Devils advocate - ML is introspective at some level and certainly can observe (with Spock-like objectivity; almost defining the term) rationality, and take the result of each move and grade it with some degree of confidence as a "good move" or "bad move" [and even contextualize the move: i.e., move : 'e4' ; context : "opening" => evaluation - "great move"]. I agree with your first point though w/r/t heuristics and more importantly pattern recognition which can be used to integrate in more heuristic knowledge in your aggregate 'decision making system' at a way more 'effective' rate (with respect to time, within the domain of the game Go).