Optimization for Simulation: LAD Accelerator

Miguel Lejeune (mlejeuneandrew.cmu.edu)
Francois Margot (fmargotandrew.cmu.edu)

Abstract: In this paper, we study the design of systems subject to stochastic events. Simulation-Optimization approaches usually search for good decision variable values for a given setting of the stochastic parameters. While relevant for solving real world problems, such approaches do not necessarily address the issue that a manager of a production system faces, namely how to assess and improve the performance of the system under current conditions. Most often, the stochastic parameters vary with time. Such systems are configurable by specifying values for stochastic parameters and values for decision variables (called manipulables) and their performance can be described by a single number. The goal of this paper is to address the problem of evaluating the performance of a system running under unknown values for the stochastic parameters. We develop a new approach called LAD for Simulation in which we execute a number of simulations with very few replications and record the mean value of directly measurable quantities (called observables). We then construct a classification model that uses as inputs the values of the observables and whose numerical output is used as a predictor for the performance of the system under this particular setting of the manipulables. We then verify whether the derived LAD classification model is accurate enough to discriminate good and bad manipulable settings. We test our approach on an “assemble-to-order” system and show that our approach generates an LAD classification model that separates very accurately the manipulable settings that generate a high and low actual performance of the system. The quality of the derived models are evaluated with respect to two classification metrics, i.e., the classification quality and the cumulative accuracy profile. In order to hedge against the risk of overfitting, we then validate the models on out-of-the sample data and using a 2-folding technique. Finally, we note that our approach is not restricted to discrete-event simulation models and can handle directly manipulables with non discrete domain.

Keywords: Simulation-Optimization, Logical Analysis of Data, Stochastic Models

Category 1: Other Topics (Optimization of Simulated Systems )

Category 2: Global Optimization (Stochastic Approaches )

Category 3: Applications -- Science and Engineering (Data-Mining )

Citation:

Download: [PDF]

Entry Submitted: 07/03/2007
Entry Accepted: 07/13/2007
Entry Last Modified: 07/14/2007

Modify/Update this entry