For Programmers: Free Programming Magazines  


Home > Archive > Matlab > March 2006 > Neural Network









You are viewing an archived Text-only version of the thread. To view this thread in it's original format and/or if you want to reply to this thread please [click here]

 

Author Neural Network
djm

2006-03-30, 7:08 pm

Hi

I have two input vectors the first one with values 0.950 -0.955 and
the second one with values 0.400 - 7990.

I am just not sure if I am defining this properly is it

net=newff([0.9500 0.9550 ; 0.400
0. 7990],[3,1],{'tansig','logsig'},'trainsc
g');

or something else ?

Also how does one figure out how many hidden nodes to use.

I would appreciate any assistance.

Regards,
Des
vijitnair@gmail.com

2006-03-30, 7:08 pm

The syntax looks fine.
PR is a "number of input vectors x 2" matrix and that is what you have
done.
I assume you have only one output vector.

You could look at the network parameters[color=darkred]
and change them[color=darkred]

and then do[color=darkred]
before you train the network.

I do not know if there is a method to determine the correct number of
hidden nodes to be used. However, this is a tradeoff between
overtraining the system (lot of hidden nodes) and undertraining the
system (lot less hidden nodes).

As a practice, I look at the performance of my network using cross
validation with different number of nodes. And choose choose the number
that seem to give decent performance on both train/test data sets.

vijitnair@gmail.com

2006-03-30, 7:08 pm

>> net
and change them TO WHATEVER SUITS YOUR DATA BEST[color=darkred]

djm

2006-03-30, 7:08 pm

Thanks for that.

Apologies for my ignorance but am I right in saying that
0.9500-0.9550 is 1 input vector and 0.400-0.7990 is the second input
vector because when
I read the matlab notes on backpropogation fundementals 5-7, I
think I could intrept the syntax below as saying that there is one
input vector with two elements

net=newff([0.9500 0.9550 ; 0.400
0. 7990],[3,1],{'tansig','logsig'},'trainsc
g');

I find this a bit confusing.
I am not sure whether I have 2 Input vectors or 1 input vector with 2
elements.

Essentially waht I want is 2 input values producing 1 target output
value

If you have any suggestions please I would welcome them.

Regards,
Des

vijitnair@gmail.com wrote:
>
>
> The syntax looks fine.
> PR is a "number of input vectors x 2" matrix and that is what you
> have
> done.
> I assume you have only one output vector.
>
> You could look at the network parameters
> and change them
>
> and then do
> before you train the network.
>
> I do not know if there is a method to determine the correct number
> of
> hidden nodes to be used. However, this is a tradeoff between
> overtraining the system (lot of hidden nodes) and undertraining the
> system (lot less hidden nodes).
>
> As a practice, I look at the performance of my network using cross
> validation with different number of nodes. And choose choose the
> number
> that seem to give decent performance on both train/test data sets.
>
>

Greg Heath

2006-03-31, 10:03 pm


djm wrote:
> Thanks for that.
>
> Apologies for my ignorance but am I right in saying that
> 0.9500-0.9550 is 1 input vector and 0.400-0.7990 is the second input
> vector


No.

> because when
> I read the matlab notes on backpropogation fundementals 5-7, I
> think I could intrept the syntax below as saying that there is one
> input vector with two elements


Yes

> net=newff([0.9500 0.9550 ; 0.400
> 0. 7990],[3,1],{'tansig','logsig'},'trainsc
g');
>
> I find this a bit confusing.
> I am not sure whether I have 2 Input vectors or 1 input vector with 2
> elements.
>
> Essentially waht I want is 2 input values producing 1 target output
> value
>
> If you have any suggestions please I would welcome them.


Use PRESTD to standardize your inputs. then use

net = newff(minmax(p),[H 1],{'tansig' 'purelin'},'trainscg').

If your output for training and test data will always be
restricted to the interval [0,1] then replace PURELIN
with LOGSIG.

Go to Google Groups and search using

greg-heath pretraining advice

Hope this helps.

Greg
[color=darkred]
> vijitnair@gmail.com wrote:

Sponsored Links







Also available: Server administration forum archive | Web Design forum archive | Software forum archive | Hardware reviews archive

Copyright 2008 codecomments.com