Embedding-Based Selection Operators for Genetic Programming

This is the authors' abstract. We don't add key points for this paper.

The field of automatic programming has made significant advances in recent years, particularly through the development of semantic-aware algorithms in Genetic Programming (GP) and the emergence of Large Language Models (LLMs) designed for code. This raises a natural question: can these two approaches be combined to achieve synergistic effects? In this paper, we explore whether the semantic information embedded in vector representations of source code, generated by dedicated LLMs, can enhance GP performance. Specifically, we integrate code representation models expected to capture both syntactic and semantic attributes of GP-generated programs into their embeddings, and propose informed selection operators that leverage these embedded attributes. Experimental results on a diverse set of benchmark problems demonstrate that embedding-based selection methods significantly improve convergence speed in GP. This integration of code embeddings into the selection process offers a promising direction for enhancing the efficiency of GP evolutionary search.