.png){kind=icon}
.png){kind=icon}
Return
Accelerated overrelaxation algorithm for optimizing twin random vector functional link regression
.png){kind=icon}
PRE
AI
DOI：10.1080/02331934.2026.2637847.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper proposes a novel twin random vector functional link regression (TRVFLR) model for efficiently addressing regression tasks. TRVFLR extends the classical RVFL framework by incorporating two nonparallel regression functions. By introducing regularization terms into the primal problem, it adheres to the structural risk minimization principle, significantly enhancing generalization performance. Moreover, TRVFLR determines adaptive upper and lower insensitive bounds for each regression function by solving two independent quadratic programming problems (QPPs) under an inequality-constrained mechanism, thereby establishing a new learning paradigm distinct from RVFL and twin support vector regression (TSVR). To handle nonlinear relationships, a kernelized TRVFLR (KTRVFLR) is further developed, mapping input data into a reproducing kernel Hilbert space (RKHS) via Mercer kernel functions. This approach replaces the random enhancement layer of RVFL with a composite kernel matrix - formed by a combination of linear and nonlinear kernels - eliminating the dependency on the number of enhancement nodes, improving robustness, and consistently outperforming existing methods such as SVR, TSVR, & varepsilon;-TSVR, and LTSVR on highly nonlinear datasets. Finally, from an optimization perspective, an accelerated overrelaxation (AOR) algorithm is designed to solve the dual QPPs. Theoretical analysis and numerical experiments demonstrate that the AOR algorithm exhibits linear convergence and significantly reduces computational costs during both training and cross-validation.
Keywords:
Regression
twin random vector functional link
twin support vector regression
nonparallel planes
Journal
O
IF:
1.8
Papers:
121
Citations:
0