Results


Generalization as Search

Classics (Collection 2)

We learn (memorize) multiplication tables, learn (discover how) to walk, learn (build UP an understanding of, then an ability to synthesize) languages. Many subtasks and capabilities are involved in these various kinds of learning. One capability central to many kinds of learning is the ability to generalize: to take into account a large number of specific observations, then to extract and retain the important common features that characterize classes of these observations. This generalization problem has received considerable attention for two decades in the fields of Artificial Intelligence, Psychology, and Pattern Recognition (e.g., [Bruner, The results so far have been tantalizing: Partially successful generalization programs have been written for problems ranging from learning fragments of spoken English to learning rules of Chemical spectroscopy. But comparing alternative strategies, and developing a general understanding of techniques has been difficult because of differences in data representations, terminology, and problem characteristics.


Teddington-4B-3-Merriman.pdf

Classics (Collection 2)

SESSION 4B PAPER 3 TO WHAT EXTENT CAN ADMINISTRATION BE MECHANIZED? Mr. J. H. H. Merriman was educated at King's College School, Wimbledon, and King's College, University of London. He obtained his B.Sc. (Hons.) in 1935 and did Postgraduate Research at King's College London obtaining his M.Sc. Engineering Department, Radio Research Branch, Dollis Hill, in 1936 and was associated with development of long distance radio communication systems. He was Officer-in-charge Castleton radio research station 1940-8, and from 1948-5 in the Office of Engineer-in- Chief G.P.O. and responsible for microwave system development and planning.


SESSION 3 PAPER 5 LEARNING MACHINES

Classics (Collection 2)

Recent activities have swung away from biology, but this will be remedied. THE application of learning machines to process control is discussed. Three approaches to the design of learning machines are shown to have more in common than is immediately apparent. These are (1) based on the use of conditional probabilities, (2) suggested by the idea that biological learning is due to facilitation of synapses and (3) based on existing statistical theory dealing with the optimisation of operating conditions. Although the application of logical-type machines to process control involves formidable complexity, design principles are evolved here for a learning machine which deals with quantitative signal and depends for its operation on the computation of correlation coefficients.


8 A Theory of Advice

Classics (Collection 2)

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."


10 And-or Graphs, Theorem-proving Graphs and Bidirectional Search

Classics (Collection 2)

And-or graphs and theorem-proving graphs determine the same kind of search space and differ only in the direction of search: from axioms to goals, in the case of theorem-proving graphs, and in the opposite direction, from goals to axioms, in the case of and-or graphs. Bidirectional search strategies combine both directions of search. We investigate the construction of a single general algorithm which covers unidirectional search both for and-or graphs and for theorem-proving graphs, bidirectional search for path-finding problems and search for a simplest solution as well as search for any solution. We obtain a general theory of completeness which applies to search spaces with infinite or-branching. In the case of search for any solution, we argue against the application of strategies designed for finding simplest solutions, but argue for assigning a major role in guiding the search to the use of symbol complexity (the number of symbol occurrences in a derivation).


8 A Further Note on Inductive Generalization

Classics (Collection 2)

In this paper, we develop the algorithm, given in Plotkin (1970), for finding the least generalization of two clauses, into a theory of inductive generalization. The types of hypothesis which can be formed are very simple. We have been guided by ideas from the philosophy of science, following Buchanan (1966). We can then look for the best such hypothesis. Although this problem is insoluble in general, some soluble subcases can be distinguished.


Report 82-36.pdf

Classics (Collection 2)

Report No. HPP 82-36 Explanatory Power for Medical Expert Systems: Studies in the Representation of Causal Relationships for Clinical Consultations J. W. Wallis, and E. H. Shortliffe Published This paper reports on experiments designed to identify and implement mechanisms for enhancing the explanation capabilities of reasoning programs for medical consultation. The goals of an explanation system are discussed, as is the additional knowledge needed to meet these goals in a medical domain. We have focussed on the gencra.lon of explanations tit are appropriate for different types of system users. This task requires a knowledge of what is complex and what is important; it is further strengthened by a classification of the associations or causal mechanisms inherent in the inference rules. A causal representation can also be used to aid in refining a comprehensive knowledge base so that the reasoning and explanations are more adequate. We describe a prototype system which reasons from ...


* Report 79 14 Induction Over Large Databases. Stanford -- KSL J.R. Quinlan, May 1979

Classics (Collection 2)

Techniques for discovering rules by Induction from large collections of Instances are developed. These are based on an Iterative scheme for dividing the Instances Into two sets, only one of which needs to be randomly accessible. Enterred) Induction Over Large Data Bases J. R. Quinlan Basser Department of Computer Science University of Sydney Abstract: Techniques for discovering rules by induction from large collections of instances are developed. These are based on an iterative scheme for dividing the Instances Into two sets, only one of which needs to be randomly accessible. These techniques have made it possible to discover complex rules from data bases containing many thousands of instances. Results of several experiments using them are reported. Introduction Since the early work with Perceptrons the study of learning systems has had an Important place In artificial Intelligence. The term'learning' of course covers a range of behavior. At one extreme there is the adjustment of ...



Buchanan22.pdf

Classics (Collection 2)

Jerold W. Wallis and Edward H. Shortliffe Developers of expert systems have increasingly recognized the importance of explanation capabilities to the acceptance of their programs; such capabilities are also critical in medical consultation system development (Gorry, 1973; Shortliffe, 1980). Good explanations serve four functions a consultation system: (1) they provide a method for examining the program's reasoning if errors arise when the system is being built; (2) they assure users that the reasoning is logical, thereby increasing user acceptance of the system; (3) they may persuade users that unexpected advice is appropriate; and (4) they can educate users in areas where users' knowledge may be weak. These diverse roles impose several requirements on the system. For example, the explanations must adequately represent the reasoning processes of the program, and they should allow the user to examine the reasoning history or underlying knowledge at various levels of detail. In addition, although the program's approach to a problem need not be identical to an expert's approach, the program's overall strategy and reasoning steps must be understandable and seem logical, regardless of the user's level of expertise.