Neural networks, how to master them, where to start? - page 13

 
Korey >> :

it turns out
The only purpose of the NS is to understand what is being explained while you are explaining it to the dumb person))))

Genius! And that's completely fair. That's why I say we should start by writing a neuron.

 
arnautov писал(а) >>

That's why I say we should start by writing a neuron.

So are we having an argument? - No!

We started by writing a neuron in Mathcad-e, training it, looking and playing with training variants (gradient, stochastic). Then we did the same for 2 and 3-layer NS.

Everything as the doctor ordered!

 
LeoV >> :
And you don't allow the following idea - in order to make a profit with this program, you don't need to change this error function in it?

Of course I do... :) I am not a phantom of anything at all...

And I also admit that a neural network is not necessary to make a profit at all... :)

And even programming is not necessary... and forex is not necessary to make a profit... :)


But as long as we are talking in this thread about making profit on forex with the help of neural networks,

But the optimization of the network itself has not been canceled...

And there are many ways to improve the network ... and ready-made software just limits some of the possibilities...

 
Solver.it >> :

But as long as we are talking about getting profit on Forex market with help of neuronetworks,

then optimising the operation of this very network has never been done away with...

And there are a lot of ways to improve the network... and off-the-shelf software just limits some of the possibilities...

So why don't you help the software developers?


For example, can you derive delta-rule for arbitrary target function? I, for example, would be very grateful.

But I doubt that you will succeed.


However, if you have about 10 grand for the licensed pro version, I think the probability of having this feature is more than 0.

 
The software that generates networks is for students and teachers,
for high school, and then there's the smart stuff for office workers.
You should know that professional stuff is what you do when you write it yourself.
 
Korey >> :
Keep it simple, what there is ready-made software that generates networks is for students and teachers,
for high school, and then there's the smart-ass plankton.
The professional stuff is when you write it yourself.

That's the funny thing: off-the-shelf (good) software usually has several levels.

Let's say, for fun, for work, for research. And usually the highest level is quite professional.

And it is far superior in functionality and versatility to any piece of hand-made by one person.


My first serious project was exactly the neural network library. It was a team of 4 people, with the final 2.

I didn't have enough time and money for guya, I quit my job, the company fell apart, the library stayed.

Number of architectures created is comparable to NS2, if not more, recurrence is built into the perseptron model by default.

And it is possible to create something that Elman and Jordaan are quietly smoking on the sidelines. The commercial project didn't work out, but it's great,

Nevertheless, I wouldn't even dream of getting a semblance of what the leading companies in the field do.


For there is no way one (or two) people can compete with research institutes and a staff of professional mathematicians.


By the way, my target function is standard.

 

In Trading Solutions itself it seems that the target function cannot be changed, but in Neuro Solutions it can. For some reason it seems to me that the sum of squares of errors when applied to financial series is not the most adequate function. It overestimates the role of large deviations, which are much more numerous in finnings than in the normal distribution.

 
TheXpert писал(а) >>

My first serious project was specifically a neural network library. There was a team of 4, with the final 2 people working on it.

The number of architectures created is comparable, if not more, to NS2, recurrence is built into the Perspectron model by default.

And you can make something that Elman and Jordaan just sit back and watch.

Respect!

By the way, I have a standard target function.

I'm also using the standard function since I haven't yet matured enough to experiment with it.

But I already have some ideas. I've already said in a thread dedicated to optimal MM, that the main load in terms of providing the profitability of the MTS, is the quality of predictions of the analytical unit, which function is taken by NS. However, it is equally important to meet the condition for the optimal MM - the horizon of the position opened and its capitalization. So, it turns out the answer to the last two important points can be given by the same NS in real time, with the same single output! - By selecting the "right" functionality, we will not only get an answer in what direction to open (sign at the output of the axon), but also get the optimum capitalization (output amplitude). However, the question of the trading horizon (optimal mean value of the transaction in pips) remains open... We still need to look into it. And we suggest(Ezhov, Shumsky) to use not (x-y)^2 as usual, but "correct" one, like -ln(1+xy), where y is the output of the NO.

 
TheXpert >> :

That's the funny thing: off-the-shelf (good) software usually has several levels.

Let's say, for pampering, for work, for research. And usually the highest level is quite professional.

And it is far superior in functionality and versatility to any piece of hand-made by one person.


My first serious project was exactly the neural network library. It was a team of 4, with 2 people working in the end.

I didn't have enough time or money for it, I quit my job, the company fell apart, the library stayed.

Number of architectures created is comparable to NS2, if not more, recurrence is built into the perseptron model by default.

And it is possible to create something that Elman and Jordaan are quietly smoking on the sidelines. The commercial project didn't work out, but it's great,

Nevertheless, I wouldn't even dream of getting a semblance of what the leading companies in the field do.


For there is no way one (or two) people can compete with research institutes and a staff of professional mathematicians.


By the way, my target function is standard.

Don't confuse a COMMERCIAL PROJECT with a PERSONAL one, so to speak....


Why would I want to compete with NS2 ??? Don't I have better things to do ? Or why should I implement a bunch of possible networks in one library?


this is kindergarten trousers on straps...

I have an extremely specific goal, and extremely specific implementation with possibility to change ANY network parameters...

I change whatever I want in the network... And I don't care about interface and other purely COMMERCIAL gadgets...


I have only one network implemented in my library, but the way I want it... and as such it's better than ANY commercial product...

(of course I'm only comparing WORKABLE products here... For if you don't know how to program, then you'll have to use someone else's implementations)

All the more so, a really good commercial product costs a pretty penny... and the funny thing is, it's not a sure thing anyway,

that it provides all the functionality I need...

 
Where can I get NS2 with a cure?