The article shows a useful neural network model for changing direction depending on an input. However, I'm not sure what this has to do with deep learning as there is only one hidden layer and even that one is only fully connected to output. Edit: also, why evolve weights when you can backpropagate?
Even random search (in the weight space) can outperform gradient-based deep reinforcement learning algorithms on some Atari games. Genetic algorithms are quite competitive. Paper: https://arxiv.org/abs/1712.06567v1
This is neat, if you want to extend it there's a version of genetic algorithm for neural networks that tries out different connection settings and hidden layer sizes -- it's called NEAT.
Correct me if I’m wrong, but doesn’t inheriting by averaging the parents’ matrices only work in single layer perceptrons?
Anyway, nice exercise! :)
Cool, I think it's Matt Stonie on the right in the competitive eating GIF.