This article addresses a classification problem relying on an ensemble of Structural Restricted Boltzmann Machines (SRBMs). Each SRBM in the ensemble is trained by imposing structural constraints on the related weight matrix, so as to enforce sparsity, and results in a probabilistic classifier. Hence, given a new instance, the ensemble gives rise to a credal classifier where the classification is carried out relying on the alpha-maxmin criterion, depending on a pessimism index \(\alpha \in [0,1]\) , and a \(\beta \) -quantile filtering of outliers. The paper presents an experimental analysis on artificial data sets to highlight the role of the parameters \(\alpha \) and \(\beta \) in the classification performances.