Local Probabilistic Sensitivity Measure By M.J.Kallen March 16 th, 2001
2 Presentation outline Definition of the LPSM Problems with calculating the LPSM Possible solution: Isaco’s method Results Conclusions
3 LPSM Definition The following local sensitivity measure was proposed by R.M. Cooke and J. van Noortwijk: For a linear model this measure agrees with the FORM method. Therefore this measure can be used to capture the local sensitivity of a non-linear model to the variables X i.
4 Problem with calculating the LPSM The derivative of the conditional expectation can only be analytically determined for a few simple models. Using a Monte Carlo simulation introduces many problems resulting in a significant error.
5 Using Monte Carlo Algorithm: 1.Save a large number of samples 2.Compute E(X|Z=z 0 ± ) and divide by 2 For good results needs to be small, but then the number of samples used in step 2 is small and a large error is introduced after dividing by 2 .
6 Alternative: Isaco’s method An alternative to calculating is proposed by Isaco Meilijson. The idea is to expand E(X|Z) around z 0 using the Taylor expansion:
7 Isaco’s method (cont.) We can then calculate the covariance:
8 Isaco’s method (cont.) The main idea in this algorithm is to now take a ‘local distribution’ Z* such that the term is equal to zero. By doing this we get
9 Choosing Z* We want to take Z* such that Z* should be as close as possible to Z, therefore we want to minimize the relative information. This results in a entropy optimization problem.
10 Relative information Definition: the relative information of Q with respect to P is given by: “The distribution with minimum information with respect to a given distribution under given constraints is the smoothest distribution with has a density similar to the given distribution.”
11 Entropy optimization (EO)
12 Solving the EO problem 1.Newton’s method 2.the MOSEK toolbox for MATLAB There are a number of ways to implement this entropy optimization problem. We have tried the following:
13 Newton’s method The implementation of Newton’s method requires a lot of work. Since you have to solve a system, a matrix has to be inverted and this introduces large errors in many cases. There are a number of reasons not to use Newton’s method for solving the EO problem:
14 MOSEK The MOSEK toolbox has a special function for entropy optimization problems, therefore the variables and constraints are easily set up. No long calculations needed, constraints can be changed in a few seconds. A much easier way of solving the EO problem is by using MOSEK created by Erling Andersen:
15 Some results Modelz0z0 Correct answer Isaco’s method (10000 samples) X,Y ~ N(0,1) Z = X+Y X,Y ~ N(0,1) Z = 2X+Y X,Y ~ U(0,1) Z = 2X+Y
16 Even worse results… Modelz0z0 Correct answer Isaco’s method (10000 samples) X,Y ~ U(0,1) Z = -ln(X)/Y
17 Attempts to fix Isaco We’ve tried many things to get better results. These attempts mostly consisted of adding and/or changing constraints. Using only the samples from a small interval around z 0. A few different approaches to this problem have been tried, but these seem to give similar results.
18 Conclusions Until now the results cannot be trusted, therefore I recommend not to use this method. We need to gain insight into what is going wrong and why it’s behaving in this way. Maybe Isaco Meilijson has an idea!