so in an unknown environment, which one should be used? Q-learning can just learn state-action pairs, but utility learning can just learn the transition model then use it to calculate state utilities. I'm having difficulty finding pros and cons of using one over the other (other than computation).