A New Approach to Machine Learning Based on Functional Decomposition of Multi -Valued Functions

This paper presents a novel design of a hardware classifier based on combining modified Ashenhurst-Curtis Decomposition and modified multiplexer-based synthesis. The PSUD classifier brings three new contributions: an approach to solve the column multiplicity problem for both types of decomposition, the way how the multiple-valued variables are encoded and used, and a modification of the Ashenhurst-Curtis Decomposition. One of the biggest challenges in Boolean function decomposition is the variable partitioning problem. Thus, we introduce a new representation of two combined classifiers for multiple-valued functions to overcome the variable partitioning problem which allows finding the minimal column multiplicity and consequently to find high quality decompositions leading to good learning. The classifier gives very good accuracy results when tested on multi-valued Machine Learning benchmarks from the UC Irvine repository.