Inferential theory of learning as a conceptual basis for multistrategy learning

Abstract

Abstract. In view of a great proliferation of machine learning methods and paradigms, there is a need for a general conceptual framework that would explain their interrelationships and provide a basis for their integration into multistrategy learning systems. This article presents initial results on the Inferential Theory of Learning that aims at developing such a framework, with the primary emphasis on explaining logical capabilities of learning systems, i.e., their competence. The theory views learning as a goal-oriented process of modifying the learner's knowledge by exploring the learner's experience. Such a process is described as a search through a knowledge space, conducted by applying knowledge transformation operators, called knowledge transmutations. Transmutations can be performed using any type of inference--deduction, induction, or analogy. Several fundamental pairs of transmutations are presented in a novel and very general way. These include generalization and specialization, explanation and prediction, abstraction and concretion, and similization and dissimilization. Generalization and specialization transmutations change the reference set of a description (the set of entities being described). Explanations and predictions derive additional knowledge about the reference set (explanatory or predictive). Abstractions and concretions change the level of detail in describing a reference set. Similizations and dissimilizations hypothesize knowledge about a reference set based on its similarity or dissimilarity with another reference set. The theory provides a basis for multistrategy task-adaptive learning (MTL), which is outlined and illustrated by an example. MTL dynamically adapts strategies to the learning task, defined by the input information, the learner's background knowledge, and the learning goal. It aims at synergistically integrating a wide range of inferential learning strategies, such as empirical and constructive inductive generalization, deductive generalization, abductive derivation, abstraction, similization, and others.Keywords. Learning theory, multistrategy learning, inference, classification of inference, deduction, induction, abduction, generalization, abstraction, analogy, transmutation. For every belief comes either through syllogism or from induction.Aristotle, Prior Analytics, Book II, Chapter 23 (p. 90) ca 330 BC. I n t r o d u c t i o nMost research in machine learning has been oriented toward the development of monostrategy methods that employ one type of inference and a single computational mechanism. Such methods include, for example, inductive learning of decision rules or decision trees, explanation-based generalization, empirical discovery, neural net learning from examples, genetic algorithm-based learning, conceptual clustering, and others. The research progress on these and related topics has been reported by many authors, among them Laird (1988), Touretzky, Hinton, and Sejnowski (1988), Goldberg (1989), Schafer (1989) of learning problems than monostrategy systems...