On-Line Learning with Imperfect Monitoring

We study on-line play of repeated matrix games in which the observations of past actions of the other player and the obtained reward are partial and stochastic. We define the Partial Observation Bayes Envelope (POBE) as the best reward against the worst-case stationary strategy of the opponent that agrees with past observations. Our goal is to have the (unobserved) average reward above the POBE. For the case where the observations (but not necessarily the rewards) depend on the opponent play alone, an algorithm for attaining the POBE is derived. This algorithm is based on an application of approachability theory combined with a worst-case view over the unobserved rewards. We also suggest a simplified solution concept for general signaling structure. This concept may fall short of the POBE.

• Type: Conference
• Year: 2003
• Where: COLT
• Authors: Shie Mannor, Nahum Shimkin
• Keywords: Average Reward, COLT 2003, Partial Observation Bayes, Worst-case Stationary Strategy