How to form the input values for the NS correctly.
| * * | * | * | * | * | * |* -----------------------1--------------- *| * | * | * | * | * |If the sigmoid enters saturation with an input value of 1, it makes no difference whether the input value is 2 or 3 - the result is about the same. The input values must be within 1. There must not be any conflicting samples - the same input value with a different output value.
... I envy those people who were taught theory at university and who have this knowledge.
>> there's nothing to envy here, in the last millennium I also studied at university, and before that subject was called more modestly, in soviet terms: TAP..., at best,
they/teachers/ will make noodles out of their ears to read hours... they won't go further than 2-3 pages of each section of a typical textbook!!!
2 StatBars Thank you very much for the articles.
Integer писал (а) >>
If the sigmoid at input value 1 enters saturation, then it makes no difference whether to apply value 2 or 3 to the input - the result will be about the same. The input values should be within 1. There should be no conflicting samples - the same value at the input with a different value at the output.What about inputs not normalised to one? Can sigmoid be used or are other functions required?
Integer писал (а) >>
Конфликтных образцов не должно быть - одинковых значений на входе с разным значением на выходе.
It turns out that it is best to have more than one value at the output (i.e. to classify the market not just up or down, but with some intermediate states as well). And more on the inputs.
You need to normalise the inputs. For example to find a sample with maximal range and normalize it, and to remove the constant component. Here is a wide field for creativity, for example you may calculate values relatively to МА or relatively to the regression line and then normalize. Each sample separately, relative to its maximal range, can also be normalized.
Yes, by the way, it's good that you brought this up. I keep doubting how it would be more correct (in your experience) to ration - one sample by itself or in total on all samples?
I have decided to rename the branch.
I am sometimes inclined to reach this conclusion too. It turns out that we just compress the data and the issue of non-normalised data is removed.
There's also the problem of weights hitting zero input values. They won't take part in the training...
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
Questions about the correct values for the NS inputs keep popping up on the forum. But unfortunately this question is still not fully answered. I have only recently taken up NS and now I understand the importance of this question. I envy those people who were taught theory at institutes and have this knowledge.
Therefore let us in this branch as fully as possible to open a question of RIGHT values and their kinds.
Just do not want to start with specifics (such as taking the differences in neighboring prices). To begin with a theory about the general requirements for the input values is desirable. And then, if it goes well, examples are also possible.