This 1984 paper from IBM’s T. J. Watson Research Center introduces YES/MVS (Yorktown Expert System for MVS operators), a pioneering attempt to move artificial intelligence out of the "consultation" era and into the world of autonomous, real-time systems control.
Moving Beyond Consultation
In the early 1980s, large data processing installations were becoming too complex for human operators to manage alone. A single mainframe might handle thousands of status variables and peak message rates exceeding 100 per minute. Existing expert systems like MYCIN were "consultation-oriented," meaning they sat idle until a human provided data.
The authors identified a gap: computer operations require a system that is data-driven, continuous, and non-monotonic. In a dynamic environment, a problem might disappear while the expert system is still calculating a solution. YES/MVS was built to handle this world where facts change in real-time.
Distributed Architecture and Real-Time Methods
To ensure the expert system didn't crash alongside the system it was monitoring, YES/MVS was partitioned into three virtual machines: an Expert machine, a Communications Control Facility (MCCF), and an Operator Interface.
The system was built on OPS5, but required significant modifications to meet real-time requirements:
- Timed Productions (
TIMED-MAKE): Added a primitive to schedule the creation of working memory elements at specific future times. - Remote Communication (
REMOTE-MAKE): Allowed rules to trigger actions or data updates across different virtual machines. - Priority-Based Conflict Resolution: Introduced an explicit priority mode to ensure critical tasks (like hardware error handling) took precedence over routine ones.
- Continuous Inference: Implemented a
WAITfunction so the inference engine would "sleep" until new data arrived rather than terminating.
Knowledge Engineering and Real-Time Data
As a classic expert system, YES/MVS was not "trained" on a dataset in the modern sense. Instead, the "data" was a knowledge base of over 500 OPS5 rules hand-crafted through interviews with operations staff, systems programmers, and MVS designers. During operation, the live stream of system messages and status variables from the subject MVS machine served as the input.
Key Operational Insights
- The "Garbage Collection" of Facts: In a continuous system, working memory fills with old data. The authors implemented rules to explicitly remove stale facts to prevent "hallucinated" rule firing.
- Avoiding the Human Bottleneck: The environment was too fast for a human to interpret every input; YES/MVS was designed to act directly.
- Managing Uncertainty: Acknowledging that a 100% accurate model of a mainframe is impossible, they used timestamps and validity flags to manage the "currentness" of data.
- Race Conditions: They identified unique race conditions where a human might solve a problem just as the AI issues a command, requiring the system to be sensitive to acknowledgement messages.
Results and Deployment
At the time of publication, the prototype was routinely scheduling large batch jobs and alerting operators to network link problems. By compiling the action part of rules into LISP macros, they achieved significant execution speedups over standard interpreted OPS5.
The system successfully automated tasks across six subdomains, including JES (Job Entry System) queue management, hardware error recovery, and performance monitoring.
Limitations and Future Work
The authors noted that MVS is too complex to simulate accurately, making testing difficult; they often had to "sabotage" a real mainframe during off-peak hours to verify reactions. The biggest hurdle was not the AI logic, but the "plumbing"—integrating the core expert system with messy, real-time communication protocols. Future plans included adding a learning component for batch scheduling and expanding the system to handle initial boot-up (IPL) and shutdown.
Summary updated