G. Annotated Bibliography
Betts,
Bill, David Burlingame, Gerhard Fischer, Jim Foley, Mark Green, David
Kasik, Stephen T. Kerr, Dan Olsen, James Thomas.
Goals and Objectives for User Interface Software, Computer Graphics,
v21, no 2,1987
Bournique,
Richard & Siegfried Treu. Specification
and Generation of Variable, Personalized Graphical Interfaces.
International Journal of Man-Machine Studies, 22, 663-684, (1985).
Some
high-level goals of a user interface language are presented. However, the definitions do not include sufficient
information for me to operationalize them. Then a BNF based language description is described,
and a way of implementing it is described. The language is used by an experimenter to implement
interfaces which are then used by subjects. Subjects are tested. Subjects cannot personalize their own interfaces.
A poor summary of a possibly good dissertation.
Brown,
John Seely, Richard R. Burton & Alan G. Bell. SOPHIE: A
Step Toward Creating a Reactive Learning Environment. International Journal of Man-Machine Studies,
7, 675-696, (1975).
Interesting,
but not particularly relevant to adaptive user interfaces. Interface does spelling correction, run-on word
correction, ellipsis, and pronoun dereferencing, but there is no adaptation,
except keeping track of unparsable inputs.
Card,
Stuart, Tom Moran, and Alan Newell. The
Keystroke-level Model for User Performance Time With Interactive Systems. Communications of the ACM, 23(7), 396-410, (1980).
Clowes,
I., I. Cole, F. Arshad, C. Hopkins, and A. Hockley. User Modelling Techniques For Interactive Systems
in People and Computers: Designing
the Interface (eds. P. Johnson and S. Cook). Cambridge University Press, New York, 1985.
A
literature survey on user models. They
propose a breakdown of user models based on location of the model in the
system.
1. None
(e.g. autopilot)
2. Backend
(e.g. heartbeat monitor which flags abnormalities)
3. Frontend
(e.g. computer aided instruction and interactive front ends)
This
distinction is neither particularly obvious nor particularly useful. Then several systems that use user models are
briefly mentioned: WURSOR, GUIDON,
GRUNDY, UMFE, DEBUGGY. Several
research directions are proposed:
AI - planning and goal recognition, discourse
modelling, belief systems
CogSci - theories of learning, mental
models
Cohen,
Ellis S., Edward T. Smith, and Lee A. Iverson "Constraint-Based Tiled
Windows", Proceedings: 1st International Conference on Computer Workstations.
IEEE Computer Society, 1985
A
description of the RTL window system.
Cohen,
Ellis S., A. Michael Berman, Mark R. Biggers, Joseph C. Camaratta, Kevin
M. Kelly. Automatic Strategies in the Siemens RTL Tiled Window Manager.
Proceedings: 2nd IEEE Conference on Computer Workstations.
IEEE Computer Society, 1987
Cohen,
Paul R. and Edward A. Feigenbaum, (eds.) The Handbook of Artificial Intelligence,
vol III. Los Altos, California:
Kaufmann. 1982.
Croft,
W. Bruce. The Role of Context and
Adaptation in User Interfaces. International
Journal of Man-Machine Studies, 21, 283-292, (1984).
Describes
two systems: POISE, which can be "trained" (by using a graphical
programming language), and a document retrieval system which adapts by
changing weights on features on an ASN (not explained here). No real connection between systems.
Evans,
T. G. "A Program for the Solution of Geometric-Analogy Intelligence
Test Questions" in Semantic Information Processing, M. Minsky (Ed.),
MIT Press, Cambridge, Mass., 1968.
Foley,
James D., Victor L. Wallace, and Peggy Chan.
The Human Factors of Computer Graphics Interaction Techniques,
IEEE Computer Graphics and Applications, Nov 4, 13-48, 1984.
Gaines,
B.R. Axioms for Adaptive Behaviour. International Journal of Man-Machine Studies,
4, 169-199, (1972).
How
to define "adaptive"
First, define a "task" as some
segmentation of the interaction between the controller and the environment.
For any task, it must be possible to say whether the controller
has performed satisfactorily. For metatheoretic reasons, the set of tasks
should be chosen so that the segmentation of an interaction into tasks
is unique, and the tasks should give adequate information about the aspects
of behavior that are of interest.
An acceptable interaction is one that
the controller eventually always satisfies.
Adapted - an adapted controller immediately
performs acceptably on interactions consisting of repetitions of a single
task.
Potentially adaptive - a controller that
will have an acceptable interaction with any one of a set of tasks is
potentially adaptive to that set.
Compatibly adapted - a controller is compatibly
adapted to a set of tasks if it is adapted to one task and potentially
adaptive to the set. Note;
in adapting to a new task, the ability to readapt to the previous
or any other may be lost.
Compatibly adaptive - above, but remains
potentially adaptive to entire set. Note:
not necessarily simultaneously.
Jointly adapted - given any sequence of
tasks from set, it remains adapted to every member of set.
Jointly adaptive - compatibly adaptive
to set (i.e. can learn entire set, one at a time), and becomes jointly
adapted to entire set during acceptable interaction with any task (learns
entire set from one task.) I don't
understand this one. How can one task train the controller for another?
Goldberg,
David E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley, 1988
A
clear explanation of the basic concepts of genetic algorithms and learning
classifier systems. Includes examples
which can be generated by hand, as well as interesting exercises for the
computer. Also includes a summary
of known application. Well worth
reading.
Goldberg,
David E. The Genetic Algorithm
Approach: Why, How, and What Next? Adaptive
and Learning Systems, K. S. Narendra, Ed. Plenum, 1986.
Greenberg,
S. and Witten, I. H. Comparison
of Menu Displays for Ordered Lists. Proceedings
of the Canadian Information Processing Society National Conference, Calgary,
Alberta, May, 1984.
Greenberg,
Saul & Ian H. Witten. Adaptive
Personalized Interfaces-A Question of Viability. Behavior and Information Technology, 4, 1, 31-45,
(1985).
An
existence proof that a self-adapting menu system increases performance. Claim is made that each user experiences increase,
but statistics are for average only. Experiment
is adaptive menu system for telephone dialer, where adaptation is based
on past frequency, and involves changing the menu hierarchy.
Hancock,
P.A., M.H. Chignell & A. Loewenthal.
An Adaptive Human-Machine System.
IEEE 1985 International Conference on Cybernetics and Society,
627-629, 1985.
Discusses
requirements of a system which adapts to human user (treated as a servomechanism)
by measuring Mental Workload (MWL) and changing task when it goes outside
of allowed bounds. Briefly discusses
relevant differences between human mechanisms and machines. Shows no examples, nor does it say anything
convincing about their ability to make this system work in practice. Very disappointing.
Hayes-Roth,
Frederick, Donald A. Waterman, and Douglas B. Lenat. Building Expert Systems. Reading, Mass.: Addison-Wesley, 1983.
Holland,
John, Adaptation in Natural and Artificial Systems. University of Michigan
Press, Ann Arbor, 1975.
Holland,
John. (1986) Escaping brittleness: The possibilities of general purpose
machine learning algorithms applied to parallel rule-based systems. In R. S. Michalski, J. G. Carbonell, and T.
M. Mitchell (eds.) Machine Learning: An artificial intelligence approach,
vol 2. Los Altos, California: Kaufmann.
Holland,
John, Keith Holyoak, Richard Nesbitt, and Paul Thagard. Induction: Processes
of Inference, Learning, and Discovery.
MIT Press, 1986.
A
basis for a general theory of induction.
Describes learning classifier systems, but doesn't go into the
mathematical analysis (for which, see Holland 1975). A gold mine of good information and theories.
See separate notes.
Holland,
John and J. Reitman (1978). Cognitive
Systems Based on Adaptive Algorithms.
In D. Waterman and F. Hayes-Roth (eds.), Pattern-directed Inference Systems. New York: Academic Press, 1978.
Innocent,
P.R. Towards Self-adaptive Interface
Systems. International Journal
of Man-Machine Studies, 16,287-299, (1982).
Presents
the idea of a self-adaptive system, which necessitates a "soft facade"
(another name for a UIMS?). Brings
up some potential problems: stability.
This
appears to be a paper with little firm content.
When I tried to pin down precisely what he was telling me, I came
up with very little in the first three sections. Section 4 was totally
blue-sky.
Langston,
Diane, and Dennis Grantham, eds. Introducing
the Andrew Toolkit, 1988.
Lenat,
D. B. 1976. AM: An artificial intelligence
approach to discovery in mathematics as heuristic search. (Doctoral dissertation. Reprinted in R. Davis and D. B. Lenat. 1980.
Knowledge-based systems in artificial intelligence.
New York: McGraw-Hill.)
Lenat,
D. B. 1977. On automated scientific
theory formulation: A case study using the AM program. In J. E. Hayes, D. Michie, and L. I. Mikulich
(Eds.), Machine Intelligence 9. New
York: Halsted Press, 251-286.
Lenat,
D. B. EURISKO: A program that learns new heuristics and domain concepts. The nature of heuristics III: Program design
and results. Artificial Intelligence,
Mar 1983
Macmillan,
Stuart. Knowledge Acquisition for
a Personal Agent. IEEE 1985 International
Conference on Cybernetics and Society, 736-740, 1985.
Another
poor summary of a dissertation. Not
enough detail to figure out what he really did.
Examples were so terse as to be incomprehensible. He proposed that the Personal Agent be composed
of many cooperating experts (Personal Knowledge Systems), but didn't explain
how they cooperated, or what they did in cases where two disagreed.
He said he talked about how to trigger changes to the system, but
it was so superficial that I didn't get anything out of it.
Minsky,
M. & S. Papert, Perceptrons, MIT Press, Cambridge, Mass., 1969.
Myers,
Brad. Issues in Window Management
Design and Implementation, in Methodology of Window Management, Hopgood,
F. R. A. ed., Springer-Verlag, New York, 1986.
Rhyne,
Jim, Roger Ehrich, John Bennett, Tom Hewett, John Sibert, Terry Bleser. Tools and Methodology for User Interface Development.
Computer Graphics, v21, no 2, 1987.
Rich,
Elaine. Users Are Individuals: Individualizing User Models. International Journal of Man-Machine Studies,
18, 199-214, (1983).
Types
of user models may be characterized along three dimensions:
1. Single
model canonical user vs. models of individual users
2. Specified
explicitly vs. inferred
3. Long
term vs. short term characteristics
Short
description of help system for scribe, which models individual users on
two dimensions: knowledge of scribe,
and knowledge of system related concepts.
Each command is related to concepts, and concepts are rated. An explanation is given in terms of concepts
rated at the users level. Users
level is determined by concepts s/he uses to pose questions, and changed
when next question shows non-comprehension of previous.
Short
description of Grundy, which recommends books based on stereotypes of
users, and which changes values and confidence of facets of those stereotypes
based on rejections of those suggestions.
Rich,
Elaine. Artificial Intelligence,
New York: McGraw-Hill, 1983.
In
an otherwise good discussion of learning, she makes the statement about
neural nets that "If you start from nothing, you will get only a
short way away from nothing" and dismisses the field entirely. She should know better than to make such rash
statements.
Roach,
J. & Wilding, M. Adapting to
Individual Users: The User Trainable
Interface. IEEE 1985 International
Conference on Cybernetics and Society, 228-235, (1985).
User
could stop program (simulation of aircraft carrier air traffic control)
at any point and change the interface.
Change applied to any time simulations was in same state (state
variables were number of planes, presence of emergency, number of missed
approaches). User could change locations of various areas,
color and audio output. Changes
were made by menu choices and locator picks.
Note: change menu is not
trainable. Interface specification
implemented as Prolog rules. No
validation. It is not clear how generalizable these techniques are.
Rumelhart,
David E. and James L. McClelland (Eds.). Parallel Distributed Processing:
Explorations in the Microstructures of Cognition. Cambridge, Mass: MIT Press. 1986.
The
bible of connectionism, but a very difficult read. I still have never understood how networks are
supposed to be trained.
Samuel,
Arthur. Some Studies in Machine
Learning Using the Game of Checkers. IBM
Journal of Research and Development, 3, 210-229, (1959).
Sayers,
Dorothy L. Busmans Honeymoon. Harmondsworth,
England: Penguin, 1937.
This,
like her other Lord Peter Wimsey novels, draws a striking portrait of
Merwyn Bunter, the perfect butler. The
perfect gentleman's gentleman, Bunter exhibits many of the fine qualities
of the top sergeant he was in the war.
He knows how to accomplish virtually any task asked of him; nothing
is too menial, little is beyond his ken.
(See pages 58 and 186) Like
a true top sergeant, he allows the officers the total responsibility for
what to do and why, himself being responsible for the how.
Initiative remains totally with Lord Peter unless it has been implicitly
or explicitly delegated. The significant
exceptions are cases where Bunter is possessed of relevant information
unknown to his master. Then, in
spite of direct orders to the contrary, he will do what he believes Lord
Peter would ask for.
Bunter
is, to me, the prototype of the perfect computer.
Senay,
Hikmet. A Knowledge-Base Approach
to Design Intelligent Interfaces. Ph.D.
dissertation, Syracuse University, 1987.
Sleeman,
D. UMFE:
A User Modelling Front-End Subsystem.
International Journal of Man-Machine Studies, 23, 71-88, (1985).
Types
of user models characterized by nature and form of information contained
and type of inference engine needed.
1. Scalar
- single number describes user. e.g.
KLM
2. Ad
Hoc - e.g. SOPHIE - what readings user has taken from circuit
3. Profile
Models - e.g. GRUNDY
4. Overlay
Models - users competence is subset of experts. Difference from profile
is use of topics that user intends to acquire (What is real difference?
How does modeler know intent?)
5. Process
Models - specification is executable, does not depend on specialized inference
engine, e.g. BUGGY (doesn't seem like much difference to me.)
(Critique
of Samuel (63) in Waterman (70))
UMFE
has a detailed user model, including inference rules to decide which concepts
the user is likely to know. First,
UMFE asks the user whether s/he knows a certain concept, then infers others
from explicit inference rules and implicit use of difficulty ratings.
When
modelling users, you need to address assumptions made about the users. UMFE assumes:
1. Users know what they
know; 2. Users knowledge is stable and context independent;
3. about internal structure;
4. concepts can be totally ordered in difficulty, and users level
can be known; 5. the main user will understand explanations better
if difficult concepts are omitted.
No
experimental validation presented.
Smith,
J. Jerrams. SUSI - A Smart User-system
Interface in People and Computers: Designing
the Interface (eds. P. Johnson and S. Cook), Cambridge University Press, New York, 1985.
Describes
the design goals of a marvelous system which will save the world for (or
from) Unix. However, only a very
small part has been implemented, and that part not experimentally verified. Many details necessary for understanding what
has been implemented are left out. From
what was presented, I would not predict eventual success.
Totterdell,
Peter & Paul Cooper. Design
and Evaluation of the Aid Adaptive Front-end to Telecom Gold in People
and Computers: Designing for Usability
(eds. M.D. Harrison and A.F. Monk), Cambridge
University Press, New York, 1986.
A
good attempt, but a negative result. Design
was simplified so much that users couldn't do any work with it.
Incorrect inferences interfered with performance.
What can be learned from this? 1.
They haven't learned how to infer plans from actions. 2. They
hadn't learned how to make effective help messages. General message: designing experiments without a good theory
is often futile.
Trevelyan,
Robert and Browne, Dermot P. A Self-Regulating Adaptive System, in Proceedings
of CHI+GI 1987 (Toronto, April 5-9), ACM, New York, pp 103-107, 1987.
Waterman,
D. A. Generalization learning techniques for automatizing the learning
of heuristics. Artificial Intelligence,
1, 121-170. 1970
Winston,
Patrick Henry, "Learning Structural Descriptions from Examples"
in The Psychology of Computer Vision, P. H. Winston (Ed.), McGraw-Hill,
New York, 1975.
Winston,
Patrick Henry, "Learning and reasoning by Analogy," Communications
of the A.C.M., vol 23, No. 12, pp 689-703, Dec. 1980.
Winston,
Patrick Henry, Artificial Intelligence, 2nd edition.
He
presents several ways of learning with a teacher. Only way presented of learning without a teacher
is Lenat, which used interestingness as a value function. The role of a teacher is to select training
examples, both correct examples and near misses.
He
quotes Martins Law: "You can't learn anything unless you almost know
it already", attributed to William A. Martin.
NeWS
manual from Sun
Windows
and Window Based Tools: Beginners Guide, Sun Microsystems, 1987
|