| Evolutionary algorithms have shown great promise in evolving novel solutions to real-world problems, but the complexity of those solutions is still limited, unlike the apparently open-ended evolution that occurs in the natural world. The power of traditional evolutionary algorithms is constrained by rapid convergence to a single solution on a sub-optimal local peak, leading to stagnation. In part, nature surmounts these complexity barriers with ecological dynamics that generate a diverse array of raw materials for evolution to build upon.;In this dissertation I focus on one ecological force that increases diversity: frequency-dependent selection that arises from competition among individuals for finite resources. I explore the benefits of incorporating this force into an algorithmic framework, focusing on how this mechanism can increase diversity to provide many evolutionary paths to a problem solution. The use of niching to simultaneously approach a single solution from many directions has not been extensively studied in previous literature. I study competition for limited resources in a digital evolution system and examine the importance of specific resource parameters on population dynamics. I show that my techniques are robust at increasing diversity over a broad range of settings for each parameter, and that the parameter settings are governed by a relatively simple set of equations.;Finally I introduce Eco-EA, a general form of an evolutionary algorithm that associates a limited resource with each trait to be evolved. I apply Eco-EA to several problems, including a real-world software engineering problem, and I show that the Eco-EA yields several advantages over traditional evolutionary algorithm approaches, including: (1) significantly more rapid evolution of targeted complex functions; (2) discovery and maintenance of a diverse set of partial solutions that together solve a problem; (3) maintenance of a selection of high-quality final solutions for the researcher to choose from, often with slightly different properties; and (4) discovery of solutions that are more evolvable when placed in new environments. |
