Learning the NPN Representation of Certain Ternary Functions
Martin Lukac, Krzysztof Podlaski, Shinobu Nagayama, Michitaka Kameyama and Tagir Nukenov
Multiple-valued functions are of high importance in theory as well in the design of algorithms in novel technologies such as quantum computers. Properties of multiple-valued functions are important for the optimization of the cost and for the minimization of resource usage. Classification of functions into distinct groups is one of the approaches allowing to efficiently build complex systems, however an efficient classification method is required. In this paper, we analyze the usage of machine learning for learning the classification of ternary two trit functions into specific categories such as NPN, Bent and maximally asymmetric. We show that a) the classification works better for certain classes than others (even within a specific group of functions), b) the learning of groups of ternary functions is consistent with their relative difficulty and finally c) the amount of data required to learn a specific class is proportional to the learning effort of learning individual components of the classification.
Keywords: NPN classification, machine learning, ternary functions, neural networks
DOI: 10.32908/ijuc.v20.260125