学術雑誌論文

学術雑誌論文

Knowledge-Based Systems. 2024, 300, p. 112197

Multiclass piecewise-linear SVMs with the restricted feasible region for the reinforcement degrees in constructing large-scale structures

Elsevier B.V.

英語

2024-09-27

https://doi.org/10.1016/j.knosys.2024.112197

Tatsumi K., Masui R., Miyahara H., et al. Multiclass piecewise-linear SVMs with the restricted feasible region for the reinforcement degrees in constructing large-scale structures. Knowledge-Based Systems 300, 112197 (2024); https://doi.org/10.1016/j.knosys.2024.112197.

In this paper, we focus on automatical determination of the reinforcement degrees at construction sites of a kind of large-scale structure to address labor shortages. In the previous work, the reinforcement degree at each section of the structure can be appropriately determined by the risk scores and natural conditions observed at the current and previous sections in the simplest and monotonic form, where a multiclass linear SVM was trained by using the empirical knowledge accumulated by experts. In our work, first, we point out that the discriminant function used in the previous work can be also regarded as a piecewise-linear one, as it uses different weights for each natural condition, and we propose new multiclass piecewise-linear SVMs in which the weights are trained according to the natural conditions. Then, by evaluating the generalization abilities of the proposed SVMs, we verify that the proposed SVMs may be too complex to the prediction problem. Secondly, we further develop the SVMs by restricting their feasible regions based on the results to adjust model complexity of them appropriately for the prediction problem. Through numerical experiments, we show the advantage of the proposed SVMs by comparing them with the existing methods.

This article is licensed under a Creative Commons Attribution 4.0 International License.

98161

Multiclass piecewise-linear SVMs with the restricted feasible region for the reinforcement degrees in constructing large-scale structures

Elsevier B.V.

英語

2024-09-27

Knowledge-Based Systems

300

112197

https://doi.org/10.1016/j.knosys.2024.112197

https://hdl.handle.net/11094/98161