Hi(adsbygoogle = window.adsbygoogle || []).push({});

This is a long story, I make it short:

I am working in a project where I need to find a matrix defined by a third degree polynomial, the solution can be found iteratively using a gradient descent technique, I am using the golden section line search already implemented in matlab (with the code described below).

The algorithm looks powerful (it finds automatically the perfect step), but unfortunately the golden section line search does not avoid being stuck in local minima. How can I implement more efficiently this. The problem is that sometime it converges and sometimes no (sometimes is not science :-).

%Initial third degree polynomial

Cest = initial_guess;

normdev=inf ;

stepsize=inf;

%Stopping condition

stopping_condition = 10^(-5) *norm(X*X'/no_samples,'fro');

while abs(normdev*stepsize) > stopping_condition

%Third degree polynomial

dnew = Cest - 1/no_samples*(X*X' - 2/sigma^2 * (Cest*Cest'*Cest-Cest*B'*Cest));

%Find the best stepsize as a minimum using the goldensection line search

stepsize = fminbnd( @(stepsize) step(stepsize,Cest,dnew,X*X',B,sigma,no_samples),-.1,.1);

%Update

Cest = Cest + stepsize*dnew;

normdev = norm(dnew,'fro');

end

function error = step(stepsize,Cest,dnew,XX,B,sigma,no_samples)

Cest = Cest + stepsize*dnew;

error = norm(Cest - 1/no_samples*(XX - 2/sigma^2 * (Cest^3-Cest*B*Cest)),'fro');

I tried :

%Quasi-Newton

stepsize = fminunc( @(stepsize) step(stepsize,Cest,dnew,X*X',B,sigma,no_samples),dnew);

But matlab get stuck (no heap memory) probably due the fact that this function is suppose to be used when we don't have a trust region.

Any suggestions ?

**Physics Forums | Science Articles, Homework Help, Discussion**

The friendliest, high quality science and math community on the planet! Everyone who loves science is here!

# Gradient descent, anything better than golden section line search

**Physics Forums | Science Articles, Homework Help, Discussion**