Searching for an arbitrary pattern using a neural network - page 4
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
Only classification. The network learns to distinguish situations (images), but it cannot know what to do in which case or what to call which image.
You are completely confused and you are misleading people. Classification may be with or without a teacher. If we teach the network with a teacher, as a rule, the output variable consists of 0 and 1 and in this case, the output variable is a call to action. (0 to sell 1 to buy) and the network will try to split the input vectors into these two classes. To be more exact, it will try to assign each input vector to one or another class. Taks.... This vector belongs to one and this one to zero.
Teaching without a teacher means initial setting of the "Number of classes" parameter. Say there is a sample of 1000 entries, and break them down to me into two classes one class will be 0 the other will naturally be 1. Just scatter them into two piles depending on distance of data. After all, if you imagine the input vector as coordinates of a point in multidimensional space, then the distance between the points is essentially what determines the grouping of the two clouds of 500 points each. Don't try to imagine multidimensional space. Imagine three-dimensional. Ordinary. As a result, we have a cloud of points that we need to partition.
In the first case, we forcibly partition them in such a way that the network response is as close to the target function as possible, while trying to get as close to it as possible while trying to impede optimization. That is, any cloud of dots can be partitioned as desired simply by changing their colour. Those that led to a profit when selling 0 and those when buying 1, it is another matter if we draw a hyperplane between them and leave zeros to the right of the plane, and ones - to the left. As an example.
In the second case, when learning takes place without a teacher, we just bluntly color these points into red and blue solely by their proximity to each other in multidimensional space. Also, this method has an option where we do not specify how many classes to divide the sample into, the network itself determines how many classes there are in the sample and the number of classes will be an important result of the optimization. I've just figured it out. Here is what I want to show you. Suppose the optimizer split our 1000 vectors into 5 classes. What should we do with them? Hoo of Hoo??? Now ta daaaaaaaaaa.... well, there's kind of a fanfare going on, you just can't hear it :-)
Once we have the 5 classes we need to classify them manually which cloud goes where. How to do it. First we should check one class for each cloud and then for another and check the cloud which has fewer mistakes. And if you take into account that there are four classes in the binary classification, I'll very calmly interpret the optimization result as an instruction to action and be like that.
The difference between the two approaches is only in one method the action instruction is prepared before optimization and in another method it is prepared after. And you can try to figure out which one is better. So.... this just came to mind....
You are confused and you are misleading people. Classification can be with or without a teacher...
Yeah... First of all, "enter" is spelled with an inflection. 2 - If all cows are horned, and an elk also has horns, it does not become a cow.
The word 'manually' is also conjugated. And it's exactly the same as "by hand" with "teacher". It's the same, but from a different angle. Without a teacher, it's just classification.
Representation of classification as a cluster of points in space and their proximity is not the point here, the actual values of the price are not of interest here. Classification is done in a different way here.
*
Generally speaking, I was saying that neuronets cannot think independently, they cannot think at all. For neural networks to be useful, they must be taught. And to teach them you need input-output (condition-result) pairs.
Anyway, the terms "with a teacher" and "without a teacher" are obsolete. Teaching with a teacher can be automated. And learning "without a teacher" is just an intriguing phrase, for gullible impressionable natures.
...
The only difference between the two approaches is that in one method the instructions for action are prepared before optimisation, in the other afterwards. Well, try to figure out which one is the right one. So.... that reminds me of....
And here you go.
Exactly. First, you have to prepare 100500 examples of different "head and shoulders" for it and teach it on these examples.
Actually, price patterns can be described by mathematics, you don't need NS for that. But an attempt to find the signs of a false pattern is exactly the task for the NS.
However controversial. If the pattern contains 3-4 bars we may use it, but what if we have dozens of them? How can mathematics help us here?
It is not clear how a neuronet changes the "focus of view" on a pattern. For example, a pattern of"Elliott Waves" consists of five waves where each wave is an independent pattern. In one large pattern one can see a variety of small shapes.
Perhaps if a neural network is trained to see all the variety of patterns, it could decompose one pattern into many shapes, and assemble many shapes into a common pattern? Or is that beyond the limits of technology?
Debatable, however. If there are 3-4 bars in a pattern, it's fine, but if there are dozens of them? What kind of mathematics would help?
So the pattern recognition algorithm should be invariant to the number of bars. It can be easily solved.
Question for connoisseurs: can a neural network be taught to scale the "view", moving between forms, generalising them into larger ones and dividing them into smaller ones, consistently identifying as a human does?
Do you personally understand how a person does it?
So the pattern recognition algorithm must be invariant to the number of bars. This is easily solved.
It's a mathematical method, not an algorithm that can detect complex patterns from any number of bars. I tried it myself, but I couldn't determine patterns of more than 4 bars mathematically.
What do you mean by "mathematically"? To compare values of OCHL parameters within a set of conditions and list variants of their relations: if(Oren[1] > Close[2] && ...)pattern = HEAD_N_SHOWLDERS;
Do you personally understand how a person does it?
I am talking about a mathematical method, not an algorithm that supposedly can determine complex patterns from any number of bars. I tried it myself, but I couldn't mathematically identify patterns of more than 4 bars.
What do you mean by "mathematically"? To compare values of OCHL parameters inside a complex of conditions and list variants of their relations: if(Open[1] > Close[2] && ...)pattern = HEAD_N_SHOWLDERS;