[FIXED] Tensorflow with genetic algorithms

Issue

I need to use a deep neural network with a genetic algorithm where I will start all the weights with random values... the net won't "learn" at least not by training but by natural selection... etc. I have used Tensorflow for training data long time ago, so this might sound stupid but should I use Tensorflow for that kind of project too?

Solution

Simple answer -- you shouldn't do it -- But if you really want to do it, I think it's possible...

TensorFlow is focused on building tensor computational graphs which can calculate tensors in parallel (which decreases computational time), but using a genetic algorithm is mostly base with crossover algorithm mutation which is much more expensive to compute than multiplying or adding tensors. Moreover, neural networks tend to have millions of parameters, which makes it much more difficult if you want to load all the parameters and create a mating pool -- which will take FOREVER just to compute the next generation.

Answered By - Ronakrit W.