How to get a "well-trained" RL agent in the first stage?

[YidingYu]

To explain the " black-box " NN is always interesting . However , it is hard to say if the " black-box " NN is well-trained (before explaining it ) . My question is that can we improve the " black-box " NN after distilling it ? IMO , the goal of RL-based method (no matter " black " or " white " box ) should be outperforming heuristic methods in most of the test cases .