Hands-On Q-Learning with Python

Adding an updated alpha value

You will recall that alpha ranges from 0 to 1. An alpha value of 0 will have an agent learn nothing, and an alpha value of 1 will have it learn policies completely specific to a deterministic environment.

It can sometimes be to our advantage to decay alpha as a task progresses, because there might be less that we need to learn about it as we progress. A high alpha value can, in some situations, lead to overfitting, or learning overly specific lessons about an environment, and decaying alpha can help to prevent that.

Here's our code from the previous section, updated with a new alpha value:

Q = np.zeros([env.observation_space.n, env.action_space.n])gamma = 0.1alpha = 0.1state = env.reset()count = 0reward = 0 ...