TL;DR
This paper introduces a novel control strategy for expert systems called "inference by information gain" (IBIG). It addresses the challenge of efficiently navigating complex knowledge bases by using a parallel, self-organizing approach that dynamically selects the most informative knowledge hierarchy to explore next, mimicking human problem-solving "restructuring."
What Problem Does This Paper Address?
Expert systems rely on two key components: a comprehensive knowledge base and an effective control strategy to guide reasoning. Traditional control strategies often struggle with the inherent complexity and ambiguity of real-world knowledge. Many systems use rigid, hypothesis-driven approaches that can lead to "dead ends" or inefficient backtracking when data is irregular or doesn't fit a single predefined path. Furthermore, expert knowledge itself is often not neatly contained within a single structure; experts commonly utilize multiple, distinct hierarchies of understanding simultaneously. This paper tackles the problem of how to effectively manage and reason over such multi-hierarchical knowledge bases in a way that is adaptive, efficient, and mirrors human cognitive flexibility.
Methods
The core of the paper's contribution lies in its novel control strategy, IBIG, which is based on Shafer's Belief theory and Dempster's rule of combination. Instead of a sequential, hypothesis-driven approach, IBIG operates in a parallel and self-organizing manner. The key concept is "information gain," calculated as the maximal difference between the current evidence distribution in the knowledge base and the potential evidence distribution. This information gain determines which consultation step to take next.
The knowledge base itself is structured as a multi-level system, allowing for the parallel representation of various knowledge relationships (e.g., causal, hierarchical, temporal) on distinct layers. For the control strategy, the paper focuses on a restricted version: a layered set of disjunctive hierarchies. This restriction ensures consistent belief distribution.
The strategy involves:
- Belief Theory Application: Shafer's belief theory is used to handle uncertain evidence, distinguishing between lack of knowledge and non-supporting knowledge. Dempster's rule combines evidence from different sources.
- Information Increment Strategy: This strategy calculates the "information gain" at each node or hypothesis set. It quantifies how much new information would be obtained by exploring a particular path. The system then selects the path with the highest information gain.
- Hierarchy Selection: In a multi-hierarchical knowledge base, IBIG calculates information gains for each hierarchy separately. The hierarchy that yields the maximal information gain is chosen for the next consultation step. If the data doesn't match a hierarchy well, the system can switch to a "competing" hierarchy, analogous to human "restructuring" in problem-solving.
Datasets
Not discussed in the paper. The paper describes an expert system (IBIG) applied to acquired speech disorders (aphasia) and mentions hypothetical knowledge bases for testing, but no specific public datasets are listed.
Key Insights and Ideas
- Parallel and Self-Organizing Control: Moving away from rigid, sequential control strategies towards a parallel, adaptive system that discovers the most informative path dynamically.
- Information Gain as a Guiding Principle: Using the concept of information gain to make decisions about data collection and knowledge exploration, offering a theoretically grounded approach.
- Mimicking Human Restructuring: The ability of the system to switch between competing hierarchies when a current path becomes less informative directly mirrors how humans might reframe a problem or switch strategies when stuck.
- Multi-Hierarchical Knowledge Representation: The importance of representing expert knowledge not as a single monolithic structure, but as a collection of distinct, complementary hierarchies.
- Belief Theory for Uncertainty: Leveraging Shafer's belief theory provides a robust framework for handling subjective beliefs and uncertain evidence, crucial for real-world expert systems.
Key Results
The paper presents a conceptual framework and implementation details for the IBIG system. While specific quantitative results on a real-world medical application are limited, they report that for a hypothetical knowledge base with 8 hierarchies, each containing approximately 300 nodes, the running time for evaluating the largest information gain is 29 seconds. The system has been successfully tested in a small number of cases for acquired speech disorders.
Limitations and Future Work
- Disjunctive Hierarchies: The current control strategy is restricted to multi-hierarchical knowledge bases with disjunctive hierarchies. This restriction is necessary to guarantee consistent belief distribution. Future work would likely involve extending this to non-disjunctive or overlapping hierarchies.
- Approximation of Belief Theory: The paper mentions using an approximation of Shafer's belief theory for combining evidence in hierarchies. While effective, further exploration of exact methods might be beneficial.
- Real-World Medical Applications: While tested on aphasia, the paper suggests further application to medical areas with larger knowledge bases is in progress, indicating a need for more extensive validation.
- Cognitive Realism: The paper draws parallels to human problem-solving, but further research could explore the cognitive validity and limitations of this computational model of human expertise.
Summary updated