Machine learning in trading: theory, models, practice and algo-trading - page 598

 
Aleksey Terentev:
1) How are things going in training? I don't see how weights are applied.
2) Are the weights of the neuron itself available?
3) As the medium, you can take the derivative of the close or fast MA of the 1-4th order. Or increments.
4) I would put the size of the hidden layer equal to the input.

weights go to adders, everything is classic, then the weight is fed to the activation function. it was just inconvenient to sign everything on the touchpad

As an "environment" - here we mean an action already performed by NS externally, for example, a transaction... i.e. system memory for n actions backwards, the same with feedback - what action led to what result

so i'll make an internal layer equal to the input one... and maybe i'll add another one afterwards

 
Yuriy Asaulenko:

Only a waste of time. It will not work on the real data.

For example: One wrong answer of NS, and it will affect all subsequent ones.


Well, not really, just ns will take into account the seriality and productivity of transactions as an additional factor... this is an option, I'm not saying that the final

That's why I am asking what other ideas you have

+ it is a built-in adaptive element... like a transistor

and the main NS performs other functions

 
Maxim Dmitrievsky:

weights go to adders, everything is classic, then the weight is fed to the activation function. it was just inconvenient to sign everything on the touchpad

As an "environment" - here we mean an action already performed by NS externally, for example, a transaction... i.e. system memory for n actions backwards, the same with feedback - what action led to what result

so i'll put an inner layer equal to the input one... and may add one more layer later

Then as an option the sum of profits in pips. A successful trade -> increase, and vice versa.
But here the nuance is that in this case we should think about adding Q-function, because it should be regarded as a reward. Or derive a formula for learning, taking into account the reward.
 
Aleksey Terentev:
Then how about the sum of profits in pips. A successful trade -> increase, and vice versa.
But here the subtlety is that in this case we should think about adding Q-function, since it must be regarded as a reward. Or derive a formula for learning, taking into account the reward.

Yes, ok :) good choice

So far the formula is simple - separate neurons, to which the past results of the trade are fed, not even neurons, but values are added to the adder. I haven't really read about coolerings yet

 
Maxim Dmitrievsky:

Yes, ok :) good choice

So far the formula is simple - separate neurons that feed the past results of the trade, not even neurons, but just values are added to the adder. I haven't really read about coolerings yet

You create two identical NS, teach one of them with traditional images, and the second one with direct transfer of weights, dosed according to results from prediction of the first one, i.e. the second one should learn only from positive experience.
 
Ivan Negreshniy:
You create two identical NS, you teach one to traditional images, and the second by direct transfer of scales, dosed depending on result from prediction of the first, i.e. the second should learn only on positive experience.

There is such a variant, at first several NS are trained on different periods, then they are combined into one... sophisticated adjustment :)

I would like it to understand when it starts to work sharply and to readjust itself

 
Maxim Dmitrievsky:

there is such a variant, at first several NS are trained on different periods, then they are combined into one... sophisticated adjustment :)

I would like it to understand when it starts to work shrewdly and to readjust itself

You obviously have the wrong structure for such purposes.
 
Yuriy Asaulenko:
You obviously have the wrong structure for such purposes.

What's the right way?

 
Maxim Dmitrievsky:

How do you do it?

I have no idea). But it's definitely not like that.
 
Maxim Dmitrievsky:

there is such a variant, at first several NS are trained on different periods, then they are combined into one... sophisticated adjustment :)

I would like it to understand when it starts to work shrewdly and rebuild itself.

Firstly - not exactly a fitting, as it is proved that the committees work better, this is explained, for example, on opposite sign deviations of individual models, leading to increased accuracy. Secondly, the proposed tandem of NS is a single, self-adapting model, which has also proven to be positive in reinforcement learning methods.