| dc.description.abstract | When modeling biological systems with a bottom-up approach, the system parameters need to be calibrated for the model behavior to match that of the physical system. The problem is typically presented in the form of an optimization problem, where the deviation between model outputs and experimental data is minimized as a function of the parameters. There are multiple methods available, but for some complicated models many either fail or take too long to find the correct solution.
In this thesis, the covariance matrix adaptation evolution strategy (CMA-ES), a stochastic global optimization method, is implemented for the purpose of parameter estimation on difficult problems. The CMA-ES is parallelized using message passing interface communication to reduce the computational time.
The method was tested on two models: An ODE model with known parameters and a hybrid DAE model with unknown parameters. The CMA-ES managed to find the correct parameters for the ODE model, albeit after a longer time than that needed by alternative methods included in Matlab (trust-region-reflective and Levenberg-Marquardt). When it came to the hybrid DAE model, the CMA-ES minimized the model error, while the Matlab methods that were tested were not applicable. However, the minimum found did not produce the desired model behavior. The most likely reasons for this are that the model was incomplete, the search space was badly defined or the objective function contained too much noise.
The results indicate that the parallel CMA-ES method is too slow for simpler model calibration problems, but that it may be a powerful tool for parameter estimation problems with noisy objective functions and long runtimes. | |
