Two novel hybrid Self-Organizing Map based emotional learning algorithms

Abstract

Emotions play an important role in human decision-making process, and consequently, they should be embedded into the reasoning process in our efforts to model human reactions. Adnan Khashman et al. have proposed an emotional backpropagation (EmBP) learning algorithm and have successfully applied it to several practical pattern recognition tasks. However, the design of the emotional input values to the EmBP is not reasonable and may thus cause the failure of its entire implementation. Aimed at improving this weakness, we propose a novel self-organizing map-based emotional neural network (EmSOM) learning algorithm. In contrast to EmBP, the emotional input values of EmSOM are determined based upon its correspondingly associated SOM blocks, and moreover, the network hierarchy has been taken into account in its design, thus improving the deficiencies of EmBP to a certain extent. Furthermore, we incorporate a sparse online SOM (SOR-SOM) algorithm into our emotional neural network learning algorithm and establish a hybrid sparse online relational SOM-based emotional neural network (Em-SOR-SOM) model, so that those advantages of SOR-SOM can be exploited to further boost the recognition performance of the model. The EmSOM and Em-SOR-SOM algorithms have been compared with SBP and EmBP, and several other recent algorithms, and their effectiveness and efficiency have been numerically confirmed by the experiments we presented on the ORL face database and three benchmark credit datasets.