Self-Improving Safety Performance of Reinforcement Learning Based Driving with Black-Box Verification Algorithms