Building expert systems
The Challenges of Real-Time AI
Computer
| Hi-index | 0.01 |
Development of an on-board real-time expert system forcontrolling engineering systems requires tradeoffs between AI andreal-time approaches. This article presents an alternative thatlies between embedding AI in real time and embedding real time inAI.Imagine that you want to fit a car with a computer thatdiagnoses the vehicle's capacity to safely execute a maneuver;control the plowing parameters on board your tractor according toenvironmental conditions; collect the data for, and interpret theresults of, a series of nondestructive eddy-current tests on thesteam generator tubes of your country's nuclear power stations; orinclude a resuscitation protocol controller in the morass ofinstruments connected to a neonatal incubator.These tasks call for a high level of expertise in performingcontrol functions, and expert system technology is a good candidatefor providing it. However, they also call for computing capacitiesthat seem incompatible at first glance with AI algorithmimplementation constraints. For example, hardware support must bepowerful and small. The implementation must be able to handlereal-time constraints as well as temporal reasoning. Thedevelopment of the expert system for this sort of engineeringcontrol must guarantee the knowledge base's reliability butnevertheless remain flexible enough for knowledge updating.Building an expert system means choosing among alternativesconcerning how the system searches for solutions. The choicedepends on the target problem. Frederick Hayes-Roth, Donald A.Waterman, and Douglas B. Lenat have classified different cases andtheir solutions. When applied to engineering systems, developmentof expert systems for control incurs additional constraints.Consequently, choosing one of these solutions involves tradeoffs.The constraints become more restrictive when the expert system isworking in a real-time environment.David J. Musliner and his colleagues have classified three basicapproaches to real-time AI: embedding AI in real time, embeddingreal time in AI, and cooperating between real time and AI. Eachapproach satisfies real-time and AI requirements to a varyingextent. Assigning priority to real-time requirements might reducethe variance of the AI tasks, or might cast the AI tasks asincremental, interruptible algorithms. Assigning priority to AIrequirements--that is, without restricting the complexity of AImethods--restricts real-time tasks to rare missions. Such astrategy also presents the possible drawback that the tasks will beinvoked by high-variance, unpredictable AI techniques. Cooperatingbetween real time and AI suggests that each part's goals might bedifferent. In this strategy, an AI task's role generally is toschedule, plan, or control the real-time subsystem tasks. Althoughsometimes flexible and powerful, this approach might lead toinefficient or unfeasible systems design.These two classifications of expert systems search methods andreal-time AI approaches have helped my colleagues and me to build aframework to provide a solution to a given problem of on-boardreal-time expert control. Our approach tries to exploit thebenefits of the three approaches to real-time AI, while avoidingtheir drawbacks. The structuration of the knowledge base helpsreduce the variance of AI algorithms. Multiple knowledge-basestructures enable a level of representation powerful enough torepresent and manage complex knowledge models. Embedding the AIsystem in a real-time operating system kernel lets us use an AIstrategy to manage planning, scheduling, and control of real-timetasks.We used this framework to develop an expert system developmentenvironment for maintaining and updating information about anevolving system as its state changes because of abnormal eventssuch as faults, or as a consequence of external actions taken uponthe system. We've successfully applied it in several engineeringdomains such as process control, medical monitoring, and on-boarddiagnosis. In this article I'll show how we apply it to a systemthat controls automobile maneuvering.