MathGroup Archive 1999

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: NonlinearRegress and numerical functions...

  • To: mathgroup at smc.vnet.net
  • Subject: [mg20197] Re: [mg20161] NonlinearRegress and numerical functions...
  • From: Daniel Lichtblau <danl at wolfram.com>
  • Date: Tue, 5 Oct 1999 04:04:26 -0400
  • References: <199910040107.VAA17261@smc.vnet.net.>
  • Sender: owner-wri-mathgroup at wolfram.com

Lars Ragnarsson wrote:
> 
> Hi
> 
> I've got a big problem making Mathematica realize that my function is
> numerical. The following problem illustrates it in a rather simplified
> way:
> 
> g[t1_] := y[t] /. DSolve[{y'[t] == -y[t] 3, y[0] == 6}, y[t], t][[1]] /.
> 
> t -> t1
> 
> data = {0, 10, 20, 100};
> 
> fit = Transpose[{data, g[data]}];
> 
> f[t1_, a_,b_] := (y[t] /.NDSolve[{y'[t] == -y[t] a, y[0] == b}, y[t],
> {t, 0, 100}][[1]]) /.t -> t1
> 
> NonlinearRegress[data, f[t, a, 6], {t}, {a, {2.9, 3.2}, 2, 4}]
> 
> It should be  rather straightforward to solve this, but NonlinearRegress
> 
> seem to evaluate the function 'f' first and the results:
> 
> NDSolve::"ndnum": "Encountered non-numerical value for a derivative at \
> 
> \!\(t\) == \!\(2.680466916142905`*^-274\)."
> 
> etc....
> 
> I've tried to use Unevaluated and Hold but nothing works!
> 
> Any help would be greatly appreciated!
> 
> Regards
> 
> Larske Ragnarsson
> 
> _________________________________________________
> Larske Ragnarsson
> Phone: +46 (0)31 7721867
> Fax:   +46 (0)31 7723622
> Chalmers University of Technology
> Solid State Electronics Laboratory
> Department of Microelectronics ED
> S-412 96 Guteborg, SWEDEN
> _________________________________________________


I'll change the notation just a bit. To start we define our function,
create a bunch of data points, and show, with NonlinearFit, that we can
recover the function.

g[a_, b_, t_] := (y[t1] /. 
	First[DSolve[{y'[t1]==-a*y[t1], y[0]==b}, y[t1], t1]]) /. t1->t
pnts = {0, 10, 20, 100};
data = {#, g[3,6,t] /. t->#}&  /@ pnts
<<Statistics`

In[5]:= NonlinearFit[data, g[a,6,t], t, {a,1}]
          6
Out[5]= -----
         3. t
        E

Now suppose instead that we have a function defined by (numerical)
solutions to a differential equation. I assume the example given was for
purposes of comparison to a known solution though of course in general
we do not have this recourse. Let me point out some issues, at least one
of which is substantial.

First, your numeric differential equation solution is not very accurate;
to obtain a result similar to the exact solution for x as large as 100
you need to fiddle with the NDSolve options (maybe this numeric error
was intentional?).

Clear[f]
f[a_?NumberQ, b_?NumberQ, t_?NumberQ] :=
	(y[t1] /. First[NDSolve[{y'[t1]==-a*y[t1], y[0]==b}, y[t1],
	{t1,0,100}, AccuracyGoal->150, MaxSteps->2000]]) /. t1->t

Let us check the relative errors at our data points.

In[9]:= Map[(f[3,6,#]-g[3,6,#])/g[3,6,#]&, pnts]
Out[9]= {0., -0.0000162022, -0.0000385361, -0.000219838}

Not bad.

Now for a more serious issue. You want to do a nonlinear function fit to
a model that is, in effect, defined by a program. While this is a
reasonable thing to want, I am not certain it can be done directly in
Mathematica (I confess to no great expertise with NonlinearFit/Regress,
so I may be mistaken). Anyway, you can instead formulate this as a
minimization problem and use FindMinimum to solve for the parameter
value. We do this explicitly below. Again, some option fiddling is
helpful in order to get a good result.

In[15]:= objfunc = Apply[Plus, Map[(f[a, 6, #[[1]]] - #[[2]])^2 &,
data]];

In[16]:= FindMinimum[objfunc, {a,2,2.1}, AccuracyGoal->50,
        MaxIterations->200, WorkingPrecision->100]
                    -53
Out[16]= {3.79385 10   , {a -> 2.99999837977173152356769137355}}

Given the disparities between f and g at the data points, this is likely
about as accurate a solution as can be obtained; to do better one would
have to get more accurate results from the model function f and perhaps
also use/request higher accuracy/precision in FindMinimum.


Daniel Lichtblau
Wolfram Research


  • Prev by Date: Re: Rebuilding polygon from CoefficientList?
  • Next by Date: Re: Calculus of variations
  • Previous by thread: NonlinearRegress and numerical functions...
  • Next by thread: Re: Re: NonlinearRegress and numerical functions...