Neural Network Creatures
This is a demonstration of some creatures that have learned to "stay high" and "move right." No gradient descent or complicated learning algorithms were applied to train the weights of the neural network that controls them. A simple hill climbing/genetic algorithm is used to find a good set of weights that get the highest cumulative reward. The code can be found here: https://github.com/JonComo/stumbly
I was surprised at how effective random weights can be in these simulations so I thought I'd share! Future steps include learning initial parameters by trial and error (done) then optimize those parameters using reinforcement learning/policy improvement. I'd appreciate any ideas on how to improve all of this. Have fun making your creatures and teaching them to walk around!