Better NN EA development - Page 64

finimej · #**631** (**permalink**) 01-27-2009, 02:46 PM

When the trainer has the class, then the EA can do the sell operation on the trained period. The EA only does the buy operation on the untrained period. So I tried to debug the code from PACO, Paco Hernández Gómez - Using a PNN (Probabilistic Neural Network) and MetaTrader 4 (MQL4) to trade Forex

in the file pnn.mq4

Quote:

int PNNClassifyVector(double vector[]) {
double length = ArrayRange(vector, 0);

double result = -99999999999999999999;
int resultClass = -1;
Print("result start =", result);
double fx[2] = {0, 0};

double classVectorCount[2] = {0, 0};

for (int i = 0; i < ArrayRange(pnn, 0); i++) {

int class = pnn[i][0];

double classVector[60];
for (int j = 0; j < length; j++) {
classVector[j] = pnn[i][j + 1];
}

classVectorCount[class]++;

fx[class] += MathExp((-1) * euclideanScalarProduct(vector, classVector) / (2 * MathPow(SIGMA, 2)));
}

for (i = 0; i < ArrayRange(fx, 0); i++) {
fx[i] *= 1 / (MathPow(2 * 3.14159265, length / 2) * MathPow(SIGMA, length)) * (1 / classVectorCount[i]);
Print("check point 1 result fxi = ",fx[i], ", result=",result, ", SIGMA=",SIGMA,", length=",length, ", classvectorcount=",classVectorCount[i]);

if ((fx[i]> result) && (fx[i] != result)){
Print("pre check point 2, fxi = ",fx[i], " > result=",result, ", i=",i,", resultClass=",resultClass);
result = fx[i];
resultClass = i;
Print("after check point 2=fxi = ",fx[i], " , result=",result, ", i=",i,", resultClass=",resultClass);
}
}
Print("check point 2 result fxi = ",fx[i], ",resultClass=",resultClass);
return (resultClass);
}

then I have the output on the trained area, see attached the picture.

when fx[1]=0, result =0, the if (fx[i]> result) gives TRUE! and let it pass the condition check, unbelievable. speakless! And this only happened on the trained period. It's never happend on the untrained period.

Can someone help to debug this?

finimej · #**632** (**permalink**) 01-27-2009, 04:54 PM

The Comez code has problem with the classification there. It only does buy operation outside trained period. Tried to debug whole day, and still not get the real problem solved.

From the russian forum here, get the modified PNN eric code. it uses the parzen window classification. The results is so so from backtest.

As for any Parzen window classificator, your will have the sigma problem, what value for the sigma is best for current priod? I optimized it, and get the result 0.03 for EURUSD M15.

finimej · #**633** (**permalink**) 01-27-2009, 07:13 PM

Link here is from codebase, he is using MACD and IMA matrix to calculate the Euclidian distance between trained vector and the test vector, good think so, need to be inproved. He is not using the parzen window classificator, but that just one step away.

What he is doing is he has a function that calcuate euclidean matric. So now we can go back to develope the PNNeric.mq4 above, to handle a matrix with iMA. Am I on the right track?

Quote:

double Euclidean_Metric(double&X_Data_Base[][v_dim_x], double Vector[v_dim_x],int num_v)
{
int i=0,i1=0,i3,i2,i4;
double Metric[1];
double t,sum;
ArrayResize(Metric,num_v);
ArrayInitialize(Metric,0.0);

for(i=0;i<num_v;i++)
{
for(i1=0;i1<v_dim_x-1;i1++)Metric[i]+=MathPow(X_Data_Base[i][i1]-Vector[i1],2);
Metric[i]=MathSqrt(Metric[i]);
}

//sort by ASCEND Metric
for(i3=0;i3<num_v-1;i3++)
for(i2=i3+1;i2<num_v;i2++)
if(Metric[i3]>Metric[i2])
{
t=Metric[i3];
Metric[i3]=Metric[i2];
Metric[i2]=t;
for(i4=0;i4<v_dim_x;i4++)
{
t=X_Data_Base[i3][i4];
X_Data_Base[i3][i4]=X_Data_Base[i2][i4];
X_Data_Base[i2][i4]=t;
}
}
sum=0;
for(i=0;i<Num_neighbour;i++)
{
sum+=X_Data_Base[i][5];
//Print(X_Data_Base[i][5]);
}
return(sum/Num_neighbour);
}