SimLab 1.1, Software for Sensitivity and Uncertainty Analysi

The aim of this paper is to present and describe SimLab 1.1 (Simulation Laboratory for Uncertainty and Sensitivity Analysis) software designed for Monte Carlo analysis that is based on performing multiple model evaluations with probabilistically selected m

ASimLab 1.1, Software for Sensitivity and Uncertainty Analysis,

tool for sound modelling

N.Giglioli, A.Saltelli,

Joint Research Centre

European Commission

Institute for Systems Informatics and Safety

Ispra, ItalyAbstract

The aim of this paper is to present and describe SimLab 1.1 (Simulation Laboratory for Uncertaintyand Sensitivity Analysis) software designed for Monte Carlo (MC) analysis that is based onperforming multiple model evaluations with probabilistically selected model input. The results ofthese evaluations are used to determine both the uncertainty in model predictions and the inputvariables that drive this uncertainty. This methodology is essential in situations where a decision hasto be taken based on the model results; typical examples include risk and emergency managementsystems, financial analysis and many others. It is also highly recommended as part of modelvalidation, even where the models are used for diagnostic purposes, as an element of sound modelbuilding. SimLab allows an exploration of the space of possible alternative model assumptions andstructure on the prediction of the model, thereby testing both the quality of the model and therobustness of the model based inference.

1. Introduction

According to Hornberger and Spear (1981) “… most simulation models will be complex, with manyparameters, state-variables and non linear relations. Under the best circumstances, such models havemany degrees of freedom and, with judicious fiddling, can be made to produce virtually any desiredbehaviour, often with both plausible structure and parameter values. ”

The problem highlighted by Hornberger is acutely felt in the critical segment of the modellingcommunity. An economist, Edward E. Leamer, suggests the following:

"I have proposed a form of organised sensitivity analysis that I call “global sensitivity analysis” inwhich a neighbourhood of alternative assumptions is selected and the corresponding interval ofinferences is identified. Conclusions are judged to be sturdy only if the neighbourhood ofassumptions is wide enough to be credible and the corresponding interval of inferences is narrowenough to be useful."

This awareness of the dangers implicit in selecting a model structure as true and working happilythereafter leads naturally to the attempt to rigorously map alternative model structures or workinghypotheses into the space of the model predictions.

Tools like SimLab favour such an attempt. It aims to provide the modeller with a general driver toperform MC-type analysis where both uncertainty in data and parameters, and that arising fromhigher order uncertainties can be accommodated.

The natural extension of an analysis aimed at quantifying the impact of different sources ofuncertainty into the output, is the analysis of how much each source of uncertainty weights on themodel prediction. One of the possible ways to quantify the importance of the input factor withrespect to the model output is to apply global quantitative sensitivity analysis methods. Here theexpression "Global Sensitivity Analysis" takes on an additional meaning, with respect of thatproposed by Leamer, in that a decomposition of the total model uncertainty is sought.

SimLab software is a tool that helps the analyst to take into account the effect of the inputuncertainties that affect the model output. It is designed to explore the space of the inputuncertainties, via either MC or other strategies, and then it uses the results to determine bothuncertainty in model predictions and how these are apportioned to the input factors (Saltelli et al.2000a). The analysis involves five steps: the selection of ranges and distributions for each factors,

The aim of this paper is to present and describe SimLab 1.1 (Simulation Laboratory for Uncertainty and Sensitivity Analysis) software designed for Monte Carlo analysis that is based on performing multiple model evaluations with probabilistically selected m

generation of a sample from the ranges and distributions specified in the first step; evaluation of themodel for each element of the sample; uncertainty analysis and sensitivity analysis.

An important feature of the software is its flexibility with respect to the model to be analysed.SimLab strives to facilitate as much as possible the troublesome task of interfacing a given model, inmost instances conceived to run on a single set of input values at a time, with the MC driver thatallows multiple execution of the model. Processing the model output, SimLab performs thenUncertainty Analysis (UA) and Sensitivity Analysis (SA).

This paper is organised as follows. Section 2 is devoted to definitions and description of themethodologies for performing UA and SA with examples of the applications. In section 3,capabilities and functionality of the software are presented. Section 4 discusses more in detailexamples of applications, to illustrate the full potentiality of the proposed software and theunderlying methodology.

2. What is UA and SA

2.1 Definitions

There are different definitions related to SA depending on the goal of the analysis and on the point ofview of the analyst.

For an engineer, it could be crucial to test the model reliability under different assumptions oflifetime, fault tree structure and the like. For a statistician it could be interesting to test the robustnessof a statistical model with respect to distributional assumptions. In econometrics it is important totest the stability of a regression model with respect to all factors excluded from it. For a modellerinvolved …… 此处隐藏：35385字，全部文档内容请下载后查看。喜欢就下载吧 ……