Context-based caching: Applications of expectations in sensor-based systems.

Expectations provide guidance for automatic perception. The research presented draws on the central thesis of the verification approach to information processing: prior expectations guide present processing. We demonstrate this approach through experiments in the robotic arena: assimilation of range-data for updating a physical model, and anticipation of ultrasonic sensor readings for driving a mobile platform in face of uncertainties. A particular problem with this approach, however, is that extensive knowledge about the domain and the task over-burdens the system. Thus, the verification approach might be limited to domains and models of a restricted size. Our solution, Context-based caching, provides a systematic method for removing this limitation by pre-fetching relevant knowledge. Context-based Caching (CbC) addresses the problem of large available knowledge bases out of which only a small, specific portion is relevant at any given time for a given scenario. Knowledge in CbC is pre-sorted and clustered into task-related meaningful entities (i.e. contexts) which are connected by relations defined as tests on the dynamic state of the system. The Context-based Caching subsystem continuously employs the state of the overall system for updating a local knowledge base, which contains the anticipated necessary information. Its application is demonstrated in a mobile system performing vision-aided pose correction by model-driven sensory anticipation. The novel idea behind CbC (the Knowledge Caching approach) is to exploit the locality of information--providing the right information at the right time.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/103138/1/9303813.pdfDescription of 9303813.pdf : Restricted to UM users only