PyTorch 1.x Reinforcement Learning Cookbook

Solving multi-armed bandit problems with the epsilon-greedy policy

Instead of exploring solely with random policy, we can do better with a combination of exploration and exploitation. Here comes the well-known epsilon-greedy policy.

Epsilon-greedy for multi-armed bandits exploits the best action the majority of the time and also keeps exploring different actions from time to time. Given a parameter, ε, with a value from 0 to 1, the probabilities of performing exploration and exploitation are ε and 1 - ε, respectively:

• Epsilon: Each action is taken with a probability calculated as follows:

Here, |A| is the number of possible actions.

• Greedy ...