We have developed analytic methods to discover barriers to student learning from data for educational technology use (see learnlab.org). Such discoveries can guide the redesign of instruction and our online experiments demonstrate enhanced learning outcomes. Our analytic methods span issues of student skill acquisition, metacognition, and motivation. Focusing on the first, I will illustrate how alternative cognitive models of learning can be evaluated by translating them to statistical models and predicting learning curve data for model comparison. We have used machine learning in a couple of ways to generate alternative cognitive models, one more practical and other more cutting edge. The second involves a computational model of student learning, SimStudent, that learns as students do by using an intelligent tutoring system. The cognitive models SimStudent acquires have been demonstrated to yield empirically-verified discoveries not present in the human-designed cognitive models behind the intelligent tutoring systems. In other words, with SimStudent is the potential to not only create intelligent tutoring systems without AI programming, but to also produce systems that are pedagogically more effective than human-built tutoring systems.