The Use of Intelligently Controlled Simulation to Predict a Machine's Long-Term Behavior

Andrew GelseyAAAI 1991

Introduction

Presented at AAAI-91, Andrew Gelsey’s research explores the intersection of classical mechanics, numerical simulation, and artificial intelligence. The work addresses a fundamental engineering challenge: transitioning from raw numerical data to a high-level understanding of a machine's long-term behavior.

Problem Statement

Numerical simulators excel at short-term predictions but are often "blind" to the big picture. While they generate vast amounts of positional and velocity data, they cannot inherently determine if a mechanism will remain stable, jam, or fail over time. Traditionally, identifying these regularities required human experts to manually analyze simulation plots—a process that is labor-intensive and difficult to scale. Gelsey proposes an "intelligent controller" to automate this high-level reasoning, characterizing behavior and determining long-term stability.

Initial state of the clock escapement mechanism showing the balance, lever, and escape wheel.

Methodology

The proposed system consists of three primary components: an automated modeler that converts CAD models into differential equations, a numerical simulator using a variable-step Runge-Kutta solver, and an Intelligent Controller.

The controller identifies "interest points" by focusing on kinetic energy minima. These points typically represent qualitative boundaries, such as direction changes or new mechanical impulses. By comparing states at these minima, the algorithm hypothesizes the machine's cycle. To verify long-term behavior without exhaustive simulation, the system employs a "jumping" technique, extrapolating the system's state far into the future and restarting the simulation to check if the behavioral hypothesis remains valid.

The full cycle of the clock escapement mechanism, illustrating the interaction between components.

Key Insights

  • Kinetic Energy as a Heuristic: Using energy minima allows the system to filter high-frequency noise and focus on the qualitative "beats" of the mechanism.
  • Evolving Regularity: The system defines regularity as a pattern that evolves predictably, allowing it to account for gradual effects like friction or a winding spring.
  • Temporal Sampling: "Jumping" ahead in time significantly reduces computational costs while ensuring stability.
  • Predicting Sudden Transitions: By using geometric "motion envelopes," the system can predict future collisions or state changes that haven't occurred yet in the simulation.

Results and Evaluation

The algorithm was evaluated on complex mechanical models, most notably a clock escapement mechanism. The system successfully automated the analysis, initially testing a simple cycle hypothesis before correctly identifying a two-minima cycle. It also handled a chiming clock model, distinguishing between the primary escapement behavior and the secondary, periodic "superimposed process" of the chime.

Limitations and Future Work

While successful in the domain of rigid-body clockwork, the system faces several limitations. The logic is currently tailored to mechanical devices with rigid-body geometry, and the descriptive language may be too simple for chaotic or highly non-linear systems. Additionally, future iterations require more robust measures for handling noisy experimental data and a more hierarchical approach for machines with many uncoupled sub-mechanisms.

Summary updated

BibTeX
@inproceedings{gelsey1991use,
  author = {Andrew Gelsey},
  title = {{The Use of Intelligently Controlled Simulation to Predict a Machine's Long-Term Behavior}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure