Learning reward functions assuming only a small subset of features are relevant, reducing complexity.