Topic: Artificial Neural Network & Genetic Algorithm Library For Deja Vu - page 2. (Read 4648 times)

I've just taken out 4 books on neural networks, genetic and simulated annealing algorithims and I'm going to try my hand at a comphrenisve bitcoin trading bot, I think I'm going to need 1 autoassociative noise filter, a "regime" decision maker, and finally individual regime based trading agents. Thankfully theres an absolute mountain of training data in the bitcoin world so I shouldn't run into major snags due to insufficent training time but we'll see I guess.

Strongly considering making a blog to document my progress but I feel like that is a bit too self-serving, so I might just have a github thats open for people to gander at.

What language are you writing in? personally I'm a C++ guy so I generally make everything I want in C++ and then have stuff I need hooked in like openCL and .json data.

No problem, he is certainly a great teacher. He is able to explain things in a very clear and simple fashion and he doesn't overcomplicate anything beyond what is necessary. I've tried watching other lectures on ANN's and all of them are very hard to follow compared to this lecture series. He also has a lot of other great lecture series related to ANN's if you look through the videos in his channel.

absolutely facinating, I've been looking for an introductory series into this for quite some time, I'm watching every video and exploring for myself, I have quite a few projects that could take advantage of this technology.

EDIT: I've watched almost all of his videos, I'm extremely grateful that you could show me this mans channel, what a brilliant teacher.

This is something I coded more than a year ago when I was attempting to build high performance trading bots. It is designed to be an extension for my Deja Vu scripting application. I was going to keep it to myself but I realized it takes a hell of a lot of computing power to evolve an ANN capable of profitable trading and I was never really able to get a safe and profitable trading bot out of it. I figured I may as well release it for other people to try their luck with it. It's written in PascalScript which probably makes it slower than just about any other ANN software out there (probably why I could never generate a useful ANN). But to be honest I've never really used any other ANN software so I wouldn't really know how fast it is compared to other ANN software. I learnt basically everything I know about ANN's from Jeff Heaton's awesome lecture series titled Neural Network Fundamentals and Neural Network Calculations.

The design I implemented was a simple recurrent neural network type of ANN with the ability to attach context neurons to any other neuron. It's basically just a feed forward network with the addition of context neurons which can remember previous outputs of any given neuron (each context neuron feeds back to only one neuron instead of the whole layer) and you can also control the number of cycles for which the context neuron will remember the data before replacing it with new output data. Of course it can generate entirely random networks and serialize the network into its corresponding weight and DNA strings which can easily be stored in text files. It also contains all the functionality required to breed nets and evolve nets in an automated fashion given some training data. The download package contains some historical price data formatted in the necessary way.

The functions which test and evaluate the performance of each net can be found in the Includes/ANN/test.pas file. I tried many different methods for evaluating the trading efficiency of each net but none of them really worked very well. The method employed in the test.pas file which comes with this library tries to simulate real trading and then measure how much profit was made, but it's quite a difficult thing to do correctly given only historical price data and I am far from a trading expert (hence why I wanted to build ANN's which would trade for me). The library will work as it is but I suggest you try modifying the evaluation functions yourself and see what works best. When you run the training script (scripts/ANN/evolve_nets.pas) you will see the scores of each net in the debug box at the bottom (anything above 1 means a profit was made). If anyone has any other questions about the library feel free to ask.