Handling constrained optimization problems and using constructive induction to improve representation spaces in learnable evolution model
Ryszard S. Michalski, James E. Gentle, Janusz Wojtusiak · 2007
This dissertation investigates two closely related problems in the learnable evolution model: the automatic improvement of the representation space using constructive induction, and the handling of constraints in optimization tasks. The former includes an investigation of the theoretical and implementational aspects of representation space transformations in the context of complex optimization problems, the development of algorithms that perform these transformations, and algorithms for creating new candidate solutions (via instantiation) in the improved representation spaces. Handling specific types of constraints is closely related to the equation instantiation task in the modified representation spaces; therefore, the same methods can be used for solving both problems. Moreover, transformations of representation spaces may help in handling constraints of other types, that is, constraints that cannot be handled directly during the instantiation process. The developed algorithms are implemented in the LEM3 and AQ21 systems and tested on a set of constrained and non-constrained benchmark optimization problems. Two exemplary applications to optimization of complex systems in the context of selected medical datasets are also presented. These applications are the optimization of AQ21 parameters, and automatic discretization of numeric attributes.