1 / 7

Kriging Metamodels for Simulation Optimization

Kriging Metamodels for Simulation Optimization. Jack P.C. Kleijnen Department of Information Management / Center for Economic Research (CentER) Tilburg School of Economics & Management (TiSEM) Tilburg University, Tilburg, Netherlands

Download Presentation

Kriging Metamodels for Simulation Optimization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Kriging Metamodels for Simulation Optimization Jack P.C. Kleijnen Department of Information Management /Center for Economic Research (CentER) Tilburg School of Economics & Management (TiSEM)Tilburg University, Tilburg, Netherlands NSF Simulation Optimization WorkshopUniversity of Maryland, College Park, May 24-25, 2010

  2. Part 1: Expected Improvement (EI)/ Efficient Global Optimization (EGO) Classic reference: Jones, Schonlau, Welch (‘98)Deterministic simulation (CAE), Kriging Problem: Plug-in estimates for predictor variance Solution: Parametric bootstrappingOrigin: Den Hertog, Kleijnen, Siem (2006) Empirical results: four test functions (Hartmann)# dimensions: 1, 2, 3, 6Bootstrap: faster (first 3 functions) or tie Future research: constrained & random outputs

  3. EI / EGO: details Local vs. global optimaExploration vs. exploitation; see Fu (2007) Expensive simulation: Kriging metamodel Deterministic vs. random simulation; see Frazier, Powell, Dayanik (2009), Ankenman, Nelson, Staum (2010) Predictor variance depends on unknown correlations θ(j) → estimate θ(j) σ²(x) = 0 if x = x(i) (i =1, …, n) (“old” data)

  4. EI / EGO: more details EI / EGO algorithm for unconstrained minimization: • Find wo = min[w(i)] (i = 1, …, n) • EI(x) = E[wo - y(x) | y(x) < wo] with y(x) ~ N(y^, s2) • Find maximizer xo of EI(x) in candidate set • Simulate xo; refit Kriging; return to 1 until EI ≈ 0 Sub 2: Bootstrap estimator s2*: • Original I/O (X, w) gives original θ^ and μ^ • Sample (w*(1), …, w*(n + 1))’ from N(μ^,θ^) • Squared Error SE = [w*(n + 1) - y*{x(n + 1)}]² • Repeat 2 & 3, B times: s²* = ΣbSE(b) / B

  5. Part 2: Robust optimization Taguchi’s worldview:Decision inputs (e.g., order quantity)Environmental inputs (e.g., demand rate) Example: EOQ Classic optimization: known demand rate a Robust optimization: a ~ N(μ, σ) Goal: minimize expected cost E(C)Constraint: keep standard dev. σ(C) below T

  6. 4 x 10 8.92 8700 Regression Regression 8.9 1-L Kriging 1-L Kriging 8600 2-L Kriging 2-L Kriging 8.88 8500 8.86 8400 8.84 C C s 8.82 8300 8.8 8200 8.78 8100 8.76 8.74 8000 1.5 2 2.5 3 3.5 4 4.5 1.5 2 2.5 3 3.5 4 4.5 Q Q 4 4 x 10 x 10 Methodology for robust opt. 1. Design: CrossSpace-filling design for decision inputs dLHS for environmental inputs e with F(e) 2. Metamodel:a. Regress: y = β0 + β’d + d’Bd + γ’e + d’Δe + ε b. Two Kriging metamodels, for μ resp. σ

  7. 4 x 10 8.92 Regression 8.9 1-L Kriging 2-L Kriging 8.88 8.86 8.84 C 8.82 8.8 8.78 8.76 8.74 8150 8200 8250 8300 8350 8400 s C Methodology continued 3. Min μ s.t. σ ≤ T : Mathematical Programming 4. Vary T: Pareto frontier 5. Quantify variability: Bootstrap metamodel Future research: Random (s, S) & constraints

More Related