I do not think that it figures out "what the win is" as the score parameter is explicitly made available to the algorithm.

In some sense this paper even demonstrates that how simple the problem actually is.

I think more important question is that what else can be modelled as such problem.