Introduction
In the early 1990s, the development of expert systems was often bottlenecked by the rigidity of Knowledge Acquisition (KA) tools. Most existing tools were "role-limiting," meaning they were hard-coded for specific problem-solving templates like classification or "propose-and-revise." If a domain expert's problem did not fit these narrow templates, the tool became ineffective. These systems lacked a fundamental understanding of why they required specific information, leading to brittle architectures that were difficult to adapt. Yolanda Gil's work on EXPECT addresses this by proposing a "reflective" architecture capable of analyzing its own knowledge base and problem-solving methods to dynamically determine what it needs to learn.
The EXPECT Architecture
The core innovation of EXPECT is its ability to treat its own goals and methods as data for introspection. The architecture is built upon two distinct pillars of knowledge representation:
- Factual Knowledge: Definitions, relations, and domain facts are represented using LOOM, a description logic system.
- Problem-Solving Knowledge: Procedural knowledge is represented as methods consisting of goals, bodies, and results.
By translating procedural goals into LOOM concepts, EXPECT can use a semantic classifier to match goals to methods based on meaning rather than simple name-matching. The system performs a "static analysis" via partial evaluation of top-level goals. It traverses the goal tree to verify if every subgoal can be satisfied. When a gap is found—such as a missing method or an undefined factual attribute—the system identifies it as a "lapse."
Reflective Knowledge Acquisition in Practice
Instead of failing when encountering missing information, EXPECT maintains an Agenda Manager. This component generates a list of specific items for the user to address, derived directly from the system's internal logic. For example, if the system needs to calculate shipping speeds but lacks data on a specific port's type, it will explicitly ask the user if the port is a seaport or an airport.
This approach was evaluated within a transportation planning domain, involving concepts like berths, ships, and locations (e.g., Los Angeles and Long Beach). The interface allows users to see the goal tree and the resulting agenda of "missing values." This dynamic expectation generation ensures that the system only asks for information that is functionally necessary for its current problem-solving tasks.
Key Insights and Implications
EXPECT demonstrates several critical advantages over traditional KA tools:
- Goals as Concepts: Representing goals in description logic allows for sophisticated, semantic goal matching.
- Knowledge Interdependence: The system recognizes that knowledge is a "web." Adding a new method might create a requirement for previously unnecessary facts, a dependency EXPECT tracks automatically.
- Analogy-based Refinement: The tool supports users in adding new methods by using existing ones as templates, facilitating easier domain expansion.
While powerful, the system faced limitations at the time, such as the need for more "relaxed" semantic matching to suggest similar methods more effectively. Additionally, as knowledge bases grow, the agenda of lapses could become overwhelming, suggesting a need for future work in prioritizing agenda items based on task criticality. Despite these challenges, EXPECT represents a significant step toward autonomous, self-analyzing AI systems that can guide their own refinement through structured interaction with human experts.
Summary updated