Learning for non-stationary Dirichlet processes

摘要

The Dirichlet process prior (DPP) is used to model an unknown probability distribution, F. This eliminates the need for parametric model assumptions, providing robustness in problems where there is significant model uncertainty. Two important parametric techniques for learning are extended to this non-parametric context for the first time. These are (ⅰ) sequential stopping, which proposes an optimal stopping time for online learning of F using i.i.d. sampling; and (ⅱ) stabilized forgetting, which updates the DPP in response to changes in F, but without the need for a formal transition model. In each case, a practical and highly tractable algorithm is revealed, and simulation studies are reported.