Hands-On Reinforcement Learning for Games

Pre-training the agent

We have already covered a number of ways to manage training performance often caused by low rewards or rewards sparsity. This covered using a technique called behavioural cloning, whereby a human demonstrates a set of actions leading to a reward and those actions are then fed back into the agent as a pre-trained policy. The winning implementation here used a combination of behavioural cloning with pre-trained image classification.

We will continue from where we left off in the last exercise and learn what steps we need to perform in order to pre-train a classifier first:

1. Firstly, we need to capture images of the environment in order to pre-train a classifier. This requires you to run the record.py script located ...