IEMAS-aided exploration of sensitivity analysis methods - PL-Grid
Transcription
IEMAS-aided exploration of sensitivity analysis methods - PL-Grid
IEMAS-aided exploration of sensitivity analysis methods implemented in MATLAB and R Włodzimierz Funika1,2, Paulina Żak1,2, Grzegorz Łaganowski1,2 University of Science and Technology, al. Mickiewicza 30, 30-059 Kraków, 2ACK Cyfronet AGH, ul. Nawojki 11, 30-950 Kraków 1AGH CGW’15 26-28.10.2015 Kraków Agenda • Sensitivity Analysis: what it is about • IEMAS • Scalarm platform • Ready-made tools in R and MATLAB • Experiments and results • Conclusions and future work 2 Sensitivity Analysis 3 1. Purpose 2. Core methodology IEMAS as the test model Immunological Evolutionary Multi-agent System for given optimization problem "We know that they work, but we do not know why” T. Back, U. Hammel and H.-P. Schwefel https://www.age.agh.edu.pl/ 4 Scalarm as data-farming platform 5 Visualisation methods in Scalarm 6 Search for tools Main criteria: 1. Freeware 2. Ease of use 3. Computation results coming from external platform 4. Built-in visualization methods 7 Ready-made tools in R and MATLAB 1. “Sensitivity” package in R 2. SAFE Toolbox in MATLAB (SAFE Toolbox due to courtesy of) F. Pianosi, F. Sarrazin, Th. Wagener http://bristol.ac.uk/cabot/resources/safe-toolbox/ 8 Research of methods Required features: Main: Method should allow for initial grading of impact of parameters. Secondary: ● ● ● ● independent from monotonicity of model accepts discrete values capability of computing variety of sample sizes computionally non intensive 9 Short review of methods ● Morris method (EET) ○ Global screening method, ○ One-step-at-a-time method (OAT) ● Sobol method (VBSA) ○ Variance based method, ○ Main effect (first-order),Total effect (total order)’ ● FAST ○ Fourier series to represent the model in the frequency domain, poor for discrete inputs (Saltelli et al., 2000) 10 Experiment specifications and parametric space 10 Inputs: ● reproduction minimum = [ 300 - 1000 ] ● newborn energy = [ 0 - 1000 ] ● transferred energy = [ 0 - 1000 ] ● amount of iterations = [ 0 - 10 ] ● immunological time span = [ 1 - 1000 ] ● bite transfer = [1 - 200] ● mahalanobis = [0.8 - 5] ● immunological maturity = [1 - 200] ● good agent energy = [1 - 1000] ● evaluation method = “rastrigin” | “schwefel” 3 outputs: ● time elapsed ● iemas fitness ● fitness calls 11 Package “sensitivity” in R in use 1/4 Morris OAT design 1000 samples µ - average direction of dependency of output and feature µ* - strength of dependency σ - the degree of nonlinearity of dependency of output and feature 12 Package “sensitivity” in R in use 2/4 Sobol Normal 1380 samples main indices - influence of only one feature while others being fixed total indices - summary of main and interaction indices interaction indices - take into account interactions of feature with others. 13 Package “sensitivity” in R in use 3/4 Sobol Martinez 330 samples main indices - influence of only one feature while others being fixed total indices - summary of main and interaction indices interaction indices - take into account interactions of feature with others. 14 Package “sensitivity” in R in use 4/4 Sobol Mara 330 samples main indices - influence of only one feature while others being fixed total indices - summary of main and interaction indices interaction indices - take into account interactions of feature with others. 15 SAFE Toolbox in MATLAB in use 1/5 Morris method (EET) Sample size: 550 16 SAFE Toolbox in MATLAB in use 2/5 Morris method (EET) Sample size: 836 17 SAFE Toolbox in MATLAB in use 3/5 Sobol method (VBSA) Sample size: 1200 18 SAFE Toolbox in MATLAB in use 4/5 Sobol method (VBSA) Sample size: 1200 19 SAFE Toolbox in MATLAB in use 5/5 FAST Sample size: 2641 20 Comparison of implementations Method impl./factor Simplicity 21 Visualization Consistency Depth Sampling size Morris - R 3 4 4 2 large Morris - SAFE 4 5 4 3 medium Sobol - R 2 2 3 5 medium Sobol - SAFE 5 N/A 3 4 large FAST 2 N/A 2 3 large Visualization - Are charts are clear and easy to interpret? Consistency - Are results the same in multiple simulations? Simplicity - How much effort does it take to adapt external model and set fitting attributes? Depth - How much information is provided with results about interactions between parameters? Sample size - What size of experiment did we deal with? Conclusions 22 Our Choice: SAFE implementation of Morris method Reasons: estimates first order effects easy to use efficient can measure the level of nonlinearity of dependency clear visualization Future work 1/2 Use of “sensitivity” package in Scalarm with visualization in HighCharts 23 Future work 2/2 Morris visualization 24 25 Thank you for attention! Questions 26