Better NN EA development - page 66

[finimej]

Thanks Barnix. Yes, what is the input parameter to PNN, and how to normalize the input parameter is the very important decision to make. (I am still searching around here). For me, the High and Low price is more important than open close price, therefore I am thinking to use the already normalized high and low, that's is the demarker indicator. I have start coding it yet.

Right! how do you smooth the indicator after the normalizing? using the average value to smooth the indicator?

Have you integrated these matrix_indicators to your PNN and backtest/optimaze your EA yet?

[solaiman]

[finimej]

I think the comez problem is at his *= there. I added some print out to check the value for the trained period, and see the output value is way too large. you can self try it. problem is not solved yet.

fx *= 1 / (MathPow(2 * 3.14159265, length / 2) * MathPow(SIGMA, length)) * (1 / classVectorCount);

Print("check point 1 result fxi = ",fx, ", result=",result, ", SIGMA=",SIGMA,", length=",length, ", classvectorcount=",classVectorCount);

if ((fx> result) && (fx != result)){ //here some time I have if 0>0 is true.

Print("pre check point 2, fxi = ",fx, " > result=",result, ", i=",i,", resultClass=",resultClass);

result = fx;

resultClass = i;

Print("after check point 2=fxi = ",fx, " , result=",result, ", i=",i,", resultClass=",resultClass);

}

[finimej]

To Signature "OnTheRoad" above,

I found this one modified Comez PNN does sell. Have check the code through yet. and this one is doing buy and sell operations. (doing weekend cleaning..)

souce Self learning probabilistic neural network (PNN) - MQL4 forum

[solaiman]

some notes.

finimej:
I think the comez problem is at his *= there. I added some print out to check the value for the trained period, and see the output value is way too large. you can self try it. problem is not solved yet.

in this part:

fx *= 1 / (MathPow(2 * 3.14159265, length / 2) * MathPow(SIGMA, length)) * (1 / classVectorCount);

we can recode as

#define pi 3.14159265

#define two_pi_sqrd 2.506628 // (2*pi)^(1/2)

fx *= 1 / (MathPow(two_pi_sqrd*SIGMA, length) ) * (1 / ClassVectorCount);

in my testing , ClassVector[0] = 4367 , ClassVector[1] = 4323

and length = 30day*288vector/day = 8640 ( I used Newbar() function for TF5m thus each 5m the trainer will generate only one vector thus my pnn.dat file for one month training is less than 5MB).

now see:

fx[0] *= 1 / (MathPow(2.506628*0.1, 8640) ) * (1 / 4367);

=> fx[0] *= 1.4392e+5198 !!!!!!!

=> fx[1] *= 1.4539e+5198 !!!!!!!

these results are beyond the double precision of mt4.

do you agree these results?

[lodol2]

hi

hi barnix

what i have to do to get the iwavelet indicator to work

it show nothing in my window only black

thanks for help

lodol

[barnix]

svm288 dll and the header file

[barnix]

SVM prediction with normalize=1 (STD)

[barnix]

SVM prediction with normalize=3 (RNG)

[barnix]

SVM prediction with normalize=0 (without normalize)