神经网络算子——一种面向符号回归问题的遗传编程新方法  

Neural network operator—novel genetic programming approach for symbolic regression

作　　者：陈勃 党隆政 陈国宏 Chen Bo;Dang Longzheng;Chen Guohong(College of Computer&Data Science,Fuzhou University,Fuzhou 350108,China;School of Economics&Management,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学计算机与大数据学院,福州350108 [2]福州大学经济与管理学院,福州350108

出　　处：《计算机应用研究》2025年第4期1158-1166,共9页Application Research of Computers

基　　金：国家自然科学基金资助项目(70973022)。

摘　　要：针对符号回归中遗传编程方法在表达式空间中随机搜索缺乏方向性,以及种群进化过程中未利用数据特征导致效率低下的问题,提出了一种称作神经网络算子的新颖变异算子。该算子通过递归神经网络学习给定数据集特征,优化种群中的表达式,使种群向误差更低的方向进化,提升种群的进化效率。实验结果表明,结合神经网络算子的遗传编程方法在公式恢复率和种群进化速度上均优于原始方法,并在宏观经济数据集上取得了较高的决定系数。结论证明,神经网络算子能够有效引导遗传编程种群进行特征导向搜索,显著提升进化效率,具有实际应用潜力。Genetic programming methods in symbolic regression suffer from a lack of direction in random search within expression space and inefficiency due to not utilizing data features during population evolution.This paper proposed a novel mutation operator called the neural network operator to address these problems.This operator used recurrent neural networks to learn features of a given dataset,optimized expressions in the population,guided the population to evolve towards lower error,and improved evolutionary efficiency.Experimental results show that the genetic programming method combined with the neural network operator outperforms the original method in both formula recovery rate and population evolution speed,and achieves a high coefficient of determination on macroeconomic datasets.The conclusion demonstrates that the neural network operator can effectively guide the genetic programming population to perform feature-oriented search,significantly improve evolutionary efficiency,and have potential for practical applications.

关 键 词：符号回归 遗传编程 变异算子 种群进化 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]