Abstract
GLIMPSE is an intelligent front-end for the GLIM statistics package that treats data analysis as a dynamic "adventure quest." It introduces a "suspending evaluator" that pauses logic when data is missing to suggest information-gathering actions, alongside a "suggested-answer" system that provides flexible guidance while respecting the user's ultimate authority.
Introduction
Statistical analysis is rarely a linear process; it is more akin to an adventure game where the "analyst-adventurer" discovers new facts that fundamentally change the state of the knowledge base. Traditional expert system front-ends typically assumed a static knowledge base—the system knew everything it needed to know at the start. GLIMPSE addresses the challenge of creating a system that can handle a shifting landscape of facts, provide multi-layered advice, and allow the user to act independently or reject the system's suggestions.
System Architecture and Methods
GLIMPSE is built using sigma-PROLOG and the APES (Augmented Prolog for Expert Systems) shell. It implements two primary novel features to facilitate dynamic interaction:
1. Suggested-Answer Facility
This is an extension of the "Query-the-User" model. When the system needs information (e.g., "Is this graph linear?"), it doesn't just wait for a value. It can offer a suggestion based on internal heuristics, such as curvature statistics. The user retains control, with the ability to accept the suggestion, reject it, or ask "WHY" to see the underlying logic.
2. Suspending Evaluator
In standard logic programming, if a goal cannot be proven because a fact is missing, the goal simply fails. GLIMPSE introduces a new state: suspension. Using meta-level assertions, the system identifies certain relations as "suspendable." If the conditions for these relations aren't met, the evaluator interrupts the process and suggests a specific action the user should take to find the missing data (e.g., "find the mean deviance of the model").
The Interaction Loop
The system distinguishes between help with the command language (syntactic help) and guidance on statistical strategy (semantic help). This allows for a flexible interaction model where the system can form its own conclusions while still allowing the user to intervene at any point.
Key Insights
- The "Adventure" Metaphor: Viewing data analysis as a quest where resources (time/money) are depleted and "blind alleys" require backtracking provides a more realistic framework for AI-assisted science than a simple decision tree.
- User as the Superior Expert: Unlike many expert systems that aim to replace the human, GLIMPSE acknowledges that the user might be better at certain tasks—like interpreting a complex scatter plot—and allows the user to override system suggestions without breaking the logic flow.
- Suspension vs. Failure: By pausing and telling the user how to get missing information rather than just failing, the system acts as a proactive collaborator.
Results and Discussion
The use of meta-level assertions for the suspending evaluator significantly reduced the number of object-level rules required. Instead of writing repetitive rules for every possible "known/unknown" state of a variable, a single rule combined with meta-logic handles all cases. This architectural efficiency supports a multi-layered advice system where novices can ask for suggestions and then ask for help on how to provide the information for those suggestions, creating a recursive support structure.
Limitations and Future Work
A primary challenge remains knowledge base volatility: once an advised action is carried out, the system cannot simply resume from the point of suspension. Because new information might change the broader knowledge state, the evaluation must be re-checked to ensure previous conclusions remain valid. Future work will need to address the scaling of meta-logic as statistical strategies become more complex and provide quantitative user-study data to compare GLIMPSE's performance against traditional tools.
Summary updated