Initial Value Problem Enhanced Sampling for Closed-Loop Optimal Control Design with Deep Neural Networks