Prediction of energy consumption in a NSGA-II-based evolutionary algorithm

A deeper understanding in how the power consumption of evolutionary algorithms behaves is necessary to keep meeting high quality results without wasting energy resources. This paper presents a black-box model for predicting the energy consumption of the NSGA-II-based Parallel Islands approach to Multiobjective Feature Selection (pi-MOFS). We analyzed the power usage of each stage in pi-MOFS when applied to a brain-computer interface classification task. Fitness evaluation showed as the most relevant stage for the case study presented in time and power consumption. The results showed a 98.81% prediction accuracy for the eight experiments designed. We believe that our findings and methodology can be used to apply pi-MOFS, NSGA-II and other EAs to current optimization problems from an energy-aware perspective.