Better NN EA - page 11

 

hi

misterD:
Hi,

do someone know some neural tools for free?

Thank you.

did u mean neural network indicator ? You can search on this forum , there's a thread discuss about that here

===================

Forex Indicators Collection

 
misterD:
Hi,

do someone know some neural tools for free?

Thank you.

You can get it from here:

https://www.mql5.com/en/forum/178272

https://www.mql5.com/en/forum/178276

 
misterD:
Hi,

do someone know some neural tools for free?

Thank you.

Try this:

https://www.mql5.com/en/forum/178276

https://www.mql5.com/en/forum/178272

 

Neural Networks & Multi-core support for MT4

Okay two real questions here. Firstly I've recently started to work with Neural Networks. I've tried various software packages (ALyuda, NeuralSolutions, Perseptio) all running in Windows and one from Stuggart University (SNNS I think). I found the windows versions too unstable or too expensive to warrant buying. The unix version was stable but a little more complicated to set up. I'm favouring Unix mainly because I have access to a Cray XT4 mainframe that has about 40Teraflops available and that runs specifically on Unix. I'm looking for suggestions for Unix neural networks that are slightly user friendly (I can use Unix but I'm not that familiar with it). Also if anyone has tried this approach what were your results?

Secondly running optimizations on my latest EA require comparison of several indicators on several time-frames. This is would be much quicker if the tester used all of the cores available rather than just one. Anyone know any platforms that support multi-core for optimizations? Cheers

 

Human brain is built for pattern recognition..

You read books, you see things, you hear stuff.. you are processing information in parallel and recognizing data all the time. Human brain is powered by 10^10 neurons with each neuron connecting each other via 10^4 synaptic links..

The architecture of computers is entirely different from human brains.

To simulate so called "pattern recognition" on computers, scientists devised Artifical Neural Networks. Now computers can recognize handwriting, speech and etc.. (only to some extent). To be able to perform as well as a human, you have to make more than 10^10 artificial neurons.

How many neurons can you possibly simulate on PC? I tried once. I constructed 10-layered NN containing 1000 neurons to predict the next EMA value from the previous 200 samples. I have 10 years of GBP/USD hourly data. I wrote program in C++. The program does not predict well at all. What's wrong? You need more neurons. But that's the limit of the processor already. Until know computers can't perform pattern recognition up to human level yet, unless you change the fundamental architecture of the processor to introduce Neural Processor which is an extract replica of human brain in Silicon.

What I am saying is.. Computers have not been able to surpass humans in pattern recognition. All they are just good at algorithmic calculations. They play chess by crunching every possible move.

On the other hand, humans are born for pattern recognition. We recognize patterns all the time. Trading is 100% pattern recognition and nothing else. If humans have problems predicting the future based on patterns, no computer will ever will!!

 

fixed spellings

Human brain is built for pattern recognition..

You read books, you see things, you hear stuff.. you are processing information in parallel and recognizing data all the time. Human brain is powered by 10^10 neurons with each neuron connecting each other via 10^4 synaptic links..

The architecture of computers is entirely different from human brains.

To simulate so called "pattern recognition" on computers, scientists devised Artifical Neural Networks. Now computers can recognize handwriting, speech and etc.. (only to some extent). To be able to perform as well as a human, you have to create more than 10^10 artificial neurons.

How many neurons can you possibly simulate on PC? I tried once. I constructed 10-layered NN containing 1000 neurons to predict the next EMA value from the previous 200 samples. I have 10 years of GBP/USD hourly data. I wrote the program in C++. The program does not predict well at all. What's wrong? You need more neurons. But that's the limit of the processor already. Until now computers can't perform pattern recognition up to human level yet, unless you change the fundamental architecture of the processor to introduce a Neural Processor which is an exact replica of human brain in Silicon.

What I am saying is.. Computers have not been able to surpass humans in pattern recognition. All they are just good at algorithmic calculations. They play chess by crunching every possible move.

On the other hand, humans are born for pattern recognition. We recognize patterns all the time. Trading is 100% pattern recognition and nothing else. If humans have problems predicting the future based on patterns, no computer will ever will!!

 

Self learning PNN Based

The original EA needs 2 EA, one trainer and one for trading. The PNN based learning is combined in this EA so it learns by itself, no need trainer. It's still in early stage

Self learning probabilistic neural network (PNN) - MQL4 forum

The attached mq4 file is only for learning purpose, do not trade with it

Files:
 

Had somebody the EAs mq4 file or the ex4?

TNX

 

Incoperating NN into my EA within 12 days time

By searching the internet twelve days ago, I found a PNN tutorial by Paco Hernandez Gomez which provides a PNN EA template, a PNN trainer EA, and a Utility library for saving and reading the data file which the PNN trainer creates (from recent historical data). Within a day, I was able to incorperate his PNN into my own EA which is built around the Derk Wehler EA template and utilities library (available on this forum). Up to this time, it's taken me seven months of constant rewritting, rethinking, and retesting to develope my EA to a point where it looses almost as much as it wins (which amounts to a waste of time). Suddenly, I now have an EA that backtests with a profit factor between 4.0 and 16.

During this past twelve days, I completed all the nifty logic I intended my EA to accomplish months ago. I got all my filters and rules perfected around the RSIOMA indicator so that I could trade during visible times of high probability and avoid the choppy market conditions of Fractile Dimension and low volitility of JuiceLevels. With the StochasticHistogram and the VoltyChannelStops_v1, I have an excellent visual view on my charts showing where to enter and exit trades. However, translating what I see in a instant from this interplay of converging lines into simple rules and filters has failed to produce reliable results that would allow me to go to work or to sleep while leave an open trade on the screen.

The more I compared results of the PNN with the results of my rules and filters, the less rules and filters I placed on the PNN. It reviews up to 60 characteristics of each vector of successful trades in it's data file as it looks for a matching entry point. The conditions for entry I write include a crossover of two lines with an angle in a certain range (three characteristics) plus my rules and filters.

I am beginning the forward testing now, and find that this EA may not work with some brokers. The signal may also be different between the demo account and the live signal from the same broker. As I discover the differences between historical data and live signal, I will review the previous post to learn more about how a PNN may learn on the fly from it's live feed.

 

What are the best neural network algos for signal filtering?Recurrent? PNN or someting else?