Metodologia de planejamento e redução do número de experimentos em problemas de busca ativa

Abstract

Many engineering problems involve the optimization of the unknown objective function. Recently, active search has emerged as a powerful tool to solve problems of this nature, whose objective function involves high evaluation costs, whether computational or experimental. This thesis proposal seeks to find an object (x) with an optimal value for a given property (y). However, direct determination of this property of interest across all available objects may not be a viable option given the resources, workload and/or time required. Thus, this proposes an active machine learning approach, called active search, to find an optimal solution, using the design of experiments for the initial search. To apply this method, two regression techniques were used, called k-nearest-neighbours and Gaussian processes. Furthermore, a stopping criterion was defined for the Gaussian regression technique to reduce the algorithm processing time. The originality of the theme lies in the proposed methodology, in the use of experimental design, no active search algorithm using regression techniques that quickly converge to a global optimum, and in the use of a stopping criterion for the algorithm based on statistical criteria. The studies were carried out with simulated data and with real data for the production of medicines, agrochemicals and application in electrical microgrids. In all cases, active search reduced the number of experiments and simulations to obtain the property of interest, compared to traditional algorithms such as Optimal Experiment Design and Kennard-Stone.