Hands-On Neuroevolution with Python

The genome encoding scheme

The deep RL neural network that we use as the controller of the game agent has about 4 million trainable parameters. Each trainable parameter is the weight of a connection between two nodes of the neural network. Traditionally, training neural networks is about finding the appropriate values of all the connection weights, allowing the neural network to approximate a function that describes the specifics of the modeled process.

The conventional way to estimate these trainable parameters is to use some form of error backpropagation based on the gradient descent of the loss value, which is very computationally intensive. On the other hand, the neuroevolution algorithm allows us to train ANNs using a nature-inspired ...