A Taxonomy of Meta-learning Techniques and Proposed Framework for Automated Landmarker Generation and Selection

Many different perspectives have been adopted regarding the form of learning labelled as metalearning, with little to no consensus as to a proper definition. As such, a general definition and taxonomy of meta-learning techniques defined, segmenting meta-learning into two categories: mono-problem and multi-problem. A further taxonomy of multi-problem metalearning methods is then described, emphasizing the relationships between these, and how they apply to a general architecture for self-adaptive learning given by Vilalta and Drissi (2002). In addition, the recent migration toward the generation of better meta-data is discussed, along with a proposed framework for the potential automated generation of meta-attributes or landmarkers for algorithm selection.