When I returned to Stanford last summer after a two-year leave of absence, serving as a program manager at the Defense Advanced Projects Agency, I was frequently asked about that experience. It was superb experience, for many reasons. As a program manager I had near-perfect vantage point from which to view the entire field of Artificial Intelligence. Not only did I become better acquainted with the most creative and active people in the field, I was also personally kept up to date on their latest research. ARPA is not just a place to go to provide a public service, but is really a central node in the research network for collecting and integrating results and disseminating them to the broader community: government, industry and the public at large. Moreover, it was my responsibility to identify new avenues of research and/or applications of research, coupled with the resources (limited, but real) to make these new activities happen -- a unique opportunity.

This paper presents the view that artificial intelligence (AI) is primarily concerned with propositional languages for representing knowledge and with techniques for manipulating these representations. In this respect, AI is analogous to applied in a variety of other subject areas. Typically, AI research (or should be) more concerned with the general form and properties of representational languages and methods than it is with the context being described by these languages. Notable exceptions involve "commonsense" knowledge about the everyday would ( no other specialty claims this subject area as its own ), and metaknowledge (or knowledge about the properties itself). In these areas AI is concerned with content as well as form. We also observe that the technology that seems to underly peripheral sensory and motor activities (analogous to low-level animal or human vision and muscle control) seems to be quite different from the technology that seems to underly cognitive reasoning and problem solving. Some definitions of AI would include peripheral as well as cognitive processes; here we argue against including the peripheral processes.

Consider a class of computing problem for which all bananas is left as an exercise for the reader, or the sufficiently short programs are too slow and all sufficiently monkey. When it has been possible to couple causal models problems of this kind were left strictly alone for the first with various kinds and combinations of search, twenty-years or so of the computing era. There were two mathematical programming and analytic methods, then good reasons. First, the above definition rules out both evaluation of t has been taken as the basis for "high road" the algorithmic and the database type of solution. In "low road" representations Second, in a pinch, a human expert could usually be s may be represented directly in machine memory as a set found who was able at least to compute acceptable A recent pattern-directed allocation, inventory optimisation, or whatever large heuristic model used for industrial monitoring and control combinatorial domain might happen to be involved.

A knowledge-based programming system can utilize a very-high-level self description to rewrite and improve itself. This paper presents a specification, in the very-high-level language V, of the rule compiler component of the CIII knowledgebased programming system. From this specification of part of itself, CIII produces an efficient program satisfying the specification. This represents a modest application of a machine intelligence system to a real programming problem, namely improving one of the programming environment's tools — the rule compiler. The high-level description and the use of a programming knowledge base provide potential for system performance to improve with added knowledge.In Hayes, J. E., Michie, D., and Pao, Y.-H. (Eds.), Machine Intelligence 10. Ellis Horwood.

The notion of an ethical machine can be interpreted in more than one way. Perhaps the most important interpretation is a machine that can generalize from existing literature to infer one or more consistent ethical systems and can work out their consequences. An ultra-intelligent machine should be able to do this, and that is one reason for not fearing it.In Hayes, J. E., Michie, D., and Pao, Y.-H. (Eds.), Machine Intelligence 10. Ellis Horwood.

ACM Computing Surveys, 14(l), March

This paper presents generalizations of Bayes likelihood-ratio updating rule which facilitate an asynchronous propagation of the impacts of new beliefs and/or new evidence in hierarchically organized inference structures with multi-hypotheses variables. The computational scheme proposed specifies a set of belief parameters, communication messages and updating rules which guarantee that the diffusion of updated beliefs is accomplished in a single pass and complies with the tenets of Bayes calculus.Proc AAAI

Discourse Processes, vol. 5, No. 1, pp. 1-32.

We are interested in constructing a computer agent whose behaviour will be intelligent enough to perform cooperative tasks involving other agents like itself. The construction of such agents has been a major goal of artificial intelligence research. One of the key tasks such an agent must perform is to form plans to carry out its intentions in a complex world in which other planning agents also exist. To construct such agents, it will be necessary to address a number of issues that concern the interaction of knowledge, actions, and planning. Briefly stated, an agent at planning time must take into account what his future states of knowledge will be if he is to form plans that he can execute; and if he must incorporate the plans of other agents into his own, then he must also be able to reason about the knowledge and plans of other agents in an appropriate way. In Hayes, J. E., Michie, D., and Pao, Y.-H. (Eds.), Machine Intelligence 10. Ellis Horwood.

PROLOG is a simple and powerful progamming language based on first-order logic. This paper examines two possible extensions to the language which would generally be considered "higher-order".t The first extension introduces lambda expressions and predicate variables so that functions and relations can be treated as 'first class' data objects. We argue that this extension does not add anything to the real power of the language. The other extension concerns the introduction of set expressions to denote the set of all (provable) solutions to some goal. We argue that this extension does indeed fill a real gap in the language, but must be defined with care.In Hayes, J. E., Michie, D., and Pao, Y.-H. (Eds.), Machine Intelligence 10. Ellis Horwood.