Philosophers and - "pseudognosticians" (the artificial intelligentsial) are coming more and more to recognize that they share common ground and that each can learn from the other. This has been generally recognized for many years as far as symbolic logic is concerned, but less so in relation to the foundations of probability. In this essay I hope to convince the pseudognostician that the philosophy of probability is relevant to his work. Formal systems, such as those used in mathematics, logic, and computer programming, can lead to deductions outside the system only when there is an input of assumptions. For example, no probability can be numerically inferred from the axioms of probability unless some probabilities are assumed without using the axioms: ex nihilo nihil fit.2

"Artificial intelligence tasks which can be formulated as constraint satisfaction problems, with which this paper is for the most part concerned, are usually solved by backtracking. By examining the thrashing behavior that nearly always accompanies backtracking, identifying three of its causes and proposing remedies for them we are led to a class of algorithms which can profitably be used to eliminate local (node, arc and path) inconsistencies before any attempt is made to construct a complete solution. A more general paradigm for attacking these tasks is the alternation of constraint manipulation and case analysis producing an OR problem graph which may be searched in any of the usual ways.Many authors, particularly Montanan i and Waltz, have contributed to the development of these ideas; a secondary aim of this paper is to trace that history. The primary aim is to provide an accessible, unified framework, within which to present the algorithms including a new path consistency algorithm, to discuss their relationships and the many applications, both realized and potential, of network consistency algorithms."See also: sciencedirect.comArtificial Intelligence 8:99-118

Further progress in the application of computers to many practical fields seems to depend heavily on the success in implementing learning and inductive processes within machines. For example, to develop a consultation system for medical or plant disease diagnosis, prognosis and decision making in general, it is very desirable, perhaps even necessary, to be able to'teach' the system through examples of correct and/or incorrect decisions, rather than by precisely describing the decision process in its full generality and then transforming this description into a computer program. A similar situation exists in computer chess. The development of computer programs playing at the master level (especially the end games) seems to be a formidable task if the programs are not eventually able to learn and improve on their decision making rules through the specific examples of games, rather than by being explicitly told all the rules. Due to easy access to human knowledge about chess and the relative simplicity of testing the results, chess is one of the most attractive testing domains for inductive inference programs.

EPISTEMOLOGICAL PROBLEMS OF ARTIFICIAL INTELLIGENCE John McCarthy Computer Science Department Stanford University Stanford, California 94305 Introduction In (McCarthy and Hayes 1969), we proposed dividing the artificial intelligence problem into two parts - an epistemological part and a heuristic part. This lecture further explains this division, explains some of the epistemological problems, and presents some new results and approaches. The epistemological part of Al studies what kinds of facts about the world are available to an observer with given Opportunities to observe, how these facts can be represented in the memory of a computer, and what rules permit legitimate conclusions to be drawn from these facts. It leaves aside the heuristic problems of how to search spaces of possibilities and how to match patterns. Considering epistemological problems separately has the following advantages: I. The same problems of what information is available to an observer and what conclusions can be drawn from information arise in connection with a variety of problem solving tasks. Recently we have found that introducing concepts as individuals makes possible a first order logic expression of facts usually expressed In modal logic but With important advantages over modal logic - and so far no disadvantages.

In an idealized picture of a scene that contains only polyhedra each line segment that is recorded can have only one of four possible "meanings". In order to understand the picture it is necessary that we be able to label each line with one of the four corresponding labels:, or A " " or "." label is associated, respectively, with a convex or concave edge that has both of its two associated planes visible. A line labelled with an arrow refers to a convex edge oriented so that only one of these two planes is visible from the camera and the other is hidden behind it. The orientation of the arrow along the line is such that the planes are to the right of the arrow. If no consistent set of line labels is possible the picture is of an "impossible"--object.

To use language one must be able to make inferences about the information which language conveys. This is apparent in many ways. For one thing, many of the processes which we typically consider "linguistic" require inference making. For example, structural disambiguation: (1) Waiter, I would like spaghetti with meat sauce and wine. You would not expect to be served a bowl of spaghetti floating in meat sauce and wine. That is, you would expect the meal represented by structure (2) rather than that represented by (3).

Machine intelligence problems are sometimes defined as those problems which (i) computers can't yet do, and (ii) humans can. We shall further consider how much "knowledge" about a finite mathematical function can, on certain assumptions, be credited to a computer program. Although our approach is quite general, we are really only interested in programs which evaluate "semihard" functions, believing that the evaluation of such functions constitutes the defining aspiration of machine intelligence work. If a function is less hard than "semihard," then we can evaluate it by pure algorithm (trading space for time) or by pure lookup (making the opposite trade), with no need to talk of knowledge, advice, machine intelligence, or any of those things. We call such problems "standard." If however the function is "semihard," then we will be driven to construct some form of artful compromise between the two representations: without such a compromise the function will not be evaluable within practical resource limits. If the function is harder than "semihard," i.e. is actually "hard," then no amount of compromise can ever make feasible its evaluation by any terrestrial device.

This paper discusses an effort in the application of artificial intelligence to the access of data from a large, distributed data base over a computer network. A running system is described that provides real-time access over the ARPANET to a data base distributed over several machines. The system accepts a rather wide range of natural language questions about the data, plans a sequence of appropriate queries to the data base management system to answer the question, determines on which machine(s) to carry out the queries, establishes links to those machines over the ARPANET, monitors the prosecution of the queries and recovers from certain errors in execution, and prepares a relevant answer. In addition to the components that make up the demonstration system, more sophisticated functionally equivalent components are discussed and proposed. The work described in this paper represents the joint efforts of an integrated, energetic group at SRI. Members of this group include Rich Fikes (now at Xerox PARC), Koichi Furukawa (now at ETL).

How can we get a computer to understand natural language? Our view of the problem has progressed over the years to a point where an answer to that question today would look quite different from one given ten or even five years ago. Originally, researchers felt that the most relevant issue was syntax. Later, most people agreed that semantics was the most relevant field of study (although few would have agreed on what semantics was). Five years ago, or so, our research was concentrated on finding an adequate meaning representation for sentences.

The program analyses carefully the initial situation. It creates some plans and tries to execute them. It analyses the situations deeper in the tree only if the plan fails. In that case it generates new plans correcting what is wrong in the old one. So, the program considers only natural branches of the tree. It can find combinations for which it is necessary to look more than twenty ply ahead. The paper describes the methods used for analyzing a situation and for modifying unsuccessful plans. Then we examine some results found by the program.Artificial Intelligence 8 (1977), 275-321