A satisficing cycle for real-time reasoning in intelligent agents

摘要

An intelligent agent must interact with dynamic entities in real time. Because it cannot predict all events that will occur, it must notice and respond to important unanticipated events. However, insuring execution of the best possible operation at each point in time conflicts with meeting deadlines, especially as event rate and number of known operations increase. The problem is: How can an agent with limited resources execute high-quality operations in constant time, despite increases in event rate and number of known operations? To support this capability in a blackboard architecture, we replace the conventional reasoning cycle with a satisficing cycle. To bound cycle time, it interrupts triggering to execute the best operation available when either it finds a “good enough” operation or a deadline occurs. To insure the availability of high-quality operations when interrupts occur, it uses dynamic control plans to order its consideration of events and known reasoning operations best-first during triggering. We have made additional extensions to the blackboard architecture to support real-time reasoning: a perceptual preprocessor to interpret, filter, and prioritize sensed events; a limited-capacity event buffer to hold recetn perceptual and cognitive events in priority order; and explicit representations of operation and event hierarchies to support ordered retrieval during triggering. In this paper we describe the extended blackboard architecture, but focus on the satisficing cycle and its support for real-time reasoning. We present experimental results that confirm the predicted performance of the satisficing cycle.