Defining Meta-Features for a Neuro-Fuzzy Meta-Learner

C. Castiello, G. Castellano, and A.M. Fanelli (Italy)

Keywords

Metalearning, Neurofuzzy systems, Metadata definition, Bias learning.

Abstract

Meta-learning is a novel paradigm aimed to overcome the limitations of traditional base-learning strategies. In par ticular, meta-learning strategies address the problem of ex ploiting past experience to dynamically identify the best bias for each task. In this paper, we present a learning framework, built up on the basis of the neuro-fuzzy inte gration, where a meta-learner is employed in cross-task ap plications to produce a form of meta-knowledge useful to improve the learning performance of a base-learner. To de rive meta-knowledge, a significant set of meta-data should be firstly collected. This work addresses the definition of meta-data, that are obtained by correlating an ensemble of meta-features (properly chosen to characterise base-level problems) with empirically determined bias configurations (presiding over the base-learner behaviour). A common en gine represented by a neuro-fuzzy learning scheme is ex ploited both at baseand meta-level of learning for deriv ing, directly from data, knowledge expressed in form of fuzzy rules.

Important Links:



Go Back