ABSTRACT
This chapter reviews several existing algorithms for online learning with expert advice are equivalent to performing marginal inference on an appropriately defined Markov random field (MRF). It discusses the spatiotemporal setting where the best expert may change over time and space. The fixed-share algorithm of models the setting where the best expert may vary over time, fixed-share can be represented as a chain MRF where there is a separate latent variable representing the identity of the best expert at each time iteration. These latent variables are chained together in the MRF by links that represent the transition dynamics of the best expert over time. The chapter aims to extend the paradigm from the single temporal MRF chain to a lattice over space and time, developing a framework that allows the best expert to "switch" over both space and time. It describes empirical results comparing MRF-based method and the more efficient Neighborhood-augmented Tracking Climate Models (NTCM) algorithm to several existing methods.
