Learning time-dependent Feedback Policies with Model-Based Policy Search