Towards Explicit Integration of Knowledge in Expert Systems: An Analysis of MYClN’s Therapy Selection Algorithm

AAAI 1986

TL;DR

This paper argues that expert systems often hide crucial knowledge integration logic within their rule bases or interpreters, making them hard to understand and maintain. The authors analyze MYCIN's therapy selection algorithm to demonstrate how explicit representation of integration techniques and their underlying assumptions can lead to more transparent, explainable, and extensible expert systems.

What Problem Does This Paper Address?

Expert systems, particularly those based on rules, face a significant challenge known as the "knowledge integration problem." This arises when multiple rules suggest different actions or conclusions that need to be combined. Current approaches either try to "compile out" this problem by carefully crafting the rule set to avoid conflicts, or they rely on a single, uniform integration mechanism (like certainty factors) built into the system's interpreter. The authors argue that both these approaches obscure the reasoning behind how recommendations are combined, making it difficult for users to understand why a system makes a particular decision and for developers to modify or extend the system. This lack of explicit integration knowledge leads to systems that are less transparent, harder to debug, and more prone to errors when assumptions are violated.

Methods

The paper's core contribution is an analysis of MYCIN's therapy selection algorithm to identify and categorize various knowledge integration techniques. They propose that instead of hiding this logic, it should be made explicit, potentially through meta-rules or by capturing the design rationale during system construction. The authors introduce several key operations for manipulating and combining preferences:

  • PARTITION: Dividing a continuous range of values (like drug effectiveness scores) into discrete categories (e.g., "first choice," "second choice"). This simplifies representation but can lose fine-grained distinctions.
  • CONDENSE: Creating a simplified representation of a preference by using a many-to-one function. This is often achieved by partitioning.
  • EXTEND: Extending an ordering on individual items to an ordering on collections of items (e.g., therapies, which are sets of drugs).
  • CONJOIN: Combining two preferences (e.g., preference for fewer drugs and preference for effective drugs) into a single, composite preference. This can result in a partial ordering.
  • LINEARIZ: Converting a partial ordering into a total ordering. This often involves making explicit or implicit assumptions about the relative importance of conflicting preferences.
  • DOMINATE: One preference takes precedence over another, used to resolve ties.
  • IGNORE: A preference is disregarded, typically when ties are rare or violating the preference has minimal impact.
  • THRESHOLD: Converting a metric into a constraint by setting a minimum acceptable value.

Datasets

Not discussed in the paper. The paper focuses on analyzing an existing expert system (MYCIN) and its algorithm, rather than training or evaluating on new datasets.

Key Insights and Ideas

  • Integration is a form of knowledge: The paper posits that how an expert system integrates conflicting recommendations is itself a form of knowledge that should be explicit, not hidden.
  • MYCIN's therapy selection is a rich example: The algorithm, designed to balance multiple, often conflicting, medical goals (drug sensitivity, efficacy, cost, etc.), provides a compelling case study for analyzing diverse integration techniques.
  • Trade-off between expressiveness and efficiency: Different representations of goals (e.g., axioms, partial orderings, metrics) offer varying levels of detail and computational cost. The choice of representation significantly impacts how integration can be performed.
  • Implicit assumptions lead to fragility: When integration logic is implicit (e.g., in a fixed table of instructions), the underlying assumptions can be violated without the system recognizing it, leading to potentially incorrect or inexplicable recommendations.
  • Explicit integration supports better explanations: By making integration techniques and their assumptions explicit, systems can provide more meaningful explanations to users and aid maintainers in understanding and modifying the system's reasoning.

Key Results

The paper doesn't present quantitative results in the traditional sense. Instead, it provides a qualitative analysis of MYCIN's therapy selection algorithm, breaking down its process into a series of explicit integration steps. It demonstrates how MYCIN uses techniques like PARTITION, CONJOIN, and LINEARIZE to combine goals like maximizing drug effectiveness and minimizing the number of drugs. The analysis highlights how these steps, while effective, can embed implicit assumptions (e.g., about the relative value of different drug ranks) that are not readily apparent to the user or maintainer.

Limitations and Future Work

Limitations: * The authors acknowledge that their analysis of MYCIN's therapy selection is not exhaustive and that computational concerns also influenced the algorithm's design. * They note that while explicit representation of integration techniques is desirable, it may impose some overhead on the design process. * The paper suggests that the "ideal sense of 'best' therapy" is non-operational, meaning a heuristic approximation is necessary, and the specific definition of "best" is determined by the design choices made.

Future Work: * The authors advocate for mechanizing the reformulation and integration techniques to automate expert system construction and documentation. * They suggest developing mechanisms for applying these techniques and reasoning about which technique to use in a given situation. * A key future direction is to record the entire knowledge integration process in a machine-understandable fashion to support explanations and system maintenance. * They propose that future work should focus on formalizing these techniques and representing the situations they apply to, aiming to create a framework for building expert systems that can explicitly represent and reason about integration.

Summary updated

BibTeX
@inproceedings{1986towards,
  title = {{Towards Explicit Integration of Knowledge in Expert Systems: An Analysis of MYClN’s Therapy Selection Algorithm}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure