Created by W.Langdon from gp-bibliography.bib Revision:1.8382
Grammar-Guided Genetic Programming (G3P) is recognised as one of the most successful approaches for grammar-obeying program synthesis that evolves programs in arbitrary languages to solve program synthesis problems based on a set of input-output examples and predefined BNF grammars. Despite its success, the restriction on the evolutionary system to only leverage input-output error rate during its assessment of the programs it derives limits its scalability to larger and more complex program synthesis problems. Alternatively, while large language models (LLMs) are showing increasing success in program synthesis from a task description, they often struggle to generate accurate code due to ambiguity in task specifications, complex programming syntax, and lack of reliability in the generated code. Furthermore, their generative nature limits their ability to fix erroneous code with iterative LLM prompting.
In this thesis, I focus on enhancing the performance of G3P for grammar-obeying program synthesis tasks. I begin by evaluating the state-of-the-art systems, including G3P and LLMs, in grammar-obeying program synthesis. Next, I explore how to map LLM-generated code into grammar-obeying forms, followed by seeding the grammar-mapped LLM-generated code into G3P initial population to improve the evolutionary process. Building on these foundations, I investigate the potential of using code similarity measures to guide the evolution of programs, leading to the development of Bi-Objective Grammar-Guided Genetic Programming (BOG3P) and Multi-Objective Grammar-Guided Genetic Programming (MOG3P), which combine input-output error rate with multiple similarity measures as objective functions. Finally, I integrate LLM generated code into the MOG3P framework in two ways: (i) using LLM-generated code as a seed for the evolutionary process through a grammar-mapping phase, and (ii) leveraging multiple similarity measures towards LLM-generated code to guide the search process throughout the evolution. Additionally, I further extend the MOG3P framework by incorporating code from multiple LLMs to enhance the diversity and effectiveness of the search process.",
Supervisors: Takfarinas Saber, Anthony Ventresque, Vivek Nallur",
Genetic Programming entries for Ning Tao