PyTorch 1.x Reinforcement Learning Cookbook

How it works...

In Step 2, the epsilon-greedy policy takes in a parameter, ε, with a value from 0 to 1, and |A|, the number of possible actions. Each action is taken with a probability of ε/|A|, and the action with the highest state-action value is chosen with a probability of 1-ε+ε/|A|.

In Step 3, we perform Q-learning in the following tasks:

• We initialize the Q-table with all zeros.
• In each episode, we let the agent follow the epsilon-greedy policy to choose what action to take. And we update the Q function for each step.
• We run n_episode episodes.
• We obtain the optimal policy based on the optimal Q function.

In Step 6, again, up = 0, right = 1, down = 2, and left = 3; thus, following the optimal policy, the agent starts in state 36, ...