When dealing with expensive-to-evaluate objective functions in optimization problems, a large class of models gathered under the name of Bayesian Optimization (BO) is efficiently able to approximate the black-box objective function and compute the optimum. This is done using a surrogate model, namely a statistical model for modelling the objective function, and an acquisition function that let us move into the feature space. The most common surrogate models are Gaussian Processes. These algorithms work well over continuous domains, however, when dealing with discrete or categorical variables, these techniques become unsuccessful and different approaches and settings are required. The Separable Bayesian Optimization algorithm (SBO) is our proposal to overcome the BO limitations.