Interval-Number Grey-Related Analysis

Overview

Higher grey relational grade r_i means closer to the ideal (reference) sequence U_0 in interval-distance. Inputs are intervals [a^-, a^+] per criterion; smaller-is-better criteria must be flagged contrary_index and are auto-converted via Eq.(6). Weights may be scalar or interval-valued; ρ controls distinguishing power (smaller ρ = sharper ranking, default 0.5). The deterministic core can be wrapped in Monte Carlo simulation (Olson & Wu §3) for trapezoidal-fuzzy inputs.

Output

grey relational grade, higher is better

Data

Interval, interval bounds complete

Weights

Needs a weight source

Size

2+ alternatives, 3-10 criteria works best

Used for

FMS / advanced-manufacturing selection under interval uncertainty, Credit-risk / data-mining feature evaluation (Olson & Wu original case), Engineering supplier ranking with interval-valued criteria, Any MCDM problem where decision-makers prefer giving min-max bounds over point estimates

How it works

1. 1

   Construct interval-number decision matrix A = ([a_{ij}^-, a_{ij}^+])_{m×n}.

   $$A=\left[\begin{array}{ccc}[a_{11}^{-},a_{11}^{+}]& \cdots & [a_{1n}^{-},a_{1n}^{+}]\\ ⋮& \ddots & ⋮\\ [a_{m1}^{-},a_{m1}^{+}]& \cdots & [a_{mn}^{-},a_{mn}^{+}]\end{array}\right]$$

   Olson & Wu 2008, p.285 Eq.(5)

2. 2

   Convert any contrary (smaller-is-better) index j into a positive index via [b_{ij}^-, b_{ij}^+] = [-a_{ij}^+, -a_{ij}^-].

   $$[b_{ij}^{-},b_{ij}^{+}]=[-a_{ij}^{+},-a_{ij}^{-}]\text{for contrary index }j$$

   Olson & Wu 2008, p.285 Eq.(6)

3. 3

   Standardize each column by its interval-vector norm ‖A_j‖ = max_i max(|a_{ij}^-|, |a_{ij}^+|); element [r_{ij}^-, r_{ij}^+] = [a_{ij}^-, a_{ij}^+] / ‖A_j‖.

   $$\Vert A_j\Vert =\underset{i}{\max }\max (|a_{ij}^{-}|,|a_{ij}^{+}|);[r_{ij}^{-},r_{ij}^{+}]=\left[\frac{a_{ij}^{-}}{\Vert A_j\Vert },\frac{a_{ij}^{+}}{\Vert A_j\Vert }\right]$$

   Olson & Wu 2008, p.284 Def.1 Eq.(1) + p.285 Eq.(7)

4. 4

   Apply interval (or scalar) weights via interval product: [c_{ij}^-, c_{ij}^+] = [c_j, d_j] · [r_{ij}^-, r_{ij}^+]. With positive intervals and r_{ij}^+ > 0, Def.4 collapses to [c_j · r_{ij}^-, d_j · r_{ij}^+]; for scalar weight w_j set c_j = d_j = w_j.

   $$[c_{ij}^{-},c_{ij}^{+}]=[c_j,d_j]\cdot [r_{ij}^{-},r_{ij}^{+}]$$

   Olson & Wu 2008, p.284 Def.4 Eq.(3) + p.285 Eq.(8)

5. 5

   Build reference (ideal) sequence U_0 by column-wise maxima of both endpoints: u_0^-(j) = max_i c_{ij}^-, u_0^+(j) = max_i c_{ij}^+.

   $$U_0=([u_0^{-}(1),u_0^{+}(1)],\dots ,[u_0^{-}(n),u_0^{+}(n)]),u_0^{-}(j)=\underset{i}{\max }{c}_{ij}^{-},u_0^{+}(j)=\underset{i}{\max }{c}_{ij}^{+}$$

   Olson & Wu 2008, p.286 Step 5

6. 6

   Compute interval distance d_{ik} = max(|u_0^-(k) - c_{ik}^-|, |u_0^+(k) - c_{ik}^+|), then grey relational coefficient ξ_i(k) with resolving coefficient ρ.

   $$d_{ik}=\max (|u_0^{-}(k)-c_{ik}^{-}|,|u_0^{+}(k)-c_{ik}^{+}|);{\xi }_i(k)=\frac{{\min }_i{\min }_k{d}_{ik}+\rho {\max }_i{\max }_k{d}_{ik}}{d_{ik}+\rho {\max }_i{\max }_k{d}_{ik}}$$

   Olson & Wu 2008, p.284 Def.2 Eq.(2) + p.286 Eq.(9)

7. 7

   Grey relational grade r_i = (1/n) Σ_k ξ_i(k); optimal plan r_i^* = max_i r_i; descending sort gives ranking.

   $$r_i=\frac{1}{n}\sum _{k=1}^n{\xi }_i(k);i^{\star }=\arg \underset{i}{\max }{r}_i$$

   Olson & Wu 2008, p.286 Eq.(10) + Step 7

Fits when / Look elsewhere when

Assumptions to verify

• Each cell is a valid interval (a^- ≤ a^+)
• Column norms ‖A_j‖ are strictly positive
• Criterion directions are correctly tagged positive_index / contrary_index
• Weights are non-negative and (for scalar form) sum to 1

Edge cases and pitfalls

• •INTERVAL-GRA kritik edge case'leri: (1) ρ resolving coefficient seçimi: Deng kanonik ρ=0.5; ρ→0 distinguishing power maksimum (en küçük farklar bile büyür) ama numerik instability artar; ρ→1 tüm coefficient'lar 1'e bastırılır, ranking discriminating gücünü kaybeder - paper ρ ∈ [0,1] standart tavsiye eder. (2) Negative interval a⁻<0 - norm hesaplamada |·| kullanıldığı için absolute değer alınır, fakat Eq.4 b⁺<0 product kuralı uygulanmalıdır; engine sign-aware interval ops kullanmazsa ranking bozulur - paper §2 Def.4 cases bunu açık eder. (3) ‖A_j‖ = 0 sütun (tüm intervaller [0,0]) → standardization tanımsız - input validation gerekir. (4) Reference U_0 coordinate-wise max - bu, gerçek bir alternatif olmak ZORUNDA DEĞİL (utopia point); rapor sırasında 'sanal optimum' uyarısı kullanıcıya verilmelidir (yanlış yorumlama önlenir). (5) Interval ortası (a⁻=a⁺) crisp dejenerasyonu → klasik GRA Deng 1989 ile tam aynı çıktı; manifest bunu garanti eder.

• Forgetting the contrary_index conversion (Eq.(6)) for cost-type criteria - ranking will silently invert.

• Confusing this method with TFN-based fuzzy GRA (Lin-Yeh 1997, Wei 2010): inputs here are intervals, not triangular fuzzy numbers; arithmetic is interval-product (Olson & Wu Def.3-4), not TFN extension principle.

• Setting ρ outside (0,1] is permitted by Eq.(9) but rarely used; ρ > 1 weakens distinguishing power; ρ → 0 makes all but the closest alternative tend to a common value.

Works with

How to cite