Connection Optimization of a Neural Emotion Classifier Using Hybrid Gravitational Search Algorithms

Artificial neural network is an efficient model in pattern recognition applications, but its performance is heavily dependent on using suitable structure and connection weights. This paper presents a hybrid heuristic method for obtaining the optimal weight set and architecture of a feedforward neural emotion classifier based on Gravitational Search Algorithm (GSA) and its binary version (BGSA), respectively. By considering various features of speech signal and concatenating them to a principal feature vector, which includes frame-based Mel frequency cepstral coefficients and energy, a rich medium-size feature set is constructed. The performance of the proposed hybrid GSA-BGSA-neural model is compared with the hybrid of Particle Swarm Optimization (PSO) algorithm and its binary version (BPSO) used for such optimizations. In addition, other models such as GSA-neural hybrid and PSO-neural hybrid are also included in the performance comparisons. Experimental results show that the GSA-optimized models can obtain better results using a lighter network structure.