Learning temporal action models from multiple plans: A constraint satisfaction approach

Abstract

[EN] Learning, as a discovery task from past observations, is interesting in engineering contexts for identifying structures and improving accuracy. Learning in planning scenarios aims at recognizing past behavior to predict action models to improve decisions. This is appealing because practical scenarios are usually complex, sometimes difficult to be described formally, which require expert knowledge and engineering that becomes impractical in real-world applications. We introduce a Constraint Satisfaction formulation for learning PDDL2.1 temporal action models in planning. Given a collection of observations on multiple plans and a set of empty operators, we automatically create a learning task that identifies which conditions+effects are necessary, together with their temporal annotation, and induces durations and costs. Our formulation encapsulates planning (causal links, threats and effect interferences) and mutex (to avoid contradictions) constraints to be fully satisfied from the observed plans. The formulation is simple, but it proves very effective and easily adaptable to different levels of expressiveness. We evaluate such effectiveness in different IPC domains and compare the quality of the learning vs. other state-of-the-art learning approaches.