Neuroevolutionary Transfer Learning for Time Series Forecasting

In this paper, we propose a neuroevolution technique specifically designed for evolving LSTM networks. The proposed technique uses a grammar-based approach to evolve LSTM neural networks for time series prediction tasks, and is based on a previous technique which was designed in order to evolve CNN networks.

We use transfer learning in order to reduce the computational time of our approach. We have compared results obtained with other state of the art time series forecasting techniques on twenty time series, which contains data generated by sensors placed on a number of Iberian pigs. Results obtained confirm the effectiveness of the strategy proposed in this work.

Overall, we showcase the potential of our proposal in producing precise and efficient deep learning models for time series prediction, as well as the adaptability of transfer learning to new datasets.