Epsilon-greedy action selection

The problem, from first principles

A bandit agent must keep trying the arm that looks best (exploitation) while occasionally trying others in case its estimates are wrong (exploration). Epsilon-greedy is the simplest rule: explore with a small fixed probability, otherwise exploit.

From theory to code

Implement epsilon_greedy_action(q_values, epsilon, rng). With probability epsilon return a uniformly random arm; otherwise return the arm with the highest estimated value.

Constraints

• q_values is a one-dimensional array of current value estimates, one per arm.
• rng is a numpy.random.Generator; draw the explore/exploit decision from it with rng.random(), and the random arm with rng.integers(...).
• On exploiting, break ties toward the lowest arm index.
• Return a Python int.

Hints

Open one at a time. Each gives away a little more than the last.