A multi-agent analogical representation for physical objects (abstract)

In a multi-agent world, several agents act simultaneously, competitively or cooperatively. In many situations, an intelligent autonomous agent must interact with the other agents or the physical environment in real time. Because it cannot predict all the events that will occur in the physical environment or result from other agents reasoning, it must notice and control its responses to unanticipated events. However, insuring execution of the best possible operation conflicts with meeting deadlines, especially as the event rate and the number of known operations increase. Rather than engineer agents to meet deadlines under particular parameter values, we aim to build autonomous agents that control their reasoning so as to guarantee real-time performance despite increases in parameter values. We propose a satisficing algorithm. To control response time, it triggers only a limited number of operations and interrupts triggering to execute the best one available whenever it triggers a "good enough" operation or a deadline occurs. To insure that it can execute high-priority operations when interrupts occur, it uses dynamic control plans to trigger operations roughly "best-first." In this paper, we describe the satisficing algorithm, informally analyse the behavior of an agent under this algorithm, and present experimental results.