Here is a real life (albeit informal) example: when I was still in the Old Country, a professional trainer told me the following (translated here to the notation of [Yoshikawa et al 2018]) about his process of "tuning up" horses that someone else (usually the client themself) has messed up:
1. Be an expert: this means (at least implicitly) knowing R*
2. Ask the horse to do a task involving Q. If the horse does it perfectly, the owner needs work. (if you like dealing with humans and also know R'*, start over by teaching the owner at step 1 and Q'; if you don't like dealing with humans or don't know R'*, pass the buck to a colleague who does both)
3. At this point determine if the horse either (a) doesn't realise you wanted Q, (b) can't do Q because they're insufficiently fit, (c) can't do Q because they don't know how, or (d) could have done Q but doesn't want to. In the cases of (a) or (d), the remedy is (for the purposes of this example) straightforward. In the cases of (b) or (c) it's more complex, leading to:
4. Walk the antimatroid A(R*) to determine which subsets Q_p the horse can physically do and Q_m the horse knows how to do; in the generic case this will give you a task Q_1 which can be trained having some subset of Q_p ⋃ Q_m as prereqs with a single novel q (unit inference?) to work on.
5. Until the owner says "good enough" (common) or can't ask for Q_n (common, see step 2) or you reach Q (rare), train the other steps of a minimal path Q_2...Q [Exercise: does such a path exist? (easy)]