Abstract: Not available
This paper, presented at AAAI-86, explores a foundational challenge in artificial intelligence: how to move beyond rigid "if-then" rules toward systems that understand the underlying principles of a domain. Authors Chidanand Apté and Se June Hong introduce a mechanism for qualitative reasoning over financial arithmetic, allowing computers to reason about money the way experts do—by understanding how variables like interest and depreciation influence the bottom line, even when exact numbers are missing.
The Problem: Brittle Expert Systems
In the mid-1980s, most expert systems were built using production rules—massive libraries of "if-then" statements elicited from human experts. While effective for specific tasks, these systems were brittle. If a situation didn't perfectly match a pre-written rule, the system failed. Furthermore, maintaining thousands of rules to ensure they didn't contradict each other was a "painstaking knowledge engineering" nightmare.
The authors argue that in domains like finance, the "underlying theory" is actually quite simple: it’s just arithmetic. A human expert doesn't need a thousand rules to know that higher interest rates lower net profit; they understand the equation. The challenge was building a system that could perform this same qualitative reasoning—understanding the direction and magnitude of change—without requiring a full set of precise numerical inputs.
Methods: The Equation Reasoner Architecture
The authors developed an architecture consisting of three main components designed to derive qualitative inferences directly from underlying arithmetic equations rather than pre-compiled heuristic rules.
- Equation Base (EB): A central repository storing financial equations in a "prefix unfocused" notation (similar to Lisp). This allows the system to treat equations as data that can be manipulated.
- Equation Reasoner (ER): The core engine that performs symbolic manipulation. It can "focus" an equation on a specific variable (solving for $x$) and then apply qualitative rules about arithmetic (e.g., "if the denominator increases, the result decreases").
- Situational Knowledge (SK): A database of domain-specific facts (e.g., "Operating income is typically much larger than interest expenses") that helps the reasoner make approximations and resolve ambiguities.
The Two-Pass Mechanism
To answer a query, the ER uses a "propagate-aggregate" approach:
- Top-Down Pass: It propagates qualitative constraints (like "increase" or "decrease") from the goal variable down to the individual components of the equation.
- Bottom-Up Pass: It aggregates the effects of controllable variables back up to the goal to determine the final impact.
- Conflict Resolution: If the math suggests two conflicting trends, the system uses SK to determine which effect dominates or resorts to numerical simulation to find the answer.
Key Insights and Ideas
- First Principles vs. Surface Rules: Reasoning from the underlying equation is more robust than writing a rule for every possible scenario. It guarantees that the system's logic is as consistent as the math itself.
- Qualitative Arithmetic: The ER doesn't just do math; it knows properties of math. For example, it knows that if a base is a positive fraction less than 1, the result decreases as the exponent increases.
- Pattern Recognition: The system identifies common mathematical "concepts" (like binomial terms) before breaking them down. This prevents "locally correct but globally incorrect" interpretations.
- The "Knowledge Engineering" Shift: The authors suggest that instead of engineering domain-specific rules, we should engineer "general knowledge about arithmetic expressions," which can then be applied to any domain governed by equations.
Key Results
The authors demonstrated the ER's utility through two primary financial scenarios using standard accounting equations (e.g., Earnings Per Share, Cash Flow):
- Ranking Alternatives: When asked how to rank financing options to maximize Earnings Per Share (EPS), the ER symbolically solves the EPS equation and deduces that the primary criterion should be minimizing "ProfitLossImpact."
- Impact Assessment: For a query about how a $5 million equipment purchase affects a firm, the ER combines equations with situational knowledge to conclude that the purchase will cause a "significant decrease" in EPS in the first year.
These results showed that a single equational representation could serve two purposes: precise numerical calculation and high-level qualitative "what-if" analysis.
Limitations and Future Work
The authors acknowledge that the language used to describe qualitative changes needs to be richer to handle more nuanced comparisons. Furthermore, not all financial decisions are mathematical; some are based on corporate policy, necessitating the integration of the ER with non-mathematical heuristic reasoners. Future directions include "caching" derived answers as new surface rules to speed up reasoning and using the reasoning trace to generate natural language explanations for human users.
Summary updated