Machine learning in trading: theory, models, practice and algo-trading - page 2300

[denis.eremin]

Funny, of course - finding the AUDNZD = AUD/USD / NZD/USD correlation is too difficult a task for the NS.

Although maybe the package is so glitchy

[Forester]

As with the multiplication table, you need more different examples.

In the case of forest multiplication, it will just be averaging (e.g., 500) the closest training examples. The exact formula will never be found.

Example for 2 training examples and 2 trees in the forest.

Trained

6*4=24

6*7=42

If you ask the model.

6*6

she will find the nearest

6*4=24

6*7=42

And it will calculate

(24+42)/2=66/2=33

For 6*5 she will answer the same way.

The more different intermediate options in training, the more accurate the answer will be.

[Aleksey Mavrin]

By the way, is there such a standard list of tests for neurons? You know, like, what tasks they should solve with a bang, half-bang, etc. I just want to run it.

[mytarmailS]

denis.eremin:

Yes

)))) so what did you want?

add hidden layers

[denis.eremin]

mytarmailS:

)))) so what did you want?

Add hidden layers.

What, the task is so difficult?

[Forester]

mytarmailS:

)))) so what did you want?

Add hidden layers.

For such a simple task, one layer, in theory, is enough.

So it's unlikely to get 99% with 10 layers

[Aleksey Mavrin]

denis.eremin:

What, the task is so difficult?

And what is the best result with what output layer activation function?

and how many output neurons? one?

[mytarmailS]

What, the problem is that complicated?

No, one layer is primitive, it's just one weight multiplication

elibrarius:
For such a simple problem, one layer, according to the theory, is enough.

So it's unlikely that 10 layers is 99%.

That's your theory.

[Aleksey Mavrin]

elibrarius:
For such a simple problem, one layer, according to the theory, is enough.

So it is unlikely to get 99% with 10 layers

If the problem can be solved with 100% accuracy manually given weights (and activation functions), then there is a minimum required number of layers and neurons. less accuracy could be much lower.

I understand this problem to be one of those.

[Maxim Dmitrievsky]

A new feature to work with the importance of signs

I'll have to try it out at my leisure

https://medium.com/bcggamma/gamma-facet-a-new-approach-for-universal-explanations-of-machine-learning-models-b566877e7812