Scholarlink

<p>As a preferable method for nonlinear system modeling, radial basis function neural network (RBFNN) achieves a great success. However, the performance of RBFNN would be degraded when the linear relationship from inputs cannot be negligible, and the structure growing algorithm faces the problem of error jumps during training when structure grows, which directly affects the training performance. Aiming to improve the accuracy of the nonlinear system modeling, a pairwise growing RBFNN with sparsely direct connections (PGRBF-SDC) is proposed in this study. Firstly, the sparse construction mechanism of input–output direct connections is designed to construct connections between input and output neurons directly and sparsely, which effectively utilizes the non-negligible linear relationship from inputs. Secondly, a pairwise growing algorithm based on error correction (PGEC) is proposed to add hidden neurons by pair adaptively, which ensures the network stability when structure grows as proved by the theoretical analysis. Finally, the model performance is tested by a nonlinear function approximation problem, two UCI benchmark problems, and the prediction of effluent biochemical oxygen demand (BOD) in the actual wastewater treatment process (WWTP). The experimental results show that compared with the existing self-organizing RBFNNs, PGRBF-SDC has superior generalization ability and faster convergence speed without sacrificing the compactness of network structure.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Pairwise Growing RBF Neural Network with Sparsely Direct Connections for Nonlinear System Modelling

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Li Wenjing,
  • Chen Zhiqian,
  • Li Zhigang,
  • Yan Zheng

摘要

As a preferable method for nonlinear system modeling, radial basis function neural network (RBFNN) achieves a great success. However, the performance of RBFNN would be degraded when the linear relationship from inputs cannot be negligible, and the structure growing algorithm faces the problem of error jumps during training when structure grows, which directly affects the training performance. Aiming to improve the accuracy of the nonlinear system modeling, a pairwise growing RBFNN with sparsely direct connections (PGRBF-SDC) is proposed in this study. Firstly, the sparse construction mechanism of input–output direct connections is designed to construct connections between input and output neurons directly and sparsely, which effectively utilizes the non-negligible linear relationship from inputs. Secondly, a pairwise growing algorithm based on error correction (PGEC) is proposed to add hidden neurons by pair adaptively, which ensures the network stability when structure grows as proved by the theoretical analysis. Finally, the model performance is tested by a nonlinear function approximation problem, two UCI benchmark problems, and the prediction of effluent biochemical oxygen demand (BOD) in the actual wastewater treatment process (WWTP). The experimental results show that compared with the existing self-organizing RBFNNs, PGRBF-SDC has superior generalization ability and faster convergence speed without sacrificing the compactness of network structure.