Lecture
Joint Lectures on Evolutionary Algorithms - December 2024
- Date
- Wednesday 11 December 2024
- Time
- Location
- Online
LLaMEA: LLM-based EA for automatically generating metaheuristics
In this talk, we introduce the Large Language Model Evolutionary Algorithm (LLaMEA), a novel framework that harnesses the generative power of Large Language Models (LLMs) to automate the creation of metaheuristic optimization algorithms. Traditional approaches to designing metaheuristics require significant domain expertise and are limited by human creativity and modular frameworks. LLaMEA disrupts this paradigm by embedding LLMs into an evolutionary loop, enabling the iterative generation, mutation, and refinement of algorithms based on runtime feedback and performance metrics. We will discuss how LLaMEA leverages GPT models to synthesize diverse algorithmic structures, optimize their hyperparameters, and adapt their logic for various optimization tasks. Using the Black-Box Optimization Benchmark (BBOB) as our evaluation suite, we demonstrate LLaMEA’s ability to produce algorithms that rival or surpass state-of-the-art methods such as Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and Differential Evolution (DE) in five-dimensional optimization problems. Key findings, including insights into mutation strategies, diversity of generated algorithms, and generalization challenges for higher-dimensional problems, will be highlighted. The presentation will also cover the computational implications of employing LLMs in algorithm design and explore future directions, such as population-based approaches, multi-LLM diversity mechanisms, and applications in combinatorial optimization and AutoML. This talk positions LLaMEA as a significant step towards automated, modular, and scalable algorithm design, offering a glimpse into the future of optimization powered by artificial intelligence.
This series is organized by a team from four universities, initiated by prof. dr. Thomas Bäck (Leiden University), prof. dr. Peter A.N. Bosman (CWI), prof. dr. Gusz Eiben (Vrije Universiteit Amsterdam), and dr. ir. Dirk Thierens (Utrecht University).
