NBI AUTOENCODER: A MULTIOBJECTIVE OPTIMIZATION STRATEGY BASED ON NONLINEAR LATENT VARIABLES

Multiobjective optimization based on the Normal Boundary Intersection (NBI) method has limitations when applied to problems with high correlation between responses. This work investigates outcome strategies, with emphasis on nonlinear approaches based on autoencoders, to restore the efficiency of NBI in scenarios of strong multicollinearity. As a case study, an experimental dataset of a cladding process was used. Three approaches were compared: conventional NBI, VRF-NBI (factor analysis) and AE-NBI (autoencoder with covariance penalty). Performance was evaluated in terms of numerical feasibility, proximity to the point of utopia and diversity of solutions. The results show that conventional NBI is ineffective, with more than 99% of unfeasibility. The latent approaches ensured 100% feasibility, with AE-NBI presenting greater precision and consistency, while VRF-NBI preserves greater diversity. It is concluded that AE-NBI offers the best compromise between stability and performance.