Periodically diluted BEGNN model of corruption perception

Abstract

This study evaluates the performance provided by a Blume-Emery-Griffiths neural network (BEGNN) for two datasets of corruption indicators, namely the Corruption Perceptions Index and the Global Corruption Barometer. Bi-lineal and bi-quadratic terms are added to the Hamiltonian of the model, as well as for the order parameters to measure the network retrieval efficiency. The network is tested for different noise levels of the patterns’ initial state during the retrieval phase in order to measure the robustness of the network and its basin of attraction. The network connectivity is diluted periodically and its performance is tested for different levels of dilution. The network is analyzed in terms of the pattern load, mixing the real corruption patterns with random patterns in order to assess the change from retrieval to non-retrieval phases.