Hands-On Q-Learning with Python

Epsilon-greedy selection

Our technique in implementing an epsilon-greedy selection is by now a familiar process:

• Choose an epsilon value
• Select a random value; if it is lower than epsilon, take a random action, and if it's higher, take the highest Q-valued action

We've begun the process of implementing an epsilon-greedy algorithm:

for view in range(total_views):    if random.random() < epsilon         #exploration option: choose a random action        ad = random.randrange(num_ads)    else:        #choose the best-performing ad so far   ...    reward = df.values[view, ad]    total = total + reward

In the generalized bandit example, epsilon-greedy uses an epsilon value to decide whether to choose the current Best arm (exploit) or a random arm (explore). The following chart ...