Domain-Specific Languages as a Method for Representation and Evolutionary Search Among Global Solution Space of Cognitively Plausible Algorithmic Behaviors
Open Access
- Author:
- Kaulakis, Ryan
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 06, 2020
- Committee Members:
- Frank Edward Ritter, Dissertation Advisor/Co-Advisor
Amulya Yadav, Committee Member
Richard Alan Carlson, Outside Member
Mary Beth Rosson, Program Head/Chair
Frank Edward Ritter, Committee Chair/Co-Chair
John Yen, Committee Member - Keywords:
- AI
ML
Cogsci
Cogarch
actr
dsl - Abstract:
- This work describes the design and implementation of a Domain-Specific Language Compiler, which permits evolutionary mechanisms to be quantified in terms of the measured behavior of the compiled ACT-R production rules and models that the compiler produces. The method permits the modeling of learning and algorithmic behaviors of arbitrary nesting and complexity. The primary aim of this work is not to solely to learn more about the example application being modeled for this work (i.e. The Block Sorting Task), but instead to prove that the mechanisms can be combined and function together. Ultimately, the goal of this is a long-term research agenda, which applies evolution to modeling real users’ algorithmic and strategic behaviors using quantitative models within cognitive architectures. To this end, a nonlinear general regression technique for matching human data with cognitive architectures is introduced. As well as a means of representing both individual behaviors and aggregate behaviors, as well as functions, which operate over these representations. Sequential problem-solving data generated by humans is used to construct programs in ACT-R, which approximate how that data was generated by the original humans. These programs will be constructed with Genetic Programming variants designed to evolve programs in ACT-R that solve the same task as the original human. These heuristically guide accurate matching of the observed data of a single human by maximizing their match percentage over the greatest subset of the human data. The resulting programs can be viewed as approximations of the algorithm used by the human to generate the original data. Further, the results from multiple humans can be aggregated in Program Space, and clustered to produce groups of programs which solve problems in similar ways. These clusters are defined as fuzzy clusters, and referred to as Strategy Groups, because they are groups that approximate some heuristic, which humans use when solving the original problem. Strategy Groups designed to be used to perform several key operations which the original unaggregated data cannot, including sampling, composition, reverse prediction, and verification. Together, this research agenda forms the context in which Domain-Specific Language Compilers are required for efficiency as well as being able to read and interpret the algorithms that they represent for humans they are used to model.