Benchmarking Global Optimization Algorithms for Core Prediction Identification
Independent thesis Advanced level (professional degree), 20 credits / 30 HE creditsStudent thesis
Mathematical modeling has evolved from being a rare event to becoming a standardapproach for investigating complex biological interactions. However, variationsand uncertainties in experimental data usually result in uncertain estimatesof the parameters of the model. It is possible to draw conclusions from the modeldespite uncertain parameters by using core predictions. A core prediction is amodel property which is valid for all parameter vectors that fit data at an acceptablecost. By validating the core prediction with additional experimentalmeasurements one can draw conclusions about the overall model despite uncertainparameter values. A prerequisite for identifying a core prediction is a global searchfor all acceptable parameter vectors. Global optimization methods are normallyconstructed to search for a single optimal parameter vector, but methods searchingfor several acceptable parameter vectors are required here.In this thesis, two metaheuristic optimization algorithms have been evaluated,namely Simulated annealing and Scatter search. In order to compare their differences,a set of functions has been implemented in Matlab. The Matlab functionsinclude a statistical framework which is used to discard poorly tuned optimizationalgorithms, five performance measures reflecting the different objectives of locatingone or several acceptable parameter vectors, and a number of test functionsmeant to reflect high-dimensional, multimodal problems. In addition to the testfunctions, a biological benchmark model is included.The statistical framework has been used to evaluate the performance of thetwo algorithms with the objective of locating one and several acceptable parametervectors. For the objective of locating one acceptable parameter vector, theresults indicate that Scatter search performed better than Simulated Annealing.The results also indicate that different search objectives require differently tunedalgorithms. Furthermore, the results show that test functions with a suitabledegree of difficulty are not a trivial task to obtain. A verification of the tuned optimizationalgorithms has been conducted on the benchmark model. The resultsare somewhat contradicting and in this specific case, it is not possible to claimthat good configurations on test functions remain good in real applications.
Place, publisher, year, edition, pages
2010. , 50 p.
Engineering and Technology
IdentifiersURN: urn:nbn:se:liu:diva-61253ISRN: LiTH-ISY-EX--10/4341--SEOAI: oai:DiVA.org:liu-61253DiVA: diva2:361267
Cedersund, GunnarLyzell, Christian