Learning Temporal Action Models Via Constraint Programming

We present a solver-independent Constraint Programming (CP) formulation for learning action models in temporal planning scenarios beyond PDDL2.1. Inspired by the CP approach for temporal planning, our formulation bases on a temporal plan trace and represents observations (as time-stamped states), actions, causal-link relationships, condition threats and effect interferences. This formulation is very expressive and supports a wide range of input knowledge. It also evidences the connection between the tasks of: i) action model learning, ii) plan validation, and iii) plan synthesis. Our experiments evaluate the quality of the learned models under different learning scenarios and in different planning domains.