Open-Universe Theory for Bayesian Inference, Decision, and Sensing (OUTBIDS).

摘要

This report describes work in Phase I of the OUTBIDS project under the DARPA MSEE program. The goal of OUTBIDS was to develop the theoretical and technological foundations for sensor data interpretation as a form of probabilistic inference. Achieving this goal requires a representation formalism for probability models of sufficient expressive power to handle the complexity of real-world sensor data. The problem involves two primary sources of difficulty: first, the underlying world generating the data typically contains many initially unknown objects interacting over time in complex ways; second, the mapping from objects and behaviors to sensor data is itself (as in the case of visual perception, for example) very complex. The core of the project is the BLOG (Bayesian LOGic) language, which combines probabilistic semantics with the expressive power of first-order logic. Unlike other attempts to combine probability and logic, BLOG supports open-universe models, which allow for uncertainty over the existence and identity of objects. We believe this is a prerequisite for any probabilistic approach to general perception. The team made substantial progress on developing and refining the BLOG language by writing a broad range of models, including two models for computer vision tasks (adaptive video background subtraction for object tracking and a simple form of 3D object recognition and scene reconstruction). We also made good progress towards an efficient inference engine, including new algorithms and initial work on compiler and parallelization technology for BLOG inference. We showed that BLOG could be extended to open-universe decision models and developed algorithms for sensor planning on this basis. We also developed a new theoretical framework for utility-directed inference and proved several foundational theorems.