Decompositional Modeling through Caricatural Reasoning

Brian C. Williams, Olivier RaimanAAAI 1994

Abstract

This paper presents Decompositional Modeling (DM), a novel approach for solving complex nonlinear algebraic systems by automating the "caricatural reasoning" used by human experts. DM identifies dominant terms in equations to create simplified "caricatures" that are mathematically tractable. By partitioning the problem space into a "patchwork" of regimes, each governed by a simplified model, DM provides both computational efficiency and physical insight. The method is demonstrated on classic analytical chemistry problems, successfully replicating expert-level region diagrams and rules of thumb.

Introduction

In scientific domains such as analytical chemistry and atmospheric modeling, physical phenomena are frequently described by systems of high-degree nonlinear equations. Solving these systems directly often presents significant challenges: equations of degree five or higher lack general closed-form solutions, and brute-force numerical methods often obscure the underlying physical intuition.

Human experts typically navigate this complexity by applying "tacit skills" to make simplifying assumptions. For instance, a chemist might assume an acid is "very weak" or a solution is "highly acidic" to ignore negligible terms. While effective, manual simplification is error-prone and does not scale to systems involving hundreds of simultaneous reactions. This research formalizes and automates this "modeler's intuition" through the Decompositional Modeling (DM) framework.

Methodology

The core of DM is Caricatural Reasoning. Similar to how a caricature exaggerates prominent physical features for recognition, DM identifies "prominent features" within equations—specifically the relative magnitudes of different terms.

The DM algorithm proceeds through five primary stages:

  1. Generation of Caricatural Assumptions: Identifying pairs of terms where one significantly outweighs the other, formalized as "much greater than" ($\gg$) assumptions.
  2. Assumption Merging: Combining local assumptions into sets that define global "regimes" of system behavior.
  3. Equation Simplification: Stripping away insignificant terms based on the assumptions to produce "caricatural equations."
  4. System Solving: Solving the simplified, tractable system.
  5. Domain Identification: Determining the mathematical boundaries where each simplification remains valid.
Table showing the mapping of initial physical equations to caricatural assumptions (A1-A4) based on prominent features.

DM utilizes two symbolic manipulation systems: SR1, a hybrid sign/real algebra, and Estimates, an order-of-magnitude algebra. It employs "qualitative resolution"—a technique analogous to propositional resolution in logic—to combine equations and derive simplified forms.

Results and Discussion

The framework was evaluated using standard problems in analytical chemistry, such as acid-base dilution and equilibrium reactions, with results validated against established textbooks like Beckwith’s Analytical Chemistry (1985).

A central concept in DM is the "Patchwork Universe." Rather than seeking a single, monolithic model for all states, DM characterizes system behavior as a collection of regimes. As system parameters shift, the model transitions from one simplified caricature to another. This approach successfully replicated the complex "region diagrams" used by chemists to analyze reaction behavior across different concentration and acidity levels.

A region diagram generated by DM, illustrating the patchwork of simplified regimes (R1-R7) for an acid-base system.

Furthermore, DM leverages a subset/superset lattice to organize assumptions. Due to the monotonicity of assumptions, if a small set is found to be inconsistent, all larger sets containing it are automatically pruned. This significantly enhances efficiency; in one test case, DM reduced 16 potential assumption sets to 9 consistent ones. The system also derived simplified concentration equations (e.g., $h^+ \approx C_a$) that align with the "rules of thumb" taught in undergraduate chemistry, but derived them rigorously from first principles.

Conclusion and Future Work

Decompositional Modeling offers a powerful bridge between symbolic logic and continuous mathematics. While currently focused on algebraic systems, future work aims to extend DM to dynamical systems (differential equations). Additionally, the authors identify the need for more robust numerical error bounding to manage the cumulative effects of multiple caricatural exaggerations in larger systems.

Summary updated

BibTeX
@inproceedings{williams1994decompositional,
  author = {Brian C. Williams and Olivier Raiman},
  title = {{Decompositional Modeling through Caricatural Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure