Writing an EA without the start function?

As we all know, the start function gets executed every new tick. some of my EAs are becoming so computation-intensive that two or three ticks pass me by before the start function is finished processing the previous tick. Now that I've advanced into machine learning, this sort of performance is not acceptable.

Can anyone show some sample code to get an EA to detect 'middle ticks' and queue them for processing?

RefreshRates()

Actually if synchronous computation doesn't work in real time why not to process the ticks asynchronously? A lightweight and quick EA or indicator gathers the ticks in a fifo queue that in turn is processed by a worker script.
BTW on quick markets ticks come in packages in which only the last one is fed to MQL programs. So the most of ticks are omitted anyway.

Well, if that's the case, I'll probably have an easier time just feeding the ticks into a MySQL database and then processing the ticks from there.

Thanks for the help.

what are you hardware specs?

You could put your EA within a continuous while loop. You have to use some kind of Sleep() function for this to not "crash" your system, but it does work.

This example would cycle your EA every 50 milliseconds:

I have 10 physical machines.

Machine #1: Trading machine.

 

1. Intel E2180
2. Gigabyte 945gcm s2c
3. 2 x 500gb Western Digital drives
4. 1 x 1TB WD drive
5. 1 x 250gb WD drive
6. 2GB ram
7. 2 x EVGA nvidia 8800gt dual monitor card

Machine #s 2 - 8: Cluster Computer

 

1. 7 x Intel Little Valley Atom
   each with 1GB ram. 6 of them have 20GB hard drives and the 7th has a 1TB hard drive

Machine #9: Supercomputer
I really went to town on this one, and built it for my financial modeling research.

 

1. Intel core i7 EE bloomfield
2. Asus P6T7 WS supercomputer
3. 24GB ram
4. 2TB WD hard disk
5. 4 x EVGA 017-P3-1298-AR GeForce GTX 295

Machine #10: Dell Inspiron 5150
My first laptop, decently powerful. I'm looking for a good replacement.



Machine #1 has 4 screens, and is what I primarily trade on.
Machine #s 2-8 are still used for some of my cluster computing.
Machine #9 was built specifically to house those 4 beasts called graphics cards.
Machine #10 was bought used in the fall of 2007. This laptop was what ran my EA when I first went live.

All told, together I would say I spent around $11-12k for this hardware. Cost a very pretty penny, but it allowed me to perform much better.

Also Wolfe, that's possible, but what happens to ticks that are missed?

What I'm trying to do is capture as many ticks as possible and feed it into the NN. The higher the quality of tick data, the better for me.

Check PM.

which NN did you use?evolving?backprop?quickprop?
IMO to manny ticks of data will leads to the chaos which the nn will had a very big mistakes/galat error

Why don't you try to use multi threading on your network..
here is what i propose..

The MQL script coding is just sending the data to your 10 computer.. one computer will compute the data and then the remaining idle computer will wait for another data, and try to harnes the power of your quad processors to heavyli use the multi threading .. but then again.. you will face the problem of serialization of tick data to your network ..

...i think you took the NN just way to serious to forecast the market movement.. aren't there simpler ways?

LOL! - I agree!! The market isnt THAT complicated!! its driven by human fear and greed and MOVES in herd waves and patterns. Most EA's have the problem of trying to make something out of nothing. Plenty of time the market is undecided and trying to trade an undecided market is blind luck. Make your move when the market moves and stay out of undecided areas (or have a stop that'll let you reverse without too much pain). I think NN's are overrated.

Fulltime, you know exactly why I'm using NNs

Yohan, I am. I've given up writing a NN in mql a long time ago. I'm writing some CUDA-compatible code to take advantage of the awesome power of the graphics cards that I bought. I'm waiting on Apple to release Snow Leopard so I can also take advantage of Grand Central Dispatch.

To explain what I'm training my networks to do.

I gave the NN 3 systems to use. 1 does well in trending, another ranging, and the last during breakouts.

The NN's job is only to choose which of the three systems to trade. That's it. I'm using NEAT (NeuroEvolution of Augmenting Topologies) to basically make the NN for me. The reason why I want as many ticks as possible is so that the NN can be as learned as possible.

Eventually, I'll add systems to the mix and allow the NN to change the lot size allotted to each system. Then I'll go further to allow the NN to change the parameters of those systems. The goal of all this is to try and eliminate me from the equation. I doubt I will ever reach it, but at least it will be an improvement on my current trading setup.

Yea RR, you know I completely endorse your approach as the only possible route to success. Your NN doesn't try to trade, it just tries to pick the correct trading system. I'm speaking to EVERYONE else that wants a computer to LEARN to trade. Most don't understand that learning computers have only now reached the point that they can play championship level chess. Poker bots that play online can only beat newbies (Pro's wipe the floor with them) and these are systems that are programmed within strict rule limits. Forex is 1,000,000X harder! - There so much random movements that a learning NN would only really become a collection of patches that try to box in all of the wrong possible decisions. Loosing it on the raw markets and expecting it to learn on its own is asking to blow your account - similar to asking a newborn to trade for you. Not gonna happen.

Its much easier to take the RR route and only ask the NN to pick the correct and proven profitable trading system to use right now. Pure NN isn't needed to trade the markets profitably. These are enough indicators that will help your EA target possible profitable trades then it only becomes a question of risk management. Limit losses with smart stops (you don't need monster stops if your entries are good enough), maximize gains and you should have a pip advantage. That's a far more realistic EA goal.