Genetic algorithms in controller design and tuning

A three-phased framework for learning dynamic system control is presented. A genetic algorithm is employed to derive control rules encoded as decision tables. Next, the rules are automatically transformed into comprehensible form by means of inductive machine learning. Finally, a genetic algorithm is applied again to optimize the numerical parameters of the induced rules. The approach is experimentally verified on a benchmark problem of inverted pendulum control, with special emphasis on robustness and reliability. It is also shown that the proposed framework enables exploiting available domain knowledge. In this case, genetic algorithm makes qualitative control rules operational by providing interpretation of symbols in terms of numerical values.