A Framework for Integrating Large Language Models into Memetic Algorithms

Area of Science:

• Artificial Intelligence
• Computational Optimization
• Machine Learning

Background:

• Memetic algorithms combine global search with local refinement for optimization.
• Current local search operators are often fixed, problem-specific, and handcrafted.
• Adaptive generation of local search operators is crucial for enhancing memetic algorithm performance.

Purpose of the Study:

• To propose a fourth-generation memetic framework integrating Large Language Models (LLMs) for adaptive local search operator generation.
• To develop a system where LLMs generate Python functions as memes within the optimization loop.
• To validate the effectiveness of LLM-driven adaptive operators in continuous black-box optimization.

Main Methods:

• Embedding LLMs into the optimization loop to generate local search operators (memes).
• Utilizing a structured state vector to provide search dynamics context to the LLM.
• Implementing a two-stage validation for generated operators (semantic similarity and implementation comparison).
• Employing a cooldown-regulated triggering mechanism and a probability-weighted selection policy for meme management.

Main Results:

• The LLM-driven memetic framework consistently outperformed non-LLM memetic baselines on the CEC 2017 benchmark suite.
• Generated operators were validated for novelty and syntactic correctness before admission to the meme library.
• The adaptive operator generation balanced periodic and stagnation-based triggering effectively.

Conclusions:

• Large Language Models can serve as effective adaptive heuristic components in population-based optimization.
• The proposed LLM-driven framework demonstrates superior performance compared to classical memetic algorithms.
• This research validates the viability of generative language models for enhancing optimization strategies.