報告書番号

ER00001

タイトル（和文）

Theory and Practice of Support Vector Machines for Non-Intrusive Monitoring Data Classification

タイトル（英文）

Theory and Practice of Support Vector Machines for Non-Intrusive Monitoring Data Classification

概要 （図表や脚注は「報告書全文」に掲載しております）

SVMについて、そのアルゴリズムの解析とパラメータ選択法の調査を通じて、SVMのパラメータ選択方法として、現状では、スパンバウンドと呼ばれる認識率評価法が計算効率、精度の面から、有効であることがわかった。また、家庭内電気機器オン・オフ推定問題への適用を通して、実問題に対して、SVMに優れた能力を発揮させるためには、カーネル関数の選択が重要であることおよび、認識精度を決定するパラメータγと、カーネル関数の自由パラメータ(RBFカーネル関数ではσ)パラメータの調整が、重要であることが明確になった。これらの結果は、手書き文字認識などのベンチマークデータにSVMを適用する際には現れなかった点である。

概要 （英文）

The estimation of the state of household electric appliances has served as the first application of Support Vector Machines in the power system research field. Thus, it is imperative for power system research field to evaluate the Support Vector Machine on this task from a practical point of view. In this report, we use the data proposed in our previous report for this purpose. We put particular emphasis on comparing different types of Support Vector Machines obtained by choosing different kernels. We report results for polynomial kernels, radial basis function kernels, and sigmoid kernels. In handwritten digit recognition research, all results for the three different kernels achieved almost same error rates. However, in the estimation of the state of household electric appliances, the results for the three differentkernels achieved different error rates. We also put particular emphasis on comparing different capacity of Support Vector Machines obtained by choosing different regularization constants and parameters of kernels. The results show that the capacity control is as important as the choice of kernel functions. This phenomenon is well-known for the Parzen density estimator. Therefore, it is of great importance to choose an appropriate value of the bandwidth parameter for a given amount of data. Similar parallels can be drawn to the solution of ill-posed problems. So, for applying Support Vector Machines to real world practical problems, the following two issues are very important to get excellent classifiers. 1. Selection of the suitable kernel function for analyzed data. 2. Control of fitted capacity to analyzed data.

報告者

キーワード