Mimicking Evolution with Reinforcement Learning

We introduce the first reinforcement learning algorithm able to mimic the evolutionary process in an open-ended nature-like environment. Our algorithm searches for policies with increasing evolutionary success by taking into account every state and action each agent goes through its lifetime. In contrast, current evolutionary algorithms discard this information and consequently limit their potential efficiency at tackling sequential decision problems. We test our algorithm in two bio-inspired environments and unravel the interestingly dark evolutionary history of these worlds.