Say I want to estimate the change in entropy between two observed microstates: "cheerios in a box" and "cheerios on the floor." Assuming I have infinite time and compute, what's the best known way to choose which macrostates are most appropriate for use in measuring the change? Should I even be thinking about macrostates, or should I measure the change in some other way?
A friend once told me the answer was something along the lines of "pick the shortest string which outputs a class that includes the observed microstate, where that observed microstate is a non-unique member of that class", i.e. where the most-precise way to communicate the location of that microstate in class-space would be to mention simply that it was member #x of said class; if it were, for example, the "only prime member" of said class, then it would be considered unique.
Does (my perhaps-faulty memory of) this answer seem legit? Are there better proposals?