Seminar
Parent Program: | -- |
---|---|
Location: | Calvin Lab Auditorium; Virtual |
Registration required to attend.
Many decision-making settings exhibit a trade-off between obtaining extra information to improve the outcome of the process and the cost of obtaining such information. The Pandora’s box problem of Weitzman (1979) is a classical model formalizing this trade-off. The algorithm is presented with a number of boxes containing unknown (stochastic) rewards, and it can select and keep one of the rewards. Boxes can be opened at a fixed cost to reveal their rewards. The problem is to determine which boxes to open, in what order, and which reward to choose, so as to maximize the reward obtained minus the costs paid. Weitzman presented a simple and elegant index-based policy for solving this problem under certain assumptions. In this talk, Shuchi Chawla will examine settings where Weitzman’s assumptions do not hold and his approach fails. She will focus, in particular, on settings where the rewards in boxes are correlated, and will present different ways of formalizing and attacking the problem. En route, she will examine issues that often arise in online decision-making, such as the trade-off between learning and optimization and benchmark choice.
No Notes/Supplements Uploaded No Video Files Uploaded