CPE 619 Experimental DesignPART IV: Experimental Design and AnalysisIntroductionIntroduction (cont’d)Slide 5OutlineTerminologyTerminology (cont’d)Slide 9Slide 10Slide 11Common Mistakes in Experiments (cont’d)Slide 13Slide 14Simple DesignsExample of Interaction of FactorsSlide 17Full Factorial Designs2k Factorial Designs22 Factorial Design22 Factorial Design (cont’d)Slide 22Slide 23Allocation of VariationAllocation of Variation (cont’d)Slide 26General 2k Factorial DesignsGeneral 2k Factorial Designs (cont’d)Slide 29Slide 30Slide 312kr Factorial Designs22r Factorial Design Errors22r Factorial Design Errors (cont’d)22r Factorial Allocation of Variation22r Factorial Allocation of Variation ExampleConfidence Intervals for EffectsConfidence Intervals for Effects (Example)Confidence Intervals for Predicted ResponsesConfidence Intervals for Predicted Responses (cont’d)Confidence Intervals for Predicted Responses ExampleConfidence Intervals for Predicted Responses Example (cont’d)Homework #6CPE 619Experimental DesignAleksandar MilenkovićThe LaCASA LaboratoryElectrical and Computer Engineering DepartmentThe University of Alabama in Huntsvillehttp://www.ece.uah.edu/~milenkahttp://www.ece.uah.edu/~lacasa2PART IV: Experimental Design and AnalysisHow to:Design a proper set of experiments for measurement or simulationDevelop a model that best describes the data obtainedEstimate the contribution of each alternative to the performanceIsolate the measurement errorsEstimate confidence intervals for model parametersCheck if the alternatives are significantly differentCheck if the model is adequate3IntroductionGoal is to obtain maximum information with minimum number of experimentsProper analysis will help separate out the factorsStatistical techniques will help determine if differences are caused by variations from errors or notNo experiment is ever a complete failure. It can always serve as a negativeexample. – Arthur BlochThe fundamental principle of science, the definition almost, is this:the sole test of the validity of any idea is experiment.– Richard P. Feynman4Introduction (cont’d)Key assumption is non-zero costTakes time and effort to gather dataTakes time and effort to analyze and draw conclusions Minimize number of experiments runGood experimental design allows you to:Isolate effects of each input variableDetermine effects due to interactions of input variablesDetermine magnitude of experimental errorObtain maximum info with minimum effort5Introduction (cont’d)ConsiderVary one input while holding others constantSimple, but ignores possible interaction between two input variablesTest all possible combinations of input variablesCan determine interaction effects, but can be very largeEx: 5 factors with 4 levels 45 = 1024 experiments Repeating to get variation in measurement error 1024x3 = 3072There are, of course, in-between choices…Chapter 196OutlineIntroductionTerminologyGeneral MistakesSimple DesignsFull Factorial Designs2k Factorial Designs2kr Factorial Designs7TerminologyConsider an example: Personal workstation designCPU choice: 6800, z80, 8086Memory size: 512 KB, 2 MB, 8 MBDisk drives: 1-4Workload: secretarial, managerial, scientificUser’s education: high school, college, graduateResponse variable – the outcome or the measured performanceE.g.: throughput in tasks/min or response time for a task in seconds8Terminology (cont’d)Factors – each variable that affects responseE.g., CPU, memory, disks, workload, user’s ed.Also called predictor variables or predictorsLevels – the different values factors can takeE.g., CPU 3, memory 3, disks 4, workload 3, user education 3Also called treatmentPrimary factors – those of most important interestE.g., maybe CPU, memory size, # of disks9Terminology (cont’d)Secondary factors – of less importanceE.g., maybe user type not as importantReplication – repetition of all or some experimentsE.g., if run three times, then three replicationsDesign – specification of the replication, factors, levelsE.g., specify all factors, at above levels with 5 replications so 3x3x4x3x3 = 324 time 5 replications yields 1215 total10Terminology (cont’d)Interaction – two factors A and B interact if one shows dependence upon anotherE.g.: non-interacting, since A always increases by 2 A1 A2 B13 6 B25 10E.g.: interacting factors since A change depends upon BA1 A2 B13 6 B25 15A1A2B1B2A1A2B1B211OutlineIntroductionTerminologyGeneral MistakesSimple DesignsFull Factorial Designs2k Factorial Designs2kr Factorial Designs12Common Mistakes in Experiments (cont’d)Variation due to experimental error is ignoredMeasured values have randomness due to measurement error. Do not assign (or assume) all variation is due to factorsImportant parameters not controlledAll parameters (factors) should be listed and accounted for, even if not all are variedEffects of different factors not isolatedMay vary several factors simultaneously and then not be able to attribute change to any one Use of simple designs (next topic) may help but have their own problems13Common Mistakes in Experiments (cont’d)Interactions are ignoredOften effect of one factor depend upon another. E.g.: effects of cache may depend upon size of program. Need to move beyond one-factor-at-a-time designsToo many experiments are conductedRather than running all factors, all levels, at all combinations, break into stepsFirst step, few factors and few levelsDetermine which factors are significantTwo levels per factor (details later)More levels added at later design, as appropriate14OutlineIntroductionTerminologyGeneral MistakesSimple DesignsFull Factorial Designs2k Factorial Designs2kr Factorial Designs15Simple DesignsStart with typical configurationVary one factor at a timeEx: typical may be PC with z80, 2 MB RAM, 2 disks, managerial workload by college studentVary CPU, keeping everything else constant, and compareVary disk drives, keeping everything else constant, and compareGiven k factors, with ith having ni levelsTotal = 1 + (ni-1) for i = 1 to kExample: in workstation study1 + (3-1) + (3-1) + (4-1) + (3-1) + (3-1) + (3-1) = 14But may ignore interaction(Example next)16Example of
View Full Document