Agentic GP: A Theoretical Framework for the Development of Genetic Programming Systems via Agentic AI

Genetic Programming (GP) systems are composed of several core components, such as solution representation, parent selectionparent selectionstrategies, and variation operators. While many GP systems exist, they are often tailored to specific problem domains. In cases where the existing systems do not generalize to novel tasks, users must assemble new GP systems. This typically involves evaluating multiple alternatives for each element of the algorithm, and if no suitable option exists for a particular component, implementing a custom one from scratch. In this chapter, we propose a theoretical framework for automating this design process. We introduce Agentic GP, a framework in which multiple AI agentsAI agentsagents, specializing in different GP components, collaborate to build GP systems from scratch based on user-defined requirements. We first describe a general version of the framework applicable to any class of AI agentsAI agentsagents, and then focus on how that framework can be instantiated using Large Language Model (LLMLLM) agentsagents. We also implemented a simple prototype of the framework using an LLMLLMagentagentdesigned to generate selection methods for a given GP system. We evaluated this prototype using a tree-based GP system on a suite of symbolic regressionsymbolic regressiontasks. While it frequently generated selection methods that resulted in runtime errors, the valid methods it produced consistently performed better than random selection but worse than established benchmark methods such as lexicaselexicaseand tournament selection. This indicates the need for further enhancements in this simple design of the agentic framework.