Neural networks. Questions from the experts. - page 14

New comment
 
alsu:

lasso, you can read about the activation functions, including SOFTMAX, here, p. 22

And not only that. Thanks for the link.

BB. Page after page is so straightforward that I'm surprised.

A lot depends on the TUTOR!!! ))

 

lasso: Очень многое зависит от УЧИТЕЛЯ!!! ))



You should not aspire to be a teacher. It is a dead-end road.
 
LeoV:

You should not aspire to be a teacher. It is a dead-end road.
Leonid, is there any way you could explain about the teacher? I had a feeling that if you teach the recognition of an up-bar-down bar, you will get nonsense. Or is this problem distant from you personally in DayTrader? My feeling is that you have to learn "bar up" - "bar down"...
 
jartmailru: Leonid, is there any way you could explain about the teacher? I had a feeling that if you teach the recognition of an up-bar-down bar, you will get nonsense. Or is this problem distant from you personally in DayTrader? My feeling is that you should learn "bar up" - "bar down"...

In short, the situation is as follows. To trade, it is enough to understand where the start of an up move and where the start of a down move are. This is enough to make profit. It is useless to strive for the teacher, because it is unrealistic, and it is not necessary, because to make a profit it is enough to receive timely signals for entries to open positions. Between these signals, the network, in fact, can do anything, up to complete mismatch with the teacher - we do not care, as long as it gives the signals for opening positions in time. )))
 
LeoV:

In short, the situation is as follows. For trading it is enough to understand where the beginning of an up move and where the beginning of a down move are. It is enough to make profit. It is useless to try to get a teacher, because it is not realistic, and even it is not necessary, because to make a profit it is enough to get signals for entries and open positions in time. Between these signals, the network, in fact, can do anything, up to complete mismatch with the teacher - we do not care, as long as it gives the signals for opening positions in time. )))

Thank you for your reply. I guess it can be understood as something like this:

- that it is possible to make a definition of points of extremes

- You still need a teacher, but the result should be measured in terms of profit, not error

- That it is necessary to give the grid a clear task to look for inputs - the follow-up is not its problem

(and, say, it is a trawl and stop topic - actually, the teacher may be trawled as well)

Hmmm... teaching in terms of maximising profits-sounds mysterious... :-)

 

Hello.

There are far more questions about NS than answers. I've read a decent amount here on the forum and in the literature, but a clear understanding of

- what you're doing, are you doing it right? -- in my head.

I haven't seen any books like "Physics for Young People with Examples" on NS........

So here's a suggestion - set a practical, uncomplicated problem

and, with the help of respected experts, bring it to a logical solution.

Then the conditions will get more complicated.

Preferably with the help of Statistica 6. So:

=================================================================

There is a set of training data (720 pairs, attached) representing:

- as input the value of the intersection point of two stochastics normalized to the range {0;1}

- on the output - the result of this signal processing.

For example:

stochastics crossed at 87.8984 --> 0.88 and the TS opens a sell and closes on a profit. Set [0.88;1].

Stochastics crossed at 33.1254 --> 0.33 and the TS opens buy and loses. Set [0.33;-1]

--------------------------

Question: What NS should be chosen (type, activation functions, number of neurons in the layers, etc.) and how to correctly train it, so that it summarizes its data and classify the sample into two classes A = 1 and B = 1.

---------------------------

I should note at once that the classes are heavily mixed, even more so than in Fig..,



but still distinguishable.

I can divide it even by non-cunning linear methods, but I can't get NS to give the best (or at least comparable in quality) result.

I hope for adequate answers and that they will be useful for the rest of the forum!

Files:
v011f_8c_frm.zip  12 kb
 
lasso:

I support you!

I myself read a lot about NS, but have not yet reached a clear understanding for myself, I can't even formalise the input data: how to input data to NS if my indicators give a signal in the form of a set {0,1,2,3} ????

 

It is not possible to teach a network using this set.

The input is one value and the output is A or B. So what? What information does this set provide? What is the network supposed to learn?

 
joo:

The network cannot be trained on this set.

The input is one value and the output is A or B. So? What information does this set contain? What should the network learn?

The input values are distributed on a straight line (segment).

The segment [OA] can be divided into two parts [OB] and [BA], with point B placed inside the segment [OA].

The NS must learn to identify this point and assign the input to one of two classes, A([OB]) or B([BA]).

1...789101112131415161718192021...25
New comment