Constructing control rules for a dynamic system: probabilistic qualitative models, lookahead and exaggeration

Abstract

Dynamic systems have sometimes to be controlled in circumstances that are very inconvenient for classical control methods. In such cases, use of qualitative controllers seems to be a promising alternative, as they can be designed without having an exact mathematical model. A method for synthesizing control rules is presented, in which a probabilistic qualitative model is used together with a decision strategy that takes into account the results of qualitative simulation. As a case study, the problem of the inverted pendulum control is considered. By combining the model simulation lookahead and a decision criterion, an interesting phenomenon of ‘exaggerated simulation’ has emerged. Experimental results show that in certain cases, using an exaggerated force in the process of simulation considerably improves the performance of generated control rules. The exaggeration helps to detect some trends of the system's behaviour that would otherwise have remained hidden. Experiments and mathematical analysis show that the improvement due to the exaggeration is comparable to that achieved by a deeper lookahead in the state transition graph.