Abstract
Abstract : Big data analysis has gained popularity over the years as a result of developments in computing and electronics. Several methods have been proposed in literature for efficiently mining data from dedicated databases and a wide range of electronic sensors. However, as the volume of data grows, diversity and velocity of the data also grows (sometimes exponentially). Neural networks have been proposed in literature for optimal big data mining; however, they suffer from problems of over-fitting and under-fitting. In this paper, an ensemble of evolutionary algorithms is proposed, namely: improved non-dominated sorting genetic algorithm (NSGA), differential evolution (DE) and multi-objective evolutionary algorithm based on dominance and decomposition (MOEAD/D). These algorithms are each combined with a convolutional neural network (CNN), and performance is evaluated using root mean square error (RMSE) and mean absolute percentage error (MAPE). The test data consists of gas sensor readings obtained from an array of 16 metal oxide semiconductor sensors. The gases being detected are Carbon Monoxide/Ethylene in air, and Methane/Ethylene in air. 4,178,504 data points were collected over an uninterrupted 12-hour period. Preliminary results show improved RMSE and MAPE values over 50 learning cycles compared to a case where the CNN learned on its own.