Experimentation serves three purposes: (a) hypothesis testing, (b) gathering of new data to constrain the theory generator, and (c) manipulation of the external system to reveal its structure. A theory formation system, EG, is described that employs experimentation and observation techniques to develop a theory of the UNIX file system and executive-level file commands. This theory formation task is more complex than previous efforts, and the goal of the project is to determine which existing theory formation methods are applicable and what new methods need to be developed. Previous techniques arc reviewed, and none of them arc found to be applicable. A new technique, based on controlled experimentation, is de3cribcd, and a hypothetical trace of EG's execution is presented. Key terms: 'Theory formation, generate-and-test, controlled experimentation, exploratory experiments, observation experiments, hypothesis-test experiments, credit assignment, new terms.

Meta-level control, in an Artificial Intelligence system, can provide increased capabilities and improved performanct. This improvement, however, is achieved at the cost of the meta-level effort itself. To ensure an overall increase in system efficiency, the savings brought about at the base level cannot be exceeded by the effort at the meta-level. This paper outlines a formalization of the costs involved in choosing between independent problem-solving methods: the cost of meta-level control is explicitly included. It is shown that when meta-level effort is related to its efficacy, there exists an amount of this effort that should optimally be expended. Too much or too little meta -le-,e1 effort can result in a loss of overall system performance.

Georgeff's controlled production systems 65 8.3.3

Diane Warner Hasling Heuristic Programming Project Computer Science Department Stanford University Stanford, CA 94305 stra:cgre explanation we want to generate.

Scientific Dateline Heuristic Programming Project May 1983 Report No. HPP-83-1 5 (Revised June 1983)

ONCOC1N's design and implementation has occurred in unison with a set of studies and analyses Intended to help us better understand the demands of physicians as computer users. Our study of physician attitudes towards computer-based clinical consultation systems 112: emphasized the importance of a system's explanation capabilities and helped convince us of the important role that Al techniques are likely to play in the development of optimal decision aids.

This is a remarkab'y exciting time to be involved professionally in the field of medical informatics. The underlying scientific principles are beginning to be identified and defined, educators are increasingly acknowledging the importance of thc field for physicians of the present and future, and the tec mology itself is growing at rates that make the future of the field both unbounded and impossible to predict. One has the sense that what was once a field for pioneers is now reaching the stage of established settlements, with a history, traditions, and a feel of permanence. It is therefore appropriate that, at the beginning of ddiberations designed to achieve significant educational goals for the field, we might start by considering the discipline itself and the characteristics that hnve tended to separate it from other traditional academic and research medical specialties. I would like to begin by assuming that certain basic points are well accepted and need not be defended here: first that medical informatics holds both realized and potential importance for the science anc practice of medicine, and second, that there is a need for all medical practitioners to be familiar both with information handling technology and with the underlying principles that make the field relevant, regardless of whether computers are involved.

Simuldtois serve two major purposes: the first is the verification of scientific thP.ories the second is xperimental result prcliction. Thr? verification function is called upon when existing t:leories are btling cxtended or new theories are being ç enerated to explain experimental data: the predictive capabilities are used to predict laboratory results in order to eliminate a great deal of experimmtal effort. An esoecially important role for a simulation pro.warn would be as par; of a larger employing art ficial innelligence techniques to develop mcd.els of a biological system bar:ed on experimental OUSerwitiOris. Such a program would accept as Input observations of a s; stem cud would!fiocii:ce as output a model for tn.t system that could account for the observations. The r.:niu!ation portion of such a program would be a crucial tool for ensuring that on:y theories mat were con::..-4 with thc data were dewloocci. It's a major research goal Cf the HOLDEN pi met to explore methods tor building a systym

The objective of this paper is to present a new method for the computer representation of empirically derived causal relationships (CR's). This method draws on the theory of multivariate linear models and path analysis. The method is contrasted with the predicate calculus methods developed by other Al researchers. The representation presented here has been used to store information on medical CR's derived empirically from a large clinical database by a computer program called RX. The principal emphasis in the representation is on capturing the intensities and variances of effects and the variation in the effects across a patient population. Once incorporated into RX's knowledge base, this information is subsequently used by RX in determining the validity of other CR's. The representation uses a directed graph formalism in which the nodes are frames and the arcs contain seven descriptive features of individual CR's: intensity, distribution, direction, mathematical form, setting, validity, and evidence. Because natural systems (such as the human body) are inherently probabilistic, linear models are useful in representing causal flow in them. Knowledge of natural systems is fundamentally probabilistic because of I) irreducible indeterminism in their component processes, 2) difficulties in accurately measuring all relevant variables, 3) variation among individuals in a population, and 4) inadequate scientific theory.

The time-sharing systems that were used to develop them are largely inadequate due to the burden placed upon computing and storage resources. Consequently, some of this research is being shifted to personal workstations which provide a sophisticated graphics interface in addition to satisfying the requirements for computational speed and memory. This paper examines the use of that graphics interface in the development of a tool for a system builder. Introduction Medical expert systems are consultation programs designed -..o give advice using both formal knowledge and the judgmental expertise of clinical specialists. When these systems first began to appear in the 1970's, many observers doubted their future role because their size and complexity placed an inordinate burdenlin - the processing and storage resources of conventional time-sharing systems.