Avoiding catastrophic failures in reinforcement learning by learning to shape rewards. - View it on GitHub