Abstract
Explanation-Based Indexing (EBI) addresses the indexing problem in Case-Based Reasoning (CBR) by leveraging domain knowledge to identify relevant features from a single problem-solving episode. By constructing and pruning explanation trees based on observed repair actions, the system distinguishes between critical diagnostic clues and coincidental noise. This approach allows for effective case retrieval even when the underlying domain theory is incomplete or imperfect.
Introduction: The Indexing Problem in CBR
In Case-Based Reasoning, systems solve new problems by retrieving and adapting solutions from past experiences. A fundamental challenge is the "Indexing Problem": determining which features of a situation are relevant for storage and future retrieval. Over-indexing on irrelevant details (e.g., the operator's name or ambient weather) leads to retrieval failures, while under-indexing results in the retrieval of inappropriate solutions.
Traditional inductive methods require large datasets to statistically isolate relevant features. In contrast, this paper introduces Explanation-Based Indexing (EBI), which uses a system's existing causal knowledge to determine relevance from a single case.
Methodology: The EBI Process
The EBI process consists of three primary stages designed to justify actions through the lens of a causal model.
- Constructing Explanation Trees: Upon encountering a symptom (e.g., a straining motor), the system identifies all initial observables that could potentially explain it. It builds trees where the root is the symptom and the leaves are the observables, connected by causal links.
- Justification through Pruning: The system examines the actions taken during the case. An observable is deemed relevant if its value helps "prune" the explanation tree—either by eliminating a competing hypothesis or by confirming the necessity of a specific action.
- Index Generation:
- Primary Indices: Features confirmed as relevant are generalized into main search keys.
- Secondary Indices: Features that the domain theory cannot explain, but are not proven irrelevant, are retained as secondary filters.
- Irrelevant Features: Features proven to have no bearing on the actions are discarded.
Case Study: Robotic Fabrication
The approach was evaluated using diagnostic cases from automated robotic fabrication cells. In a typical scenario, an operator hears a motor "loading up" and performs a sequence of recovery actions such as checking air gauges or manually rotating tools.
Key Results
A qualitative analysis of a motor fault case demonstrated the system's ability to filter noise. For instance, "Motor Temperature" was correctly identified as irrelevant to the action of rotating the tool, despite a theoretical link in the causal model. Conversely, features like "Tool changed less than 2 hours ago" were extracted as primary indices. By using domain theory, the system generalized specific data points (e.g., a timestamp of 1:00 PM) into functional ranges (e.g., < 2 hours since last change), enhancing the utility of the case for future retrieval.
Discussion and Insights
EBI shifts the focus from validation (proving a solution is correct) to justification (explaining why an action was reasonable). This makes the process more robust to incomplete domain knowledge. Furthermore, the system can identify "knowledge gaps" when an operator's actions contradict the domain theory, providing a roadmap for future model refinement.
However, limitations remain. The system is constrained by its existing causal links; it cannot index by features it does not recognize as causally relevant. Additionally, the current binary logic of "indicates" links may need to evolve toward probabilistic reasoning to handle real-world uncertainty more effectively.
Summary updated