Abstract: To address the challenges of Neural network Architecture Search (NAS), a NAS Guided by Hybrid Heuristic Information (GHHI-NAS) algorithm was proposed. First, a heuristic information construction module integrating prior knowledge with local search feedback was designed to generate multidimensional dynamic heuristic indicators, and a hybrid update strategy was combined to guide architecture search, effectively mitigating the global exploration deficiency and local optima traps caused by unidirectional updates in traditional NAS. Second, an adaptive Covariance Matrix Evolution Strategy (CMA-ES) framework was adopted, enhanced by a hybrid fitness evaluation function to escape small-model traps during early iterations. Finally, a fitness-sharing strategy was implemented to smooth evaluation noise and improve population diversity. Additionally, a penalty-based Monte Carlo exchange sampling method was introduced to reduce performance degradation caused by random sampling. Experimental results demonstrated that GHHI-NAS achieved validation accuracies of 97.55% and 83.44% on CIFAR-10 and CIFAR-100, respectively, along with a test error rate of 24.7% on ImageNet. It also exhibited outstanding performance on NAS-Bench-201, matching or slightly surpassing state-of-the-art Evolutionary NAS (ENAS) algorithms while requiring only 0.12 GPU-Days, thereby achieving low search costs and high-performance architecture discovery.

Key words: Evolutionary Algorithm (EA), Neural network Architecture Search (NAS), hybrid heuristic information, adaptive Covariance Matrix Adaptation Evolution Strategy (CMA-ES), Monte Carlo swap sampling

摘要： 针对神经网络架构搜索（NAS）任务，提出一种混合启发信息指导NAS（GHHI-NAS）算法。首先，GHHI-NAS算法核心通过设计融合先验知识与局部搜索反馈的启发信息构造模块，生成多维动态启发指标，并配合混合更新策略指导架构搜索，有效克服传统NAS因更新方向单一导致的全局探索不足及局部最优陷阱问题。其次，算法使用自适应协方差进化策略（CMA-ES）作为更新框架，辅以混合适应度评价函数，指导算法在初期跳出小模型陷阱；最后，通过适应度共享策略平滑评价噪声并提升种群多样性。为进一步降低采样带来的性能损失，还提出了带惩罚机制的蒙特卡洛交换采样方法。实验结果表明，GHHI-NAS在CIFAR-10和CIFAR-100上取得97.55%和83.44%的验证准确率，在ImageNet数据集上取得24.7%的测试错误率，在NAS-Bench-201上也取得杰出的表现，接近甚至略优于最新的进化神经网络架构搜索（ENAS）算法，同时搜索时间仅为0.12 GPU-Days，实现了较低的搜索开销和较高水平的测试性能。

关键词: 进化算法, 神经网络架构搜索, 混合启发信息, 自适应协方差策略, 蒙特卡洛交换采样