NeuroEvolution: Using Genetic Algorithm for optimal design of Deep Learning models

Convolutional Neural Networks (CNN) have proved to be influential in image classification techniques. However, CNN has not proved much useful in time series prediction. This is because of the causal nature of time series data. With proper modifications in the architecture of CNN the model can perform well on time series prediction as well. However, the modification and design of CNN architecture require expertise and knowledge of data. In order to tackle this problem, a genetic algorithm solution is proposed, that designs its own optimal CNN architecture and was proved useful in predicting time series data. The robustness and accuracy of the model were tested on our dataset and compared with an RNN architecture. It was found that the CNN model performed well irrespective of the duration of prediction and these results are comparable to state of the art optimal RNN models.