We propose to clone the behavior of ``blindfolded'' experts that are compelled to employ non-trivial exploration to solve the task, which leads to better generalization.