Trying to get optimization to work with custom cost function

GrazyMos · **Joined:** Tue Sep 25, 2012 8:07 pm **Posts:** 2

I wish to replace my custom gradient descent algorithm with an optimized algorithm. The solution contains a cost function which can calculate the cost and the gradient.

Code:

public void function1_grad(double[] x, ref double func, double[] grad, object obj)
{
   NNCostFunction costFunction = (NNCostFunction)obj;
   ILArray<double> parameters = ILMath.array(x);
   ResultCostFunction result = costFunction.costFunction(parameters);
   func = result.Cost; //double
   grad = result.Gradient.ToArray(); //double[]
}

One concept I don't understand is why the gradient is not set as output. When trying to optimize the function it runs the functions one time only (while not changing the initial parameters), while the stopping criteria is set to 50 iterations.

Code:

double epsg = 0;
double epsf = 0;
double epsx = 0;
int maxits = 50;
alglib.minlbfgsstate state;
alglib.minlbfgsreport rep;

alglib.minlbfgscreate(x.Length, 3, x,  out state);
alglib.minlbfgssetcond(state, epsg, epsf, epsx, maxits);
alglib.minlbfgsoptimize(state, function1_grad, null, costFunction); //cost function only is called once
alglib.minlbfgsresults(state, out x, out rep);
System.Console.WriteLine(parameters[r(0,20),0]); 
parameters.a = ILMath.array(x);
System.Console.WriteLine(parameters[r(0, 20),0]); //results are similar while they should differ

My guess is that the gradient from the cost function is not used. The question is how can I modify my code so it does use the gradient?

Thanks

Sergey.Bochkanov · **Joined:** Fri May 07, 2010 7:06 am **Posts:** 927

GrazyMos wrote:

Code:

public void function1_grad(double[] x, ref double func, double[] grad, object obj)
{
   NNCostFunction costFunction = (NNCostFunction)obj;
   ILArray<double> parameters = ILMath.array(x);
   ResultCostFunction result = costFunction.costFunction(parameters);
   func = result.Cost; //double
   grad = result.Gradient.ToArray(); //double[]
}

One concept I don't understand is why the gradient is not set as output. When trying to optimize the function it runs the functions one time only (while not changing the initial parameters), while the stopping criteria is set to 50 iterations.
Thanks

Gradient is not set to out in order to avoid excessive reallocation of dynamic arrays. You have preallocated array grad, and you should write into it - without reallocating its memory. BTW, that's the reason why algorithm stops immediately - you function allocates new array and stores gradient into it, but optimizer is not aware of this change and still looks to the old place (filled with default zeros).

I know that C# programmers tend to think that you can easily allocate dynamic structures due to efficient garbage collector :) But optimizers should be as efficient as possible, and memory allocation is a bottleneck even in C#.

GrazyMos · **Joined:** Tue Sep 25, 2012 8:07 pm **Posts:** 2

Thanks, spot on!

Adding a simple array copy solved the problem

Code:

Array.Copy(result.Gradient.ToArray(), grad, grad.Length);

forum.alglib.net

Forum rules

Trying to get optimization to work with custom cost function

Who is online