Reinforcement Training (i think)!!!! Much easier to use and better for playing games, because you just have to give it a loss (if greater score is better then make loss 1/score or something), and train it! Press space twice to start In this example, it trains a neural network with 2 inputs (mouse x and y), 2 outputs (where to go), 2 hidden layers, and 3 neurons per hidden layer. I think I could have made this a lot better if I removed ReLU on the neurons, but I already removed sigmoid on the output, and most stuff needs both. And again, this does not use backpropagation or any of the complicated stuff because I don't understand it, but instead uses random hill climbing, which works fine for small networks.
