INTRODUCTION: Machine learning techniques have shown to be a powerful technique for inferring complex patterns and data. It is possible to use them as soft sensors, with the idea of reconstructing a variable difficult to measure or expensive to collect. In this work, we focus on the multi-criteria analysis of the number of features required to rebuild a variable and its trade-off with the learner complexity. – OBJECIVES: The general objective of this work is to evaluate the performance of {Logistic, Linear} regression trained with multi-objective optimization and/or multi-criteria ideas to compare their performance in machine learning benchmarks competitions. As specific objectives are defined: Identifying relevant competitions and data sets to the strategic objectives of the PPG program to be used in this research; Modifying algorithms for learning process to be evaluated in this research; Comparing obtained results with benchmark results to evaluate their performance. – MATERIALS AND METHOD: Matlab is used as the primary coding platform. Machine learning structures such as linear regression, decision tree, and support vector machine are employed. Criteria to select the most relevant features is the minimum-redundancy / maximum relevance technique. As a case study, a well-known benchmark of the area is used. Such a benchmark is based on a sulfite recovery unit (SRU). – RESULTS: An analysis of the number of features required to reconstruct the desired variables was performed. The underlying idea was to compare the trade-off between complexity (number of features) and reconstruction capabilities (performance) via the minimum-redundancy, maximum relevance criteria. – FINAL CONSIDERATIONS: The general and specific objectives of the project were achieved, with similar results as the ones reported in the literature.
