The annotation and characterization of tissue-specific cis-regulatory elements (CREs) in non-coding DNA represents an open challenge in computational genomics. Several prior works show that machine learning methods, using epigenetic or spectral features directly extracted from DNA sequences, can predict active promoters and enhancers in specific tissues or cell lines. In particular, very recently deep-learning techniques obtained state-of-the-art results in this challenging computational task. In this study, we provide additional evidence that Feed Forward Neural Networks (FFNN) trained on epigenetic data and one-dimensional convolutional neural networks (CNN) trained on DNA sequence data can successfully predict active regulatory regions in different cell lines. We show that model selection by means of Bayesian optimization applied to both FFNN and CNN models can significantly improve deep neural network performance, by automatically finding models that best fit the data. Further, we show that techniques applied to balance active and non-active regulatory regions in the human genome in training and test data may lead to over-optimistic or poor predictions. We recommend to use actual imbalanced data that was not used to train the models for evaluating their generalization performance.