Maximizing the stability of a majority perceptron using Tabu search

Quantization of the synaptic weights is a central problem in hardware implementation of artificial neural networks using numerical technology. The author reports on perceptrons performing a particular linear threshold function, called majority function, whose parameters are restricted to only three values: -1, 0, +1. Maximizing the stability of a perceptron computing such a function is NP-hard and requires the use of heuristic methods. Different approaches based on Tabu search (TS) are proposed to solve this problem, and comparative experiments are reported