Learning approach that provides guarantees even when the true target function is outside the hypothesis class.