Safe reinforcement learning in high-risk tasks through policy improvement | IEEE Conference Publication | IEEE Xplore