That depends on whether you think that games should be simulations. Even if you think they should be simulations, you still might only want to reward optimal play over lucky play. If I'm using a simulation to train someone, I only want them to be rewarded for ideal play.

This is incorrect, as far as the research into training for expertise goes. You actually want simulations to be like the real world: complex, ambiguous, sometimes with delayed feedback and presenting multiple forms of the problem. Of course, it also takes a good mentor to debrief with and untangle what was bad luck and what was bad decision making. (See the Oxford Handbook of Expertise if this stuff interests you.)

Note that this is also why some people prefer games that don't have chance -- to experience what that is like.