随机权神经网络增量构造学习方法研究进展  

Recent advances in incremental learning methods for random weight neural network

作　　者：代伟[1,2] 南静 DAI Wei;NAN Jing(Artificial Intelligence Research Institute,China University of Mining and Technology,Xuzhou 221116,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]中国矿业大学人工智能研究院,江苏徐州221116 [2]中国矿业大学信息与控制工程学院,江苏徐州221116

出　　处：《控制与决策》2023年第8期2231-2242,共12页Control and Decision

基　　金：国家自然科学基金面上项目(61973306);江苏省自然科学基金优秀青年基金项目(BK20200086).

摘　　要：随机权神经网络(random weight neural network,RWNN)在解决数据定性和定量分析方面具有强大的潜力,其最显著的特征是隐含层参数随机生成.这一特征使得RWNN相比于基于梯度下降优化微调节点参数的神经网络具有诸多优势,如结构简单、易于实现和低人工干预等.RWNN的隐含层和输入层之间的参数是在一个固定区间内随机生成,而隐含层和输出层之间的输出权值则通过解析法进行求解.增量构造方法从一个小的初始网络开始,逐渐添加新的隐含层节点以提升模型品质,直到满足预期性能目标.基于此,重点从基础理论、增量构造学习方法和未来开放研究方向等方面切入,全面综述增量RWNN的研究进展.首先介绍RWNN的基本结构、理论和分析;进一步重点介绍RWNN在增量构造学习方法上的各种改进及应用;最后指出RWNN增量构造学习未来开放的研究方向.Random weight neural network(RWNN)has strong potential for solving qualitative and quantitative data analysis problems,and its most prominent feature is the random generation of parameters in the hidden layer.This feature makes RWNN has many advantages over neural networks based on gradient descent optimization for fine-tuning node parameters,such as simple structure,easy implementation,and low human intervention.The parameters between the hidden layer and the input layer of RWNN are randomly generated from a fixed interval,while the output weights between the hidden layer and the output layer are solved using an analytical method.The incremental construction method starts from a small initial network and gradually adds new nodes to the hidden layer to improve the quality of the model until the expected performance goal is met.This paper provides a comprehensive review of the research progress of incremental RWNN by focusing on basic theory,incremental construction learning method,and future open research directions.First,the basic structure,theory and analysis of RWNN are introduced,and the improvements and applications of RWNN in the incremental construction learning methods are further highlighted.Finally,future open research questions and promising directions of RWNN are pointed out.

关 键 词：随机权神经网络 增量构造学习方法 前馈神经网络 随机方法 数据分析 无限逼近性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]