DEAP and ConvnetJS - Genetic Algorithm vs Neural Networks - Same or Different?

BradT7 commented 6 years ago

Hey again,

Quick question, are DEAP and ConvnetJS essentially doing the same thing? Or is there a difference between Genetic Algorithm and Neural Networks?

The reason I ask is I stumbled upon these threads using convnetjs in gekko, and if they are indeed different methods of testing, perhaps they could be used together somehow to complement each other?

Here's the links to the findings: Gekko Bot Resources Which was linked from: Bestone Strategy which is what gave me faith in the possibility of gekko actually being profitable, would love to mimic these results! And this is where I found you! Integrating Machine Learning AI instead Technical Analysis to Gekko

If they're different then I thought you might find this interesting, would be interested in your thoughts on how the two differ!

I know I'm getting ahead of myself here as I'm still running some preliminary GA tests with japonicus, and doubt I could figure out how to implement this git I came across, but maybe you could understand it better and make use of it, maybe even blend them together, assuming convnetjs differs from deap's methodology.

Thanks!

Gab0 commented 6 years ago

Hi @BradT7, thats a nice material that I'm reading on the following hours. Genetic Algorithm and Neural Networks are both methods of da famous "machine learning", both inspired by biologic life, but much different on purpose and workings. GA simulates biological evolution, and we use it here to evolve the parameters of a gekko strat, to find a good-maybe-best set of settings.

Neural networks simulates the workings of neurons in the brain, widely used on classification of input data. In our context of crytocoin trading botz it would be inside a strategy, maybe taking candles as input and spitting 'buy', 'sell' or 'do nothing'. A strat that includes a Neural Network (could be easily implemented with the lib ConvnetJS) sadly would be hard on CPU time, so probably not suitable for use with japonicus. NN's need to be trained prior to use and that would should be sufficient for pr0fit.

In my opinion NNs could be good as cryptobot brains but implementation will not be a simple one. I don't believe on current NN strats on github, but I'm checking those you posted.

Keep testing some strats, if you find some good, please tell x] .. Cheers!

Gab0 commented 6 years ago

Just tested Gekko-bot-resources/bestone-tulind strategy. Seems to never trade.

BradT7 commented 6 years ago

Thanks for the great feedback Gab, that does make more sense now to me! That's awesome you got it running! Too bad the results seem poor at the moment. I'll keep plugging away at these different GA's. Just a lot of "hurry up and wait!" going on now :p I'll keep you posted!

Gab0 commented 6 years ago

Thanks. Note that most of strats from github have settings hardcoded on them. That way does not work with japonicus. Indicators inside strategies must get their settings from strategy.settings AKA this.settings on the strategy init function. And then feed those settings ranges on configStrategies.py from japonicus.

Gab0 commented 6 years ago

@stfnschfr Can you explain the general workings of your ConvNet strat? And, for example, on that 11-02-17 to 01-30-18 run, you trained the network for that specific time period, or the network was previously trained on a larger dataset? Thanks. I have tried to write some NN strategies without practical results.