Automating Prior Work Replication Through Experiment Configuration Evolution
Department
Computer Science
Major
Computer Science
Research Advisor
Tauritz, Daniel R.
Advisor's Department
Computer Science
Funding Source
Missouri S&T Opportunities for Undergraduate Research Experiences (OURE) Program; Computer Research Association's Committee on the Status of Women in Computing Research
Abstract
Replicating prior work to validate claims and perform comparisons is a critical but arduous task in many research fields. This project surveyed a large set of papers published at the prestigious 2011 ACM Genetic and Evolutionary Computation Conference and found that this task is particularly onerous in the field of Evolutionary Computing due to ambiguity in terminology, a lack of commonly accepted standard algorithms, and a pervasive tendency in literature to not fully specify algorithmic configurations nor provide source code. In response, a system is being proposed to automatically replicate configurations for Evolutionary Algorithms (EAs) based on published experimental data, with accelerations possible if partial configuration information is available. The proposed system will employ a meta-EA to evolve various configurations for a base EA, including the operators used and parameter values, to most accurately match the published results of a given EA.
Biography
Jacob is a senior in Computer Science at Missouri S&T. He is an undergraduate research assistant in S&T’s Natural Computation Laboratory. Jacob spends most of his free time in theatre or working with ACM SIG-Game to develop the next MegaMinerAI competition. MegaMinerAI is a 24 hour AI programming competition held twice a year welcoming programmers and strategists of all skill levels to come and compete.
Research Category
Research Proposals
Presentation Type
Poster Presentation
Document Type
Poster
Location
Upper Atrium/Hallway
Presentation Date
10 Apr 2012, 1:00 pm - 3:00 pm
Automating Prior Work Replication Through Experiment Configuration Evolution
Upper Atrium/Hallway
Replicating prior work to validate claims and perform comparisons is a critical but arduous task in many research fields. This project surveyed a large set of papers published at the prestigious 2011 ACM Genetic and Evolutionary Computation Conference and found that this task is particularly onerous in the field of Evolutionary Computing due to ambiguity in terminology, a lack of commonly accepted standard algorithms, and a pervasive tendency in literature to not fully specify algorithmic configurations nor provide source code. In response, a system is being proposed to automatically replicate configurations for Evolutionary Algorithms (EAs) based on published experimental data, with accelerations possible if partial configuration information is available. The proposed system will employ a meta-EA to evolve various configurations for a base EA, including the operators used and parameter values, to most accurately match the published results of a given EA.
