Libraries: MLP Neural Network Class
The grid works strangely.
During the learning process, the error first decreases, then starts to increase.
Is that the way it's designed? Or am I doing something wrong?
Is this how it should be? (The output is 0,0,0,0,0,0 and a huge error).
Hello Yury,
How can I make an Expert Advisor using this MLP class?
Thanks.
Maybe I'm doing something wrong or the code doesn't work correctly
I want to teach the NS the multiplication table and count 2x3, I do this:
#property copyright "Yurich" //+------------------------------------------------------------------+ #include <class_NetMLP.mqh> void OnStart(){ double vector[2]; // Input vector int snn[]={2,2,1}; // Network structure double out[1]; // Array for network responses double inpdata[];// Array of input training data double outdata[];// Array of output training data CNetMLP *net; int epoch=1000; int AFT=0; net=new CNetMLP(ArraySize(snn),snn,2,AFT); ArrayResize(inpdata,20); ArrayResize(outdata,10); for(int i=0;i<10;i++){ for(int j=0;j<10;j++){ inpdata[j*2] = (i+1)/10.0; inpdata[j*2+1] = (j+1)/10.0; outdata[j] = inpdata[j*2] * inpdata[j*2+1]; // Print("inpdata[",j*2,"]=",DoubleToString(inpdata[j*2])," / inpdata[",j*2+1,"]=",DoubleToString(inpdata[j*2+1])); } net.Learn(10,inpdata,outdata,epoch,1.0 e-8); vector[0] = 0.2; vector[1] = 0.3; net.Calculate(vector,out); Print("MSE=",net.mse," , out =",out[0]*100); } Print("MSE=",net.mse," Epoch=",net.epoch); } //+------------------------------------------------------------------+
in the log I have:
2012.10.07 22:46:43 TestMLPs (EURUSD,D1) 1824 bytes of leaked memory 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) 3 objects of type CLayerMLP left 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) 1 object of type CNetMLP left 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) 4 undeleted objects left 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=3.215934174267907 e-005 Epoch=1001 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=3.215934174267907 e-005 , out =23.81042803092551 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=2.506540371444645 e-006 , out =22.233366741152 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=1.524148111498897 e-006 , out =20.42036901380543 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=1.519171222235065 e-006 , out =18.89110154263913 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=1.047462369320528 e-006 , out =16.63410153653344 2012.10.07 22:46:43 TestMLPs (EURUSD,D1) MSE=9.477321159986828 e-007 , out =14.24605748950336 2012.10.07 22:46:42 TestMLPs (EURUSD,D1) MSE=6.585902193183645 e-007 , out =11.66913117122246 2012.10.07 22:46:42 TestMLPs (EURUSD,D1) MSE=2.237858920539329 e-007 , out =8.906822741170629 2012.10.07 22:46:42 TestMLPs (EURUSD,D1) MSE=2.540333890146069 e-007 , out =6.033412338430783 2012.10.07 22:46:42 TestMLPs (EURUSD,D1) MSE=2.26424262746638 e-007 , out =2.942888766617119
Maybe I'm doing something wrong or the code doesn't work correctly
I want to teach the NS the multiplication table and count 2x3, I do this:
In fact, you train the network with 10 examples. If you want to pass all 100 examples to the network, you need to take the training out of the data preparation cycle. It is also important to determine the number of neurons and the criterion for stopping training - 1000 epochs is too short.
#include <class_NetMLP.mqh> void OnStart() { double vector[2]; // Input vector int snn[]={2,2,1}; // Network structure double out[1]; // Array for network responses double inpdata[]; // Array of input training data double outdata[]; // Array of output training data // network creation CNetMLP *net; int epoch=1000000; int AFT=0; net=new CNetMLP(ArraySize(snn),snn,2,AFT); // preparing data for training ArrayResize(inpdata,200); ArrayResize(outdata,100); int m=0, k=0; for(int i=1; i<=10; i++) for(int j=1; j<=10; j++) { inpdata[m++]=i/10.0; inpdata[m++]=j/10.0; outdata[k++]=(i*j)/100.0; } // network training net.Learn(100,inpdata,outdata,epoch,1.0 e-8); Print("MSE=",net.mse," Epoch=",net.epoch); // network check for(int i=1; i<=10; i++) { vector[0]=i/10.0; vector[1]=i/10.0; net.Calculate(vector,out); Print(i,"*",i,"=",DoubleToString(out[0]*100,1)); } // deleting the network delete net; }
2012.10.08 13:46:59 test_nn (EURUSD,M15) MSE=4.22005256254196 e-005 Epoch=1000001 2012.10.08 13:46:59 test_nn (EURUSD,M15) 1*1=1.3 2012.10.08 13:46:59 test_nn (EURUSD,M15) 2*2=3.4 2012.10.08 13:46:59 test_nn (EURUSD,M15) 3*3=7.6 2012.10.08 13:46:59 test_nn (EURUSD,M15) 4*4=14.8 2012.10.08 13:46:59 test_nn (EURUSD,M15) 5*5=25.0 2012.10.08 13:46:59 test_nn (EURUSD,M15) 6*6=37.2 2012.10.08 13:46:59 test_nn (EURUSD,M15) 7*7=50.2 2012.10.08 13:46:59 test_nn (EURUSD,M15) 8*8=64.3 2012.10.08 13:46:59 test_nn (EURUSD,M15) 9*9=82.2 2012.10.08 13:46:59 test_nn (EURUSD,M15) 10*10=96.9
In fact, you are training the network with 10 examples. If you want to pass all 100 examples to the network, you need to take the training out of the data preparation cycle. It is also important to determine the number of neurons and the criterion for stopping training - 1000 epochs is too short.
thanks, I have figured it out, I will experiment with your code some more
Only one request:
CNetMLP *net=new CNetMLP(number of layers, network structure array, input vector size, activation function type: 0 - sigmoid, 1 - hyperbolic tangent).
do it this way: CNetMLP *net=new CNetMLP(network structure array, activation function type: 0 - sigmoid, 1 - hyperbolic tangent).
i.e. your code will calculate the parameters "number of layers" and "input vector size" from the network structure array by itself, imho it will add clarity and readability of the code.
Hallo Yuri,
zunächst einmal vielen Dank für dieses Stück Code teilt die Gemeinschaft.
Ich habe Ihren Code ein Fachberater zu bauen Diagrammwerte vorherzusagen, aber es scheint, dass es ein Fehler in der class_netmlp.mqh ist.
sobald ich versuchte, 3 zu verwenden oder mehrere Eingangswerte ist der Ausgang nicht ganz richtig mehr scheint .... können Sie Sie mir dieses Problem Festsetzung helfen?
Comparing the results of multiplication table training your network loses noticeably. On ALGLIB the network 2,5,1 for 100 epochs of training(https://www.mql5.com/ru/forum/8265/page2) gives better answers than yours with 1000000 epochs. The speed of computing 10000000000 epochs is not pleasing either.
Apparently the learning method is not very efficient. But still - thanks for your work, it is easier to understand in small code than in ALGLIB. But then we still need to move there.
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
MLP Neural Network Class:
CNetMLP provides multilayer perceptron (MLP).
The feature of the class is that input vector and network structure notions are separated, i.e. input vector and network structure descriptions are not connected to each other.
The size of the input vector can have any value within reasonable limits. Input data should be normalized, i.e. the data should be within the range -1 .. 1 or 0 .. 1. Various activation functions are applied for the network depending on the type of the used data: hyperbolic tangent should be used for -1..1 data range, while sigmoid is used for 0..1 data range.
The network has a layer-by-layer structure with a direct signal transmission. Tne network structure is described by a one-dimensional array, where the value of the array element determines the number of neurons in the appropriate layer. The number of layers and neurons is not limited. The network may consist of a single neuron.
Each neuron has multiple inputs, defined by its place in the network, and one output. If you need the network to give out N responses, the last layer should contain N neurons. The learning algorithm is iRprop. Input and output training data are located in one-dimensional arrays vector by vector. The learning process is limited either by the number of learnnig epochs or by a permissible error.
Author: Yury Kulikov