In the initial stage of prediction modeling, the primary focus was on model accuracy, often overlooking the interpretability of results. However, as those models gained popularity in various fields, an increasing demand for accurate and interpretable models emerged. In response to this demand, a series of models was developed to reconcile those two inherently contradictory aspects. An example is the e-AutoMFIS, based on principles of fuzzy logic and ensemble technique, which aims to reduce the dimensionality of problems and make multivariate predictions. The e-AutoMFIS outperforms more conventional modes in accuracy metrics in some cases. Its interpretability, coupled with those gains in accuracy, makes it a valuable tool for deeper analyses. However, some aspects require improvement, such as the configuration of model parameters, still carried out through exhaustive search, and random subsampling, which may require manual intervention for results optimization. Faced with those challenges, this dissertation proposes the optimization of e-AutoMFIS through the development of e-AutoMFIS2. This work details the architecture of e-AutoMFIS2, describing the modifications implemented in the subsampling process and the application of a multiobjective optimization genetic algorithm for parameter selection. Additionally, the potential benefits of those changes are discussed and results are evaluated through case studies, comparing them with its predecessor and other traditional models in the literature regarding accuracy and interpretability.