A method for comparison of optimization algorithms used in compartmental modelling

Abstract

A protocol is proposed to compare the performance of five minimization algorithms used in compartmental modeling. Kinetics are computed from the model being tested, and noise added to them with different signal-to-noise ratios. These noisy curves undergo compartmental analysis in the case of good a priori and poor a priori knowledge of the parameters to be estimated. The protocol makes it possible to guide the choice of an algorithm and to test numerical identifiability of a model. It also emphasizes the unreliability of the conventional confidence intervals computation.<>