Since its birth in the 1950ies, portfolio optimization has suffered from errors regarding the esti-mation of the input parameters (Michaud, 1989). To overcome the resulting underperformance, recent advances in Machine Learning mitigate the impact of estimation errors by directly opti-mizing portfolio weights from raw input data, e.g., using deep neural networks. However, these initial approaches still lack one important practical aspect by neglecting the (portfolio) weight constraints faced by real world asset management companies (e.g., short sale restrictions, in-dustry exposure limitation, factor exposure targets, diversification requirements, or upper bounds on transaction costs). We strive to improve on existing approaches by allowing for the implementation of such constraints. At the conclusion of this project, in addition to a scientific paper, we plan to provide a software toolbox in R and/or Python that implements our findings.