Coevolutionary feature selection and reconstruction in neuro - evolution for time series prediction

Feature reconstruction of time series problems produces reconstructed state-space vectors that are used for training machine learning methods such as neural networks. Recently, much consideration has been given to employing competitive methods in improving cooperative neuro-evolution of neural networks for time series predictions. This paper presents a competitive feature selection and reconstruction method that enforces competition in cooperative neuro-evolution using two different reconstructed feature vectors generated from single time series. Competition and collaboration of the two datasets are done using two different islands that exploit their strengths while eradicating their weaknesses. The proposed approach has improved results for some of the benchmark datasets when compared to standalone methods from the literature