need help in undestanding neural network codes

Hi, I am writing a code to create an artificial neural network with some input,hidden and output layers. I also need a convergence chart showing the MSE vs the iteration run . First look at this example and please explain some sentences to me: p = [0 1 2 3 4 5 6 7 8]; t = [0 0.84 0.91 0.14 -0.77 -0.96 -0.28 0.66 0.99]; netj1 = newff(p,t,10); netj1 = init(netj1); netj1.trainParam.show = 10; netj1.trainParam.epochs = 100; netj1.trainParam.goal = 1e-12; netj1 = train(netj1,p,t);

z=sim(netj1,p);
trainerror=t-z;
perf = mse(trainerror);
**********************************
question 1 
what does * netj1.trainParam.show* mean?

question 2 Is netj1.trainParam.epochs = 100; maximum number of epochs?

question 3 netj1.trainParam.goal = 1e-12; why should a goal be determined? what value is more appropriate?

question 4 when i run this code a window pops up . in the window title it is written "Neural Network Training (nntraintool)". It has some data such as algorithms, progress, and so on. When i click on performance it shows MSE for test, train, and validation data. I can not understand where didi it choose the train, test and validation data? What should i do if i have an arbitrary set of data and i want 1/5 of them be test data and the rest train data? how can i have its plot for MSE versus run comparing train and test?

question 5 what does perf = mse(trainerror); really mean? where is perf used? in the plot maybe?

i'm sorry that my questios are too long and basic!

Thanks for reading it

:)

Answer 1

[Greg Heath]

close all, clear all, clc;

p = [0 1 2 3 4 5 6 7 8]; % 9 one-dimensional inputs

t = [0 0.84 0.91 0.14 -0.77 -0.96 -0.28 0.66 0.99]; % 9 one-dimensional outputs

rng(0) % Initialize the RNG so the run can be duplicated

netj1 = newff(p,t,10); % ( 10 hidden nodes are too many for only 9 points)

% netj1 = init(netj1); % Delete. Newff is self initializing

disp('SOME NEWFF DEFAULTS:')

dividefcn = netj1.divideFcn % dividerand (random division of trn/val/tst points)

datasplitratios =netj1.divideParam % trn/val/tst = 0.7/0.15/0.15

show = netj1.trainParam.show % Show training plot every 25 epochs

maxepochs = netj1.trainParam.epochs %1000 = maximum number of epochs

MSEgoal = netj1.trainParam.goal % 0 (Unrealistic for RW problems)

performancefunction = netj1.performFcn % 'mse'

netj1.trainParam.show = 10; % Show training plot every 10 epochs

netj1.trainParam.epochs = 100; % 100 (Unrealisticly low for RW problems)

netj1.trainParam.goal = 1e-12; % (Unrealisticly low for RW problems)

netj1.trainParam.goal = var(t')/100; % 5.4e-3 is More appropriate

% netj1 = train(netj1,p,t); % % z=sim(netj1,p); % % trainerror=t-z; % % perf = mse(trainerror);

[ netj1 , tr, z, trainerror ] = train(netj1, p, t ) % No semicolon for display

% Study all of the useful info in the 1st two outputs

perf = tr.perf(end) % The mse performance measure

normperf = perf/var(t') % 0.0089 < normalized mse goal = 0.01

% See the demo in the documentation for datadivision details

% doc nnet

% Neural Network Toolbox

% Getting Started

% Getting Started

% Fitting a function

% Using Command-Line Functions

Hope this helps

Thank you for formally accepting my answer

Greg