In contemporary guitar music, guitar tablatures represent an efficient way of writing down playing instructions for a guitarist. Contrary to classical sheet music notation, tablatures describe one distinct, ideally easy way of playing a melody. This makes tablatures an effective and intuitive notation system, but makes transcribing sheet music to a suitable tablature a tedious task requiring expert knowledge. The project at hand aims to automate this procedure and predict the optimal tablature for any input melody. We show two different approaches, both based on Machine Learning: In our first approach, guitar frettings are directly predicted based on previously played frettings. This is accomplished by a Long Short-Term Memory Recurrent Neural Network, enhanced by a notion of intention of the musical outcome. Our second approach predicts the difficulty of a fretting in terms of a cost function, rather than predicting the ideal fretting directly. The cost function is based on conditional probabilities in the training data and estimated by a Feed-forward Neural Network. The agreement between our best approach and published tablature is measured at 72.9% median validation accuracy, higher than what we achieve by applying a heuristic similar to previous approaches. Subjective evaluation shows that all generated tablatures are playable and that in many cases, divergence from the published tablature is well justified.