Discussion of article "Matrix Utils, Extending the Matrices and Vector Standard Library Functionality"

New comment
 

New article Matrix Utils, Extending the Matrices and Vector Standard Library Functionality has been published:

Matrix serves as the foundation of machine learning algorithms and computers in general because of their ability to effectively handle large mathematical operations, The Standard library has everything one needs but let's see how we can extend it by introducing several functions in the utils file, that are not yet available in the library.

The Multilayer perceptron will have 2 input nodes/neurons one for the Legs Height and the other for the body diameter on the input layer, meanwhile the output layer will have 3 nodes representing the 3 outcomes Dog, Cat, and mice.

Now say that we feed this MLP with the value of 12 and 20 for height and diameter respectively, we expect the neural network to classify this to be a dog right? what the One hot encoding does is that it put the value of one in the node that has the correct value for the given training dataset in this case at the node for a dog the value of 1 will be put and the rest will carry the values of zero. 

Since the rest of the values have zero's we can calculate the cost function by substituting the values of one hot encoded vector to each of the probabilities that the model gave us, this error will then be propagated back to the network in their respective previous nodes of a prior layer.

Author: Omega J Msigwa

 

thanks for the article, its really inspiring.

btw, what i found to be hard to deal with is the training NN based ea on Metatreader.

the use of matrices and vectors do save me a lot of efforts, but still training lots of variables on metatreader optimizer is hard to do. i've tried to write the weights & biases to CSV file to be loaded by training agent on metatrader optimizer, yet its just stopped at step less than 100. over than that, its look like the agent is just breaking the csv file. its no longer contain the weights & biases value as it suppose to be.

do you might have any idea to train such numbers of variables.?

 
adicahyanto #:

thanks for the article, its really inspiring.

btw, what i found to be hard to deal with is the training NN based ea on Metatreader.

the use of matrices and vectors do save me a lot of efforts, but still training lots of variables on metatreader optimizer is hard to do. i've tried to write the weights & biases to CSV file to be loaded by training agent on metatrader optimizer, yet its just stopped at step less than 100. over than that, its look like the agent is just breaking the csv file. its no longer contain the weights & biases value as it suppose to be.

do you might have any idea to train such numbers of variables.?

Its hard to tell without seeing the code and everything involved. I do not train the NN on the strategy tester I prefer to train trading parameters there. All my NN are self trained

 
adicahyanto #:

thanks for the article, its really inspiring.

btw, what i found to be hard to deal with is the training NN based ea on Metatreader.

the use of matrices and vectors do save me a lot of efforts, but still training lots of variables on metatreader optimizer is hard to do. i've tried to write the weights & biases to CSV file to be loaded by training agent on metatrader optimizer, yet its just stopped at step less than 100. over than that, its look like the agent is just breaking the csv file. its no longer contain the weights & biases value as it suppose to be.

do you might have any idea to train such numbers of variables.?

If you are using the fast genetic algorithm of the tester you must be careful as it has different "motivations" depending on the way you set the reward up .

Like mr Msigwa said ,its preferable to train your nets with back propagation . You can train more weights and the network's only reward is increasing "accuracy" 

New comment