scientific article; zbMATH DE number 108002

Authors

Publication date 23 January 1993

belief revision counterfactuals degrees of belief update

Classical first-order logic (03B10) Other nonclassical logic (03B60) Proceedings, conferences, collections, etc. pertaining to computer science (68-06) Logic in artificial intelligence (68T27) Knowledge representation (68T30)

Review text

The articles of this volume will not be indexed individually. This is a collection of papers devoted to belief revision. The main problem is outlined in the first paper of the book (written by \textit{P. Gärdenfors} himself): If we get a new piece of knowledge that contradicts our previous knowledge, then we have to revise our knowledge. The word ``knowledge may be a little misleading here, because we are talking about the statements about which we are not sure; therefore, the book under review primarily uses the word ``belief instead of ``knowledge. There are two main ways to do the revision: the first (``foundations) approach is that in case of inconsistency, we must update the axioms that form the basis of our beliefs. In the majority of areas of knowledge, however, there is no clear distinction between axioms and conclusions, and even when there are axiomatic descriptions, these descriptions use different sets of beliefs as axioms. Therefore, it is more reasonable to update the whole set of conclusions without worrying much about whether they are axioms or corollaries of these axioms. The comparison between these two approaches is described in some detail in a paper by \textit{J. Doyle}. A modification of the first approach is proposed by \textit{S. O. Hansson}, who remarks that sometimes not only the basic beliefs \(B\), but also the logic \(L\) that allows us to make corollaries has to be changed (he proposes and analyzes postulates for changing the pair \((B,L)\)). All the other papers are devoted to the second approach. In this approach, we must assign to every set of beliefs \(K\) and a statement \(\phi\) that is inconsistent with \(K\), a ``revisedset of beliefs \(K+\phi\). Gärdenfors and others has proposed several reasonable conditions that such an operation + must satisfy (e.g., \(K+\phi\&\psi\subseteq (K+\phi)+\psi\)). These conditions are enumerated in the first paper. Some of them turn out to be inconsistent, some of them are consistent with each other. One of the ways to describe + is to order the beliefs from \(K\) according to our ``degree of belief in them, and if necessary, eliminate the statements in which we believe the least. For an arbitrary operation +, we can say that \(\phi\) is ``more believed than \(\psi\) if whenever \(\neg(\phi\vee\psi)\) is added to \(K\), we retain \(\phi\) but not \(\psi\). In this sense, this ordering is equivalent to an operation +. Here, the ordering appears as a trick. However, in knowledge bases used in expert systems, degree of belief is already used and is usually expressed by a number from the interval \([0,1]\). What revisions we get if we use these degrees of belief is described in a paper by \textit{D. Dubois} and \textit{H. Prade}. A paper by \textit{B. Nebel} finds another equivalent description of revisions: in default theory. Nebel also describes computational complexity of revision. For propositional case revision is practically of the same complexity as propositional satisfiability; addition postulates can make this problem slightly harder. Relations between different languages of revision are also analyzed by \textit{H. Rott}. \textit{C. B. Cross} and \textit{R. H. Thomason} analyze the relation between revision and counterfactuals ``if \(\phi\) was true, then \(\psi\) would be true. This relation is easy: we can retain \(\psi\) in \(K+\phi\) iff the above counterfactual is true; but since historically different postulates were used for counterfactuals and revisions, this translation is not always 1- to-1. A paper by \textit{J. R. Galliers} describes a realistic situation when our knowledge consists of the knowledge sets of several communicating intelligent agents. In this case, an inconsistency discovered by one of them must be communicated to the others and must lead to a consistent set of revisions. \textit{H. Katsuno} and \textit{A. O. Mendelzon} apply revision ideas to updating a knowledge base. The difference is that unlike revision, in updating what we have known before is still true, but this knowledge describes the situation in the previous moment of time. Now, some properties have changed, so we must update our knowledge (assuming that if a change in some property has not been announced, and cannot be deduced from announced changes, then this property remains true). Mathematically, this leads to a similar but slightly different problem. A paper by \textit{M. Morreau} applies these ideas to planning, i.e., to choosing such changes that the resulting (updated) knowledge has some desired properties.

Recommendations

Cited in

(only showing first 100 items - show all)

This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4023583)