Abstract
Presented at the 1982 National Conference on Artificial Intelligence (AAAI-82), this research introduces a proactive expert system designed for the Vax/VMS operating system. Unlike traditional reactive help systems that trigger only upon user error or explicit request, this system monitors user command streams to identify "correct but inefficient" workflows. By utilizing an expectation-based parser and a catalog of "bad plans," the system volunteers unsolicited advice to guide users toward more elegant and efficient command sequences.
Introduction
In the early 1980s, human-computer interaction was largely governed by reactive support structures. Help systems typically intervened in two scenarios: when a user committed a syntax error or when a user explicitly invoked a HELP command. This paradigm, however, ignored a significant demographic of "intermediate" users—those who could navigate the system without making errors but did so through laborious or suboptimal methods.
For instance, a user might manually print several files individually using $PRINT FILE1.MEM`, `$PRINT FILE2.MEM, and so on, unaware that the system supports wildcard operations like $PRINT *.MEM`. Similarly, a user might execute a `$COPY followed by a $DELETE` to move a file, rather than using the more efficient `$RENAME command. This paper describes a system that bridges the gap between basic error recovery and human expert consultation by "volunteering" advice when it detects these inefficient patterns.
System Architecture and Methodology
Implemented in Franz Lisp, the system acts as an intelligent intermediary between the user and the VMS operating system. The architecture relies on three core pillars:
1. The Catalog of "Bad Plans"
The researchers curated a library of common inefficient command sequences frequently observed in novice and intermediate users. These "bad plans" serve as the templates against which live user behavior is compared.
2. Expectation-Based Parsing
Drawing inspiration from the work of Riesbeck and Schank, the system employs an expectation-based parser to monitor the command stream. When a "head" command—the initial action in a potential sequence—is detected, the system activates a corresponding node in a KL-One-like knowledge network. This creates an "expectation" that subsequent commands may complete a known inefficient plan.
3. Dynamic Knowledge Networks
As the user enters further commands, the system dynamically updates the state of the knowledge network. It must navigate three specific technical hurdles to remain accurate:
- Non-contiguity: Inefficient sequences are rarely perfectly sequential; users often intersperse unrelated commands (like checking the time or directory) between the steps of a plan.
- Ambiguity: A single command might serve as the starting point for multiple different plans, requiring the system to maintain multiple hypotheses until more data is available.
- Extensional Knowledge: To provide meaningful advice, the system must understand the state of the file system. For example, it must verify if a wildcard expansion would actually target the same files the user is currently processing manually.
User Modeling and Proactivity
A central challenge in designing a system that volunteers advice is avoiding "user annoyance." To mitigate this, the authors implemented a rudimentary user profile. The system tracks the history of commands issued by a user; if a user demonstrates knowledge of an efficient command (e.g., using RENAME correctly), the system notes this in the profile and ceases to offer advice on that specific topic. This ensures the system remains a helpful coach rather than a repetitive nuisance.
Results and Prototype Performance
The authors successfully developed a functional prototype for the Digital Command Language (DCL). The system was capable of recognizing five complete high-level goals and several "short-cut" strategies, such as the use of wildcards or combined command targets.
Crucially, the system generated context-aware advice. Rather than simply pointing the user to a general manual page, the system explained why the user's recent actions were inefficient and provided the specific syntax required to achieve their goal more effectively in that exact context.
Limitations and Future Outlook
While pioneering, the system faced several limitations typical of early AI research:
- Scalability: The catalog of plans was hand-coded, a labor-intensive process that limited the system's scope to a small fraction of possible OS interactions.
- Error Handling: The system focused exclusively on inefficient but correct commands, ignoring commands that resulted in syntax errors. This meant it could not assist users who were genuinely "stuck."
- User Profile Depth: While the concept of a dynamic user profile was introduced, the authors noted that the prototype made only marginal use of this data.
Conclusion
This 1982 study represents an early and significant step toward the "intelligent assistants" and "proactive coaching" features found in modern software. By treating a sequence of system commands as a "sentence" in a language of actions, the authors demonstrated that natural language processing techniques could be effectively applied to goal recognition and user support. The system's move from reactive to proactive help prefigures modern developments in IDE linting and automated workflow optimization.
Summary updated