A hybrid method for feature selection in the context of alternate test

Abstract

Machine-learning test strategy has been developed in the last decade as an alternative to costly specification-driven tests for Analog, Mixed-Signal and RF circuits (AMS-RF). The concept is simple: powerful algorithms are used to map simple measurements onto specifications. But the proper execution requires an information-rich input space. This paper presents an efficient hybrid algorithm to select the best subset of signatures (or features) among a large number of candidates and shows how it can be applied to eventually propose the development of new ones.