From: Givemore on
Does anyone know how to formulate an objective function from Nueral Network simulation results to input in the GA for optimisation. I have a trained network which can simulate an ouput based on three inputs (x1, x2, x3). X1 ranges from 0.1:0.6, x2ranges from 5:15, x3 from 20:100. I want to use the trained network to define an objective function and then use GA to find the optimum parameter combination. I am stuck on defining the objective function and would appreciate any help.
From: Alan Weiss on
On 7/14/2010 1:00 AM, Givemore wrote:
> Does anyone know how to formulate an objective function from Nueral
> Network simulation results to input in the GA for optimisation. I have a
> trained network which can simulate an ouput based on three inputs (x1,
> x2, x3). X1 ranges from 0.1:0.6, x2ranges from 5:15, x3 from 20:100. I
> want to use the trained network to define an objective function and then
> use GA to find the optimum parameter combination. I am stuck on defining
> the objective function and would appreciate any help.

Take a look at how objective functions are documented:
http://www.mathworks.com/access/helpdesk/help/toolbox/gads/brdvu8r.html
The main point: put your three inputs into one vector
x = [x1 x2 x3]
Use x as the input to your fitness function.

More examples:
http://www.mathworks.com/access/helpdesk/help/toolbox/gads/exampleindex.html

Alan Weiss
MATLAB mathematical toolbox documentation
From: Greg Heath on
On Jul 14, 1:00 am, "Givemore " <give...(a)gmail.com> wrote:
> Does anyone know how to formulate an objective function from  Nueral Network simulation results to input in the GA for optimisation. I have a trained network which can simulate an ouput based on three inputs (x1, x2, x3). X1 ranges from 0.1:0.6, x2ranges from 5:15, x3 from 20:100. I want to use the trained network to define an objective function and then use GA to find the optimum parameter combination. I am stuck on defining the objective function and would appreciate any help.

On the few GA nets that I designed, the R^2 statistic
was adequate;

For training target t and naive constant
(i.e., independent of input x) output y00,
the mean-square error is obtained from

y00 = repmat(mean(t,2),1,N);
MSE00 = mse(t-y00)

For training target t, input x, and output y,
the mean-square error and corresponding R^2
statistic is obtained from

y = sim(net,x);
MSE = mse(t-y)
R2 = 1-MSE/MSE00

Hope this helps.

Greg
From: Givemore on
Greg Heath <heath(a)alumni.brown.edu> wrote in message <ded9b354-8cc4-4231-aa62-ad909f849670(a)r27g2000yqb.googlegroups.com>...
> On Jul 14, 1:00 am, "Givemore " <give...(a)gmail.com> wrote:
> > Does anyone know how to formulate an objective function from  Nueral Network simulation results to input in the GA for optimisation. I have a trained network which can simulate an ouput based on three inputs (x1, x2, x3). X1 ranges from 0.1:0.6, x2ranges from 5:15, x3 from 20:100. I want to use the trained network to define an objective function and then use GA to find the optimum parameter combination. I am stuck on defining the objective function and would appreciate any help.
>
> On the few GA nets that I designed, the R^2 statistic
> was adequate;
>
> For training target t and naive constant
> (i.e., independent of input x) output y00,
> the mean-square error is obtained from
>
> y00 = repmat(mean(t,2),1,N);
> MSE00 = mse(t-y00)
>
> For training target t, input x, and output y,
> the mean-square error and corresponding R^2
> statistic is obtained from
>
> y = sim(net,x);
> MSE = mse(t-y)
> R2 = 1-MSE/MSE00
>
> Hope this helps.
>
> Greg
Hi Greg,
Thanks for the input, I am however looking at optimising a process with the variables as listed above. The suggestion you gave will be good for optimising the NN model, I think.
Givemore
From: Greg Heath on
On Jul 17, 2:37 am, "Givemore " <give...(a)gmail.com> wrote:
> Greg Heath <he...(a)alumni.brown.edu> wrote in message <ded9b354-8cc4-4231-aa62-ad909f849...(a)r27g2000yqb.googlegroups.com>...
> > On Jul 14, 1:00 am, "Givemore " <give...(a)gmail.com> wrote:
> > > Does anyone know how to formulate an objective function from Nueral Network simulation results to input in the GA for optimisation. I have a trained network which can simulate an ouput based on three inputs (x1, x2, x3). X1 ranges from 0.1:0.6, x2ranges from 5:15, x3 from 20:100. I want to use the trained network to define an objective function and then use GA to find the optimum parameter combination. I am stuck on defining the objective function and would appreciate any help.
>
> > On the few GA nets that I designed, the R^2 statistic
> > was adequate;
>
> > For training target t and naive constant
> > (i.e., independent of input x) output y00,
> > the mean-square error is obtained from
>
> > y00 = repmat(mean(t,2),1,N);
> > MSE00 = mse(t-y00)
>
> > For training target t, input x, and output y,
> > the mean-square error and corresponding R^2
> > statistic is obtained from
>
> > y = sim(net,x);
> > MSE = mse(t-y)
> > R2 = 1-MSE/MSE00
>
> > Hope this helps.
>
> > Greg
>
> Hi Greg,
> Thanks for the input, I am however looking at
optimising a process with the variables as listed
above. The suggestion you gave will be good for
optimising the NN model, I think.
> Givemore

I am not sure I understand you correctly:

You want to add input variable subset
selection to the weight optimization?.

If so, it seems that there should be a combined
objective function that includes a penalty for
having more and larger weights.

Search the FAQ and archives of comp.ai.neural-nets
with

"weight elimination"

Hope this helps.

Greg

P.S. Make sure you standardize your inputs

help zscore