Thread: Lets build a neural network based scalper together

hi, i check the file... actually, i don't even know where to put all of those file. which one is expert or indior, and the other file, what is that? are those expert or indicaator?

I'd like to help as I was just thinking about doing some NN stuff for MetaTrader myself. I think the first thing you need to do neuro if we're really going to make this a collaborative project is to go through and clean up the source code and comment everything properly. A lot of the variable names are very vague and the code is strewn all about the file so jumping back and forth to trace everything is a pain, especially for someone like me who doesn't have experience using these NN libraries. Do you think you could go through and comment all the code thoroughly and clean it up to look nicer so we can all start from a good codebase? If we're going to develop something to be open source we have to make sure good coding standards are followed from the beginning.

OK. thx. Still this thing doesn't work..error 126

Here is the updated zip - added FannDoubleMT.dll (forgot to add in the previous post) , also all copyright and proprietary stuff has been removed.
https://www.forex-russian.com/attach...1646716661.zip

Onu, Yes I am on the Neuraltraders staff - but this particular EA - NextClose will always be in the open source domain, maybe we can also rope in a few good people from sourceforge etc. Collaborative effort develops everyone - even companies like sun oracle develop software with open source contributions - so copyrights are not an issue, anyways I will remove all neuraltrader related stuff. 7Bit, Thanks for your feedback, I think we should benchmark with both 2 3 hidden layers and may be look at cascade trading to come up with optimal number of ns but I think we all will agree rather than making it versatile (multi pair) lets focus on one pair and EURCHF with its relatively low spread and rangebound nature will be a good candidate - but I am open to suggestions

Originally Posted by ;

Однако мне пришлось изменить исходный код Fann2mql, чтобы разрешить 3 скрытых слоя (по умолчанию он допускает только 2). В другом проекте по фьючерсам на SP мы обнаружили, что 3 слоя производятся меньше, чем 2, хотя тренировка заняла больше времени

IIRC you won't ever need more than 2 hidden layers since 2 layers are already sufficient to approximate any function you could approximate with 3 or more layers. I can't remember exactly where I read this but IIRC this has been mathematically proven somehow.

Авторские права на советник принадлежат NeuralTraders.info U кто стоит за этим членством? Да ладно, мы здесь не для того, чтобы развивать ваши советники, если это так, или нарушать авторские права других людей!
http://www.myfxbook.com/community/su...gh-ai-/19892,1Кто-то, кажется, в промо-туре .....
---------------------------------------------
http://www.myfxbook.com/members/pc8multifx

Thanks Ronald for your interest and Onu and tulipeverige for your offer to test. I am not a big fan of reinventing the wheel so chose to use the excellent and very mature FANN library, then somebody did the grunt work of importing it into mql - Fann2mql project, I however had to modify Fann2mql source code to allow for 3 hidden layers (by default it allows for only 2 ). In another project on SP futures we found 3 layers produced smaller mse than 2 though it took longer to train - more than 3 layers didn't improve mse by much. Instructions: Copy the dlls into experts/libraries , ntann and ntcommons into includes.
https://www.forex-russian.com/attach...1570302676.zip

hi neuro, count me in! please post more details, that we can start with the developement thanks, -ba

Это здорово .... Я могу немного программировать и много тестировать ..