Learning Verifiable Control Policies Using Relaxed Verification