Maintaining Diversity in an SVM integrated Case Based GA for Solar Flare Prediction

Abstract

Unusually intense solar flares may cause serious calamities such as damages of electric/nuclear power plants. It is thereupon highly demanded, but is quite difficult, to predict intense solar flares due to the imbalanced character of the available data. To cope with this problem, we have heretofore developed and applied a Case Based Genetic Algorithm (CBGALO) that contains a Local Optimizer, which is a Support Vector Machine (SVM). However, the prediction performance significantly depends on input data for learning. Hereupon, CBGALO is further extended by a Case Based automatically restartable Good combination searching GA for both learning features and input data (CBRsGcmbGA). Even the powerful but computationally expensive Deep Learning cannot automatically (evolutionarily, in our approach) search the learning data. Our approach solved this problem a little better by the case-based approach. However, it became obvious that even this work suffers from the typical GA effect in falling into local optima. To improve the results, we hence developed newly a diversity maintenance approach that inserts good individuals with large Hamming distance into the case base as elite individuals in GA’s population. In 2 out of 3 classes of solar flares, the performance of our new approach became as high as the best ones among the conventional world top records. Namely, even in those ≥ C class solar flares, our approach applying the Hamming distance to increase diversity had as high a performance 0.662 as compared with the conventional world top record 0.650.