Separable Nonlinear Least-Squares Parameter Estimation for Complex Dynamic Systems

Figure 3

The plot presents a comparison of NLS and SLS. In panel (a) boxplots of the losses (5) and (7) (on a log scale) evaluated at the optimal parameter estimates are displayed. For high-quality prior information, the NLS and SLS loss distributions are close. As the quality of prior information degrades, NLS losses start to assume higher values compared to SLS, and their variability increases, as evidenced by the elongation of boxplots. In panel (b) the computation times are compared. The NLS computation times tend to be longer than those of SLS and increase as the quality of prior information increases. In both panels, the performance of SLS does not vary with the quality of prior information, in concordance with the experimental design.