I also think the armorphous idea of "fun" can be 'metric'ed" to some degree.
What is "fun" in a game? Can fun be measured? I think it can be measured as the amount of engagement in the game. A game activity should have engagement as much as possible, and it should provide a decent range of win/loss conditions, based on how much you engage. Something is "easy" if merely engagement itself is guaranteed to win, whereas something is "hard" if you need engagement plus additional win conditions to ensure victory. Optimized fun is any state which maximizes engagement, in the sense that you feel like what you did mattered. If something doesn't require engagement at all to win, it's not a game.
e.g. if you're jumping over a pit in a side-scroller, then that's engagement, and the level of fun is dependent on things such as how quickly you jump and whether you can "glide" the controls in the air. the level of difficulty is dependent on such things as how accurate you need to be when you take off, and whether you need to be able to control the thing in the air, and if so, how sensitive it is. If you're guaranteed to make the jump no matter what, it's (not) so "fun", if you lose control for too long (e.g. the jump is too slow) that's not fun either, or if it's too difficult and fiddly then that becomes less fun as well. So you can in fact produce metrics for all those things.
e.g. something is not fun if you don't need to engage enough, or you need to engage too much to win. There should be a good sense that your options of how you engaged lead to the victory or loss condition. So in terms of this physics game, it's tuning the controls to jump a car over a ravine, and the car is a sci-fi one with some glide control from spoilers/wings. I believe we can in fact tune how "fun" that activity is via machine learning (I was using AI as the broad term, because it can really be by any means). e.g. how much is success or failure dependent on the specific controls you press. That can be tuned for directly.