Experiments in adaptive rule-based control

In this paper, it will be shown how artificial intelligence techniques can be applied in control engineering. Our initial interest was the development of an adaptive rule-based controller, where the control rules are derived through observing human skills on a simulator then applied to automatically control a physical system. Experimentation has shown that it is impossible for a human to control even a simple physical apparatus. Both on the simulator and on the physical system the efficiency and robustness of our controller has been compared with the performance obtained from another rule-based controller, one where the rule was derived from considering the systems' dynamic equations only. Trials conducted on the simulator indicated whose rule was based on theory alone. Our controller coped with changing system parameters, the other could not. However, trials conducted on the physical system showed that the two rule-based automatic controllers could control the physical system equally effectively. The results obtained from the automatically controlled physical system were used for trial-and-error learning, passive-learning. Here, sensor data was post-processed off-line, using a version of the ID3 rule-induction algorithm called c4.5. Feedback from these passively learned rules allowed a direct comparison to be made with the human generated rules.