|
Thursday, May 5, 2005 |
|
|
|
|
Sense and Respond Systems |
|
|
This talk presents theory, prototypes and applications of distributed sense and respond systems consisting of networks of sensors, event processing agents and responders. A sense and respond system is specified by a set of guard-action rules where a guard is a predicate on the history of system states, and an action is an operation of a responder. Sense and respond systems are useful in applications such as healthcare, homeland security and finance. Sense and respond systems are distributed systems in which processes share a clock and have a common concept of time. By contrast, processes in traditional distributed systems do not share variables and have no concept of common time; algorithms in the literature for taking “global snapshots” and theories of how processes “learn” depend in fundamental ways on the absence of shared clocks. In this talk I’ll revisit these concepts in the context of systems with clocks though clocks may drift. Another aspect of the theory deals with incremental computations – how processes update their state when each new event arrives, where the complexity of the update is independent of the length of the history. I will look at incremental algorithms for problems in areas such as statistics. One way of developing incremental algorithms is in terms of delta-dataflow networks in which the data flowing along edges of the network deal with deltas – changes in state. In the prototype development part of the talk I’ll describe a system called the “Event Web” built as a thin layer on top of Web Services. I’ll describe experiments executing delta-dataflow networks on parallel machines, an event-processing notation for specifying systems, and experience with specifying, deploying and maintaining sense and respond systems. I’ll discuss applications dealing with credit card networks, health care, and security.
|
|