Many months have passed since the onset of COVID-19, and 5 medical laboratories have created and trialed 5 vaccines with no signicant adverse side eects after careful clinical trials. From the trials it is believed that some vaccines are more eective at preventing contraction of COVID-19 for some patients, but no one knows which. Doctors across the world must now determine how to select the right vaccine for any given patient, as patients stream into general practice clinics. How might you as Chief Health Officer use machine learning to help fight COVID-19 by recommending appropriate vaccines?

(a) Formulate the vaccine selection task as a non-contextual multi-armed bandit problem. Explain what your arms are, what rounds correspond to, and what are your rewards. [6 marks]

(b) Suppose in earlier clinical trials various measurements were taken of patients prior to test vaccination. Given access to these measurements, and eectiveness of the vaccinations in trials, how might you decide whether contextual multi-armed bandits might oer an improvement over your non-contextual solution? [4 marks]

(c) Suppose 100 machine learning companies approach you with their MAB solutions, that have never been previously tested. What drawback is there to trying all of the 100 approaches out on real patients in the wild? [4 marks]

(d) Now suppose the prior clinical trails were completely randomised, such that for a patient that arrived, a uniformly random vaccine was given to them, and the result recorded. How could you use this data to evaluate all 100 MAB learners? [4 marks]