Determining Representativeness of Training Plans: A Case of Macro-operators

Most learning for planning approaches rely on analysis of training plans. This is especially the case for one of the best-known learning approach: the generation of macrooperators (macros). These plans, usually generated from a very limited set of training tasks, must provide a ground to extract useful knowledge that can be fruitfully exploited by planning engines. In that, training tasks have to be representative of the larger class of planning tasks on which planning engines will then be run. A pivotal question is how such a set of training tasks can be selected.
To address this question, here we introduce a notion of structural similarity of plans. We conjecture that if a class of planning tasks presents structurally similar plans, then a small subset of these tasks is representative enough to learn the same knowledge (macros) as could be learnt from a larger set of tasks of the same class. We have tested our conjecture by focusing on two state-of-the-art macro generation approaches. Our large empirical analysis considering seven state-of-the-art planners, and fourteen benchmark domains from the International Planning Competition, generally confirms our conjecture which can be exploited for selecting small-yet-informative training sets of tasks.