Re: Keeping global parameters in an evaluated function
- To: mathgroup at smc.vnet.net
- Subject: [mg113267] Re: Keeping global parameters in an evaluated function
- From: "Carl K. Woll" <carlw at wolfram.com>
- Date: Thu, 21 Oct 2010 07:02:39 -0400 (EDT)
On 10/20/2010 3:10 AM, Davide Mancusi wrote:
> Hello everyone,
>
> I have a function that depends on some variables and a parameter (say
> c). My goal is to find the global minimum of the function for a given
> value of the parameter, say c=0. However, the function has a lot of
> local minima for c=0, and is thus very difficult to hit the global
> one. For c=1, however, the function is much smoother, and the
> minimization algorithms work pretty well.
>
> What I'm trying to do is to change the value of the parameter
> continuously while Mathematica minimizes the function. By doing so I
> hope to trap the minimization algorithms close to the global minimum
> using the smooth version of the function, and then slowly let c go to
> 0.
>
> I have come up with the following toy model (I use NMinimize in the
> real problem, but that shouldn't be important):
> f[x_] := (x - c)^4
> c=1;
> iterations=1;
> FindMinimum[ f[x], {x, 1001}, StepMonitor :> {c = c/2; iterations+
> +} ]
> Print[{c // N,iterations}]
> If I run this I get
> {4.48416*10^-32, {x -> 1.}}
> {1.45519*10^-11,37}
> The algorithm finds the minimum at x==1 while I would like it to find
> it close to 1.45519*10^-11.
>
> I think the problem lies in the fact that f[x] is evaluated only once,
> at the beginning of the minimization algorithm, and the global value
> of c gets substituted. Thus, even if I change c later, that doesn't
> affect the function being minimized.
>
> Can anyone suggest a way to update c from within the minimization
> routine?
>
> Thanks in advance,
> Davide
>
>
You might try using calculus and homotopic continuation. For example,
suppose you're trying to minimize
f[x, c]
for a given value of c. Let x[c] be the value of x that minimizes f for
a given value of c. Then, x[c] should be a stationary point of f. That
is, let:
g[x_, c_] = D[f[x, c], x]
Then x[c] should satisfy:
g[x[c], c]==0
If we differentiate this equation with respect to c, we obtain an ODE
that x[c] should satisfy, which can be solved by NDSolve. As a concrete
example, suppose:
f[x_,c_]:=(x^4+c x^2-c^2 x)
Then the derivative with respect to x is:
In[208]:= g[x_,c_]=D[f[x,c],x]
Out[208]= -c^2+2 c x+4 x^3
The ODE that needs to be solved is:
In[209]:= eqn=D[g[x[c],c],c]==0
Out[209]= -2 c+2 x[c]+2 c (x^\[Prime])[c]+12 x[c]^2 (x^\[Prime])[c]==0
The minimizer for c=10 is:
In[210]:= min10=ArgMin[f[x,10],x]
Out[210]= Root[-25+5 #1+#1^3&,1]
Use NDSolve to find the minimizer for values of c from 1 to 10:
In[211]:= min=x/.NDSolve[{eqn,x[10]==min10},x,{c,1,10}][[1]]
Out[211]= InterpolatingFunction[{{1.,10.}},<>]
Check:
In[212]:= ArgMin[f[x, 5.], x]
min[5]
Out[212]= 1.40072
Out[213]= 1.40072
In[214]:= ArgMin[f[x,2.],x]
min[2]
Out[214]= 0.682328
Out[215]= 0.682327
So, min is an interpolating function that gives the minimizer x[c] as a
function of c. To get the minimum value, one can do:
f[min[c], c]
for a particular value of c.
Carl Woll
Wolfram Research