Abstract: Early expert systems relied on rigid production rules that often struggled with the non-linear nature of human expertise. This paper introduces WHEEZE, a system that combines frame-based data structures with an agenda-driven control mechanism. By representing knowledge as frames and managing reasoning through a prioritized agenda, WHEEZE achieves a flexible diagnostic process capable of both data-driven and goal-directed reasoning.
Introduction
In the late 1970s, the field of artificial intelligence was dominated by rule-based expert systems such as MYCIN. While successful, these systems were often criticized for their "rigidity." They primarily utilized backward-chaining logic, which forced the system to pursue a single goal until completion, regardless of new, potentially more relevant evidence. This led to several practical issues: the pursuit of irrelevant questions, an inability to pivot hypotheses, and the need for complex, non-intuitive data structures to manage control flow.
The WHEEZE Architecture
WHEEZE addresses these limitations by reimagining the underlying architecture of expert systems. It replaces flat production rules with hierarchical frames and introduces an agenda-based control structure.
Frame-Based Representation
In WHEEZE, every diagnostic assertion is represented as a frame. These frames contain "slots" that store not only the logic for confirming a hypothesis but also the strategic information needed to guide the search. Key slots include Manifestation (supporting evidence), SuggestiveOf (hypotheses to consider if the current one is confirmed), and ComplementaryTo (alternative hypotheses to consider if the current one is refuted).
This structure allows for a more nuanced calculation of certainty through custom functions, moving beyond simple binary logic. As seen in the figure above, a frame can encapsulate manifestations, suggestions, and even specific findings to be reported to the user.
Agenda-Based Control
The reasoning process is managed by a prioritized agenda. Instead of following a hard-coded search tree, the system evaluates the most promising goal on the agenda. If a hypothesis is confirmed, it adds its "suggested" next steps to the agenda; if refuted, it adds "complementary" alternatives.
As shown in the conceptual network above, the relationships between conditions like Obstructive Airway Disease (OAD) and Restrictive Lung Disease (RLD) are explicitly mapped. This allows the system to perform "surprise pivots"—for instance, if evidence for RLD is being gathered but a specific measurement (like a low FEV1/FVC ratio) strongly indicates OAD, the agenda mechanism allows the system to immediately shift its focus.
Key Insights and Results
The primary advantage of the WHEEZE architecture is the decoupling of domain knowledge from control strategy. By adjusting importance weights within the frames, developers can alter the system's behavior (e.g., making it more exhaustive or more focused) without rewriting the underlying diagnostic logic.
Evaluation on ten standard pulmonary function test cases demonstrated that WHEEZE achieved functional parity with its predecessors, PUFF and CENTAUR. However, WHEEZE offered a significantly cleaner and more maintainable knowledge base. Complex rule sets were collapsed into elegant, self-contained frames, and the need for manual "hard-coding" of question sequences was eliminated.
Conclusion
WHEEZE represents a significant step toward more flexible AI architectures. By integrating frames and agendas, it provides a framework that more closely mirrors the dynamic reasoning of human experts. While future work is needed to address computational efficiency and broader scalability, the system demonstrates the power of uniform representation in managing complex diagnostic tasks.
Summary updated