Search processes guided by objectives are ubiquitous in machine learning. They iteratively reward artifacts based on their proximity to an optimization target, and terminate upon solution space convergence. Some recent studies take a different approach, capitalizing on the disconnect between mainstream methods in artificial intelligence and the field's biological inspirations. Natural evolution has an unparalleled propensity for generating well-adapted artifacts, but these artifacts are decidedly non-convergent. This new class of non-objective algorithms induce a divergent search by rewarding solutions according to their novelty with respect to prior discoveries. While the diversity of resulting innovations exhibit marked parallels to natural evolution, the methods by which search is driven remain unnatural. In particular, nature has no need to characterize and enforce novelty; rather, it is guided by a single, simple constraint: survive long enough to reproduce. The key insight is that such a constraint, called the minimal criterion, can be harnessed in a coevolutionary context where two populations interact, finding novel ways to satisfy their reproductive constraint with respect to each other. Among the contributions of this dissertation, this approach, called minimal criterion coevolution (MCC), is the primary (1). MCC is initially demonstrated in a maze domain (2) where it evolves increasingly complex mazes and solutions. An enhancement to the initial domain (3) is then introduced, allowing mazes to expand unboundedly and validating MCC's propensity for open-ended discovery. A more natural method of diversity preservation through resource limitation (4) is introduced and shown to maintain population diversity without comparing genetic distance. Finally, MCC is demonstrated in an evolutionary robotics domain (5) where it coevolves increasingly complex bodies with brain controllers to achieve principled locomotion. The overall benefit of these contributions is a novel, general, algorithmic framework for the continual production of open-ended dynamics without the need for a characterization of behavioral novelty.
If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu
Doctor of Philosophy (Ph.D.)
College of Engineering and Computer Science
Length of Campus-only Access
Doctoral Dissertation (Open Access)
Brant, Jonathan, "Open-ended Search through Minimal Criterion Coevolution" (2020). Electronic Theses and Dissertations, 2020-. 21.