Multi-Armed Bandits Quiz

Answers and explanations appear as you take the quiz.

1. The central tradeoff in a multi-armed bandit problem is between:
   • Exploration and exploitation
   • Bias and variance
   • Precision and recall
   • Training and inference
2. In bandit problems, 'regret' refers to:
   • The memory used by the algorithm
   • The cumulative reward gap versus always playing the best arm
   • The number of arms available
   • The time to train a model
3. What does the epsilon-greedy strategy do?
   • Always picks the arm with the highest mean
   • Never explores any arm
   • Mostly exploits the best arm but explores randomly with probability epsilon
   • Pulls every arm an equal number of times forever
4. The Upper Confidence Bound (UCB) method follows the principle of:
   • Pessimism in the face of uncertainty
   • Ignoring uncertainty entirely
   • Random selection only
   • Optimism in the face of uncertainty
5. Thompson sampling chooses an arm by:
   • Sampling from each arm's posterior belief and playing the best sample
   • Always picking the arm pulled least
   • Picking arms in fixed round-robin order
   • Minimizing the training loss
6. How does a contextual bandit differ from a standard one?
   • It ignores rewards entirely
   • It uses features of the current situation to inform the choice
   • It has only one arm
   • It cannot learn over time
7. In UCB, which arm usually receives a larger exploration bonus?
   • An arm whose sample count is largest
   • An arm with no uncertainty by definition
   • An arm that has been sampled relatively few times
   • An arm with a negative reward only
8. What does logarithmic regret growth indicate over a long run?
   • The algorithm receives zero reward after each round
   • The best arm changes at every step
   • Exploration stops before the first observation
   • The cumulative gap from always choosing the best arm grows much slower than the number of rounds