Discussion of article "Neural networks made easy (Part 2): Network training and testing" - page 4

New comment
 

Good time, Dmitry.

If the topic is not closed I want to give you an idea (and it comes out of your jewellery application " Custom event"), here is its essence:

During the operation of the already"Trained Neural Network", add"Objects" to the graph with the mouse. Each/defined object is a signal for the Neural Network.

For example...

- when adding a"Rectangle", the Neural Network receives a signal - Here (in the opinion of the "Master") the Neural Network module for actions in the "sidewall" should be enabled;

- when adding"Arrows" - Here (according to the "Host") there should be an Entry into position (direction depends on the arrow type);

- ...

and so on.

Receiving "human" corrections in the process, the Network makes changes and takes into account User Corrections in its next recommendations.

Understandably, this is hard on the machine.....

Files:
xidkt9.JPG  31 kb
 
hello, really a very useful article.
 
I have question,

how does the EA open a buy or sell position after training?
I want to know the accuracy of this neural network
 
KrisnaMT5 #:
hello, really a very useful article.
 
I have question,

how does the EA open a buy or sell position after training?
I want to know the accuracy of this neural network

After training you can use neural network like indicator to generate signals to make trades.

 

Good afternoon!

Please help me to understand these questions:

1. Is it a never-ending learning process?

2. How can the robot trade using the results of its learning?

3. Where can I see what conclusions the robot has reached as a result of learning?

4. Can the robot learn and trade in parallel?

I would be grateful for your help.

 

There are errors in MT -> tools -> experts:

1. Fractal - 2022.08.26 08:05:33.239 Fractal(EURUSD,H1) OnInit - 131 -> Error of reading EURUSD_PERIOD_H1_ 5fr_ea.nnw prev Net5004

2. Fractal_2 - 2022.08.26 08:05:42.389 Fractal_2 (EURUSD,H1) CNet::feedForward - 418 -> Different size of inpur layer(238) and input data(240) - 20 messages per second.

Also Fractal is trained much slower than Fractal_2. In Fractal the number of Error, Undefine and foracast increases with each passed %, while in Fractal_2 Undefine starts from 100% and gradually decreases.

Does this need to be fixed or is it normal?

 
uNique8 #:

There are errors in MT -> tools -> experts:

1. Fractal - 2022.08.26 08:05:33.239 Fractal (EURUSD,H1) OnInit - 131 -> Error of reading EURUSD_PERIOD_H1_ 5fr_ea.nnw prev Net5004


This is informing that the pre-trained net is not loaded. This is normal at first startup. You should pay attention to this message if you have already trained a model and at the second run the Expert Advisor failed to load it.

 
uNique8 #:

There are errors in MT -> tools -> experts:

2. Fractal_2 - 2022.08.26 08:05:42.389 Fractal_2 (EURUSD,H1) CNet::feedForward - 418 -> Different size of inpur layer(238) and input data(240) - 20 messages per second


Check what you are feeding to the model input. This message indicates different array sizes.
 
uNique8 #:

Good afternoon!

Please help me to understand these questions:

1. Is it a never-ending learning process? - The learning process ends when goals are reached. It can be a finite number of iterations or reaching a desired error level

2. How can the robot trade using the results of its learning? - The model can be used as a generator of trading signals

3. Where can I see what conclusions the robot has reached as a result of training? - What do you mean by "what conclusions"?

4. Can the robot learn and trade in parallel? - Potentially yes. It depends on the algorithm of the EA construction.

I would be grateful for your help.

Answers above in the text.
 

Could you help me solve this?



 
Guilherme Mendonca #:

Could you help me solve this?



Hi, remove "const" in feedForward and calcHiddenGradient method parameters


123456
New comment