ll the range of influence which fundamental models exercise over inductive inference (see section 3.3.3). An independent research effort in the United States attempts to integrate fundamental models with specific abstractive, or generalizing, techniques (Srinivasan, 1980; Srinivasan and Sandford, 1980). However, unlike the Czech group the American team is still at the stage of theory development -a working system has yet to be implemented in hardware. Abductive inferences have scarcely been touched by the AI community, but nevertheless some tentative first steps do exist. A few papers on "nonmonotonic logic" were delivered at the First Annual National Conference on Artificial Intelligence at Stanford University in August 1980 (e.g., Balzer, 1980), and much discussion followed. However, this attempt to deal with the invention of new or revised knowledge structures is hampered (and finally undermined) by their lack of a general theory of abductive inference -with one notable exception, the recent work of Frederick Hayes-Roth (1980). Hayes-Roth takes a theory of abductive inference developed by Lakatos (1976) for mathematical discovery and makes two of the low-level members of the family of abductive inferences which Lakatos identifies operational. Still, this work is only a preliminary step toward implemented systems of mechanized abductive inference and, unfortunately, it seems to represent the extent of theory-based At work on abductive inference to date. In summary, state-of-the-art AI treatments of analytic and inductive inference provide no fundamental models as a theoretical foundation to support the detailed knowledge structures and inference techniques upon which the treatments are built. Yet these models are an essential and integral element of analytic and inductive inferences. State-ofthe-art AI virtually lacks treatments of abductive inference. However, model-based analytic and inductive inference systems and an abductive inference system are all necessary prerequisites for machine learning systems. There appears to be a growing acceptance within the AI community of the above problems and that overcoming these gaps in current treatments of analytic, inductive, and abductive inference is an important future research direction for the entire field. For example, at the First National Conference on Artificial Intelligence, Peter Hart admitted that the fact that rule-based expert systems lack a fundamental model to ground the detailed rules makes them superficial and inflexible. Charles Rieger at the University of Maryland is beginning to address the question of layering models under rule-based systems. Several recent AI initiatives with respect to inductive and abductive inference have already been noted. A concerted and serious attack on the problem of developing a theory of abductive inference for machine intelligence could pay enormous dividends. First, machine learning systems cannot possibly possess a full learning capability unless they can perform abductive inferences. Second, a successful mechanization of abductive inference would require the solution of problems which must also be solved for the successful mechanization of analytic and inductive inference. These problems include: (1) how to represent the fundamental models of the processes which underlie the detailed occurrences of domains, (2) how to inferentially relate these to more detailed knowledge structures such as laws, principles, generalizations, and classification schemes, and (3) how to map the representations of a domain occurrence in one "language," say, that of the model, onto its representation in another "language," say, that of a set of diagnostic rules. Since an investigation of abductive inferences seems to hold many keys to solving the problem of machine learning, and since recent developments in AI seem to promise receptivity to such an investigation, the development of a theory of abductive inference for machine intelligence appears to be the preferred research direction for work on machine learning systems.
6.2.3 Two Barriers to Machine Learn&g Two points from the above discussion must be emphasized. First, state-of-the-art At work on hypothesis formation is almost totally devoid of research on abductive inference. However, machine systems must have this capability in order to be true learning systems. Second, current AI work on analytic and inductive inference tends to proceed in the absence of relevant theories, and this seems to be the reason why state-of-the-art A1 treatments fail to give fundamental models their proper role in inference systems. However, adequate theories of all three types of inference are a necessary foundation for successful machine learning systems. Both of these barriers to machine learning -the abductive inference barrier and the theory barrier -must be bridged before machine intelligence systems can be given a full learning capability. The abductive inference barrier has already been fully treated, but some additional discussion of the "theory barrier" is useful here. Historically, technology has developed in two distinct patterns -empirical and theoretical. Empirical technology is a "black box" approach. Given the problem of producing action A from some set of inputs (I_ ..... lj), it leaves the real-world process connecting (11 ..... I/) with A unanalyzed. Because a theoretical model of the process is not available, rules for producing A must be obtained exclusively by empirical discovery. For instance, gunpowder was discovered and utilized by people who did not have a theory of combustion adequate to explain chemical explosive action. Various steelmaking technologies were developed by medieval European and Arabian smiths in the complete absenceof howandwhytheirtech of anunderstandingniquesworked.givenproblems,
