Epistasis-based FSA: Two versions of a novel approach for variable selection in multivariate calibration.

Variable Selection in large datasets is a commonly procedure in multivariate calibration, which is a field of study from chemometrics. Selecting the most informative variables becomes an important step to build mathematical models through statistical techniques in order to predict some property of interest from an analyzed sample. Recombination-based search methods such as Genetic Algorithms (GAs) have been widely used as variable selection techniques to solve several optimization problems. However, previous works from literature have emphasized the schemata disruption problem caused by genetic operators. Therefore, this paper proposes two versions of an epistasis-based implementation (EbFSA) as a novel approach for variable selection in multivariate calibration problems, where each version is deterministic and performs a different strategy. The use of epistasis concepts becomes important to assess the genes (variables) interdependence. Based on our experimental results, we are able to claim EbFSA can select the most informative variables and overcome some state-of-the-art algorithms. • A novel approach for variable selection in multivariate calibration problems is presented. • We present two versions of an Epistasis-based implementation (EbFSA). • A comparison with traditional algorithms is performed. • We obtained advantages by reducing the number of variable as well as the prediction error. • We experimentally demonstrated that EbFSA can be a good choice for the variable selection problem. [ABSTRACT FROM AUTHOR]