Introduction
In the mid-1980s, a primary challenge in artificial intelligence was problem-solving efficiency. Early systems often struggled with expansive "search spaces," leading researchers to develop "macro-operators"—sequences of successful steps stored as single "shortcuts." While effective for specific tasks, these linear macro-operators were brittle; a shortcut learned for a 5-step problem could not generalize to a similar 50-step problem. In 1986, Patricia Cheng and Jaime Carbonell introduced FERMI (Flexible Expert Reasoner with Multi-Domain Inference), a system designed to overcome these limitations by inducing iterative logic and loops from experience.
The Problem: Linear Macro-Operator Limitations
Traditional systems like STRIPS or SOAR improved performance through "chunking," but these chunks were rigid. Repetitive processes, such as variable substitution in algebra, required separate macro-operators for every possible number of repetitions. The core issue was a lack of generality. FERMI was developed to analyze successful problem-solving traces, recognize iterative patterns, and synthesize generalized "While-Loops" or "If-Then" structures capable of handling arbitrary numbers of iterations.
Methodology: Learning to Loop
FERMI employs a three-step process to transform experience into generalized knowledge:
- Pattern Detection via Subgoal Abstraction: Rather than searching for identical rule sequences, FERMI identifies repeating patterns in the changes to subgoals. This higher level of abstraction allows the system to detect a loop even when low-level actions vary slightly due to the problem state.
- Formation of Conditional Macro-operators: To manage variations within a loop, FERMI utilizes an "agenda" structure where rules are grouped into "buckets." If a primary rule is inapplicable, the system tries subsequent rules in the bucket, effectively implementing internal "If-Then-Else" logic.
- Iterative Generalization: FERMI analyzes variable changes across cycles—for instance, noting that each iteration reduces the number of remaining equations. It then constructs a generalized iterative operator with a counter and "up-front" conditions to verify solvability before execution.
Evaluation and Results
FERMI was evaluated across several rule-based domains, including linear algebra, fluid statics, classical mechanics, and DC circuits.
- Linear Algebra: The system successfully learned a generalized rule for solving $N$ equations in $N$ unknowns. By inducing a substitution loop, it could solve systems of any size after observing a few examples.
- Physics (Fluid Statics): In problems involving multi-layer liquid containers, FERMI "invented" the summation formula for pressure differences ($\Delta P = \sum \rho_i g \Delta h_i$) by recognizing the iterative nature of the calculation.
- Search Reduction: The induction of iterative macro-operators led to a massive reduction in search time. Instead of step-by-step reasoning, the system could directly apply a learned iterative shortcut.
Discussion and Future Work
While FERMI demonstrates significant progress in learning iterative logic, the authors note limitations in handling highly chaotic or non-linear patterns. Future research directions include learning from failure to refine "up-front" conditions, analogical transfer of iterative operators across different domains (e.g., applying algebra strategies to circuit analysis), and moving toward the automatic synthesis of increasingly complex subroutines. By moving beyond linear chunks, FERMI provided a foundational step toward more flexible and generalizable AI problem-solvers.
Summary updated