An Expert Systems Approach to Realtime, Active Management of a Target Resource

The application of expert systems techniques to process control domains represents a potential approach to managing the increasing complexity and dynamics which characterizes many process control environments. This thesis reports on one such application in a complex, multi-agent environment, with an eye toward generalization to other process control domains. The application concerns the automation of large computing system operation. The requirement for high availability, high performance, computing systems has created a demand for fast, consistent, expert quality response to operational problems, and effective, flexible automation of computer operations would satisfy this demand while improving the productivity of operations. However, like many process control environments, the computer operations environment is characterized by high complexity and frequent change, rendering it difficult to automate operations in traditional procedural software. These are among the characteristics which motivate an expert systems approach to automation. JESQ, the focus of this thesis, is a realtime expert system which continuously monitors the level of operating system queue space in a large computing system and takes corrective action as queue space diminishes. JESQ is one of several expert systems which comprise a system called Yorktown Expert System/MVS Manager (YES/MVS). YES/MVS automates many tasks in the domain of computer operations, and is among the first expert systems designed for continuous execution in realtime. The expert system is currently running at the IBM Thomas J. Watson Research Center, and has received a favorable response from operations staff. The thesis concentrates on several related issues. The requirements which distinguish continuous realtime expert systems that exert active control over their environments from more conventional session-oriented expert systems are identified, and strategies for meeting these requirements are described. An alternative methodology for managing large computing installations is presented. The problems of developing and testing a realtime expert system in an industrial environment are described