Learning Safe Numeric Action Models

Powerful domain-independent planners have been developed to solve various types of planning problems. These often require a model the acting agent's actions, given in some domain description language. Yet obtaining such an action is notoriously hard task. This task even more challenging mission-critical domains, where trial-and-error approach learning how act not option. In used generate plans must be safe, sense that generated with it applicable and achieve their goals. Learning safe models for has recently explored domains which states are sufficiently described Boolean variables. this work, we go beyond limitation propose NSAM algorithm. runs time polynomial number observations and, under certain conditions, guaranteed return models. We analyze its worst-case sample complexity, may intractable domains. Empirically, however, can quickly learn most problems domain.