Copy Link
Add to Bookmark
Report
AIList Digest Volume 4 Issue 082
AIList Digest Saturday, 12 Apr 1986 Volume 4 : Issue 82
Today's Topics:
Bibliography - Technical Reports #5
----------------------------------------------------------------------
Date: WED, 10 JAN 84 17:02:23 CDT
From: E1AR0002%SMUVM1.BITNET@WISCVM.WISC.EDU
Subject: Technical Reports #5
%R DCS-TR-90
%I Rutgers University, Department of Computer Science
%D 5/80
%T Knowledge-based learning, an Example
%A C. V. Srinivasan
%D 1/82
%K AI04
%X How may a machine "learn" from examples of situations that are
presented to it? What may constitute the "knowledge" of a set of such
situations? How should the examples be presented to the machine? Are
there general principles which a machine can use to acquire the
knowledge automatically by examining the examples presented to it, and
to use the knowledge so obtained to solve problems in a domain? These
are the general concerns of my research.
%R CBM-TR-138
%D 5/84
%T Hardware Fault Diagnosis & Expert Systems
%A Allen Ginsberg
%D 5/84
%K AI01 AA04 AA21
%I Rutgers University, Department of Computer Science
%X Recent research in Expert Systems has begun to deal with problem
domains that do not fit into the "classification problem" mode, the
latter being the sort of problems that have been most amenable to
Expert System technology. Hardware Fault Diagnosis(HFD) is an example
of such a problem domain. Problems in HFD typically involve a
"localization" problem as a component, i.e., @i[where] is the location
of the fault? This paper takes a critical look at some current work
in HFD, viz. Genesereth, Davis, with a view towards determining the
differences between classification and localization problems that are
likely to necessitate new approaches to knowledge representation and
acquisition if Expert Systems are to be successful in such a domain.
%R CBM-TR-139
%I Rutgers University, Department of Computer Science
%D 5/84
%T Localization Problems and Expert Systems
%A Allen Ginsberg
%K AI01
%X Expert systems approaches to problem solving have recently had
enormous success and influence in the field of AI. The most
successful of these systems tend to deal with a certain kind of
problem type which have been called "classification problems." Very
recently, we have seen the emergence of a number of expert systems
that deal with a different category of problem, a category that I will
call "localization problems." The purpose of this paper is to
characterize this class of problems, contrast it with the
classification problem category, give some examples of localization
problems, and suggest some new avenues for expert system research
dealing with problems in this category.
%R CBM-TR-140
%I Rutgers University, Department of Computer Science
%D 5/84
%T Investigations in the Mathematical Theory of Problem Space
Representations and Problem Solving Methods
%A Allen Ginsberg
%K AI09
%X In this paper I address the issue of how a system that has the ability
to do problem solving and planning - in the sense of being in
possession of generalized schemas or templates for carrying out these
activities - can know whether a particular type of planning or, if you
will, problem solving strategy, is a "good" one to employ in solving
problems in a particular domain? It seems to me that, in general, in
order to make such judgements in a reasonable fashion a problem solver
must either be in possession of some general theoretical facts
concerning the nature and structure of problem types, i.e., a theory
of problem types, or at the very least, have been programmed by
someone having such a theory. This paper is a step in the direction
of constructing such a theory.
.sp 1
The structure of the paper is as follows. First I discuss the nature
of problem solving and planning in general, and give a preliminary
description of a particular planning template. Next I describe and
illustrate a mathematical framework within which one can formulate
problem representations. Finally I deal with the question of what
facts about the structure of a problem representation are relevant to
the determination of whether or not the aforementioned planning
template is applicable to the problem at hand.
%R CBM-TR-141
%D 5/84
%I Rutgers University, Department of Computer Science
%T Representation & Problem Solving: Theoretical Foundations
%A Allen Ginsberg
%X The word "representation" and its cognates is probably the most
popular word in AI today. If anything qualifies as "the fundamental
assumption of AI," it is probably the view that intelligence is
essentially the ability to construct and manipulate symbolic
@i[representations] of some "reality" in order to achieve desired
ends. Furthermore, probably every researcher in AI would agree that
the key to AI's success lies with the general area known as "knowledge
representation." This point of view has been buttressed not only by
the failures of early "general purpose" AI systems, but much more so
by the recent success of expert systems. The philosophy behind the
expert systems approach is one that has, rightfully come to infect the
entire field of AI: intelligence essentially depends upon the ability
to @i[represent] and store a potentially vast amount of knowledge in ways
that enable it to be easily accessed and utilized in the performance
of various tasks. The key concept here is @i[representation].
%X Given the fact that AI has come to embrace these doctrines, and the
likelihood that there is a good deal of truth in them, it is incumbent
upon us to examine their foundations, for better or for worse. It
would be nice to have answers to questions such as What is a
representation?, When are two or more representations representations
of the same or different real world situations?, What are the ways in
which representations can be "manipulated?" It would be even nicer if
the answers to such questions were provided by a general formal theory
of representation. In this paper I attempt to lay some of the
groundwork for such a theory, with emphasis on the role of
representation in problem solving.
%R CBM-TR-142
%D 5/84
%T A Model for Automated Theory Formation for Problem Solving Systems"
%A A. Ginsberg
%X The goal of this paper is to contribute towards the understanding and
eventual mechanization of the processes whereby an @i[intelligent]
problem solver @i[learns] to improve its performance in a given task
domain by formulating and using @i[theories] regarding that domain. In
order to achieve this goal it is necessary for us, as designers of
such a system, to have a fairly good idea of a) the various sorts of
knowledge that are required for a problem solver to acquire new
knowledge that will hopefully improve performance, and of b) how each
of these types or sources of knowledge comes into play in this
process. In this paper I give an abstract description of the domains
of knowledege required for theory formation, and also illustrate the
ideas with a concrete example. The type of system contemplated in
this paper incorporates ways of structuring background knowledge that
are natural and will, I believe, prove to be useful in designing
self-improving AI programs.
%X In this paper I address the issue of how a system that has the ability
to do problem solving and planning - in the sense of being in
possession of generalized schemas or templates for carrying out these
activities - can know whether a particular type of planning or, if you
will, problem solving strategy, is a "good" one to employ in solving
problems in a particular domain? It seems to me that, in general, in
order to make such judgements in a reasonable fashion a problem solver
must either be in possession of some general theoretical facts
concerning the nature and structure of problem types, i.e., a theory
of problem types, or at the very least, have been programmed by
someone having such a theory. This paper is a step in the direction
of constructing such a theory.
%X The structure of the paper is as follows. First I discuss the nature
of problem solving and planning in general, and give a preliminary
description of a particular planning template. Next I describe and
illustrate a mathematical framework within which one can formulate
problem representations. Finally I deal with the question of what
facts about the structure of a problem representation are relevant to
the determination of whether or not the aforementioned planning
template is applicable to the problem at hand.
%R CBM-TR-143
%D 5/84
%I Rutgers University, Department of Computer Science
%T A Knowledge Representation Framework for Expert Control of
Interactive Software Systems
%A Apte, C.
%A S. Weiss
%K AI01 AA08
%X Expert problem solving strategies in many domains make use of
detailed quantitative or mathematical techniques coupled with
experiential knowledge about how these techniques can be used to solve
problems. In many such domains, these techniques are available as part
of complex software packages. In attempting to build expert systems
in these domains, we wish to make use of these existing packages, and
are therefore faced with an important problem: how to integrate the
existing software, and knowledge about its use, into a practical
expert system. We define a framework of a @i[hybrid model] for
representing problem solving knowledge in such domains. A hybrid
model consists of a @i[surface] and a @i[deep] model. The surface
model is the production rule-based expert subsystem that is driven by
domain specific control and interpretive knowledge. The deep model is
the existing software, reorganized as necessary for its interpretation
by the surface model. We present an outline of a specialized
form-based system for acquisition and representation of expert
knowledge required for this hybrid modeling.
%R CBM-TR-144 (THESIS)
%I Rutgers University, Department of Computer Science
%D 9/84
%T A Framework for Expert Control of Interactive Software Systems
%A C.V. Apte
%K AI01 AA08
%X Expert problem-solving strategies in many domains require the use of
detailed mathematical techniquers coupled with experiential knowledge
about how and when to use the appropriate techniques. In many of
these domains, such techniques are made available to experts in large
software packages. In attempting to build expert systems for these
domains, we wish to make use of these existing packages, and are
therefore faced with an important problem: how to integrate the
existing software, and knowledge about its use, into a practical
expert system. The expert knowledge is used, in dynamic selection of
appropriate programs and parameters, to reach a successful goal in the
problem-solving. This kind of expert problem-solving is achieved
through two interacting bodies of knowledge; problem domain knowledge,
and knowledge about the programs that comprise the software package.
%X This thesis describes the framework of a @i[hybrid expert system] for
representing problem-solving knowledge in these domains. This hybrid
system may be characterized as consisting of a @i[surface] model and a
@i[deep] model. The surface model is a production-rule based expert
subsystem that consists of heuristics used by an expert. The deep
model is a collection of methods, each parameterized by a set of
controlling and observed parameters. The method and their results are
reasoned about using their parameter sets. The existing software is
reorganized as necessary to map it into the deep model structure of a
hybrid system. This framework has evolved out of an effort to build an
expert system for performing well-log analysis (ELAS - @i[Expert Log
Analysis System]). A generalized expert-system building methodology
based upon principles drawn from ELAS is introduced. The use of
@i[method-abstractions] in assembling a hybrid system is discussed.
The notion of @i[worksheet-reasoning] is defined, and discussed.
%R CBM-TR-145 (THESIS)
%D 10/84
%T Shift of Bias for Inductive Concept Learning
%A Paul E. Utgoff
%K AI04
%X We identify and examine the fundamental role that bias plays in
inductive concept learning. Bias is the set of all influences,
procedural or declarative, that causes a concept learner to prefer one
hypothesis to another. Much of the success of concept learning
programs to date results from the program's author having provided the
learning program with appropriate bias. To date there has been no
good mechanical method for shifting from one bias to another that is
better. Instead, the author of a learning program has himself had to
search for a better bias. The program author manually generates a
bias, from scratch or by revising a previous bias, and then tests it
in his program. If the author is not satisfied with the induced
concepts, then he repeats the manual-generate and program-test cycle.
If the author is satisfied, then he deems his program successful. Too
often, he does not recognize his own role in the learning process.
.sp 1
Our thesis is that search for appropriate bias is itself a major part
of the learning task, and that we can create mechanical procedures for
conducting a well-directed search for an appropriate bias. We would
like to understand better how a program author does about doing his
search for appropriate bias. What insights does he have? What does
he learn when he observes that a particular bias produces poor
performance? What domain knowledge does he apply?
.sp 1
We explore the problem of mechanizing the search for appropriate
bias. To that end, we develop a framework for a procedure that shifts
bias. We then build two instantiations of the procedure in a program
called STABB, which we then incorporate in the LEX learning program.
One, called "constraint back propagation" uses analytic deduction. We
report experiments with the implementations that both demonstrate the
usefulness of the framework, and uncover important issues for this
kind of learning.
%R CBM-TR-146
%I Rutgers University, Department of Computer Science
%D 5/85
%T A Framework for Representation of Expertise in Experimental Design
for Enzyme Kinetics
%A Von-Wun Soo
%A Casimir A. Kulikowski
%A David Garfinkel
%K AA10 AI01
%X In this paper, we present part of our current research on expert
systems in enzyme kinetics. Because of the richness and diversity of
the problem solving knowledge required in this domain, we have found
it to be an excellent vehicle for studying issues of knowledge
representation and expert reasoning in AI. Biochemical experimental
design, the focus of this paper, is a major problem solving activity
of the enzyme kineticist that has not been explored by expert systems
researchers. Their problem solving expertise can usually be described
as the application of a sequence of methods. In designing a
complicated biochemical experiment, the experimenter has several
methods to choose from at any stage. These methods are represented as
computer programs which can be organized into a hierarchy. This paper
proposes a structure for these problem solving methods and an expert
consultation system for experimental design.
.sp 1
We have found that problem solving expertise in experimental design
can be divided into three phases. In the first phase, we deal with
problems of selecting the experimental methods that satisfy an
experimenter's goal, given certain postulated models. The
experimental conditions and optimal design points can be derived if
the model is given and the goal and the assumptions of the optimal
design criterion are satisfied. In the second phase, we deal with the
problems of preparing an enzyme assay. The interactions among
experimental conditions and other influencing factors must be
carefully controlled so that the correct concentration of a given
species can be calculated. In the third phase, we face the problem of
analyzing and interpreting the experimental data and recommending
further refinement of the experiment.
------------------------------
End of AIList Digest
********************
