This paper presents our submission to the competition on LLM-designed evolutionary algorithms for the GNBG benchmark suite. We study a simple alternative to recent automated algorithm-design frameworks: instead of placing the language model inside a specialized evolutionary loop, we treat optimizer construction as an ordinary coding task. In our setup, a general-purpose coding agent iteratively modifies and evaluates a candidate optimizer inside a small repository that provides a fixed interface, benchmarking harness, and lightweight experiment log. The main question is whether such an agent, without any bespoke search controller, can still discover competitive metaheuristics. Our results indicate that the answer is positive: the resulting optimizer performs well on a substantial subset of GNBG II. These findings suggest that recent progress in language models and tool use may already be sufficient to support practical automated metaheuristic design.
Jan Iłowski, Marcin Małek, Wojciech Achtelik et al.· Proceedings of the Genetic a...· 0 citations
Optimization algorithms are traditionally represented by their descriptions, source code, and comparative results in benchmark environments. In recent years, however, it has become more popular to extract numerical features describing the behavior of algorithms during their operation. Such metrics can distinguish between algorithms, assess their novelty, or assist in selecting the best algorithm for a given problem. Unlike existing trajectory-based approaches that use partial trajectory information to derive hand-crafted behavioral features, we propose to represent complete execution trajectories as graphs and learn their embeddings using graph neural networks. In this paper, we explore two graph construction schemes and enrich graph vertices with evaluation and topological features. We compare the proposed approach with existing behavioral features and introduce additional experiments demonstrating the utility of behavioral analysis. Through extensive experiments, we show that the learned embeddings can effectively distinguish between algorithms even when evaluated across diverse benchmark functions. We introduce quantitative measures to assess how configuration influences behavioral expression and demonstrate that partial trajectory features can predict optimization success, enabling informed early stopping decisions. The resulting embedding space provides a foundation for constructing diverse algorithm portfolios, evaluating parameter significance, and enabling future analyses of relationships between problem characteristics and algorithm performance.
Bartłomiej Walczak, Hubert Guzowski, Wojciech Achtelik et al.· Annual Conference on Genetic...· 0 citations