Hands-On Q-Learning with Python

Alpha

The goal of the learning parameter alpha is to smooth out the agent's update curve, protecting it against extreme outliers.

In machine learning terms, a well-chosen alpha value is intended to prevent an RL model from overfitting. In other words, it keeps the model from learning too much from random noise and outliers and from making overly-specific generalizations about the specific data it has already seen.

We choose an alpha parameter for this problem based on the level of generalization we should expect based on the experiences it has had of its environment so far.