Abstract
The Quantity Lattice is a reasoning system designed to bridge the gap between simple integer arithmetic and complex symbolic algebra for "commonsense" reasoning tasks. By integrating qualitative relationships with quantitative intervals, the system enables AI to reason about physical processes with partially specified data. This paper explores the architecture of the Quantity Lattice, its five core reasoning techniques, and its application in domains like geology and semiconductor manufacturing.
Introduction
In the mid-1980s, AI researchers faced a significant hurdle in arithmetic reasoning. While calculators handled precise numbers and symbolic systems managed complex algebra, everyday "commonsense" reasoning—essential for fields like geology or engineering—often fell between these extremes. Researchers frequently encountered partial information, such as knowing a distance is "less than 120 miles" or a value is "positive." Existing systems were either too specialized or computationally prohibitive for large-scale networks. The Quantity Lattice was developed to provide a fast, expressive framework that combines qualitative logic with numeric bounds.
Methodology: The Quantity Lattice Architecture
The Quantity Lattice represents mathematical expressions as nodes in a directed graph (digraph), where relationships are represented as arcs. The system employs five integrated reasoning techniques to maintain consistency and derive new knowledge.
Qualitative Reasoning and Graph Search
At the core of the system is a transitivity table used to find qualitative relationships through graph search. For instance, if the system knows $A = B$ and $B \leq C$, it can infer $A \leq C$ by looking up the intersection in the table.
Integrated Reasoning Techniques
Beyond graph search, the system uses:
- Numeric Constraint Propagation: Updates interval bounds (e.g., $X \in [1, 5]$) as new relationships are asserted.
- Interval Arithmetic: Performs operations on ranges, such as $[3, 4] + [1, 4] = [4, 8]$.
- Relational Arithmetic: Infers relationships based on identity values (e.g., if $X > 0$, then $Y + X > Y$).
- Constant Elimination Arithmetic: Simplifies algebraic relationships by identifying common terms across expressions.
The integration of these techniques allows an inference in one (e.g., a new numeric bound) to trigger a qualitative inference in another, creating a robust reasoning loop.
Evaluation and Key Results
The system was evaluated using real-world domain models rather than synthetic benchmarks:
- Geologic Interpretation: Modeling deposition, erosion, and uplift.
- Semiconductor Fabrication: A 27-step simulation involving over 1,300 quantities.
In the semiconductor simulation, the system demonstrated remarkable efficiency. Asserting a new relationship took an average of 0.01 seconds, while querying took only 0.003 seconds on a Symbolics 3600. This performance is attributed to the system's focus on "on-demand" inference; in the semiconductor test, only 0.27% of potential relationships were actually calculated, significantly reducing computational overhead.
Discussion: Efficiency and Limitations
The Quantity Lattice achieves near-linear complexity ($O(R)$) by exploiting the sparsity of commonsense problems. Most expressions involve only a few variables, allowing for localized interactions. Additionally, the system records "justifications" for every inference, facilitating explanations and the retraction of assumptions—a critical feature for planning and diagnosis.
However, the system has limitations. It is not a full algebraic engine and cannot simplify complex expressions like $X = (X + Y) - Y$ without explicit intermediate steps. Furthermore, constraints primarily propagate "upward" from arguments to expressions, requiring users to explicitly assert inverse relationships for "downward" propagation. Despite these trade-offs, the Quantity Lattice provides a successful balance between power and efficiency for large-scale qualitative simulations.
Summary updated