Hands-On Genetic Algorithms with Python

Controlling the CartPole with a neural network

To successfully carry out the CartPole challenge, we would like to dynamically respond to the changes in the environment. For example, when the pole starts leaning in one direction, we should probably move the cart in that direction, but possibly stop pushing when it starts to stabilize. So, the reinforcement learning task here can be thought of as teaching a controller to balance the pole by mapping the four available inputs—cart position, cart velocity, pole angle, and pole velocity—into the appropriate action at each time step. How can we implement such mapping?

One good way to implement this mapping is by using a neural network. As we saw in Chapter 9, Architecture Optimization of Deep Learning ...