"New Neural" is an Open Source neural network engine project for the MetaTrader 5 platform. - page 10

New comment
 
sergeev:

Is working with NS only about selecting its topology?

Of course :) . The main thing is the topology, everything else is a dance with tambourines.

We need to createa constructor of networks, not limiting ourselves by some presets. And if possible, universal.

We don't owe anything to anyone. The network constructor is much more convenient and obvious, and it's better to do it in anything other than MQL5.

This is firstly.

Second, the network concept is not related to the constructor concept at all. That is, after creating the network libraries, nothing prevents you from adding the renderer on top. THEN.

Third, the training method has little effect on behavior. The behavior is ultimately influenced by the target function. And it does not matter what method is used to achieve it.

 
TheXpert:
the network concept is not related to the constructor concept at all. I.e., nothing prevents you from bolting a visualizer on top after creating the network libraries. POTUS.


By the way, I'm not talking about the visualizer at all.

What's the name of the topic ... Neural Network Engine Project at MQL5

What is meant by "engine"?

I think that the "engine" is not just a single topology with a chosen method of training.

By "engine" I mean a kernel of possibilities of working with principles of NS. This engine will allow to construct topologies and use different methods of learning (as far as we can).


As for the visualizer for this stage of creation of the "engine" it is not necessary yet. The visualizer is just a means to view the resulting topology and it doesn't play any role at all in the end.

 

I don't mind that we start with some quite specific topology and just one teaching method.

But I'd really like the base classes to be extensible to new topologies. In the best expected case we should have something like bricks and cement, so that we can build any kind of structures.

 
sergeev:

I'm not against starting with a very specific topology and only one teaching method.

But I'd really like the basic classes to be extensible to new topologies. In the best case we can expect, we should get something like bricks and mortar, so that we can build any kind of structure.

In my opinion, it is in the form of bricks, as a basis (or as separate modules) and it is necessary to build.

papaklass:
Start with something simple. Get a result, and then get into the thick of it. This way you'll never come to a consensus. Choose a leader and follow his decisions. This, however, does not preclude discussion. But if the decision is made, it must be implemented.

And with this I completely agree, you need something simple and universal at the first stage. Then you can implement more serious things.

TheXpert:

We don't owe anything to anybody. The net constructor is much more convenient and obvious, and it is better and more obvious to do it in anything but MQL5.

Just the implementation in pure MQL5 is interesting, at least to me.
 

I will state my imho in full:

TheXpert:

-- Type of project (the way you interact with the user)

A set of libraries. With a possible add-on to visualize the construction and operation of networks.

The set of libraries includes

-network implementations

-Implementation of a committee, in which you can easily combine networks into a coherent architecture.

-Implementation of pre-processing and post-processing of data.

-Implementation of an advisor template, using libraries that implement three functions -- training, direct run of the trained network (committee) and auto-training (for the tester).

-- Networks to be implemented in the project

spoke up earlier

-- Pre-processing of data and everything connected to it

this includes first of all normalization.

Then, this can also include pre-processing of inputs for redundancy and dependency.

Also filters, compression, etc.

-- Architecture, interfaces

I'll show you in a separate post. Already started a little earlier.

-- implementation, connectivity.

The implementation in the form of libraries.

- The implementation of the interface, i.e. the library returns to the user a pointer to the interface with basic functions and minimal network and topology setup.

- Full functionality.

- ability to create your own customized networks with the ability to connect them to the committee

- ability to create your own custom pre-processors with the ability to work seamlessly with all networks.

--Testing, debugging.

I don't have a good idea of the whole process yet. But we need to

- testers. Those who will write test cases for networks and run, looking for bugs

- reviewers. Those who will watch the revisions (and possibly merge into the main branch), you can cross

 
TheXpert:

I am stating my imho in full:


That makes more sense.
 
Interesting:
That makes more sense.
Alex's doc makes it even clearer.
 
TheXpert:
Alex's doc makes it even clearer.

I did it especially for you (and everyone else), so that you wouldn't have to search the wilderness of the forum for planning.

You have to keep the basic documents and regulations at least that way.

1...34567891011121314151617...100
New comment