Mastering Machine Learning Algorithms

Q-learning in the checkerboard environment

Let's repeat the previous experiment with the Q-learning algorithm. As all of the constants are the same (as well as the choice of a ε-greedy policy and the starting point set to (0, 0)), we can directly define the function that implements the training for a single episode:

import numpy as npdef q_step(epsilon):    e = 0        i = x_start    j = y_start        while e < max_steps:        e += 1                action = select_action(epsilon, i, j)            if action == 0:            if i == 0:                x = 0            else:                x = i - 1            y = j                        elif action == 1:            if j == width - 1:                y = width - 1            else:                y = j + 1            x = i                        elif action == 2:            if i == height - 1:                x = height - 1            else:                x = i + 1            y = j                        else:            if j == 0:                y = 0            else:                y = j - 1            x = i                        reward = tunnel_rewards[x, y]         if is_final(x, ...