Learning First-Order Rules for Derived Predicates from Plan Examples
Rao Dong · Chinese Journal of Computers · 2010
Derived predicates are a natural way to depict indirect effects of domain actions,and their truth values in the current state are inferred from that of other predicates via domain rules.However,domain rules designed by human experts cannot be guaranteed to be correct or complete.So it is often difficult to explain why an observed plan is valid under imperfect domain rules.Combining inductive learning with analytical learning,in this paper,we develop an algorithm called FODRL(First-Order Derived Rules Learning)to automatically discover first-order rules for derived predicates from observed plans under an initial domain theory.FODRL is based on the pure inductive learning system FOIL(First-Order Inductive Learning),which learns a new rule that covers partial positive examples but avoids all negative examples once a time,until all positive examples are covered.However,better than FOIL,FODRL uses activation sets of derived predicates to expand search steps so as to improve the accuracy of learned rules.An activation set is a minimal set of basic facts or predicates which can make a derived predicate hold true under domain rules.The learning process is divided into two steps:first,extract training examples from observed plans;then,learn first-order rules for derived predicates which can best fit training examples and the initial domain theory.We experiment in two derived planning domains,PSR and PROMELA.The results show that,with the guidance of an initial domain theory,the rules learned by FODRL are more accurate than those from FOIL,even FOCL(a descendant of FOIL).