Presented at AAAI-86, AGNESS (A Generalized Network-based Expert System Shell) represents a significant architectural shift in the design of expert systems. Moving away from the rigid production rules that defined early AI frameworks like MYCIN and PROSPECTOR, AGNESS introduced a more flexible, efficient, and transparent network-based approach to knowledge representation and inference.
The Problem: Rigidity and Computational Overhead
In the mid-1980s, most expert system shells relied on "if-then" production rules. While effective for certain tasks, these systems were often computationally expensive because they required exhaustive searches to determine which rules should "fire" at any given time. Furthermore, they were typically restricted to specific data types—usually Bayesian probabilities—and offered limited explanation facilities, making it difficult for users to understand the reasoning behind an AI's conclusion.
Network-Based Knowledge Representation
The core innovation of AGNESS is its use of a computation network structured as a Directed Acyclic Graph (DAG). Instead of a flat list of rules, knowledge is represented as nodes (propositions) connected by edges (dependencies).
This architecture offers several advantages:
- Pre-wired Logic: By defining dependencies at design time, the system eliminates the need for runtime rule-searching. When a value changes, the system simply propagates the update through the pre-defined edges.
- Construct Sharing: AGNESS increases efficiency by allowing different parts of the network to share common sub-conditions, reducing storage requirements.
- Flexible Inference: Unlike its predecessors, AGNESS allows experts to define custom inference methods, such as Mycin-style confidence factors or unique logical connectives, rather than being locked into a single mathematical model.
To ensure logical consistency, the system employs domain constraints. These constraints define which types of objects can validly occupy specific roles within the network, preventing nonsensical matches (e.g., attempting to match a person with a university course).
The Merit System and User Interaction
One of the most sophisticated features of AGNESS is its Merit System, designed to optimize how the system asks questions to the user during backward chaining. Instead of asking questions in a fixed or random order, AGNESS calculates a "Merit" value for each potential query. This value is a ratio of the expected change in the final result to the "cost" (in time or effort) of asking the user. By using partial derivatives of assignment functions, the system mathematically ensures that the most impactful questions are prioritized, minimizing user fatigue.
Furthermore, AGNESS provides an advanced explanation facility. While early systems could only answer "Why" and "How," AGNESS supports eight recursive query types. This allows users to explore the network structure, see where data was used in the past, and even perform "what-if" analysis to see how different answers might affect the final outcome.
Evaluation and Results
The authors evaluated AGNESS across two distinct domains:
- A Dating Service: A "toy" domain used to demonstrate the system's ability to handle complex matching based on social and physical compatibility.
- ECG Exercise Tests: A real-world medical application involving the evaluation of electrocardiogram results to diagnose heart conditions.
In the medical domain, the expert system built with AGNESS achieved a level of performance that surpassed the human doctors originally tasked with analyzing the ECG data. The system demonstrated that a network-based shell could handle diverse data types—including strings and complex frames—while maintaining high computational efficiency.
Conclusion
AGNESS successfully addressed the limitations of rule-based shells by providing a more general and mathematically grounded framework for expert systems. While the authors noted that future work was needed to handle deep causal reasoning and improve graphical interfaces for system builders, AGNESS proved that network-based propagation was a viable and superior alternative for complex, data-intensive AI applications.
Summary updated