Relative Sparsity and Optimality-Based Reward Learning With Applications to Medical Decisions and Toxicology