Jul 2026· Proceedings of the Genetic and Evolutionary Computation Conference Companion· 0 citations· 26 references
TL;DR
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.
Abstract
Neural Architecture Search (NAS) aims to automate the design of neural network architectures, reducing reliance on manual engineering. Evolutionary approaches provide flexible search mechanisms but suffer from high computational cost due to expensive fitness evaluations. In this paper, we propose a Surrogate-Assisted Linear Genetic Programming (SA-LGP) framework for automated design of Graph Neural Network (GNN) architectures for node classification. Unlike conventional NAS methods based on fixed-length or graph-structured encodings, our approach employs variable-length linear genetic programs to construct architectures composed of diverse message-passing layers and regularization operators. To reduce computational cost, we introduce a pairwise surrogate model that predicts whether an offspring architecture is likely to outperform its parent using genotype-derived features and early training signals. The surrogate acts as a filter within the evolutionary loop, enabling full evaluation only for promising candidates. A two-stage evaluation protocol is used, consisting of a cheap early-training phase and a full-training phase for selected architectures. Experiments on benchmark graph datasets show that SA-LGP discovers compact and competitive GNN architectures while significantly reducing the number of expensive evaluations, demonstrating the effectiveness of surrogate-assisted evolutionary search.
Experimental results on benchmark graph datasets show that SA-DCGP discovers compact and high-performing architectures while significantly reducing computational overhead, demonstrating the effectiveness of surrogate-assisted evolutionary search for graph-based representation learning.
Maciej Krzywda, Szymon Łukasik, Amir H. Gandomi· Proceedings of the Genetic a...· 0 citations
Designing effective graph neural networks for node classification still depends heavily on human intuition, repeated trial-and-error, and expert knowledge of architectural components. To reduce this manual burden, we introduce a self-evolving neural architecture search framework inspired by Neural Cellular Automata (NCA), in which a population of NCA-based controllers autonomously generates and improves candidate GNN architectures over successive generations. Each controller encodes layer-wise architectural decisions, including layer activation, message-passing operator type, and hidden dimensionality, thereby defining a dynamic search space of graph models. The sampled architectures are trained and evaluated on node classification tasks, while their quality is assessed through a fitness function that jointly considers validation accuracy, validation loss, architectural depth, and parameter count relative to a baseline model. In this way, the proposed framework promotes not only predictive performance but also structural efficiency. Evolution proceeds through elitist retention, tournament selection, and mutation of controller parameters, enabling progressive adaptation of the architecture distribution without hand-crafted design rules.
Maciej Krzywda, M. Wermiński, Szymon Łukasik et al.· Proceedings of the Genetic a...· 0 citations
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.
Maciej Krzywda, Szymon Łukasik, Amir H. Gandomi· Proceedings of the Genetic a...· 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.
The graph neural networks (GNNs) have been successfully applied to non-Euclidean graph data mining tasks, attracting widespread attention. At present, to achieve promising performance, many researchers use neural architecture search (NAS) on GNNs to enhance applicability and accuracy without requiring significant manual intervention or rich domain knowledge. However, most existing graph NAS (GNAS) algorithms struggle with high-dimensional data and often overlook the characteristics and distribution of different solutions within the search space, which may result in local optima and underutilization of the search potential. This article proposes Ensemble-GNAS, a novel framework that integrates evolutionary GNAS with ensemble learning to address these limitations. Within this framework, we develop two key strategies: an evolutionary GNAS method based on a niche strategy to enhance local search capability and the diversify candidate networks and an ensemble fusion strategy based on a tree-structured Parzen estimator (TPE) to optimize the weights of candidate networks in an ensemble model. The proposed approach is evaluated on: 1) node classification datasets, including citation networks for paper classification and cancer-specific gene interaction networks for driver gene identification and 2) link prediction and graph classification datasets. The experimental results indicate that the proposed approach outperforms the state-of-the-art methods in both classification accuracy and performance stability. Moreover, the ablation study verifies the contribution of the integrated niche and TPE strategies within the overall framework. The source code of Ensemble-GNAS can be obtained from https://github.com/WilfongGuo/EnsembleGNAS.
Weifeng Guo, Pengyu Wang, Ying Bi et al.· IEEE Transactions on Neural...· 0 citations
Sparse neural networks are essential for deploying deep learning models on resource-limited and latency-sensitive platforms, where efficiency must be improved without compromising predictive performance. While the Graph-Constrained Neural Multi-Objective Evolutionary Algorithm (GCNMOEA) has demonstrated advantages in sparsity promotion, runtime efficiency, and population diversity, its performance is hindered by reduced accuracy, unstable early-stage convergence, and lower hypervolume under moderate sparsity. To address these limitations, this study proposes Enhanced GCNMOEA, an improved variant that integrates accuracy-aware graph mutation, a two- phase dominance–decomposition selection mechanism, multifidelity evaluation, and adaptive diversity scheduling. These enhancements collectively strengthen structural preservation, reduce fitness noise, and improve convergence behaviour. Experimental evaluations on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 across LeNet-5, ResNet-18, and MobileNetV2 architectures demonstrate that the proposed variant achieves consistently higher accuracy, improved hypervolume, and reduced IGD variance, while maintaining the sparsity and runtime benefits of the baseline algorithm. The results confirm that Enhanced GCNMOEA deliverreal-worldsuperior Pareto-front quality and improved robustness, making it a strong candidate for real world edge and embedded neural network applications.
Collen Channer, Syed Sajjad Hussain Rizvi, Peter Ndajah et al.· Journal of Independent Studi...· 0 citations