Midterm Exam 1 ISYE 4301 | Answered Latest 2025-26 Total points: 95 points + 15 bonus. - GT. - ISYE 4301 (ISYE4301) - Stuvia US
- The term inside the square root will be negative, which makes the calculation of ε∗ infeasible. False,
there is no way to make that term negative. (1 point) Q5 (10 points) Consider an online learning framework with 5 experts, with decisions {0, 1}, and T = 500 time periods. We would like to use the deterministic weighted majority algorithm and use this side information in fine-tuning the learning parameter. Answer the following two questions:
- If we do not know the number of mistakes made by the same expert, what ε should we use as seen in
class? Please provide a numerical value.
- Suppose we have magical powers and, before running the online learning algorithm, we know that the
best expert will make mistakes M ∗ = T /2 time periods. Is it possible to guarantee always sublinear regret using WMA?
- What if we know that M ∗ = 0?
Answer:
q q
- In this case we use εalt = ln n
T ln 5 ∼ 0.0567. (2 point) = 500 √
- Now, the best regret you can guarantee is M ∗ + O( T ln), which is at least T /2 and is linear, not
sub-linear. (2 points) √ √
- Yes, because the regret of M ∗ + O( T ln) becomes O( T ln). (1 point)
Q6, Q7, and Q8 will use the following weather prediction setup that you should read carefully In Q6, Q7, and Q8, we will explore the performance of various online algorithms on the problem of predicting Atlanta’s weather (sunny or rainy) at 7 pm every day. We will do this prediction on the advice of 3 weather forecasters: AccuWeather, weather.com and NOAA. The following table shows the forecasts of the 3 experts