Quasi-exact logic functions through classification trees

Over the last few years machine learning (ML) has evolved becoming pervasive in many scientific and industrial fields. Such success is ignited by the possibility of applying algorithms that can solve complex problems through generic rule formulations. With such premise, ML surely represents a new opportunity to improve existing design paradigms. In this paper we propose a novel ML-driven synthesis methodology that allows to describe generic Boolean functions through a representative subset of core expressions using Classification Trees (CTs). Obtained circuits are able to mimic Boolean functions to a certain degree of accuracy, hence the name quasi-exact logic functions. The proposed synthesis flow enables a smart hardware mapping of quasi-exact logic functions by means of reduced and ordered decision diagrams. Experiments conducted on a subset of open-source benchmarks demonstrate that CTs are indeed able to cover rather complex Boolean functions with a very high degree of accuracy, 88% on average, still requiring 3 x less area over standard multi-level circuit counterparts.