Evolutionary Feedforward Neural Network Algorithm and Its Application on The Example of Human Action Recognition

Abstract

An optimization method is suggested in this paper to improve the learning process in the artificial neural network algorithm (ANN). There are three phases included in the optimisation process: the breaking ANN process, an evolutionary algorithm (EA), and the combining ANN process into a new powerful feedforward artificial neural network (FANN). ANN breaks into multiple ANNs to facilitate and accelerate the optimization process in the EA phase. EA enhanced the ANNs, starting with a selection step by sorting ANNs based on the obtained accuracy for each ANN; the crossover step exchanges the characteristics between the best and worst ANN, and the rest of the ANNs are enhanced by the mutation process. Distinguishing the type of human activity is an important topic and one of the issues that concerns society in several aspects, like health care, criminal cases, etc. The system is applied to human action datasets. A multi-dataset of human actions has been applied to the system. On the desired recognition task, the system demonstrated high performance and efficiency in a range higher than 90% for each different motion type.