- author = "Eric O. Scott",
- title = "Enhancing Evolutionary Algorithm Performance with Knowledge Transfer and Asynchronous Parallelism",
- school = "Department of Computer Science, George Mason University",
- year = "2022",
- address = "Fairfax, VA, USA",
- month = "Summer",
- note = "honorary mention ACM SIGEVO Best Dissertation Award 2023",
- keywords = "genetic algorithms, genetic programming, cartesian genetic programming, CGP",
-
URL = "
https://mars.gmu.edu/handle/1920/13201",
-
URL = "
https://mars.gmu.edu/bitstream/handle/1920/13201/Scott_gmu_0883E_12924.pdf",
-
URL = "https://www.proquest.com/openview/00365f4a46ea7e36102be2fca9ec7872/1",
- size = "317 pages",
- abstract = "Search and optimization problems are widespread throughout business and engineering, but these computational problems are often complex enough that they must be approached with heuristic algorithms. Evolutionary algorithms (EAs) offer a very general framework for solving many of these tasks, but their computational complexity can be difficult to manage when they are applied to mature and large-scale classes of problems. From my experience working on EA applications, I became concerned about two EA efficiency challenges that have been inadequately addressed to date: 1) parallelization strategies for evolutionary algorithms routinely suffer from idle CPU resources that go unused, and 2) customizing evolutionary algorithms with the domain-specific prior knowledge that they require in order to solve useful problems often requires costly and time-consuming research programs. In response to the idle-resources problem, I have engaged in a detailed study of asynchronous steady-state evolutionary algorithms (ASEAs) and their ability to better use large clusters of CPU resources. I studied issues of speedup, evaluation-time bias, and excess computational effort in ASEAs, concluding in particular that evaluation-time bias is a less severe problem for these algorithms than many practitioners have assumed. Next, I have addressed the problem of prior knowledge in EAs by engaging in a broad preliminary study of evolutionary knowledge transfer (EKT) and multi-task optimization. Motivated by natural examples of innovation engines that repurpose solutions to past tasks to find solutions to complex future tasks, I studied issues of transferability in different problem classes, negative transfer, and representation-based knowledge transfer approaches for evolutionary algorithms. My contributions in this area include proofs of a new set of no-free-lunch theorems for various types of transfer optimization, and several novel algorithms to address EKT challenges, including a many-source population-seeding algorithm that avoids negative transfer fairly easily, a multi-task Cartesian genetic programming approach, and a representation-learning algorithm that is able to learn and transfer genotype-phenotype maps across problem classes.",
-
notes = "MITRE Corporation
ProQuest Dissertations & Theses 2022.2932519
Manuel Lopez-Ibanez, University of Manchester, UK ACM SIGEVO Best Dissertation Award 2023, https://doi.org/10.1145/3629719.3629721
Supervisors: Sean Luke and Kenneth A. De Jong",
