Automatic synthesis of control programs by combination of learning and problem solving methods

This paper outlines an approach for generating a series of optimal control actions in processes, which cannot (or can only partly) be modelled mathematically, by combining learning with problem solving methods. This approach has a suitable real time behaviour and can handle a larger number of different discrete control actions. A training data set is generated by a problem solver which generates for any points of the state space an optimal control action. "Optimal" means that an evaluation of that action is obtained from an empiric evaluation function. For the choice of that function some apriori knowledge of the current process is desirable. Classification algorithms (e.g. CAL5 and DIPOL) were used to split the state space into boxes with a unique control action attached. Optimal control trajectories are automatically generated by applying the resulting classifier recursively. A simulation of the roll axis stabilization of a communication satellite in orbit was taken to illustrate thi s approach. The main result is that the controller based on the learned classifier properly operates and it delivers better results in most cases than an appropriate "classical" PD- controller.