Artificial neural networks. - page 6

[Vladimir]

There is no insanity. There is a deep understanding of networks. It would take a long time to state it here, and there is no point in doing so. So I will limit myself to stating my opinion briefly. Those who do not agree, speak up, give examples, even Renaissance Technologies if you want. Let's laugh together.

[Sergey Petruk]

gpwr:

There is a deep understanding of networks. It takes a long time to set it out here, and there's no point in doing so.

I don't quite understand your point - is self-optimising algorithms not in the realm of neural networks?, or is the very activity of building networks nothing, just a waste of resources?

[Denis Lazarev]

gpwr:

There is no insanity. There is a deep understanding of networks. It would take a long time to state it here, and there is no point in doing so. So I will limit myself to stating my opinion briefly. Those who do not agree, speak up, give examples, even Renaissance Technologies if you want. Let's laugh together.

It is probably impossible to give examples now, but in the future, with the evolution of the terminal, programming language, tools, it will be possible to make a full-fledged AI, many great men have been wrong about progress, some have changed their minds in time, many have failed because of it, so we should wait for

gpwr:

2 mathematicians, 2 physicists, 2 neurobiologists and 2 economists

(there's no guarantee that these neural networks don't already exist, they're probably working quietly in someone's terminal))

[Dmitry]

I also want to believe in the good news that neural networks can do a lot, and it is not for nothing that they are used in the latest modern developments. I will hope and try to create a Grail based on neural networks, but I don't know how long it will take.

[JohnyPipa]

Reshetov:

• We won't. Because if we do, then:

It's a vicious circle!

If we know what to feed the NS inputs, then the NS is useless.

And if we supply our inputs with stuff, we will not need NS.

Conclusion: neural networks are bullshit? !!!! =)

[Dmitry]

I think the really important information should be fed to the NS:

1)last quotes, e.g. last 10 hour bars (so NS can detect levels and paternas);

2)Current time (so the NS knows when to expect volatility spikes, because news and market openings happen in round time values);

3)Macroeconomic indicators (preferably in real time, how and from where to enter them in MT5 - do not know);

4) Good and bad news by different regions (I do not know how to implement it, I can enter the bad/good news ratio, but I do not know how to feed them to MT5 and from where to feed them automatically, without human intervention);

[Vladimir]

lazarev-d-m:

It may not be possible to give examples now, but in the future, with the evolution of the terminal, the programming language, the tools, it will be possible to make full-fledged AI, many great men have been wrong about progress, some have changed their minds in time, many have failed because of it, so we have to wait until

(They will not get together and make something really working))), in addition, there is no guarantee that these neural networks do not already exist, they are probably working quietly in someone's terminal

Progress in artificial neural networks started in 1943 and in 69 years they have not even learned to distinguish cats from dogs. There are areas where they have been used successfully, areas where the data does not need significant non-linear transformation. For example, predicting electricity consumption based on time of day and air temperature. Prediction of GDP growth based on factory orders, personal income, unemployment, etc. In the market, networks must make decisions based on past prices. To give N past prices themselves as inputs to the network is no good so establishing a non-linear relationship between noisy data is silly. Patterns do exist on the market but they are so distorted that even by conventional methods of non-linear time and price transformation they cannot be identified. For example, out of N past prices the most important moments can be only one or two segments, namely the segments where these prices reach support and resistance levels, trend lines or change their direction. How prices move between these important sections is in most cases not important. That is, the dimensionality N of the data is significantly reduced to 2-3x by such a non-linear transformation. The network itself will never learn to do such a non-linear transformation of prices. It is up to us to do it. But if we know what to do with prices, we already know the strategy in advance (breakdown of levels for example) and why do we need the network? If we know that a 1-2-3 pattern in a trend leads to continuation of the trend, why do we need the net? Maybe in the future computers will be so powerful that there will be new types of networks, more similar to biological ones (maybe in 20-30 years).

Read also here, where the disadvantages of networks are well described: https://en.wikipedia.org/wiki/Neural_network

For example this passage:

To implement large and effective software neural networks, much processing and storage resources need to be committed. While the brain has hardware tailored to the task of processing signals through a graph of neurons, simulating even a most simplified form on Von Neumann technology may compel a NN designer to fill many millions of database rows for its connections - which can consume vast amounts of computer memory and hard disk space. Furthermore, the designer of NN systems will often need to simulate the transmission of signals through many of these connections and their associated neurons - which must often be matched with incredible amounts of CPU processing power and time. While neural networks often yield effective programs, they too often do so at the cost of efficiency (they tend to consume considerable amounts of time and money).

[Stanislav Korotky]

lazarev-d-m:

It is probably impossible to give examples now, but in the future, with the evolution of the terminal, programming language and tools it will be possible to create a full-fledged AI, many great people have been wrong about progress, some have changed their minds in time, many have failed because of this, so we have to wait

And the terminal and MQL5 have nothing to do with it, because no one forbids the AI in general and the MU in particular, has long been prohibited to implement it (fully) using other tools, including adding it as a library to MT.

The problem is not to create a grid and teach it (how to do it is clear, because there is plenty of information and even ready-made implementations). The problem with meshes is quite different - it is in fact voiced here - the whole effectiveness of a mesh is "buried" in the choice of input data on which we train it. And to select correct data (in a market context), to transform it correctly (for example, to convert it into a multidimensional space, which forms, according to one theory, the attractors of a stochastic process hidden behind the outward manifestations of changes in quotes) - this is the most valuable know-how, about which I haven't encountered any useful information in practice. Everything suggested in the above branch has already been tried, of course, without success. For example, the macro indicators don't need the grid, because if you know how to interpret them correctly, you can trade hands. The news is not an indicator because, firstly, the reaction on them will be ex post facto, and we kind of want to predict the movements and make decisions before the news hit the market, and secondly, the reaction on the news is unpredictable - for example: an earthquake in Japan - the yen seems to be bad, but in fact the demand for it increased, and also the news is often negative, but less negative than expected and it is perceived as a reason for raising the market, and so on. In general, everything has been tried. Those who have succeeded in something (if any) do not sit here. And they do not share their experience.

[Vladimir]

Continuing the discussion about networks in the marketplace. Take the visual cortex of our brain as an example. Only V1 layer of this cortex contains 140 million neurons, while there are only 6 layers. All these neurons process information in parallel and are connected to each other through millions of adaptive synapses. Computers at the disposal of traders can have up to 1000 CPU cores. That is, to mimic the visual cortex, each core must compute hundreds of thousands of neurons in real time. Training such a network would take around a year. And all this to see the world and recognise images. Even if we are able to train this network successfully, it will still not achieve the same accuracy of object recognition that we do, because we are using more than just visual information. For example, one of the most difficult tasks for artificial networks is shadow recognition. This is a no-brainer for us, as we are familiar with the properties of light. But the network does not know these properties of light, and unless we teach it to do so, it cannot detect them. The network is also not familiar with transparency of objects and so on. Take the market where there is much more noise than in visual information and objects (price patterns) are much more distorted. And we want a couple dozens of neurons looking at past prices to set the market patterns. Ridiculous, isn't it?

[St.Vitaliy]

gpwr:

Continuing the discussion about networks in the marketplace. Take the visual cortex of our brain as an example. Only V1 layer of this cortex contains 140 million neurons, while there are only 6 layers. All these neurons process information in parallel and are connected to each other through millions of adaptive synapses. Computers at the disposal of traders can have up to 1000 GPU cores. That is, to mimic the visual cortex, each core must compute hundreds of thousands of neurons in real time. Training such a network would take around a year. And all this to see the world and recognise images. Even if we are able to train this network successfully, it will still not achieve the same accuracy of object recognition that we do, because we are using more than just visual information. For example, one of the most difficult tasks for artificial networks is shadow recognition. This is a no-brainer for us, as we are familiar with the properties of light. But the network does not know these properties of light, and unless we teach it to do so, it cannot detect them. The network is also not familiar with transparency of objects and so on. Take the market where there is much more noise than in visual information and objects (price patterns) are much more distorted. And we want a couple dozens of neurons looking at past prices to set the market patterns. Ridiculous, isn't it?

I will give you another example.

I went to RBC site, there are 137 news for today and only 3 of them had real impact on the ruble exchange rate. And it is very likely that there are effects that are not mentioned in the news.

That is, it is not only necessary to learn how to filter the input stream very carefully, but it may well be that in the input stream there is no information describing the reasons for these or those price changes.

In general, you should not confuse the warm and the soft. Machines were originally created to replace monotonous/studied human labour. If we take the trading sphere, the vast majority of participants do not understand what they are doing, which is reflected in the results. There are no unambiguous principles that guarantee results in the future, if you take the M3 dollar and the dynamics of SP500, you can see that even the buy&hold strategy is losing money. What then should the machines do - also lose monotonically...

Although art for art's sake, also has a right to life.