Guiding evolutionary search towards innovative solutions

The main goal of this work is to develop a method that, operating on top of an Evolutionary Algorithm, increases its likeliness of finding innovative solutions. This likeliness is laid out to be increased with the diversity of the solutions found, provided that they are of sufficient quality. The developed method needs to be applicable in a scenario in which the search is required to be started from a single, fixed solution. Therefore, a scheme is envisioned in which the search is performed in a sequential fashion, zooming in on a locally-optimal solution, and then exploring for a new potentially high-quality region based on a memory of solutions encountered earlier in the search. Two exploration criteria, one using an archive of earlier solutions as memory and the other deriving from a surrogate model trained on earlier solutions, were established to be worthwhile for integration into quality-based search. The resulting schemes were applied to a real-world airfoil optimization task, showing both to perform better than the baseline method of multiple standard optimization runs. The model-based approach delivers the best results, in the sense that it finds more solutions, more diverse solutions, and better-quality solutions than the baseline method.