Escolar Documentos
Profissional Documentos
Cultura Documentos
Notes
1
Editor’s note: See Shanker, Stuart (1993), Review of Boden 1990, Minds and Machines 3, pp.
109–113.
2
Editor’s note: See Shelley, Cameron, Review of Robert M. French (1995), The Subtlety of Sameness:
A Theory and Computer Model of Analogy-Making, in Minds and Machines 7, pp. 292–296.
References
Boden, Margaret A. (1990), The Creative Mind: Myths nnd Mechanisms, London: Abacus; New
York: Basic Books.
Evans, Thomas G. (1968), ‘A Program for the Solution of Geometric-Analogy Intelligence-Test
Questions’, in Marvin L. Minsky, ed., Semantic Information Processing, Cambridge, MA: MIT
Press, pp. 271–353.
Falkenhaimer, Brian; Forbus, Kenneth D.; and Gentner, Dedre (1990), ‘The Structure-Mapping
Engine’, Artificial Intelligence 41, pp. 1–63.
Langley, Pat; Simon, Herbert A.; Bradshaw, Gary L.; and Zytkow, Jan M. (1987), Scientific Discovery:
Computational Explorations of the Creative Proceess, Cambridge, MA: MIT Press.
Reitman, Walter (1965), Cognition and Thought: An Information-Processing Approach, New York:
John Wiley.
Robert Cummins and John Pollock (eds.), Philosophy and AI: Essays at the Inter-
face, Cambridge, MA: MIT Press, 1991, xi + 304 pp., $29.95 (cloth), ISBN 0-262-
03180-9; $15.00 (paper), ISBN 0-262-53135-6.
Philosophy and computer science are both concerned at least in part with abstract
symbolic systems, so it should come as no surprise that they have much to gain
from one another. In particular, the interface between artificial intelligence (AI)
and those branches of philosophy concerned with cognition ought to be especially
fruitful. It is this interface that Philosophy and AI endeavors to explore.
AI has changed drastically in the six years since this book was published, and
much earlier work has not aged well. While most of the technical work in the book
is by now outdated, the philosophical implications of several essays are both deep
enough to stand the test of time and sufficiently insightful to inspire new research.
The book is certainly worth reading for these insights alone, despite the fact that
much of the technical content is now far from the cutting edge.
The book is concerned almost exclusively with the relations between AI and
analytic epistemology. This is a narrower focus than one ought automatically to
expect. AI has sophisticated relations with ontology and descriptive metaphysics,
and today it is no longer unusual for AI to take existential and pragmatistic consid-
erations seriously. Of course, such errors of omission do not diminish the positive
qualities of the essays that have been included. If there is a purely epistemic objec-
tion to be raised, it is the absence of essays devoted primarily to induction. There
is far too much discussion of non-monotonic reasoning and too much lamenta-
tion for Tweety, who, though a bird (alas!), does not fly. Such issues, though still
problematic, are no longer as important as they once were.
With one exception, all of the models discussed in the book fall squarely within
the classical serial-symbol-processing approach to AI. Though this approach has
by now suffered serious (some would say insurmountable) setbacks, several of
the essays once again have philosophical implications sufficiently significant that
they are worth reading even by those (including myself) who regard serial symbol-
processing as inadequate. Still, it is difficult not to regard many of the essays as
belonging to a past that now (only 6 years later!) seems utterly remote. This is
a serious problem faced not just by these authors, but by anyone working with
computers.
The introduction by Cummins and Pollock aims to explain how philosophers
and AI researchers come to interact. Such relations are far more fragile than they
ought to be, and it is unfortunate that many philosophers remain suspicious of the
computer. The essay’s most interesting points concern the computational modeling
of philosophical theories. The Cummins and Pollock essay offers serious reflection
on what it means to make such models, and is certainly required reading by anyone
concerned with how computers might be used profitably for philosophical research.
The title of the first essay, “Plans and Resource-Bounded Practical Reasoning”,
by Michael E. Bratman, David J. Israel, and Martha E. Pollack, clearly expresses its
content. The essay makes the point that planning must be adaptive and dynamic and
so should remain at least partly abstract. Much of the essay makes interesting use of
the notion of the planner’s “surviving options”. This suggests a genetic algorithm in
which a population of competing plans evolves, but no such approach is discussed.
Unfortunately, an essay about resource-boundedness is precisely where one would
expect a discussion of the limits of seriality and the virtues of parallel processing.
But such a discussion is not offered.
The second essay, “Cross-Domain Inference and Problem Embedding”, by
Cummins, deals with how one problem may be embedded in another analogous to
it. What is most interesting about this essay is not so much its use of the notion
of task-embedding, but the example used to illustrate the notion. The example
involves discussing how two agents can develop a communicative convention.
The essay is intriguing reading for those interested in computational solutions to
coordination problems (particularly linguistic ones).
The third essay, “The Foundations of Psychology: A Logico-Computational
Inquiry into the Concept of Mind”, is in my opinion the most stimulating and
powerful essay in the book. The book is well worth buying for this essay alone.
In it, Jon Doyle lays the foundations for “rational psychology”: the mathematical
study of the totality of possible minds. In our world, we have many kinds of minds.
At least one of these kinds is occasionally instantiated by human beings, and it
seems that there is a one-to-one mapping from kinds of organisms to kinds of
minds. But if AI is successful, we may have either one kind of mind (multiply
realized) or many inorganic kinds of minds. The taxonomy of intelligences is one
that can be studied mathematically, and is one whose species can be classified up
to isomorphism much as in abstract algebra. While some might object that Doyle’s
vision is too grand or Platonic, his mathematical approach to the problem of the
nature of mind is fresh and inspiring. Though the consequences of his conclusion,
that possible minds are narrowly realized theories, are almost impossible to accept,
the study of his arguments is very enjoyable, and his essay is a paradigm case of
how AI and philosophy can profitably interact.
The essay “Memory, Reason, and Time: The Step-Logic Approach”, by Jen-
nifer J. Elgot-Drapkin, Michael Miller, and Donald Perlis, proposes an approach
to reasoning in which the dynamics of the inference process are taken at least as
seriously as the conclusions. The model presented consists of a knowledge base
modified by an inference cycle. The inference engine described is clearly a dis-
crete dynamical system. Dynamical-systems approaches to the mind have recently
gained great currency, and this essay suggests that such approaches can be suc-
cessfully applied to reasoning. Inference is an iterative optimization process. The
questions of method raised by the essay far surpass it. If the essay is correct, then
the logical methods the essay uses are surely ineffective; genetic algorithms have
been used to breed cellular automata capable of universal computation, so why not
use a genetic algorithm to breed a cellular automaton whose dynamics approximate
those of the step-logic system?
The fifth essay, “Artificial Intelligence and Hard Problems: The Expected Com-
plexity of Problem Solving”, by Clark Glymour, Kevin Kelly, and Peter Spirtes,
attempts to blunt complexity arguments against serial processing in AI by replying
that statistical analyses of problem complexity show that serial solutions often fare
better than worst-case analyses suggest. But the success of all kinds of massively
parallel approaches to problem solving (not just connectionism!) and the accep-
tance of massively parallel approaches by more and more AI workers shows that
the battle for serial processing was lost long ago. So, while the technical virtues of
the essay remain, its relevance has certainly diminished.
In “Normative and Descriptive Ideals”, Henry Kyburg shows that the norma-
tive/descriptive distinction is rather muddled. Philosophers once used this distinc-
tion to separate their work from psychology, but, given something like Doyle’s
program for rational psychology, this hardly seems necessary. The problems raised
by Kyburg are good test cases for Doyle’s approach; I think Doyle’s fresh approach
to psychology is far more productive than criticizing old positivist distinctions; little
new ground is broken here.
In “Ampliative Inference, Computation, and Dialectic”, Ronald P. Loui talks
about non-monotonic reasoning, and claims that what defeasible reasoners need are
References
Dennett, Daniel C. (1971), ‘Intentional Systems’, Journal of Philosophy 68, pp. 87–106; reprinted in
Daniel C. Dennett, Brainstorms, Montgomery, VT: Bradford Books, pp. 3–22.
Thagard, Paul (1992), Conceptual Revolutions, Princeton, NJ: Princeton University Press.
Peter Ludlow, ed., High Noon on the Electronic Frontier: Conceptual Issues in
Cyberspace, Digital Communication series, Cambridge, MA: MIT Press, 1996,
xxii + 536 pp., $30.00 (paper), ISBN 0-262-62103-7.