan to store, for each disease
category, a representation for the contexts (or frames) in which the
disease’s attributes apply. We conceive of a disease category as a
"sphere" (embodying a structural model of the disease) in a hyperspace
defined by dimensions on attributes. The significance of (or "belief in")
the model of disease is modified by interactions between attributes.
During acquisition, and after contexts have been defined, we can simulate
clinical situations for the benefit of our expert, who indicates his level
of belief in the disease model in the given situation. This we plan to do
124

with the help of plasma panels, for graphical presentations of the
relevant contexts comprising each situation. This approach is especially
convenient for specifying typical "default" situations, and for modeling
the time course of disease (in terms of modifications on attributes).

Presently, while we continue interviews with our expert, we are
formalizing our diagnostic model and expect shortly to finish an in-depth
report.

D) Current list of project publications

C1] Chang Shi-Kuo, O’Brien M., Read J., Borovec R., Cheng W. H., Ke Jd. S.
(1976) Design considerations of a database system ina clinical
network environment. Accepted for 1976 National Computer Conference.

[2] Chang S. K. (1975) Preliminary report on the implementation of a
relational data base management system with structurally decomposed
relations. MISL internal report M.D.C. 1.1.3.

[3] Chang S. K. and McCormick B. H. (1975) An intelligent coupler for
distributed database systems. MISL internal report M.D.C. 1.1.7.

[4] Malone J. (1975) User’s guide to uniclass cover synthesis. MISL
report M.D.C. 4.4.1,

[5] Malone J. (1975) Addendum to AQVAL/1 (AQ7), part 1: User’s guide and
program description. MISL report M.D.C. 4.4.1.

[6] Manacher G. K. (1975) On the feasibility of implementing a large
relational data base with optimal performance on a minicomputer.
Proceedings of the International Conference on very large data bases,
Framingham, Mass,

[7] McCormick B. H. and Wilensky J. (1975) Clinical knowledge
acquisition: design of a relational data base in ophthalmology. To
be published in Proceedings Second Annual Medical Informations
Systems Conference, Urbana, I1l.

[8] McCormick B. H., Goldberg M. F., and Read J. S. (1974) Clinical
decision-making: design of a data base in ophthalmology. Proceedings
First Annual Medical Information Systems Conference, Urbana Ill.

[9] Michalski, Ryszard S. (1975) On the selection of representative
samples from large relational tables for inductive inference. MISL
internal report M.D.C. 1.1.9.

[10] Murata T. (1976) Equations governing the behavior of E-nets. MISL
internal report.

[11] Murata T. (1975) State equation, controllability, and maximal
Matchings of petri nets. MISL report M.D.C. 1.1.10.

[12] Murata T. and Church R. W. (1975) Analysis of marked graphs and petri
nets by matrix equations. MISL report M.D.C. 1.1.8.
125

{13] Vere S. A. (1975) Induction of concepts in the predicate calculus.
Proceedings of the Fourth International Joint Conference on
Artificial Intelligence vol 1., Tbilisi, U.S.S.R.

[14] Vere S. A. (1975) Relational productions systems. MISL internal
report M.D.C. 1.1.5.

E) Funding status

Year 02 -- 6/30/75 - 6/29/76 : $248,793.
Year 03 -- 6/30/76 - 6/29/77 : $228,000.
II) Interactions with the SUMEX-AIM Resource
A) & B) Collaboration, cross-fertilization
Most of our interaction of late has involved the Glaucoma Network
fostered by the Rutgers Computers in Biomedicine group. This network has
made it especially convenient for our expert in glaucoma, Dr. Jacob

Wilensky, to maintain close contact with investigators around the country.

In addition, monitoring of SUMEX-AIM system messages has helped us
keep abreast of developments in other projects. We have come to rely on
this facility as a vital source of up-to-date information.

C) Critique of resource services

In our view SUMEX-AIM services are excellent. We have been very

pleased with the prompt and personal attention given to our requests by
the resource staff.