Presentation on theme: "Introduction to parameter optimization"— Presentation transcript:
1 Introduction to parameter optimization Soils & Environmental Exposure AssessmentIntroduction to parameter optimizationSabine Beulke, Central Science Laboratory, York, UKKinetic Evaluation according to Recommendations by the FOCUS Work Group on Degradation KineticsWashington, January 2006PSD, 6-7 March 2003
5 Automatic optimization Stops when:Convergence criteria are metComparison between RSS for actual and previous runs. Convergence reached if difference is smaller than user-specified differenceTermination criteria are metFor example, when maximum number of runs has been carried out (user-specified)Good fit not guaranteed!
7 Non-uniqueness Parameter correlation Inadequate model Parameters strongly relatedEffects on RSS of changes in one parameter can be compensated by changes in another parameterInadequate modelFor example, selection of bi-phasic model not warranted if data follow SFO
8 Global versus local minimum RSS as a function ofchanges in 2 parametersFrom:The optimisation may find a local “valley” in the RSS surface, but not the absolute, global minimum.Different parameter combinations may be returned for different starting values.Good fit not guaranteed!
9 FOCUS recommendations Always evaluate the visual fitAvoid over-parameterisationAim at finding reasonable starting valuesAlways use different starting valuesConstrain parameter ranges if appropriatePlausibility checks for parameters and endpointsStepwise fitting where necessaryBe aware of differences between software packages
11 Goodness of fit - statistical criteria 2 testwhereC = calculated valueO = observed value= mean of all observed valueserr = measurement error percentageIf calculated 2 > tabulated 2 then the model is not appropriate at the chosen level of significanceError percentage unknown Calculate error level at which 2 test is passed
12 Goodness of fit - statistical criteria Confidence in parameter estimatesCalculate e.g. from ModelMaker outputA parameter is significantly different from zero if p (t) < alphaOthers (e.g. model efficiency, F-test)