In the mid-1980s, the field of artificial intelligence faced a significant architectural hurdle: the divide between qualitative logical reasoning and quantitative numerical computation. While early expert systems like MYCIN and PROSPECTOR excelled at symbolic logic, they struggled to incorporate the heavy mathematical modeling required in engineering and physical sciences. The Model Expert System (MES) was developed as a foundational framework to bridge this gap, providing a structure where mathematical formulas and logical inference coexist within a unified reasoning chain.
The Challenge of Hybrid Workflows
The authors of MES argue that fields such as oil log interpretation—the analysis of data from borehole sensors—require a "sandwich" process that traditional rule-based systems could not handle gracefully. This workflow consists of three distinct phases:
- Preliminary Inference: Using qualitative logic to select appropriate mathematical parameters based on the geological context.
- Numerical Calculation: Executing complex formulas using those selected parameters.
- Further Inference: Interpreting the results of those calculations to reach a final geological conclusion.
Existing systems were primarily designed for the first and third steps, often treating numerical computation as an external or poorly integrated afterthought. MES was designed to provide a uniform framework where math and logic could interact without "impairing the advantages of the rule-based system."
Methodology and Architecture
To manage this hybrid complexity, the authors introduced several novel structural components that moved beyond simple "If-Then" logic.
The "TASK" Concept
Instead of merely searching for the truth or falsehood of a specific fact, the MES architecture revolves around "Tasks." A task is defined as a procedure to evaluate an object. Solving a task can trigger both logical rules and numerical subroutines, making the goal-seeking behavior of the system more flexible than standard inference engines.
Hybrid Knowledge Representation
MES utilizes two specific types of rules to handle different aspects of problem-solving:
- Task Transformation Rules: These define the logic for decomposing a complex geological problem into smaller, manageable sub-tasks.
- Procedure Rules: Unlike standard rules, these allow the system to call external FORTRAN subroutines. This enables the system to perform heavy-duty numerical modeling as part of its standard reasoning process.
Handling Uncertainty
To address the inherent noise and uncertainty in geological data, MES implements Dempster-Shafer Evidence Theory. Rather than relying on simple probabilities, the system uses "Support" and "Plausibility" metrics. This allows the system to represent not only the degree of belief in a conclusion but also the degree of ignorance (what the system does not know), which is critical in high-stakes industrial environments.
Real-World Application: WELIES
While MES serves as the underlying "tool system" or engine, its capabilities were demonstrated through WELIES, a comprehensive oil log interpretation system. Developed in collaboration with the Daqing Petroleum Administrative Bureau—one of China's largest oil fields—WELIES was tested against real-world well-logging data. This modularity between the engine (MES) and the application (WELIES) was a precursor to modern AI shells and specialized frameworks.
Key Insights
- Inference as a "Side Effect": In MES, evaluating a geological property is treated as a side effect of solving a task. The system gathers data and performs calculations only when they are functionally necessary to move the task forward, optimizing computational resources.
- Backward-Chaining Procedures: The system utilizes backward-chaining for its mathematical components. It only triggers a complex FORTRAN calculation when it realizes it needs a specific numerical value to satisfy a higher-level logical rule.
- Operational Deployment: Unlike many theoretical papers of the era, MES was implemented on a PERKIN-ELMER 3220 minicomputer and used as a practical tool for petroleum engineering, proving the viability of the framework in a rigorous industrial setting.
Limitations and Legacy
As a product of the 1980s, MES was limited by the processing power of its era and the use of FORTRAN 7. Furthermore, the paper focuses on architectural innovation rather than modern quantitative benchmarking (such as statistical accuracy vs. human experts). However, its successful integration of numerical modeling with logical inference remains a relevant blueprint for modern "neuro-symbolic" or hybrid AI systems that attempt to combine deep learning with symbolic reasoning.
Summary updated