Automated Trading and Strategy Testing Forum
New article Growing Neural Gas: Implementation in MQL5 is published:
The article shows an example of how to develop an MQL5-program implementing the adaptive algorithm of clustering called Growing neural gas (GNG). The article is intended for the users who have studied the language documentation and have certain programming skills and basic knowledge in the area of neuroinformatics.
Thanks for clear explanation on the subject.
Would it be possible to share some practical code for Reconstruction future price, for instance from optimal signals.
The idea is:
1. Input (Source): multiple currencies (18)
2. Destination: Optimal signal of the currency we would like to predict (pic: 2. Optimal_Signals)
3. Find a neuro-connections between Source and Destination and explode it in trading.
Another question about NN reconstruction:
Is it possible instead of Random Samples use our samples instead , like on pic 2:
Our brain can reconstruct the picture in less then a second, let's see how much time it takes for NN do the same, just a joke, it is not challenge.
Random generated Samples are not very interesting to see as there is no meaning behind or use, however if we can draw dots ourselves with some meaning behind it , it would be much better fun. :-0)
Thank you for the great job, this is very exciting.
Can you download the Expert Advisor for this acticle? Can this be tested using Strategy Tester?
I am looking forward to implement it and use it instead of a fixed neuronal network.
i ask if anyone of the community success to develop EA using GNG?
i ask if the result is good ?
Well.. after what? five months o_O no one is willing, so maybe I can comment on that myself.
In the implementation you've seen in this article the NN is a self-adapting variant of a so called radial-based functions net. If you compare an EA based upon GNG algorithm to the same type of EA based on a non-adaptive type of clusterization neural network you most probably are going to obtain better results with GNG rather than without it. So, to answer your latter question, yes, the result is good in the sense I just explained.
Turning to the former one, I myself have developed EAs with a GNG inside, and it worked moderately well, yay. However, for everyday use I prefer other algorithms which as a rule are non-neural. I can back it up by reminding that ANN always presents a "black box" meaning that you do not really understand what is going on in there when it processes input data. That implies that ANN would be an algorithm of choice only in a situation when you'd have a completely unstructured set of data with absolutely unknown intrinsic dependencies which you would like for the ANN to somehow extract. Note: without any promise for an accurate result. In any other case, i.e. when you have some ideas about how the dependencies in your dataset may be organized, you'd wanna first try other, more deterministic "white box" ways of structuring it. Thousands of them.