Learning planning models

In AI planning, planners typically require a precise description of the input model. Creation of such model is a difficult task, so methods that automatically generate models from input plans are also created. However, a lot of them make assumptions about the model or are imprecise. In this thesis, we present a new method called LOUGA (Learning operators using genetic algorithms), which uses genetic algorithms to learn models. Unlike other methods, LOUGA does not make any assumptions about the model and works precisely even with small amount of predicates in input plans. In the first step, LOUGA generates all such pairs operator-predicate, that the operator can add or remove the predicate from the world. Every such pair is represented by one gene. In the second step, evolution is being performed until such model that explains input data without errors is found. In this paper, we have also proved empirically that LOUGA works faster and more precisely than existing method ARMS.