Artificial Intelligence Frontiers in Statistics: AI and Statistics III.

Linster, Bruce G.

This volume contains a diverse selection of refereed papers that were presented at the Third International Workshop on Artificial Intelligence and Statistics, Fort Lauderdale, Florida, in January 1991. This collection of essays is very impressive in terms of the breadth of topics covered as well as the depth of analysis for each subject. Although this book is clearly not aimed at economists, students of the dismal science can benefit from this collection in two ways. First, some of the papers demonstrate how expert systems can help researchers do statistical analysis. Also, a number of essays show how statistical ideas can be applied within expert systems. Although seeing how research can be enhanced is important, the most interesting part of the book deals with the applications of statistical ideas.

The book begins with detailed descriptions of various expert systems. One paper, for instance, describes an expert system for experimental design, while others describe aids in developing linear models. The editor's paper in this section is particularly interesting. Professor Hand's proposal that metadata, which is information about data, be made explicit and available to the software so it can better guide the researcher is novel and potentially very important. The author applies his ideas to measurement scales as one form of metadata.

One particular area of Artificial Intelligence that has been significantly influenced by statistical ideas is belief networks. Part Two presents five very technical papers on the subject that will not be easily understandable by most economists. (I include myself in this group.) The papers in this section describe the relationship between belief networks and knowledge-based systems along with some specific examples.

Most readers of this journal have a profound interest in how learning takes place, and Part Three deals with the subject as it applies to artificial intelligence and learning algorithms. One paper, for example, describes a method for determining causation with background information as well as statistical data. There is also a paper that suggests a Bayesian tree learning algorithm, while another discusses a system for generating probabilistic networks. The discussions of learning algorithms in computing machines are enlightening and provide some insight into human learning.

Another part of this volume that will be generally inaccessible to many economists is the section on the interface between neural networks and Artificial Intelligence. Some advanced probabilistic and statistical analysis is used in Part Four. The papers address interesting topics and are generally understandable.

Part Five was the most interesting for this reader. In this section three papers are offered that deal with text manipulation. One paper discusses a statistical technique for matching the sentences in text of different languages. Another describes a probabilistic approach to text understanding. Finally, there is a very interesting paper describing a system that uses Artificial Intelligence to determine the subject of some piece of text. These chapters stimulated ideas for research involving other uses of machine learning that may be more traditionally economic in nature.

Finally, the last section explores other areas where Artificial Intelligence and statistics are used together. Here a wide array of topics are discussed. The first paper in this section describes how causal influences can be distinguished from spurious covariances, while the second discusses the application of stochastic complexity to classification problems. In a very interesting paper, the problems associated with the aggregation of expert opinions is discussed. This section concludes with papers on nonmonotonic reasoning, probabilistic logic, and a decision theoretic approach to controlling the cost of planning.

It should be apparent that this book is not for every economist's book shelf. However, there are certainly a number of interesting papers here for economic researchers who would like to use expert systems in statistical analyses, as well as for those interested in machine learning algorithms and other Artificial Intelligence applications.

Bruce G. Linster United States Air Force Academy