Engineering Skills

製品開発エンジニアがデータ解析のノウハウを垂れ流します

決定的スクリーニング計画

実験計画法の中の、決定的スクリーニング計画について書きます。

決定的スクリーニング計画

まずスクリーニング計画とは、実験の初期段階で重要な因子を特定するための実験を少ない試行数で特定する計画です。決定的スクリーニング計画はモデル式として多項式二次など曲面性を仮定して最適化を行う実験計画法(Design of Experiment, DOE)の一種です。

決定的スクリーニング計画(Definitive screening design, DSD)はJonesとNachtsheim[1]に提案され、Xiaoら[2]によってカンファレンス行列を用いて3水準の決定的スクリーニング計画が定式化されることが示されました。この計画は取り上げたn個の因子に対して2n+1回の実験を行い、すべての主効果が交互作用および2次効果と交絡しないことが特徴です。

カンファレンス行列[math] \displaystyle C [/math]は下記のように定義され、[math] \displaystyle CC^T = (n-1) I_n [/math]となります。

1.対角成分は0

2.各要素は-1か+1

3.各列が直交する

[math] \displaystyle n [/math]次のカンファレンス行列は[math] \displaystyle n \textrm{ } mod \textrm{ } 4 \equiv 2 [/math]の場合対称(symmetric)行列になり、[math] \displaystyle n \textrm{ } mod \textrm{ } 4 \equiv 0 [/math]の場合対称にはならず交代行列になります。また[math] \displaystyle n [/math]が偶数であることはカンファレンス行列が存在する十分条件ではなく、Belevitch[3]により[math] \displaystyle n [/math]次のカンファレンス行列が存在する場合、n-1は二つの整数の二乗和であることが示されています。


カンファレンス行列の一例は下記です。

\begin{pmatrix} 0 & +1 & +1 & +1 \\ -1 & 0 & -1 & +1 \\ -1 & +1 & 0 & -1 \\ -1 & -1 & +1 & 0 \end{pmatrix}

\begin{pmatrix} 0 & +1 & +1 & +1 & +1 & +1 \\ +1 & 0 & +1 & -1 & -1 & +1 \\ +1 & +1 & 0 & +1 & -1 & -1 \\ +1 & -1 & +1 & 0 & +1 & -1 \\ +1 & -1 & -1 & +1 & 0 & +1 \\ +1 & +1 & -1 & -1 & +1 & 0 \end{pmatrix}

肝心の実験計画ですがカンファレンス行列[math] \displaystyle C [/math]が定義されるときに、決定的スクリーニング計画は下記のように表されます。

\begin{pmatrix} C \\ -C \\ 0 \end{pmatrix}

実験配置イメージ

今度は因子数=2を例に、中心複合計画(Central Composite Design, CCD)Doehlert計画(Doehlert Design, DD)、決定的スクリーニング計画(DSD)決定的スクリーニング計画(DSD)の実験配置イメージを示します。

f:id:OceanOne:20210922015427p:plain
Fig.1 2要因の実験計画例

実験回数の比較

中心複合計画(Central Composite Design, CCD)Box-Behnken計画(Box-Behnken Design, BBD)Doehlert計画(Doehlert Design, DD)、決定的スクリーニング計画(DSD)について、要因数毎の実験回数を比較します。各実験計画の実験回数は中心点の実験回数を[math] \displaystyle C_0 [/math]、要因数を[math] \displaystyle k [/math]とすると

中心複合計画:[math] \displaystyle 2^k + 2k + C_0 [/math]

Box-Benken計画:[math] \displaystyle 2k(k-1) + C_0 [/math]

Doehlert計画:[math] \displaystyle k^2+k + C_0 [/math]

決定的Screening計画:[math] \displaystyle 2k + C_0 (k=odd)[/math]

と、上記のようになります。

中心点の実験回数を[math] \displaystyle C_0 = 1 [/math]、としてテーブル化してみます。

f:id:OceanOne:20210918033402p:plain
Table.1 実験回数の比較

横軸を因子数、縦軸を実験回数でグラフ化してみます。

f:id:OceanOne:20210922014719p:plain
Fig.2 実験回数の比較

2次以上の非線形モデリングを行える実験計画のなかでは圧倒的に実験回数が少ないことが分かります。

相関マップの比較

中心複合計画(Central Composite Design, CCD)Box-Behnken計画(Box-Behnken Design, BBD)Doehlert計画(Doehlert Design, DD)、決定的スクリーニング計画(DSD)について、各実験点の相関係数をマップ化したものを比較してみます。相関マップは、効果間における相関係数の絶対値を示したものです。相関係数の値は小さいほうが望ましく、本例では青ほど良い(無相関)です。下記マップでは主効果の他、交互作用と2次の効果まで計算しています。

CCDから、BBDDDDSDに向かって交互作用や2次効果の間の相関(交絡)は増加していきますが、どの実験も主効果間と主効果と交互作用/2次効果間の交絡はないことが分かります。そして、決定的スクリーニング計画は最小の実験回数でその特性を保持しています。

f:id:OceanOne:20210922020212p:plainf:id:OceanOne:20210922020308p:plain
Fig.3-(a) 相関マップの比較(左から、CCD/BBD)
f:id:OceanOne:20210922020336p:plainf:id:OceanOne:20210922020128p:plain
Fig.3-(b) 相関マップの比較(左から、DD/DSD

まとめ

決定的スクリーニング計画について述べました。今回も、こちらのツールに実装しています。メニューバーからDOE>Make Definitive Screening Design fileで計画を作成できます。

[1] Jones, B., and C. J. Nachtsheim. 2011. A class of three levels designs for definitive screening in the presence of second order effects. J. Quality Technolo., 43, 1–15.
[2] Xiao, L., D. K. J. Lin, and F. Bai. 2012. Constructing definitive screening designs using conference matrices. J. Quality Technol., 44, 2–8.
[3] V. Belevitch, 2n-terminal networks for conference telephony, Electrical Commun. 27 (1950), 231–244.