Beruflich Dokumente
Kultur Dokumente
Book Review
I
n 1992, Glenn Shafer completed a essential simplicity. One might have
draft of three chapters for an first learning about probability. Shafer assumed that a deep understanding of
intended book on join-tree compu- argues, and the length and depth of the conventional paradigm of the
tation methods for combining evi- the book convinces, that this basic
measure-theoretic approach to proba-
dence and other aspects of computa- structure is a powerful way of thinking
bility was necessary for a good under-
tion in expert systems. However, he about probability and a way that nat-
standing of the book, but this is not
found his attention increasingly dis- urally links probability and causality.
the case. Appendixes do present an
tracted by the possibilities provided He distinguishes the traditional events
overview of the conventional
by probability trees for understanding of probability theory, which he calls
paradigm, and there are links through-
probability and causality—so much Moivrean events and which correspond
out the book between this and the
so, in fact, that instead of finishing to sets of leaf nodes (or, equivalently,
new approach presented here, but the
the first book, he wrote a different one of paths from the root to a leaf), from
book can be read without a deep grasp
on this second topic. It is this second Humean events, which are steps, or
of the conventional paradigm. It will
book that is the subject of this review, chains of steps down through the tree.
be interesting to see how future gener-
and it is easy to see why the power The distinction, which is clear when
presented in the context of a probabil- ations come to regard the ideas in this
and breadth of the ideas seduced
ity tree but less clear from the sample book. The author might have been
Shafer from his original purpose. (Do
space perspective, is fundamental to especially astute in publishing it in an
not despair, however, those three draft
the discussion of causality presented AI series, so that perhaps more suscep-
chapters have now also appeared,
in the book. tible minds, less encumbered by the
although without the others that were
paraphernalia of the traditional views,
originally intended to accompany
them, in Probabilistic Expert Systems, are exposed to the ideas. This, howev-
er, is not the reason he gives for pub-
Society for Industrial and Applied The Art of Causal Con- lishing it as part of an AI series. The
Mathematics, 1996).
The author describes The Art of
jecture, Glenn Shafer, book could have appeared in various
Causal Conjecture as presenting “a new MIT Press, Cambridge, subject-matter series, ranging from
mathematical and philosophical statistics to computer science, but
foundation for probability” (p. xv)
Massachusetts, 1996, Shafer chose an AI series because AI, he
and seeking to show how “this foun- 511 pp., ISBN believes, is currently in the best posi-
tion to put the book’s ideas to use.
dation can help us understand causal- 0-262-19368-X. (Note the word use; it implies that he
ity” (p. xv). This is a large claim. Does
the book live up to it? intends the book not simply to be a
First, of course, there is nothing purely theoretical treatise on a novel
new under the sun; so, the question way of looking at probability.) He
He takes the philosophical orienta-
arises about the extent to which the argues that the debate about causal
tion that probability is subjective— reasoning has been especially intense
presented foundation is new. The cen-
not in the sense that it depends on in the AI and expert system communi-
tral notion of the book is the probabil-
someone’s state of mind but, rather, ty, so ideas that combine issues of
ity tree. These trees have nodes that
that it depends on the perspective probability and causality fit most suit-
represent the points at which events
with more than one outcome can from which one sees the world. Prob- ably into this milieu. Speaking as a
occur and branches that represent each abilities thus describe the position of statistician, I think it is revealing, true,
possible outcome, with attached the observer relative to the world as and something of an indictment of
numbers signifying the probability of well as describe the world. Objective the discipline of statistics that the
each outcome. This structure is probabilities fit nicely into this inter- author can say that causal ideas have
old—Shafer shows that it was familiar pretation because these probabilities not been incorporated into the basic
to Huygens in the seventeenth centu- describe the view of the world from theory of statistics.
ry—and indeed it is one to which the the position of someone who can see The theory described in the body of
Copyright © 1998, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1998 / $2.00 SPRING 1998 131
Book Review
132 AI MAGAZINE