Robust optimization of simulation models - Tilburg University仿真模型的鲁棒优化的蒂尔堡大学.pptVIP

Robust Optimization ofSimulation models through Metamodels Jack Kleijnen 1, Gabriella Dellino 2, Carlo Meloni 3 1 Tilburg University, 2 University of Siena, 3 Polytechnic of Bari Seminar presented at theDepartment of Statistics, Stanford, May 28, 2010 */17 Problem type Central poblem:Design ‘optimal’ product or production system Solution: Simulation model of product or system Problem: Simulation requires much CPU time Solution: ‘Metamodel’ of simulation modelType 1: Polynomial regressionType 2: Kriging (Gaussian process)Note: Many more types (CART, MARS, NN, etc.) Problem: Decision versus environmental inputsEnvironment: Uncontrollable */17 */17 Overview of seminar Methodology for simulation optimization with uncertain inputs(Static, deterministic models: Ben-Tal, Bertsimas) Methodology integrates Taguchi’s view: Decision environmental inputs Design of experiments metamodeling: RSM or Kriging Mathematical Programming: Pareto frontier: Change T Confidence interval: Bootstrap Example: Economic Order Quantity (EOQ) Future research */17 */17 Taguchi’s worldview Decision factors: d(j) (with j = 1, …,k)Example: Q in EOQ model; bumper in car design Environmental factors: e(g) (with g = 1, …, c)Example: demand rate; driver Track record in production engineering (Toyota) Critique: statisticians; see panel report Nair (1992) Real-life versus simulation experiments:Number of inputs, values, combinations: LHS Taguchi’s scalar loss function versus Pareto frontier */17 DOE metamodel type I: RSM Box Wilson (‘51): Classic optimization of real-life systemsTrack record: Myers, Montgomery, Anderson-Cook (2009) Myers et al.: Taguchian robust optimization of real-life systems Myers et al.’s Taguchian metamodel: y = b0 + b’d +d’Bd + γ’e + d’Δe + ε (1)Optimize d; interaction between d e Implications of (1):E(y) = b0 + b’d +d’Bd + γ’E(e) + d’ΔE(e) (2)var(y) = (γ’ + d’Δ)Ω(e)(γ + Δ’d) + var(ε) (3) Notes:  (γ + Δ’d): gradient of y relative to e; see (1)