A probabilistic L-system-inspired encoding for architecture search, where each individual is a small set of production rules and repeated rewriting generates a symbol sequence that is translated into a GNN.
Abstract
Designing graph neural networks (GNNs) for node classification still involves many manual choices. We study a probabilistic L-system-inspired encoding for architecture search, where each individual is a small set of production rules and repeated rewriting generates a symbol sequence that is translated into a GNN. Instead of evolving a fixed layer sequence directly, evolution optimizes the rule contents, their probabilities, the rewriting depth, and a few global architectural parameters. The derived sequences may contain graph message-passing operators, normalization, activation, dropout, and structural components such as Jumping Knowledge. We evaluate the method on nine node-classification benchmarks under multiple search settings that vary selection, crossover, and mutation. The results show that a compact rule-based representation can discover competitive architectures across different graph types while keeping the final model easy to inspect at the rule level.
A self-evolving neural architecture search framework inspired by NCA, in which a population of NCA-based controllers autonomously generates and improves candidate GNN architectures over successive generations, enabling progressive adaptation of the architecture distribution without hand-crafted design rules.
Maciej Krzywda, M. Wermiński, Szymon Łukasik et al.· GECCO Companion· 0 citations
A pairwise surrogate model is introduced that predicts whether an offspring architecture is likely to outperform its parent using genotype-derived features and early training signals, enabling full evaluation only for promising candidates within the evolutionary loop.
Maciej Krzywda, Szymon Łukasik, Amir H. Gandomi· GECCO Companion· 0 citations
Discovering optimal graph neural network (GNN) architectures for various tasks is both labor-intensive and time-consuming. To reduce human effort, graph neural architecture search (GNAS) has recently been utilized to automatically identify effective GNN architectures for specific tasks, achieving competitive or even superior performance compared to manually designed architectures. However, existing GNAS methods fail to identify optimal architectures in the presence of structural and semantic noise, where structural noise refers to missing or redundant edges within the graph structure, and semantic noise denotes inaccurate node representations derived from ambiguous node features such as textual vagueness or semantic ambiguity. In this paper, we address this problem for the first time via theoretical analyses and empirical evaluations. We discover that existing differentiable GNAS methods typically select architectures based on task-relevant information hidden in the graph, being highly sensitive to structural and semantic noise, which results in suboptimal selection of GNN architectures under noise. To handle the structural and semantic noise, we propose Curriculum-GraphLLM, a novel graphLLM framework for joint optimization of architectures, structures, and texts for denoised graph neural architecture search. The core idea is to jointly optimize GNN architectures, graph structures, and textual semantics as a unified denoising process during architecture search. Specifically, we first develop a dynamic topology updating mechanism to adaptively adjust the graph structure. Then, we jointly optimize the GNN architecture and graph structure through a curriculum-based iterative updating approach. To further deal with semantic noise on text-attributed graphs (TAGs), we introduce LLMs as an auxiliary text modeling module to refine textual semantics and guide the co-optimization of text representations, graph structures, and GNN architectures. We conduct extensive experiments to show that our proposed Curriculum-GraphLLM achieves consistently competitive or superior performance compared with existing baselines, especially under structural and semantic noise.
Xin Wang, Haibo Chen, Linxin Xiao et al.· IEEE Transactions on Pattern...· 0 citations
Evolutionary program search guided by Large Language Models (LLMs) has emerged as a powerful paradigm for automated scientific discovery. However, current approaches are fundamentally constrained by three bottlenecks: structurally blind parent selection, sparse whole-program evaluation rewards, and static mutation operators that fail to adapt during search. We present GAE (Graph-Augmented Evolution), a framework that resolves these limitations through a tightly coupled, three-pillar architecture. First, a relational graph neural network (GNN) parses programs into typed computation graphs, producing structure-aware embeddings. Second, an RL-optimized meta-controller leverages these embeddings to replace blind evolutionary sampling with a directed policy, dynamically selecting optimal parents and mutation directions based on reward history. Third, an online GRPO fine-tuning loop continuously updates the LLM mutation operator at test-time using group-normalized evaluation rewards, directly aligning the model's generation distribution with high-fitness structural edits. We evaluate GAE on a challenging scientific discovery task: symbolic regression for complex nonlinear oscillator systems. By transforming stochastic search into a directed, self-improving trajectory, GAE efficiently discovers closed-form physical equations, consistently matching or outperforming static LLM-driven baselines and achieving state-of-the-art out-of-distribution performance.
GraphIR, an architecture-aware intermediate representation that supplements executable programs with a mutation-aligned candidate state, achieves the best overall search performance while maintaining comparable model size and favorable end-to-end NAS efficiency when integrated into OpenEvolve.
Zhen Liu, Wanqi Zhou, Shuanghao Bai et al.· 1 citation
Ensemble-GNAS is proposed, a novel framework that integrates evolutionary GNAS with ensemble learning to address limitations of existing graph NAS algorithms and outperforms the state-of theart methods in both classification accuracy and performance stability.
Weifeng Guo, Pengyu Wang, Ying Bi et al.· IEEE Transactions on Neural...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
