DeepMind proposes novel way to train 'safe' reinforcement learning AI

Alphabet's DeepMind proposes a new technique that "safely" trains reinforcement learning models, so as to avoid undesirable states.