Hands-On Reinforcement Learning for Games

Tuning hyperparameters

The collection of training parameters that we investigated at the top of the last code example are called hyperparameters, so named to differentiate them from normal parameters or weights we use in deep learning. We have yet to look at deep learning in detail, but it is important to understand why they are called hyperparameters. Previously, we played around with the concept of a learning rate and discount factor but now we need to formalize them and understand their effect across methods and environments.

In our last example, both the learning rate (alpha) and discount factor (gamma) were set to .5. What we need to understand is what effect these parameters have on training. Let's open up sample code Chapter_4_2.py ...