[R] nls, reach limit bounds

[spencerg]

      Have you plotted the data?  There are two standard sources of 
non-convergence problems like this:  First, there may not be enough 
information in your data to estimate all four parameters.  Second, if 
that's not the case, then  your starting values are not appropriate. 


      I routinely use "optim" or "nlminb" to find a sensible solution, 
because these general purpose optimizers are more likely to converge 
than "nls".  To do this, I write a function with a name like 
"SSElogistic" to compute the sum of squares of residuals for your 
model.  I like to use "optim" with "hessian=TRUE".  Then I compute 
"eigen(fit$hessian, symmetric=TRUE)", with "fit" = the object returned 
by "optim".  If the smallest eigenvalue is negative, it says that optim 
found a saddle point.  If the smallest eigenvalue is less than, e.g., 
1e-8 times the largest, it says that the smallest eigenvector is very 
poorly estimated.  Round those numbers off grossly to 1 significant 
digit, and they will likely suggest which parameter you can fix and drop 
from the model. 


      Hope this helps. 
      Spencer Graves


UyenThao Nguyen wrote:
> Hi,
>
> I am trying to fit a 4p logistic to this data, using nls function. The function didn't freely converge; however, it converged if I put a lower and an upper bound (in algorithm port). Also, the b1.A parameter always takes value of the upper bound, which is very strange. Has anyone experienced about non-convergent of nls and how to deal with this kind of problem?
>
> Thank you very much.
>
>
>
> ########################################################################3
>            y           x
> 1  0.8924619 -0.31875876
> 2  1.1814749 -0.21467016
> 3  1.6148266  0.06069784
> 4  2.2091363  0.54032947
> 5  2.7019079  1.04921802
> 6  3.0679585  1.60745502
> 9  0.9436973 -0.31875876
> 10 1.2201133 -0.21467016
> 11 1.6470043  0.06069784
> 12 2.2090048  0.54032947
> 13 2.6864857  1.04921802
> 14 3.0673523  1.60745502
>
> new.cont=nls.control(maxiter = 10000, tol = 1e-05, minFactor = 1e-08,
>             printEval = FALSE, warnOnly = FALSE)
>
>
> b0.A=.9*min(DAT$y)
> b1.A=1.1*max(DAT$y)-b0.A
> b2.A=-1*mean(DAT$x)
> b3.A=1
>
>
> b0.A
> b1.A
> b2.A
> b3.A
>
> nls.mdl.A=nls(y~b0 + b1/(1+exp(-b2-b3*x)),data=DAT,start = list(b0=b0.A, b1=b1.A, b2=b2.A, b3=b3.A), lower=-10, upper=10, algorithm="port",trace=T,control=new.cont)
>
> ##################################
>
>
>
> 	[[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
>