This feels like exactly the kind of problem domain that belongs in (and can be solved by) RL?