A0933
Title: Adaptive partial monitoring in non-stationary environments
Authors: Henry Reeve - Nanjing University (China) [presenting]
Abstract: A framework is introduced for sequential decision making in a non-stationary stochastic environment in which the outcome distribution may change over time and the rewards may not be directly observable. The regret of a policy contrasts performance with the expected reward of a dynamic oracle capable of selecting an optimal sequence of actions for the non-stationary stochastic environment. An algorithm is provided which leverages e-processes to provably adapt to distributional changes in settings where the reward attained from a given action is not directly observed. It is found that the optimal regret depends upon a fascinating interplay between the level of observability, the noise level, the complexity of the action space, and the degree of non-stationarity.
