This week on 9/22 we will have Aaron Roth speaking at the theory seminar. Details included below:
Title: High Dimensional Calibration for Rational Decision Making
Abstract: It has been known since the work of Foster and Vohra that decision makers that best respond to -fully calibrated- forecasts take actions that guarantee them no swap regret. Unfortunately the computational and data complexity of full distributional calibration is exponential in the dimension of the problem.
Using game theoretic techniques, we show how efficiently make high dimensional calibrated forecasts tailored to particular decision makers, sufficient to give them no swap regret guarantees, and then extend these to guarantees for decision makers in large action settings. We mention applications to mechanism design and to uncertainty quantification in machine learning.
Joint work with Georgy Noarov, Ramya Ramalingam, and Stephan Xie