How can i stop the neural network after just one iteration ?

Hi i am a master student and i am developping a matlab code for evolutionary neural network so the training algorithm is done with a genetic algoritm and i dont need to train my neural network with any training function available ,that's why i set the epochs to 1 . This is the code

net=NEWFF(minmax(trainInput'),[hn,1],{'tansig','purelin'},'traingd');
net.trainParam.show = NaN;
net.trainParam.lr = 0.05;
net.trainParam.mc = 0.1;
net.trainParam.epochs = 1;
net.trainParam.goal = 1e-2;
[net,tr]=train(net,trainInput',trainOutput');

I need a function which can stop the neural network after just one iteration that s mean i don t want the neural network to calculate a new weights and ierates .Dis This partial code realize this ?

closeloop is it useful ?

[Joakim Lindblad]

Guess I'm not the only one ending up on this ancient question...

To stop after one iteration, use an output function which returns true:

'OutputFcn',@(varargin) true

See https://se.mathworks.com/help/nnet/ref/trainingoptions.html#input_argument_namevalue_d119e54101

This can be useful to do, e.g., after resuming from a checkpoint.

If you know how to stop after zero iterations, plz post.

Answer 1

[Greg Heath]

To design an I-H-O feedforward multilayer perceptron (with H hidden nodes) using Ntrn input/target training pairs with dimensions I and O, respectively,

1. Create the input and target matrices Xtrn and Ttrn with dimensions

 [ I Ntrn ] = size(Xtrn) 

 [ O Ntrn [ = size(Ttrn)

2. Standardize ( help mapstd, help zscore) the columns to have zero mean and unit variance. The resulting matrices are xtrn and ttrn.

3. This data will generate

 Neq = Ntrn*O

training equations which will be used to obtain

 Nw = (I+1)*H+(H+1)*O

unknown weights (including H+O bias weights).

4. Although

 H <= Hub = (Neq-O)/(I+O+1)

insures that

 Neq >= Nw,

the stronger condition

 H << Hub

is preferred in order to mitigate noise, measurement error and insufficient sampling variety.

5. The weights and biases are typically stored in matrices IW, b1, LW and b2 with sizes

 [ H I ]   = size(IW) 

 [ H 1 ]  = size(b1) 

 [ O H ] = size(LW) 

 [ O 1 ] = size(b2).

6. For an arbitrary input x, the hidden node and output signals are given by

 h = tanh( IW * x + b1 ); 

 y = LW * h + b2;

7. When the input xtrn yields the output ytrn, the corresponding error is

 etrn = ttrn-ytrn;

8. The corresponding degree-of-freedom adjusted mean-square error is

 MSEtrna = sum( etrn(:).^2 ) / (Neq-Nw)

9. A reasonable design objective is to choose (H,IW,b1,LW,b2) to minimize MSEtrna.

10. I would use the genetic algorithm to find the weights given an integer value for H that satisfies H < Hub (or better yet, H << Hub).

11. Notice that no functions from the NN Toolbox are necessary.

Hope this helps.

Greg

Answer 2

[Greg Heath]

>Hi i am a master student and i am developping a matlab code for evolutionary 
>neural network so the training algorithm is done with a genetic algoritm and i
>dont need to train my neural network with any training function available ,that's
>why i set the epochs to 1.

Since you are not going to use any of the Toolbox training functions, do not use the function train and don't initialize any of it's parameters

This is your code

 net=NEWFF(minmax(trainInput'),[hn,1], 'tansig','purelin'},'traingd');

Erroneously capitalized newff, omitted a left brace and initialized a training algorithm that won't be used.

If you are going to use a genetic algorithm you should not use TRAIN. Therefore, it makes no sense to initialize training parameters.

Once input and target matrices are normalized and edited, the only NNET TBX functions needed are net.IW, net.LW, net.b, sim and mse.

I recommend starting with a a general genetic algorithm outline. Then determine where the NN creation, weight updating and simulation commands should be placed.

It might be worthwhile to search online, including the archives of comp.ai.neural-nets and comp.soft-sys.matlab for sample code.

Hope this helps.

Greg

Answer 3

[Mariem Harmassi]

Hi firstly i want to think you for your help , i used a gentic algorithm as follow [x fval exitflag output population scores] = ga(@my_fun, nvars, options); and edit a matlab file my_fun.m where i declared the objective function of the genetic algorithm my_fun which will recieve as input the gneration of chromosomes : in this file a need to train my neural network just one time (one iteration) to have the mse and decide on the performance of the chromosome used.

[Greg Heath]

No.

Do not use TRAIN (capitals used for emphasis; use lower case in code).

Standardize inputs and targets (help ZSCORE or help MAPSTD)

After you generate the new population of weights (IW,LW and b), use SIM to get outputs y and objective function MSE (normalized mean-square-error if targets are standardized).

P.S. You erroneously posted your comment to my answer in a new answer. Please use the space provided for comments.

[Mariem Harmassi]

1. Ok for the capitals i am just new in Matlab but even i use sim the problem persists how can i control the number of iterations ?
2. i tried to set the network weights baut it doesn t work this is the code :

ainInput=[1 0.4 0.2 0.7]
trainOutput=[0.8 ]
chrom=[0.5 0.6 0.8 0.6 0.7 0.9 0.8 0.4 0.5 0.9 0.9;
0.1 0.7 0.6 0.7 0.9 0.5 0.9 0.2 0.4 0.5 0.9]
X=chrom(1,:)
net=newff(minmax(trainInput'),trainOutput',[6,1]);
et.IW

ans = 

[6x0 double]
[]
[]

>> net.LW

ans = 

[] [] []
[1x6 double] [] []
[] [0x1 double] []

>> net.b

ans = 

[6x1 double]
[ 0]
[0x1 double]

for the input layer it seems there is no connection between the four input trainInput and the hiiden layer ? what s wrong ?

[Greg Heath]

You appear to be very confused. I suggest the following:

1. Try to duplicatethe results from demos and/or examples in the NN Toolbox. 2. Check out some of my posted code via the search words

heath newff close clear

Do not attempt a genetic solution until you understand these basics.

Hope this helps.

Greg