SEIR lsqcurvefit question ！！！A poor fit！

xdata=1:30；ydata=【8730000 21556 1975 49

8730000 30453 2744 51

8730000 44132 4515 60

8730000 59990 5974 103

8730000 81947 7711 124

8730000 102427 9692 171

8730000 118478 11791 243

8730000 137594 14380 328

8730000 152700 17205 475

8730000 171329 20438 632

8730000 185555 24324 892

8730000 186354 28018 1153

8730000 186045 31161 1540

8730000 189660 34546 2050

8730000 188183 37198 2649

8730000 187518 40171 3281

8730000 187728 42638 3996

8730000 185037 44653 4740

8730000 181386 46472 5488

8730000 177984 48467 6723

8730000 169039 49970 7465

8730000 158764 51090 9419

8730000 150539 70548 10844

8730000 141552 72436 12552

8730000 135881 74185 14376

8730000 126363 75002 16157

8730000 120302 75891 18266

8730000 113564 76288 20659

8730000 106089 76936 22888

8730000 97481 77150 24734】

[Boxn Hen]

A poor fit with this program.Help me and I'm appreciated for that!

[John D'Errico]

A poor fit from a curve fitting tool (one that is known to work, has been heavily used and tested over the years, so is effectively a good tool) is almost always a result of either

1. A poor choice of starting values for the curve fit, resulting in either a poor local suboptimal solution, or divergence to a bad place in the solution space.
2. A poor choice of model, that lacks the ability to fit the desired data. Not all models can fit any set of data. Sometimes this can be simply the result of an incorrectly coded model.

My point is, these problems reflect not on the program used to do the fit, but on how that curve fitting tool is used or misused.

Which of those cases apply here is impossible to know for us, as we see only an arbitrary list of numbers and an unreadable block of code where we have no idea if the model chosen is a reasonable one or if the code was even written correctly to implement a model that may in fact be reasonable.

[Boxn Hen]

can I put the images of the code to have a look?

[Boxn Hen]

I have learned something after reading your comment.Thank you!But I still want to solve this problem,

[Alex Sha]

Hi，are you sure the values of y1 are all constant like 8730000?

[Boxn Hen]

yes，it's the population of Wuhan city.

[Bjorn Gustavsson]

Then that's most likely where you have the problem. You see, in the simplest case SEIR is respectively the Succeptible(sp?), Exposed, Infected, and Recovered (or Removed). You want to model the evolution of these 4 groups for Wuhan. Presumably under the condition that there are no influx or outflux of people, meaning that the sum S+E+I+R are the same through your simulation. You'll have to figure out how that should be reflected in your function and data.

[Boxn Hen]

I think I have solved this problem:The SEIR model's structure depends.I do have a pretty fit now although it can't make every times.Hahaa.Thanks for telling me so much.

[Bjorn Gustavsson]

The reason for poor fits at times is surely John's second point, even though SEIR is a "reasonably good" model, it is also a very simple model. In Wuhan I guess additional factors are for exampl countermeasures against the epidemic-spread (introduced at a couple of moments in time, and likewise relaxed later), the actual spreading being slightly different in reality than the average spread in the SEIR-model (for example slightly different fraction of contamination in different parts of town?) and the recovery-characteristic in SEIR-models are a simplistic linear recovery-rate that is "un-biological".

[Boxn Hen]

I think so.And I mean if there are not so many parameters changing all the time or a good choice of starting values to fit the data ,the large accuracy error exists forever or you can't get a resonable answer for the medical knowledge.John's answer is great indeed.