jspaaks / inverse-modeling-2017 Goto Github PK

currently we aggregate two sources of information into one and optimize on that, but it's better to also show how to make a pareto front

for finding just the optimum, it is kind of OK, but for the metropolis part, the incorrect sampling really affects the shape of the distribution. This can be avoided by setting F equal to 0 (disabling the first distance, dist1).

We should update the text, snippets, and pictures in the syllabus, as well as the scripts in /exercises and in /solutions

Currently, there are some differences between our DiffEvo version (which is based on Jasper's homework2 assignment), and what Storn and Price describe. For example,

• Storn and Price jump either from a random member of the population, or from the best-scoring member
• Storn and Price mutate a point with 1 difference vector, not 2
• Storn and Price apply a crossover mechanism to increase diversity in the population

this is one of the most relevant topics for students, but currently we cover it only with lectures.

manual calibration (respsurf.m)

Gauss-Newton and Levenberg-Marquardt (Run_Optimization.m)

affects exercises 6 through 12

I think they are the optima (and standard deviation) of T and S calculated based on SSR residuals (as in Burt & Barber example), but it's not mentioned anywhere that I can see.

The caption talks about the standard deviation being 0.01 [m], which is information that you need to calculate the magic numbers.

Currently it's 1-D and that makes it more difficult to see that it is just another benchmark function.

code:

manual calibration assumes d=60 (text, mancal_sluginj.m, respsurf.m)

Gauss-Newton and Levenberg-Marquardt assume d=10 (Exercise1.m)

affects exercises 6 through 12