Safe Learning and Repairing of Numeric Action Models for Planning

摘要: Automated planners often require a model of the acting agent’s actions, given in some planning domain description language. Yet obtaining such an action model is a notoriously hard task. This task is even harder in mission-critical domains, in which a trial-and-error approach for learning how to act is not an option. In such domains, the action model used to generate plans must be safe, in the sense that plans generated with it must be applicable and achieve their goals. The challenge of learning safe action models for planning has been recently addressed for domains in which states are sufficiently described with Boolean variables. In this work, we go beyond this limitation and propose the Numeric Safe Action Model (N-SAM) learning algorithm. N-SAM runs in time that is polynomial in the number of observations and, under certain conditions, is guaranteed to return safe action models. Experimental results show that N-SAM is able to quickly learn a safe action model that can solve the majority of problems in a given domain. Finally, we describe an application of N-SAM to to repair action models that are observed to be incorrect.