Adaptive simplex-GA hybrid for rule learning and parameter identification of complex systems

This dissertation puts forward an adaptive simplex-GA hybrid for rule learning and parameter identification of complex systems. A supervisory architecture is designed for optimization problems of complex systems. The upper supervisory layer uses domain knowledge to help gradually reduce the search space while the lower layer hybrid GA performs the actual optimization process within the range that the supervisory layer specifies. The proposed approach is applied to a real complex system—metabolic modeling and the results are better than those without using the supervisory architecture. An adaptive simplex genetic algorithm is also presented in which the percentage of simplex is self adaptive during the run. A set of rules is designed to adjust the simplex percentage by the feedback of fitness value. The new algorithm has been tested on both the sin function maximization problem and the real metabolic modeling problem, with results that are better than the fixed percentage approach. To further explore the application of GA on data mining area, an entropy-based adaptive GA approach for rule learning problems is put forward. Mutation inversion probability is self adaptive during the run and the trained classifier gives the final classification by entropy-based voting. This algorithm outperforms several other traditional data mining techniques on three testing databases.