"New Neural" is an Open Source neural network engine project for the MetaTrader 5 platform. - page 4

 
Figar0:

(Of course, everything is clear here, except one thing - what does it have to do with the NS?)

It seemed to me that the tests will take place, and to analyze the course of the tests will be different people, moreover, in different parts of the world - to exclude an additional variable in the work on the project, I propose to make it a constant component.
 

I don't quite understand how the neural network will be implemented: in the form of an advisor, libraries, or something else.

I propose the following:

1. Make a database of classic EAs/libraries. For each of these EAs / libraries, assign an id, name or count hash so that we can distinguish them from each other.

2. When optimizing Expert Advisors/Libraries from this database, the optimization results will try to be loaded from the database (centralized or distributed). If there are no optimization results for this EA in the database, the EA is optimized as usual and its optimization results are loaded into the database.


 
Radioamator:

I don't quite understand how the ATC on two MAs is connected to a neural network. I understand that there will be some tricky way to optimize classic EA. I suggest the following:

1. Make a base of classic Expert Advisor. As a classic EA, we can take an EA which was generated by the wizard. For each of these EA, assign an id, name or hash to be able to distinguish the EA from each other.


In the wizard the EAs are raw, there are problems in the modules of signals and I can not take the choice presented there right now to the category of elementary trading rules. If this situation is not solved before testing, there will be many undefined factors, for example - the network fails - or the EA's code? On MA the code can do here if not every second then every third and check.
 

gpwr:

Google "sparse coding" and "compressed sensing" and the work of Olshausen and Fields on Sparse Nets and their followers. It's a treasure trove. Restricted Boltzman Machines (RBM), which is the basis of Deep Belief Nets (DBN), and Convolutional Networks have also gained a lot of popularity because of their versatility.

Can you talk about the latter in a nutshell? And where are they used? Links are good, but I just can't do it right now.
 

I think we should not go into the "thicket", Figar0 is right. You shouldn't chase new things, otherwise this race will never end.

We need to stop at the classic types of networks, those voiced by TheXpert. And when the library (what to call the final version of the project?) will be brought to a working condition, you can make improvements to infinity.

 
It would be nice to make a library for quick work with matrices... I have some doubts, that it will work well in MQL5...
 
And yes, I propose to allow the use of system DLLs in the project
 
TheXpert:
And yes, I propose to allow system DLLs in the project

This is going to be a real problem.

We are specifically planning to make the library completely in source code and include it in the terminal, so that we can write safe experts.

The inclusion of DLLs kills the mass-market, although it opens up a narrow niche of special solutions.

 
Renat:
The key word is systemic, this is imho normal.
 
TheXpert:
The key word is system, which is fine, imho.

There are no "safe system" DLLs.

All of them are dangerous and all of them are destructive - it's elementary to cause a stack failure followed by an attack.