Classical reinforcement learning formulations compress all evaluation criteria into a single scalar reward value