PLANNER TAP Compiler SCHEDULER

ion for Real-time Intelligent Control: Extended Abstract Robert P. Goldman, David J. Musliner, Mark S. Boddy, Kurt D. Krebsbach Automated Reasoning Group Honeywell Technology Center 3660 Technology Drive Minneapolis, MN 55418 fgoldman,musliner,boddy,[email protected] Introduction In this paper we discuss two abstraction techniques we use in CIRCA planning (Musliner, Durfee, & Shin 1993; 1995). CIRCA agents construct and execute plans for controlling real-time systems that interact with a dynamic environment including uncontrolled, exogenous events. In these environments, catastrophic failure is possible if a timely control action is not taken in certain situations. Control plans for these environments must provide guarantees that failures will not occur. Therefore, the CIRCA planning problem is one of generating a timed, discrete event controller (Ostro & Wonham 1990) that attempts to achieve goals while delivering performance guarantees. In this generation process we abstract temporal information, using a special-purpose temporal prover to summarize information about the latency of various events. We also use a technique we call Dynamic Abstraction Planning (Goldman et al. 1997) to minimize the (nontemporal) feature space of the controllers. In this position paper we present the two abstraction methods used in CIRCA state-space planning. We start by reviewing the CIRCA architecture, which couples a deliberative planning component with a scheduler and a real-time executive. We then present the state-space planning problem, which is the responsibility of CIRCA's AI Subsystem. We then discuss temporal abstraction in state-space planning, and then present Dynamic Abstraction Planning (DAP), a planning technique that dynamically and locally abstracts the feature space of the controller NFA. We relate these techniques to methods used in AI planning, Model (automaton) Minimization and Markov Decision Processes. CIRCA CIRCA is designed to support both hard real-time response guarantees and unrestricted AI methods that can guide those real-time responses. Figure 1 illustrates the architecture, in which an AI subsystem (AIS) reasons about high-level problems that require its powerful but potentially unbounded planning methods, while a separate real-time subsystem (RTS) reactively executes the AIS-generated plans and enforces guaranteed response times. The AIS and Scheduler modules cooperate to develop executable reaction plans that sensor data Environment