Abstrakt
Planning domain models are critical input to current automated planners. These models provide description of planning operators that formalize how an agent can change the state of the world.
It is not easy to obtain accurate description of planning operators, namely to ensure that all preconditions and effects are properly specified. Therefore automated techniques to learn them are important for domain modelling.
In this paper, we propose a novel method for learning planning operators (action schemata) from example plans. This method, called LOUGA (Learning Operators Using Genetic Algorithms), uses a genetic algorithm to learn action effects and an ad-hoc algorithm to learn action preconditions.
We show experimentally that LOUGA is more accurate and faster than the ARMS system, currently the only technique for solving the same type of problem.