Experimental Results of EDGE-DNNTuner: Evaluating the Effectiveness of the Symbolic Component in Guiding Neural Architecture Search

This dataset collects the experimental results obtained from the evaluation campaign of EDGE-DNNTuner, a neuro-symbolic Neural Architecture Search (NAS) framework designed to enhance the discovery of Pareto-optimal solutions. The study investigates the contribution of the symbolic reasoning component integrated within the tuner, specifically assessing its capability to effectively guide the exploration process in multi-objective optimization scenarios. The experimental campaign was conducted on three widely adopted benchmark datasets: CIFAR-10, CIFAR-100, and TinyImageNet. Across all test cases, EDGE-DNNTuner consistently demonstrated a strong ability to steer the search toward high-quality regions of the solution space. Notably, the symbolic component enabled a random search strategy to outperform Bayesian optimization, under identical conditions in terms of search space and maximum number of iterations. Results highlight that the integration of symbolic reasoning not only improves the quality of the identified Pareto fronts, but also enhances the overall efficiency of the search process, reducing the need for complex optimization strategies. These findings provide evidence of the effectiveness of hybrid neuro-symbolic approaches in advancing the state of the art in NAS. The dataset includes performance metrics, configurations, and evaluation logs necessary to reproduce and further analyze the reported results.本数据集收录了EDGE-DNNTuner的评估实验结果，该框架是一款旨在增强帕累托最优解探索效率的神经符号化神经架构搜索（Neural Architecture Search, NAS）框架。本研究探究了集成于该调优器内的符号推理组件的作用，具体评估其在多目标优化场景中有效引导搜索过程的能力。 本次实验基于三类广泛使用的基准数据集展开：CIFAR-10、CIFAR-100与TinyImageNet。在所有测试用例中，EDGE-DNNTuner始终展现出强劲的搜索导向能力，能够将搜索过程引导至解空间的高质量区域。值得注意的是，在搜索空间与最大迭代次数完全一致的条件下，该符号组件使得随机搜索策略的表现优于贝叶斯优化。 实验结果表明，符号推理的集成不仅提升了所识别帕累托前沿的质量，还增强了搜索过程的整体效率，减少了对复杂优化策略的依赖。上述发现验证了混合神经符号方法在推动神经架构搜索领域现有技术水平方面的有效性。 本数据集包含复现与进一步分析报告结果所需的性能指标、模型配置与评估日志。