Obermayer, 2016. Non-Deterministic Policy Improvement Stabilizes Approximated Reinforcement Learning.