| View previous topic :: View next topic |
| Author |
Message |
Thorsten Kiefer
Guest
|
 Posted: Tue Apr 17, 2007 1:47 pm Post subject: Re: Should Brain / Turing Machines / DNA Computers / Neural |
 |
|
I think, neural nets can be trained with quantum physics.
You give an input. Then you compute the superposition of
all weights. Then you compute the super position of all
possible outputs. Then you mask out all weights which do
not yield the target output. And then you have a
global minimum assignment for the weights, for which the
error is zero.
Regards
Thorsten |
|
| |
|
 Back to top |
Guest
|
| Posted: Tue Apr 17, 2007 7:45 pm Post subject: Re: Robot Forth |
|
|
On Apr 16, 8:37 pm, menti...@myuw.net wrote:
| Quote: | --http://mind.sourceforge.net/js.html-- JavaScript for AI.
|
I've played with that... I also played with Eliza, which was
supposedly
named after the authors pet dog because it was so stupid. Eliza seems
to give more alive-seeming responses.
Rich |
|
| |
|
| Back to top |
John W. Kennedy
Guest
|
| Posted: Wed Apr 18, 2007 3:14 am Post subject: Re: Robot Forth |
|
|
aiiadict@gmail.com wrote:
| Quote: | On Apr 16, 8:37 pm, menti...@myuw.net wrote:
--http://mind.sourceforge.net/js.html-- JavaScript for AI.
I've played with that... I also played with Eliza, which was
supposedly
named after the authors pet dog because it was so stupid. Eliza seems
to give more alive-seeming responses.
|
The original Eliza was named for Miss Doolittle.
--
John W. Kennedy
A proud member of the reality-based community.
* TagZilla 0.066 * http://tagzilla.mozdev.org |
|
| |
|
| Back to top |
Evertjan.
Guest
|
| Posted: Wed Apr 18, 2007 12:39 pm Post subject: Re: Robot Forth |
|
|
John W. Kennedy wrote on 18 apr 2007 in comp.lang.javascript:
| Quote: | aiiadict@gmail.com wrote:
On Apr 16, 8:37 pm, menti...@myuw.net wrote:
--http://mind.sourceforge.net/js.html-- JavaScript for AI.
I've played with that... I also played with Eliza, which was
supposedly
named after the authors pet dog because it was so stupid. Eliza seems
to give more alive-seeming responses.
The original Eliza was named for Miss Doolittle.
|
No, Doolittle was the name of the professor,
I seem to remember from pre-assembler days,
memory being expensive in those days.
03 ."Eliza" ."Doolittle" = not .
--
Evertjan.
The Netherlands.
(Please change the x'es to dots in my emailaddress) |
|
| |
|
| Back to top |
Andrew Haley
Guest
|
| Posted: Wed Apr 18, 2007 1:45 pm Post subject: Re: Robot Forth |
|
|
In comp.lang.forth Evertjan. <exjxw.hannivoort@interxnl.net> wrote:
| Quote: | John W. Kennedy wrote on 18 apr 2007 in comp.lang.javascript:
aiiadict@gmail.com wrote:
On Apr 16, 8:37 pm, menti...@myuw.net wrote:
--http://mind.sourceforge.net/js.html-- JavaScript for AI.
I've played with that... I also played with Eliza, which was
supposedly named after the authors pet dog because it was so
stupid. Eliza seems to give more alive-seeming responses.
The original Eliza was named for Miss Doolittle.
No, Doolittle was the name of the professor, I seem to remember from
pre-assembler days, memory being expensive in those days.
|
Have I been trolled? The "girl" was Eliza Doolittle, the professor
'Enry 'Iggins, as any fule kno.
Andrew. |
|
| |
|
| Back to top |
Evertjan.
Guest
|
| Posted: Wed Apr 18, 2007 8:34 pm Post subject: Re: Robot Forth |
|
|
Andrew Haley wrote on 18 apr 2007 in comp.lang.javascript:
| Quote: | In comp.lang.forth Evertjan. <exjxw.hannivoort@interxnl.net> wrote:
John W. Kennedy wrote on 18 apr 2007 in comp.lang.javascript:
aiiadict@gmail.com wrote:
On Apr 16, 8:37 pm, menti...@myuw.net wrote:
--http://mind.sourceforge.net/js.html-- JavaScript for AI.
I've played with that... I also played with Eliza, which was
supposedly named after the authors pet dog because it was so
stupid. Eliza seems to give more alive-seeming responses.
The original Eliza was named for Miss Doolittle.
No, Doolittle was the name of the professor, I seem to remember from
pre-assembler days, memory being expensive in those days.
Have I been trolled? The "girl" was Eliza Doolittle, the professor
'Enry 'Iggins, as any fule kno.
|
Told you that memory was expensive.
--
Evertjan.
The Netherlands.
(Please change the x'es to dots in my emailaddress) |
|
| |
|
| Back to top |
Bill Marcum
Guest
|
| Posted: Thu Apr 19, 2007 12:19 am Post subject: Re: Robot Forth |
|
|
["Followup-To:" header set to comp.lang.forth.]
On 18 Apr 2007 07:39:04 GMT, Evertjan.
<exjxw.hannivoort@interxnl.net> wrote:
| Quote: |
No, Doolittle was the name of the professor,
I seem to remember from pre-assembler days,
memory being expensive in those days.
03 ."Eliza" ."Doolittle" = not .
You might be thinking of Doctor Dolittle, who talked to animals. Rex |
Harrison portrayed both Henry Higgins and Doctor Dolittle in movies.
--
Teachers have class. |
|
| |
|
| Back to top |
Charlie Springer
Guest
|
| Posted: Thu Apr 19, 2007 4:56 am Post subject: Re: Robot Forth |
|
|
On Wed, 18 Apr 2007 00:39:04 -0700, Evertjan. wrote
(in article <Xns991661D575A30eejj99@194.109.133.242>):
| Quote: | No, Doolittle was the name of the professor,
I seem to remember from pre-assembler days,
memory being expensive in those days.
|
Rex Harrison payer both Higgins and Dr, Doolittle, the guy who talked to
animals.
-- Charlie Springer |
|
| |
|
| Back to top |
Ivan F. Villanueva B.
Guest
|
| Posted: Thu Oct 25, 2007 2:09 am Post subject: Re: Software depository for AI / NeuralNet / ALife |
|
|
On 2007-09-20, fdecom <francescodecomite@tvcablenet.be> wrote:
| Quote: | On 19 sep, 04:10, pen...@catholic.org wrote:
Can someone enlighten me as to where I can find files (binary and/or
source code) pertaining to areas such as Artificial Intelligence /
Artificial Life / Genetic Algorithm / Neural Net ?
Is there a depository for such thing ?
Thanks a bunch !
www.sourceforge.net : try your keywords
|
sourceforge is just one of many Free Software Hostings. Probably the best
list of Free Software Hostings is at:
http://en.wikipedia.org/wiki/Comparison_of_Free_Software_Hosting_Facilities
BTW, does anyone know a meta search page for free software?
If there is no one and you would like to use one, would you be
interested in co-financing a weekend for 3 or 4 programmers who will do it?
Send me a private message if so.
--
Ivan F. Villanueva B.
Open General Artificial Intelligence -- http://ogai.org
FFII.org Deutschland -- http://www.ffii.de
FFII.org Espana -- http://es.ffii.org |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Tue Jan 08, 2008 5:55 pm Post subject: Re: Asking NN Backpropagation |
|
|
On Jan 3, 7:20 am, "navri nalhad" <navri_bint...@mathworks.com> wrote:
| Quote: | I'm a student from Universiti Sains Malaysia, Malaysia.
Right now, I’m doing a small project related with the
preparation of adsorbent to absorb air pollution. Here I’m
trying to study it usingneuralnetwork (NN). However, I’m
facing difficulties with programming NN in Matlab (here I’m
not using NN toolbox). I try to write the program based on
backpropagation NN, however the input data was not fit with
the target (output) data (R value <<<). And also I have
difficulty on how to optimize the input data toward maximum
output. (Below, I give the data of our experiment)
I hope someone could help me to this matter.
Data:
In preparation of adsorbent we have 4 different variables,
we varied each variable and we have 40 samples of prepared
adsorbent.
So this is the (4) input data: (M: metal loading, R: weight
ratio, P: period, T: type of material). Here, T value was
not real value, only two different type of material, coded
with 300 and 600.
M = [5 15 5 15 5 15 5 15 0 20 10 10 10 10 10 10 10 10 10 10
5 15 5 15 5 15 5 15 0 20 10 10 10 10 10 10 10 10 10 10];
R = [1.5 1.5 3.5 3.5 1.5 1.5 3.5 3.5 2.5 2.5 0.5 4.5 2.5
2.5 2.5 2.5 2.5 2.5 2.5 2.5 1.5 1.5 3.5 3.5 1.5 1.5 3.5 3.5
2.5 2.5 0.5 4.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5];
P = [12 12 12 12 24 24 24 24 18 18 18 18 6 30 18 18 18 18
18 18 12 12 12 12 24 24 24 24 18 18 18 18 6 30 18 18 18 18
18 18];
T = [300 300 300 300 300 300 300 300 300 300 300 300 300
300 300 300 300 300 300 300 600 600 600 600 600 600 600 600
600 600 600 600 600 600 600 600 600 600 600 600];
After we prepared the adsorbent samples, we test the
activity of these 40 adsorbents and we found out the
activity of these sorbents as follow:
Here we have two outputs (target):
X = [50.96 95.75 46.33 83.39 74.13 86.48 80.30 89.57 35.52
97.29 58.68 61.77 77.22 106.56 95.75 94.20 95.75 95.75
95.75 94.20 57.14 84.94 52.51 74.13 75.67 91.11 84.94 78.76
47.87 86.48 69.49 49.42 63.32 95.75 86.48 88.03 88.03 86.48
88.03 88.03];
Y = [11.65 14.87 7.03 15.87 12.05 16.88 11.85 17.48 0 15.87
12.86 11.85 11.65 16.88 15.67 15.47 15.67 15.47 15.67 15.67
10.85 13.46 7.43 13.66 11.05 14.26 10.25 15.87 0 12.46
11.85 10.45 13.86 15.27 14.67 14.67 14.46 14.67 14.67
14.67];
|
If you use the NN Toolbox, the input and target matrices
p and t would have the dimensions
size(p) = [Nvarin Ncases] = [4 40]
size(t) = [Nvarout Ncases] = [2 40]
If you do not use the NN TB, go to comp.ai.neural-nets and
search for free code (e.g., SNNS,...) or free introductory
copies of cheap code (e.g., EASYNN,...).
I have crossposted this reply to c.a.n-n.
Good Luck
Greg |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sat Jan 12, 2008 9:17 pm Post subject: Re: Creating Custom Transfer Function in NNToolbox |
|
|
On Jan 11, 7:36 am, trt <erenay...@gmail.com> wrote:
| Quote: | Hi all,
I'm having problems with defining my own transfer function
for a RBF Network.
|
RBFNN is the acronym for Radial Basis Function
Neural Network.
Although the function NEWRBE could be used, it
uses all of the Ntrn training input vectors as
centers of the hidden node Gaussian radial basis
functions.
This is usually overkill and, for large data
sets, ridiculous.
Unfortunately, MATLAB does not have a function
to prune the hidden nodes of NEWRBE down to a
practical size.
It is better to use the function NEWRB that
iteratively constructs an RBF network with H
(H<= Ntrn) hidden nodes.
A nice project would be to combine the
functions NEWRB and STEPWISEFIT so that a
practical value for H could be obtained
using either a the NEWRB forward search or
a backward elimination starting from NEWRBE.
| Quote: | I want the network to classify an input vector as belonging > to class 1, 2 or 3. I'm trying to achieve this by making
the output layer produce 1, 2 or 3 as the output.
|
THIS IS NOT RECOMMENDED.
See the comp.ai.neural-nets FAQ w.r.t coding
inputs and outputs.
Use three output nodes with unipolar binary
{0,1} targets. The targets for each input are
t1 = [1 0 0]', t2 = [0 1 0]' and t3 = [0 0 1]'
for classes 1, 2 and 3 respectively.
For example, if the training subset sizes are
Ntrn1, Ntrn2, and Ntrn3 = Ntrn-(Ntrn1+Ntrn2),
a reordering of training inputs could yield
the target matrix
t = [repmat(t1,1,Ntrn1),repmat(t2,1,Ntrn2),...
repmat(t3,1,Ntrn3)];
However, reordering is not necessary.
For an RBF network with size I-H-O, the sizes
of the training set matrices are
size(p) = [I Ntrn]
size(t) = [O Ntrn]
Search the archives in Google groups using
greg-heath rbfnn design newrb
for hints on how to use NEWRB.
| Quote: | When I create and train my network in the following way:
...
net1 = newff(minmax(P), [9,9,1],
{'tansig','tansig','purelin'},'traincgp');
|
This not an RBF NN.
It is a Feed Forward Multilayer Perceptron
Neural Network (FFMLPNN) sometimes ambiguoulsy
referred to as just MLP, MLPNN or FFNN.
Hope this helps.
Greg
P.S. Crossposted to c.a.n-n
-----SNIP |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sat Jan 12, 2008 10:17 pm Post subject: Re: Creating Custom Transfer Function in NNToolbox |
|
|
On Jan 11, 7:36 am, trt <erenay...@gmail.com> wrote:
| Quote: | Hi all,
I'm having problems with defining my own transfer function
for a RBF Network.
|
RBFNN is the acronym for Radial Basis Function
Neural Network.
| Quote: | I want the network to classify an input vector as belonging
to class 1, 2 or 3. I'm trying to achieve this by making
the output layer produce 1, 2 or 3 as the output.
|
THIS IS NOT RECOMMENDED.
See the comp.ai.neural-nets FAQ w.r.t coding
inputs and outputs.
Use three output nodes with unipolar binary
{0,1} targets. The targets for each input are
t1 = [1 0 0]', t2 = [0 1 0]' and t3 = [0 0 1]'
for classes 1, 2 and 3 respectively.
Then the 3 outputs will be estimates of
input conditional class posterior probabilities .
The input should be assigned to the class with
the maximum posterior.
For example, if the training subset sizes are
Ntrn1, Ntrn2, and Ntrn3 = Ntrn-(Ntrn1+Ntrn2),
a reordering of training inputs could yield
the target matrix
t = [repmat(t1,1,Ntrn1),repmat(t2,1,Ntrn2),...
repmat(t3,1,Ntrn3)];
However, reordering is not necessary.
For a single hidden layer NN with size I-H-O,
the sizes of the training set matrices are
size(p) = [I Ntrn]
size(t) = [O Ntrn]
| Quote: | When I create and train my network in the following way:
...
net1 = newff(minmax(P), [9,9,1],
{'tansig','tansig','purelin'},'traincgp');
|
This not an RBF NN.
It is a Feed Forward Multilayer Perceptron Neural
Network (FFMLPNN) sometimes ambiguously
referred to as just MLP, MLPNN or FFNN.
I cannot tell if you are designing the net
correctly without knowing
size(P)
size(T)
Unless you have a priori info to the contrary,
there is no reason to use two hidden layers.
Go to Google groups and search on
greg-heath pretraining advice
| Quote: | ...
[net tr] = train(net1,P,T,[],[],val,test);
a = sim(net1,Ptest);
|
I don't know if val, test and Ptest were
created correctly.
Correction:
a = sim(net,Ptest);
In fact, just replace net1 with net in all
of your commands.
| Quote: | outputs I get as the result of the simulation are real numbers. I
think it should be better to replace the purelin transfer function
with my own function instead of adjusting the results of simulation
after training with purelin.
-----SNIP |
| Quote: | Would you please help me to define my own transfer function or maybe
recommend another way for reaching my goal.
Any help or recommendation would be great.
Thank you.
|
Use
unipolar binary target coding
one hidden layer with H nodes
default training function TRAINLM
net = newff(minmax(P), [H 3], {'tansig','purelin'});
or
net = newff(minmax(P), [H 3], {'tansig','logsig'});
[net tr] = train(net,P,T,[],[],val,test);
a = sim(net,Ptest);
The assigned class index for each column of
a is the row index with the largest output.
Hope this helps.
Greg
P.S. crossposted to comp.ai.neural-nets |
|
| |
|
| Back to top |
trt
Guest
|
| Posted: Sat Jan 12, 2008 10:35 pm Post subject: Re: Creating Custom Transfer Function in NNToolbox |
|
|
On 12 Ocak, 21:17, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Jan 11, 7:36 am,trt<erenay...@gmail.com> wrote:
Hi all,
I'm having problems with defining my own transfer function
for a RBF Network.
RBFNN is the acronym for Radial Basis Function
Neural Network.
Although the function NEWRBE could be used, it
uses all of the Ntrn training input vectors as
centers of the hidden node Gaussian radial basis
functions.
This is usually overkill and, for large data
sets, ridiculous.
Unfortunately, MATLAB does not have a function
to prune the hidden nodes of NEWRBE down to a
practical size.
It is better to use the function NEWRB that
iteratively constructs an RBF network with H
(H<= Ntrn) hidden nodes.
A nice project would be to combine the
functions NEWRB and STEPWISEFIT so that a
practical value for H could be obtained
using either a the NEWRB forward search or
a backward elimination starting from NEWRBE.
I want the network to classify an input vector as belonging > to class 1, 2 or 3. I'm trying to achieve this by making
the output layer produce 1, 2 or 3 as the output.
THIS IS NOT RECOMMENDED.
See the comp.ai.neural-nets FAQ w.r.t coding
inputs and outputs.
Use three output nodes with unipolar binary
{0,1} targets. The targets for each input are
t1 = [1 0 0]', t2 = [0 1 0]' and t3 = [0 0 1]'
for classes 1, 2 and 3 respectively.
For example, if the training subset sizes are
Ntrn1, Ntrn2, and Ntrn3 = Ntrn-(Ntrn1+Ntrn2),
a reordering of training inputs could yield
the target matrix
t = [repmat(t1,1,Ntrn1),repmat(t2,1,Ntrn2),...
repmat(t3,1,Ntrn3)];
However, reordering is not necessary.
For an RBF network with size I-H-O, the sizes
of the training set matrices are
size(p) = [I Ntrn]
size(t) = [O Ntrn]
Search the archives in Google groups using
greg-heath rbfnn design newrb
for hints on how to use NEWRB.
When I create and train my network in the following way:
...
net1 = newff(minmax(P), [9,9,1],
{'tansig','tansig','purelin'},'traincgp');
This not an RBF NN.
It is a Feed Forward Multilayer Perceptron
Neural Network (FFMLPNN) sometimes ambiguoulsy
referred to as just MLP, MLPNN or FFNN.
Hope this helps.
Greg
P.S. Crossposted to c.a.n-n
-----SNIP
|
Thank you Greg, for your detailed and clear explanations.
Unfortunately, RBF was a different project of mine, I wrote it wrong,
my network is a MLP as you said.
I will try to use 3 outputs as you said. Thank you. |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sun Jan 13, 2008 4:04 am Post subject: Re: Creating Custom Transfer Function in NNToolbox |
|
|
On Jan 12, 5:35 pm, trt <erenay...@gmail.com> wrote:
| Quote: | On 12 Ocak, 21:17, Greg Heath <he...@alumni.brown.edu> wrote:
On Jan 11, 7:36 am,trt<erenay...@gmail.com> wrote:
Hi all,
I'm having problems with defining my own transfer function
for a RBF Network.
RBFNN is the acronym for Radial Basis Function
Neural Network.
Although the function NEWRBE could be used, it
uses all of the Ntrn training input vectors as
centers of the hidden node Gaussian radial basis
functions.
This is usually overkill and, for large data
sets, ridiculous.
Unfortunately, MATLAB does not have a function
to prune the hidden nodes of NEWRBE down to a
practical size.
It is better to use the function NEWRB that
iteratively constructs an RBF network with H
(H<= Ntrn) hidden nodes.
A nice project would be to combine the
functions NEWRB and STEPWISEFIT so that a
practical value for H could be obtained
using either a the NEWRB forward search or
a backward elimination starting from NEWRBE.
I want the network to classify an input vector as belonging > to class 1, 2 or 3. I'm trying to achieve this by making
the output layer produce 1, 2 or 3 as the output.
THIS IS NOT RECOMMENDED.
See the comp.ai.neural-nets FAQ w.r.t coding
inputs and outputs.
Use three output nodes with unipolar binary
{0,1} targets. The targets for each input are
t1 = [1 0 0]', t2 = [0 1 0]' and t3 = [0 0 1]'
for classes 1, 2 and 3 respectively.
For example, if the training subset sizes are
Ntrn1, Ntrn2, and Ntrn3 = Ntrn-(Ntrn1+Ntrn2),
a reordering of training inputs could yield
the target matrix
t = [repmat(t1,1,Ntrn1),repmat(t2,1,Ntrn2),...
repmat(t3,1,Ntrn3)];
However, reordering is not necessary.
For an RBF network with size I-H-O, the sizes
of the training set matrices are
size(p) = [I Ntrn]
size(t) = [O Ntrn]
Search the archives in Google groups using
greg-heath rbfnn design newrb
for hints on how to use NEWRB.
When I create and train my network in the following way:
...
net1 = newff(minmax(P), [9,9,1],
{'tansig','tansig','purelin'},'traincgp');
This not an RBF NN.
It is a Feed Forward Multilayer Perceptron
Neural Network (FFMLPNN) sometimes ambiguoulsy
referred to as just MLP, MLPNN or FFNN.
Hope this helps.
Greg
P.S. Crossposted to c.a.n-n
-----SNIP
Thank you Greg, for your detailed and clear explanations.
Unfortunately, RBF was a different project of mine, I wrote it wrong,
my network is a MLP as you said.
I will try to use 3 outputs as you said. Thank you.- Hide quoted text -
- Show quoted text -
|
See my last response. it is not a duplicate.
Hope this helps.
Greg |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Fri Feb 29, 2008 12:08 pm Post subject: Re: Neural Network - Training/Learning???? |
|
|
This reply was corrected for the unpardonable sin of top-posting
and cross-posted to the relevant newsgroup comp.ai.neural-nets.
On Feb 26, 11:00 pm, Kiran <k3bha...@gmail.com> wrote:
| Quote: | On Feb 27, 5:50 am, "pavan kumar" <avspa...@gmail.com> wrote:
I am usingNeuralNetwork toolbox in matlab. I am a beginner
reg Nnets. I have a question regarding "Training" and
"Learning". Whats the difference between these two? I saw
that training is for entire network and learning for each
node....is there anything more than this to the difference?
Training and learning are the same things but seen from different
point of views.
Just remember..."You train and the system learns"
Its like with any other teaching procedure wherein the teacher trains
and the student learns.
|
No. (That includes the rudeness of top-posting!)
A teacher teaches and a student learns.
Training is the combined process of teaching and learning.
You teach the net the target responses for a training sample
of Ntrn inputs that adequately characterize the salient
properties of the I/O relation to be modeled.
The net learns a set of weights to minimize an objective
function that characterizes the success of the learning.
Design is the combined process of training and validation.
For more details. Go to Google Groups and search on
greg-heath design train validate
BOTTOM LINE
Training and Learning are used interchangeably so often
that you shouldn't worry too much about it. It is more
important to understand the distinction and purposes
of training validation and testing.
Hope this helps.
Greg |
|
| |
|
| Back to top |
|
