A logical theory of nonmonotonic inference and belief change (Q2727768)

The logic of belief change and nonmonotonic logic are both creatures of the 1980s. The literature on the two subjects has grown through individual journal papers, dedicated issues of journals, books collecting papers from workshops, and single-authored books. Until the present, the books focused on one of these or other of the two subjects, touching only tangentially on the other, although it has been known since 1991 that there are close formal relations between the two. In the last year two books have appeared that seek to treat both subjects together. One of them is the volume reviewed here. The other, by Hans Rott, is reviewed separately. The author does not merely present a digest of work that has already entered the mainstream. He gives the subject an individual shape based on concepts that he has himself developed in a series of recent papers. The general vision behind the book may be described as follows. The operations of belief change, contraction and revision, may be applied to the `epistemic states' of an agent. However, an epistemic state should not be seen simply as a belief set (i.e. a set of propositions closed under classical consequence) as in the AGM approach; nor a base for a belief set, as in some other approaches; nor even one of these equipped with a relation over its elements. It is a more complex item, consisting of an indexed family of belief sets, with a relation over them. Intuitively, the elements of the indexed family represent all the belief sets that the agent currently accepts as possibilities for adoption, and the relation indicates preferences between them. Each epistemic state determines a unique set of actual beliefs, namely the intersection of all its most preferred belief sets. The operations of contraction and revision are applied to such epistemic states, also giving epistemic states as values, thus satisfying the principle of categorical matching (unlike e.g. the AGM operations), and thereby facilitating the investigation of iterated belief change. Likewise, the process of nonmonotonic inference is viewed as carried out modulo such an epistemic state, which in this context can be seen as a generalization of the familiar notion of a preferential model (the generalization consists of putting belief sets in place of the worlds alias maximal consistent sets of propositions of the preferential model). Nonmonotonic inference thus appears as a static counterpart of the dynamic process of belief change. In effect, when inferring something from a premise we are identifying an element of a possible revision of our background beliefs with that premise as input, but without committing ourselves to revising the beliefs themselves. With this general vision guiding the volume, the detailed development unfolds on lines that look quite different from previous presentations of either of the two subjects. In Part I, the author reviews the notion of multiple-conclusion (alias Scott) consequence operations, and develops a general theory of epistemic states as they are defined above. In Part II, he applies this general theory to the static context, i.e. nonmonotonic reasoning. The application begins with a very weak system of `basic inference relations'. This covers preferential reasoning with both sceptical and the credulous policies for aggregating extensions (representing inference proper and abduction respectively). Stronger systems are then built up one by one. In Part III the general theory of epistemic states is applied to the dynamic context, i.e. belief change. Taking epistemic states as the objects of attention radically changes the whole perspective of the subject. The two basic belief change operations are taken to be contraction and expansion. As applied to epistemic states, the latter is quite non-trivial (in contrast to the situation under the AGM account, where expansion is no more than set-theoretic addition followed by closure under the background consequence operation). Revision is defined as usual via the Levi identity, i.e. as the composition of contraction and expansion, but its role is significantly diminished. The presentation is abstract. At some points, the level of abstraction may perhaps go beyond what is needed. For example, the first substantive chapter develops the general theory of multiple-conclusion consequence operations, which can be used as an alternative way of expressing the general theory of epistemic states. But the applications to nonmonotonic reasoning in Part II of the book use the usual single-conclusion (Tarski) consequence operations, and multiple-conclusion consequence operations are needed only in Part III (specifically in chapter 12 on contraction). The style is demanding. Definitions are often introduced with only brief intuitive explanation, and properties are established while leaving their strategic interest implicit. Moreover, the reader is expected to keep everything in memory as he proceeds through the book; there is very little in the way of reminders or precise back-references to crucial definitions and earlier results. The skimpiness of the publisher's index also makes such checks difficult. This reviewer admits to having had considerable difficulty following the maze of development, particularly after interruptions in reading. He suspects that students could be frustrated if they attempt the volume without the help of an instructor who has already mastered it thoroughly. In brief, the book is based on a novel and coherent vision, developed according to a definite overall plan to yield non-trivial results, but it cannot be described as reader-friendly.