The last several years have seen a great expansion and diversification of research directions and approaches in machine learning, and a simultaneous interest in developing systems that integrate various learning strategies. Such a situation creates a need for analyzing and clarifying the relationships among different strategies and approaches, and building a theoretical basis for the implementation of multistrategy learning systems. This research attempts to develop a theoretical framework for describing diverse learning processes, based on a general notion of inference (hence, inference-based theory). According to this theory, a system learns from input information by trying to understand it i.e., to relate it to its background knowledge (BK). Based on this theory the researcher plans to develop a multistrategy task-adaptive learning (MTL) system, that synergistically integrates different learning strategies. Given an input and a goal of learning, an MTL learner applies the strategies) that are most appropriate according to the relationship between the input and learner's BK in the context of the learner's goal. The MTL methodology is intended to integrate ultimately such learning strategies as empirical learning, constructive induction, explanation-based learning, abduction, learning by analogy and abstraction.
