Abstract
Building expert systems for complex diagnostic tasks has traditionally been hindered by the "knowledge acquisition bottleneck," where experts struggle to translate intuition into rigid rules and precise numeric weights. MOLE (a successor to the MORE system) addresses this by utilizing "smart" heuristic assumptions—such as the principles of exhaustivity and exclusivity—to infer missing information and refine knowledge bases. By prioritizing the logical structure of associations over arbitrary numeric support values, MOLE reduces the burden on domain experts and streamlines the creation of robust diagnostic systems.
Introduction: The Knowledge Acquisition Bottleneck
In the mid-1980s, the development of expert systems was often stalled by the difficulty of extracting knowledge from human experts. Previous tools, like MOLE’s predecessor MORE, functioned primarily as "knowledge acceptors," requiring experts to provide complete and consistent data, including precise "support values" or weights of evidence. However, human experts are often inconsistent when assigning these numbers. MOLE was developed to transition from a passive acceptor to an active "knowledge acquirer" that understands the underlying logic of diagnosis, allowing it to identify gaps and ask targeted questions rather than pestering experts for arbitrary details.
Methodology: Heuristic Classification and Smart Assumptions
MOLE operates on the principle of heuristic classification, using two fundamental assumptions to guide its reasoning:
- Exhaustivity: Every abnormal finding or symptom must have at least one explanation (a hypothesis or cause).
- Exclusivity: Explanations should follow the principle of Occam’s Razor—hypotheses should not be multiplied beyond necessity. If a single hypothesis can explain all observed symptoms, it is preferred over a combination of multiple hypotheses.
Static and Dynamic Analysis
MOLE employs a dual-phase approach to build and refine its knowledge base:
- Static Analysis: The system examines the initial network of associations provided by the expert. Using its heuristics, MOLE distinguishes between "covering" associations (where a hypothesis explains a symptom) and "circumstantial" ones. It assigns default support values, relieving the expert of the need to provide precise initial weights.
- Dynamic Analysis: During this interactive phase, the expert provides test cases with known "correct" diagnoses. If MOLE’s internal logic fails to reach the correct conclusion, it performs a "gap analysis." It identifies exactly where the network is failing and asks the expert highly specific questions, such as whether a certain condition might mask a symptom in a particular context.
Key Insights: Structure Over Numbers
A central thesis of the MOLE project is that the structure of knowledge—the web of relationships between symptoms and causes—is significantly more important than the specific numeric weights assigned to those links. If the structural logic is sound, the system can often infer the necessary weights. MOLE also utilizes the "Best Relative" principle, where a hypothesis is accepted if it is the strongest candidate relative to its competitors, even if it doesn't meet an absolute threshold. This mirrors human investigative reasoning, where a cause is often identified by ruling out all other possibilities.
Applications and Results
MOLE has been applied to several real-world diagnostic domains, demonstrating its versatility:
- Steel Rolling Mill Diagnosis: Identifying causes for steel sheets being produced too narrow.
- Computer Hardware: Diagnosing disk faults and network issues.
- System Tuning: Assisting with Micro-Vax performance tuning.
- Power Plants: Developing diagnostic systems for power generation facilities.
In practice, MOLE proved to be a significant advancement over MORE. It allowed experts to start with an "under-specified" network, which the system then helped to flesh out. In the steel rolling mill application, MOLE identified eighteen instances where additional distinguishing features were needed but only prompted the expert for them when they became relevant to a specific diagnostic failure, making the acquisition process far more efficient and less tedious for the human expert.
Limitations and Future Directions
While highly effective for diagnostic tasks, MOLE is specifically tailored for heuristic classification. It may face challenges in domains that lack a clear causal or "covering" structure, such as certain recommendation systems. Additionally, extremely under-determined networks may still require significant expert intervention to resolve temporal ambiguities (e.g., determining the direction of causality). Future research aims to extend MOLE’s heuristics to non-diagnostic tasks, such as configuration problems (e.g., selecting computer components based on specifications), to test the broader applicability of its core assumptions.
Summary updated