Artificial intelligence, or AI, is largely an experimental science—at least as much progress has been made by building and analyzing programs as by examining theoretical questions. MYCIN is one of several well-known programs that embody some intelligence and provide data on the extent to which intelligent behavior can be programmed. As with other AI programs, its development was slow and not always in a forward direction. But we feel we learned some useful lessons in the course of nearly a decade of work on MYCIN and related programs. In this book we share the results of many experiments performed in that time, and we try to paint a coherent picture of the work. The book is intended to be a critical analysis of several pieces of related research, performed by a large number of scientists. We believe that the whole field of AI will benefit from such attempts to take a detailed retrospective look at experiments, for in this way the scientific foundations of the field will gradually be defined. It is for all these reasons that we have prepared this analysis of the MYCIN experiments.

The complete book in a single file.

"Analysis of the job shop scheduling domain has indicated that the crux of the scheduling problem is the determination and satisfaction of a large variety of constraints. Schedules are influenced by such diverse and conflicting factors as due date requirements, cost restrictions, production levels, machine capabilities and substitutability, alternative production processes, order characteristics, resource requirements, and resource availability. This paper describes ISIS, a scheduling system capable of incorporating all relevant constraints in the construction of job shop schedules. We examine both the representation of constraints within ISIS, and the manner in which these constraints are used in conducting a constraint-directed search for an acceptable schedule. The important issues relating to the relaxation of constraints are addressed. Finally, the interactive scheduling facilities provided by ISIS are considered." Expert Systems 1(1):25-49.

Proceedings of AAAI-84.

In Proc. First Conference on AI Applications, pp. 410–416.

This paper describes a device-independent diagnostic program called dart. The resulting generality allows it to be applied to a wide class of devices ranging from digital logic to nuclear reactors. Although this generality engenders some computational overhead on small problems, it facilitates the use of multiple design descriptions and thereby makes possible combinatoric savings that more than offsets this overhead on problems of realistic size.

A formalism for reasoning about actions is proposed that is based on a temporal logic. It allows a much wider range of actions to be described than with previous approaches such as the situation calculus. This formalism is then used to characterize the different types of events, processes, actions, and properties that can be described in simple English sentences. In addressing this problem, we consider actions that involve non-activity as well as actions that can only be defined in terms of the beliefs and intentions of the actors. Finally, a framework for planning in a dynamic world with external events and multiple agents is suggested.

The am program was constructed by Lenat in 1975 as an early experiment in getting machines to learn by discovery. In the preceding article in this issue of the AI Journal, Ritchie and Hanna focus on that work as they raise several fundamental questions about the methodology of artificial intelligence research. Part of this paper is a response to the specific points they make. It is seen that the difficulties they cite fall into four categories, the most serious of which are omitted heuristics, and the most common of which are miscommunications. Their considerations, and our post-am work on machines that learn, have clarified why am succeeded in the first place, and why it was so difficult to use the same paradigm to discover new heuristics. Those recent insights spawn questions about “where the meaning really resides” in the concepts discovered by am.

This volume brings together current work on qualitative reasoning. Previous publication has been primarily in scattered conference proceedings. The appearance of this volume reflects the maturity of qualitative reasoning as a research area, and the growing interest in problems of reasoning about physical systems. Anyone concerned with automated reasoning about the real (physical) world should read and understand this material.

Much artificial intelligence research is based on the construction of large impressive-looking programs, the theoretical content of which may not always be clearly stated. This is unproductive from the point of view of building a stable base for further research. We illustrate this problem by referring to Lenat's am program, in which the techniques employed are somewhat obscure in spite of the impressive performance.

"By studying the problem-solving techniques that people use to design algorithms we can learn something about building systems that automatically derive algorithms or assist human designers. In this paper we present a model of algorithm design based on our analysis of the protocols of two subjects designing three convex hull algorithms. The subjects work mainly in a data-flow problem space in which the objects are representations of partially specified algorithms. A small number of general-purpose operators construct and modify the representations; these operators are adapted to the current problem state by means-ends analysis. The problem space also includes knowledge-rich schemas such as divide and conquer that subjects incorporate into their algorithms. A particularly versatile problem-solving method in this problem space is symbolic execution, which can be used to refine, verify, or explain components of an algorithm. The subjects also work in a task-domain space about geometry. The interplay between problem solving in the two spaces makes possible the process of discovery. We have observed that the time a subject takes to design an algorithm is proportional to the number of components in the algorithm's data-flow representation. Finally, the details of the problem spaces provide a model for building a robust automated system." Information Processing and Management 20(l-2):97-118.