MC-Question 8. Consider the multi-armed bandit problem with 2 arms and adversarial losses (or equivalently adversarial...

  

Transcribed Image Text:

MC-Question 8. Consider the multi-armed bandit problem with 2 arms and adversarial losses (or equivalently adversarial rewards). We would like to use the Thompson sampling algorithm for this setting. What do you think about the normalized regret of this algorithm (RT/T)? (a) RT/T will converge to zero as Thomson sampling is a randomized al- gorithm. (b) RT/T will not converge to zero if the losses are chosen carefully because Thompson sampling is designed for stochastic rewards. (c) RT/T will converge to zero as Thompson sapling has lower regret than EXP3. (d) RT/T will not converge to zero as Thompson sampling is a deterministic algorithm and we are considering adversarial losses. MC-Question 8. Consider the multi-armed bandit problem with 2 arms and adversarial losses (or equivalently adversarial rewards). We would like to use the Thompson sampling algorithm for this setting. What do you think about the normalized regret of this algorithm (RT/T)? (a) RT/T will converge to zero as Thomson sampling is a randomized al- gorithm. (b) RT/T will not converge to zero if the losses are chosen carefully because Thompson sampling is designed for stochastic rewards. (c) RT/T will converge to zero as Thompson sapling has lower regret than EXP3. (d) RT/T will not converge to zero as Thompson sampling is a deterministic algorithm and we are considering adversarial losses.

Answer rating: 100% (QA)

a The normalized regret will converge to zero as T since Thomson sampling is a randomized algorithm why this is the case The reason is that for any gi... View the full answer