Market etiquette or good manners in a minefield - page 45
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Convergence of weights during training (for illustration, I took 100 epochs, but 24 is enough)
Abscissa - epoch number, Ordinate - value of a particular weight (there are 24 of them in total)
And on the trained grid, I rarely have the values of weights beyond +/-1, so for now I think that the mutual arrangement of weights (after training) is what's important, and their absolute values are probably enough to place in a single range. So, in all likelihood, GTR rules here.
Like Einstein's General Theory of Relativity?
Exactly -:)
You know what the nets remind me most of? People... Why go far, take for example the way I learn from you - it's a trivial adjustment of my "scales"... -:)
And the picture is very similar - from accidentally hearing, reading, through the shock of total incomprehension and finally "click" ... you know which direction to go in.
I recall a pun which said that even if you put together a panel of fools you would never find a good solution.
Why did I want a committee? Because I'm used to thinking in terms of turkeys - like if one turkey is wrong or overdrawn, then two turkeys is more reliable, and three or more is quite good. However, when I look at Reshtov's Expert Advisors (he is a good man himself), which have a dozen of inbuilt indices, I often think of this pun...
Genetics is a bitch. I can make a case for it.
>> I look forward to it.
Looking forward to it.
Okay. Genetics is a universal algorithm.
Any of the current neural network learning algorithms (other than genetics and the like)
is specialised and a priori better because genetics doesn't take into account domain specifics.
It's about the same as looking up the area of an arbitrary figure using the Monte Carlo method.
And in general it's not I, but you should substantiate your statement that "genetics rules".
Because I repeat the gradient descent method is specialized, genetics is universal.
to YDzh
What do you get out of it with genetics? I wasted two months of my time on it. First of all, 8 parameters are nothing. To optimize code blocks separately is absurd - the network must be a single whole, otherwise it's not a network at all. Suppose you have 8 parameters more than enough, then after optimization you have to take some result (out of ten thousand) and here it is... HERE IT IS...)
Well, you have no way to choose! Of course, you can, of course, trust Nadezhda Petrovna and choose the highest balance, or the lowest drawdown, but it would be a coin flip, and this lady is fickle.
Ok. Genetics is a universal algorithm.
Any of the current neural network learning algorithms (other than genetics and the like)
is specialized and a priori better because genetics doesn't take into account the peculiarities of the domain.
It's about the same as looking for area of arbitrary figure by Monte Carlo method.
And in general it is not I but you who should substantiate your statement that "genetics rules".
As I repeat the method of gradient descent is specialized, genetics is universal.
To my taste, genetics is better applied where the expected result is difficult to formulate. When you consider optical recognition, or XOR, then it is clear: here is the input, here is the data that I want to get at the output. And in the case of Forex, what do I want to get as an output? Profit. And what are we waiting for on a single tick or bar? I do not know. Or we must develop a strategy, according to which the NS should act in my understanding - and then adapt it to this strategy of behavior. Genetics allows NS to "optimally realize itself" on input data in order to make a profit. By this criterion, it gets the opportunity to reproduce further. IMHO this is a more flexible approach.
And in the case of forex, what do I want to get out of it? Profit. On a single tick or bar, what are we waiting for? I do not know.
You don't have to be so quick... You know -:)
For me the next increment sign on the output is enough, and I will make a profit myself.
You don't have to be so quick... into the dumps, you know -:)
For me the sign of the next increment is enough, and I will make a profit myself.
Well, let's say on small timeframes this sign won't give you anything. Because of the spread, of course.
Well, let's just say that on small timeframes this sign won't do you any good. Because of the spread, of course.
Why do I need timeframes? Who invented them, let him trade on them.