Rafael Morales

Modelling Learners of a Control Task with Inductive Logic Programming: A Case Study

We present results of using inductive logic programming (ILP) to produce learner models by behavioural cloning. Models obtained using a program for supervised induction of production rules (ripper) are compared to models generated using a well-known program for ILP (foil). It is shown that the models produced by foil are either too specific or too general, depending on whether or not auxiliary relations are applied. Three possible explanations for these results are: (1) there is no way of specifying to foil the minimum number of cases each clause must cover; (2) foil requires that all auxiliary relations be defined extensionally; and (3) the application domain (control of a pole on a cart) has continuous attributes. In spite of foils limitations, the models it produced using auxiliary relations meet one of the goals of our exploration: to obtain more structured learner models which are easier to comprehend.