Learning: The Construction of A Posteriori Knowledge Structures

This paper is a critical examination of both the nature of learning and its value in artificial intell igence. After examining alternative definitions it is concluded that learning is in fact any process for the acquisition of synthetic a posteriori knowledge structures. The suggestion that learning will not prove useful in machines is examined and it is argued. that i ts main application in practical Al sys terns is in providing a means by which a system can acquire know1 edge which is not readily formalizable. Finally some of the imp1 ications of these conclusions for future Al research are explored. In recent years machine learning seems to have undergone someth i ng of a renaissance. One manifestation of this is the recent publication of a book surveying the field (Michalski, Carbonell and Mitchel 1, 1983) Most of this book is devoted to reviewing what has been accomplished but, in what he clearly i ntended to be a provocative paper, Simon (1983) has rai sed a number of fundamental quest ions regard i ng the nature and value of machine learning. In this paper I attempt to provide answers to these questions. In particular I shal 1 try to define what learning is, why it is of great importance in artificial intelligence, how it relates to other branches of the subject and what these answers imply regarding future research in machine learning. 2. What Is Learninq ? When people use the term 'learning' in ordinary conversation they run little risk of being misunderstood. It is therefore somewhat surprising that A.I. researchers have had so much difficulty in arriving at a satisfactory definition. The usua 1 explanation of this phenomena is the claim that the everyday use of 'learning' is very general and imprecise and *82039% tion Sciences, University of actually refers to a heterogeneous collection of behaviors. There is much truth in this but the fact that people apply the same term to all these behaviors suggests that they have something fundamental in common. The most widely accepted broad definition of learning within the A.I. community appears to be one relating it to improved performance. For example :-" Learning is any change in a system that al lows it to perform better the second time on repetition of the same task or another task drawn from the same population " Simon (1983) This is a functional definition. That is it …