Tw o important issues in machine learning are explored: the role that memory plays in acquiring new concepts; and the extent to which the learner can take anactive part in acquiring these concepts. This chapter describes a program, called Marvin, which uses concepts it has learned previously to learn new concepts. The program forms hypotheses about the concept being learned and tests the hypotheses by asking the trainer questions. Learning begins when the trainer shows Marvin an example of the concept to be learned. The program determines which objects in the example belong to concepts stored in the memory. A description of the new concept is formed by using the information obtained from the memory to generalize the description of the training example. The generalized description is tested when the program constructs new examples and shows these to the trainer, asking if they belong to the target concept. 1.

When Newell introduced the concept of the knowledge level as a useful level of description for computer systems, he focused on the representation of knowledge. This paper applies the knowledge level notion to the problem of knowledge acquisition. Two interesting issues arise. First, some existing machine learning programs appear to be completely static when viewed at the knowledge level. These programs improve their performance without changing their "knowledge." Second, the behavior of some other machine learning programs cannot be predicted or described at the knowledge level. These programs take unjustified inductive leaps. The first programs are called symbol level learning (SLL) programs; the second, non-deductive knowledge level learning (NKLL) programs. The paper analyzes both of these classes of learning programs and speculates on the possibility of developing coherent theories of each. A theory of symbol level learning is sketched, and some reasons are presented for believing...

Abstract not found

Solutions to the problem of learning from examples will have far-reaching benefits, and therefore, the problem is one of the most widely studied in the field of machine learning. The purpose of this study is to investigate a general solution method for the problem, the automatic generation of decision (or classification) trees. The approach is to provide insights through in-depth empirical characterization and evaluation of decision trees for one problem domain, software resource data analysis. The purpose of the decision trees is to identify classes of objects (software modules) that had high development effort or faults, where "high" was defined to be in the uppermost quartile relative to past data. Sixteen software systems ranging from 3000 to 112,000 source lines have been selected for analysis from a NASA production environment. The collection and analysis of 74 attributes (or metrics), for over 4700 objects, capture a multitude of information about the objects: development effort...

ions are made over the structural information (relations)

Marvin is a program which is capable of learning concepts from many different environments. It achieves this by using a flexible description language based on first order predicate logic with quantifiers. Once a concept has been learnt, Marvin treats the concept description as a program which can be executed to produce an output. Thus the learning system can also be viewed as an automatic program synthesizer. The ability to treat a concept as a program permits the learning system to construct objects to show a human trainer. Given an initial example by the trainer, Marvin creates a concept intended to describe the class of objects containing the example. The validity of the description is tested when Marvin constructs an instance of the concept to show the trainer. If he indicates that the example constructed by the program belongs to the concept which is to be learnt, called the 'target', then Marvin attempts to generalize the description of its hypothesized concept. If the example does not belong to the target then the description must be made more specific so that a correct example can be constructed. This process is repeated until the description of

Abstract not found

Abstract not found

In this paper we compare the two main lines of research in learning most specific generalizations (MSG's) in a unifying framework. By reducing them to each other we show that even in some simple subset of first-order logic, the MSG grows exponentially in the number of examples. We then review two polynomial approaches, learning most specific Hornclauses without existential variables and learning most specific ij-determinate Hornclauses within this framework. We also show that ij-determinate Hornclauses are a maximal subset of Hornlogic, which is polynomial learnable, as the relaxation from ij-determinate Hornclauses to determinate Hornclauses lead to exponentially longer MSGs. Keywords: Inductive Logic Programming, Most Specific Generalization, Least General Generalization. 1 Introduction Recently the limited expressiveness of attribute-based descriptions has lead to an increased interest in learning from logical descriptions which, however, leads to an increased complexit...

This paper traces the development of the main ideas that have led to the present state of knowledge in Inductive Logic Programming. The story begins with research in psychology on the subject of human concept learning. Results from this research influenced early efforts in Artificial Intelligence which combined with the formal methods of inductive inference to evolve into the present discipline of Inductive Logic Programming.