Evolvability is an important factor in evolution’s ability to create the engineering marvels that are ubiquitous in the natural world. It allows individual species to avoid extinction, invade new niches, and quickly adapt to new environments. For these reasons, producing evolvability in artificial evolution has been a long-standing research goal, both to better understand evolvability from a biological motivation and to harness evolvability in evolutionary algorithms to solve challenging engineering problems. Despite concerted effort by the research community, the conditions that promote evolvability are poorly understood and successful efforts to produce evolvability in artificial evolution are rare and have occurred only in narrowly tailored situations. The assumption by the research community is that we have yet to truly identify the main drivers of evolution in nature and that, as we increase our understanding of the natural drivers of evolution, we will be able to implement them and observe significant increases in evolvability in simulations of evolution and evolutionary algorithms. Doing so will improve our understanding of evolvability and reap tremendous societal benefits in all the various fields that harness evolution via evolutionary algorithms to solve challenging engineering problems.
Although evolvability is important both in natural and artificial evolution, there is no overall agreement on its definition. The two common definitions are (1) evolvability as the ability of populations to quickly adapt to unseen environments, hereafter called adaptive evolvabilityand (2) evolvability as a tendency of an individual’s offspring to exhibit phenotypic diversity, hereafter called phenotypic-variation evolvability. During my Ph.D. I wrote two papers that present methods of producing both types of evolvability in artificial evolution.
My first paper shows that the evolution of a hierarchical structure improves adaptive evolvability. Hierarchy is a recursive composition of modules and is a prevalent organizational property in both man-made and natural systems. Before showing how this common organizational structure could increase adaptive evolvability, the paper investigates why hierarchy evolves in the first place. It shows that a cost for network connections promotes the evolution of hierarchy, shedding light on the biological mystery of the evolutionary origins of hierarchy and providing a practical tool for encouraging hierarchy in networks evolved with evolutionary algorithms. Additional experimental results show that such hierarchy increases adaptive evolvability. I should note that this paper was a thorough and exhaustive examination of one way of promoting adaptive evolvability (encouraging hierarchy via a connection cost), representing many years of experimentation, analysis, and writing effort. It could have been broken down into smaller publishable units, but I chose instead to present one detailed, experimentally extensive, exhaustive journal paper on the subject.
My second paper introduces a new search algorithm called Evolvability Search that significantly increases phenotypic-variation evolvability by directly selecting for it. The algorithm quantifies phenotypic-variation evolvability by measuring the propensity of phenotypic variation among an organism’s offspring and then incorporates that quantification into a fitness function of an evolutionary search algorithm. That is, individuals whose offspring posses high phenotypic variation are rated highly by the fitness function and selected for reproduction. Results from experiments in two evolutionary robotics domains confirm that Evolvability Search produces high-performing solutions with an elevated phenotypic evolvability.
Overall, my Ph.D. study produced two papers that introduce methods of enhancing adaptive and phenotypic evolvability, contributing to the long-standing goal of evolving increasingly complex behaviors. In addition to showing methods that increase evolvability, my papers shed light on the fields of robotics and biology. For example, knowing the evolutionary origin of hierarchy will accelerate future research into evolving more complex, intelligent computational brains for robots. Evolvability Search, on the other hand, will make generating phenotypic-evolvability easy and straightforward, facilitating our ability to study, understand, and harness evolvability.
|Advisor:||Clune, Jeff M.|
|Commitee:||Douglas, Craig, James, Caldwell, Ruben, Gamboa, Thomas Jr, Bailey|
|School:||University of Wyoming|
|School Location:||United States -- Wyoming|
|Source:||DAI-B 78/12(E), Dissertation Abstracts International|
|Subjects:||Artificial intelligence, Computer science|
|Keywords:||Divergent search, Evolutionary algorithms, Evolvability, Evolving artificial neural networks, Hierarchy|
Copyright in each Dissertation and Thesis is retained by the author. All Rights Reserved
The supplemental file or files you are about to download were provided to ProQuest by the author as part of a
dissertation or thesis. The supplemental files are provided "AS IS" without warranty. ProQuest is not responsible for the
content, format or impact on the supplemental file(s) on our system. in some cases, the file type may be unknown or
may be a .exe file. We recommend caution as you open such files.
Copyright of the original materials contained in the supplemental file is retained by the author and your access to the
supplemental files is subject to the ProQuest Terms and Conditions of use.
Depending on the size of the file(s) you are downloading, the system may take some time to download them. Please be