Statistical Machine Translation (SMT) can be viewed as a generate-and-select process, where the selection of the best translation is based on multiple numerical features assessing the quality of a translation hypothesis. Training a SMT system consists in finding the right balance between these features, so as to produce the best possible output, and is usually achieved through Minimum Error Rate Training (MERT). Despite several improvements, training remains one of the most time consuming step in the development of SMT systems and is a major bottleneck for experimentations. Building on recent advances in stochastic optimization and online machine learning, this paper studies a possible alternative to MERT, based on standard and well-understood algorithms. This approach is shown to deliver competitive solutions, at a much faster pace than the standard training machinery.