loading page

A Multi-Objective Evolutionary Algorithm based on Sub-Populations For a Real-World Variable Selection Problem
  • Anderson Soares
Anderson Soares

Corresponding Author:[email protected]

Author Profile

Abstract

Variable selection is the focus of several research in applications in which hundreds of thousands of variables are available. In this work we analyzing a real-world problem of protein content estimation using Near-InfraRed spectroscopy. Instrumental methods are capable of generating nearly instantaneous results and can be operated by non-technical personnel with good precision. However, the accuracy of the method is dependent upon its calibration to associate the energy absorbed in hundreds of frequencies by the protein concentration in a sample. The main challenge for the calibration task is to select a suitable number of measures (or variables). The selection of a subset of uncorrelated variables is fundamental for establishing correct correlations and reducing prediction error. This paper proposes a multi-objective evolutionary algorithm based on sub-populations for variable selection in multivariate calibration problems. The multi-objective problem is decomposed into single objectives with a tournament among them to evaluate the best solutions for each objective considered. This approach is compared in this paper with a mono-objective evolutionary algorithm and deterministic classical variable selection methods. The results show that the proposed formulation has a smaller prediction error when compared to the mono-objective formulation with a lower number of variables. Finally, a noise sensitivity study obtained by the multi-objective formulation shows a significantly better result when compared to the other algorithms.