Making Robust Decision from Data
Anil Aswani
University of California at Berkeley
Wednesday, May 9, 2018
4:15 - 5:15 PM
Location: 160-323
Abstract:
Though machine learning has found success in decision-making contexts, these
methods are fragile to model mismatch and malicious interference. This is a
major impediment to the deployment of automated decision-making in
safety-critical systems like those found in healthcare or physical
infrastructure. This talk describes three methods we have developed for robust
decision-making in different scenarios. The first is a framework for combining
robust control with machine learning, and applications to energy efficiency and
robotics are highlighted. The second is algorithms to solve inverse
optimization (and inverse reinforcement learning) with noisy data. This problem
arises when estimating utility functions or modeling human-automation systems,
and we show it is NP-hard and that existing approaches are statistically
inconsistent. We develop a polynomial time algorithm that is asymptotically
optimal as more data is collected. Then we discuss applications of our inverse
optimization approach to a clinical trial on personalized goal-setting through
smartphone apps to increase physical activity, and to studying an incentive
design problem in the Medicare Shared Savings Program where we show that an
investment sharing plan could potentially save Medicare an additional $85
million per year. The third is an approach for bandit models where repeated
application of an action causes habituation and a decrease of that action's
rewards, while refraining from an action causes recovery and an increase of
that action's awards. Though such problems are PSPACE-complete, we define a
class of models called ROGUE bandits for which we can construct policies that
achieve logarithmic regret. We describe an application of ROGUE bandits to a
personalized healthcare problem of choosing an optimal sequence of daily
messages to encourage an individual to increase their physical activity.