A New Architecture for Automated Modelling

AAAI 1998

TL;DR

This paper introduces AIM, an automated modeling system that tackles the complexity of existing approaches by separating ideal component models from their corrections. This novel architecture allows for the generation of parsimonious models by iteratively refining them based on behavioral differences, addressing the need for flexible yet structured automated modeling.

What Problem Does This Paper Address?

Developing automated modeling systems has historically faced a trade-off between flexibility and complexity. Existing systems fall into two camps: model composition systems, which rely on large, intricate libraries of pre-defined model fragments and their applicability conditions, and model induction systems, which are more flexible but require complete knowledge of the referent system's behavior and struggle to define model boundaries. This paper aims to bridge this gap by proposing an architecture that is simpler than compositional libraries but more sophisticated than basic induction systems, enabling the automatic generation of accurate and concise models without requiring exhaustive prior knowledge.

Methods

AIM employs a component-centered architecture where knowledge is partitioned into a "component library" (physical parameters) and an "m-component library" (idealized models of components). These m-components are described by bond graph fragments and possess "intra-actions" that define how effects propagate within a component. The core novelty lies in the introduction of "corrections," which represent the retraction of simplifying assumptions made in the ideal m-components.

The modeling process itself is a two-stage approach:

  1. Model Boundary Determination: This stage identifies the relevant components and their interactions for a given task definition. It uses a boundary analysis queue to iteratively expand the model's scope based on the task's specified ports and effects, ensuring that all relevant causal relationships are captured.
  2. Model Refinement: This is the iterative core of AIM. It starts with an initial, potentially over-simplified model and progressively adds corrections. The algorithm assesses the behavioral impact of each potential correction by comparing the behavior of the model with and without the correction. Corrections that cause a "significant" change in behavior are added, and the process repeats until no remaining corrections significantly alter the model's behavior. This "steepest ascent hill climbing" search aims to identify the most impactful corrections first, leading to parsimonious models.

Datasets

Not discussed in the paper. The paper focuses on a single illustrative example: a water-filled syringe.

Key Insights and Ideas

  • Separation of Ideal Models and Corrections: This is the central architectural innovation. By decoupling the base model of a component from the specific assumptions that can be relaxed, AIM gains significant flexibility. This allows for a much simpler base knowledge representation compared to complex model composition libraries.
  • Behavioral Comparison for Refinement: Instead of relying on explicit rules or complex applicability conditions, AIM uses the behavioral difference between models as the primary driver for refinement. This makes the process more adaptive and less dependent on the structure of the knowledge base itself.
  • Parsimony as a Goal: The paper explicitly aims to generate "parsimonious" models, meaning adequate models with the fewest possible simplifications (corrections). This is achieved by only including corrections that demonstrably improve the model's behavior relative to the task.
  • Intra-actions for Model Boundary: The use of intra-actions within m-components is crucial for determining the model boundary. They define how effects propagate within a component, guiding the expansion of the model to include relevant interactions.

Key Results

The paper demonstrates AIM's capability by generating models for a water-filled syringe. The results show that AIM successfully produces parsimonious models that are adequate for the specified task. For instance, in one scenario, AIM generated a model that was adequate (performance index < 0.1), while precursor models were not. The paper highlights how different physical parameters (e.g., steel vs. rubber cylinder, constant vs. varying force) necessitate different sets of corrections, showcasing AIM's adaptability. Table 3 provides performance indices comparing AIM-generated models to a complex reference model, indicating that AIM's models achieve adequate behavior.

Limitations and Future Work

  • Insignificant Correction Assumption: AIM's stopping criterion relies on the assumption that insignificant corrections do not significantly affect the model's behavior. While this generally holds, the paper acknowledges that this assumption is not always perfectly true, especially in non-linear systems, and can occasionally lead to models that are not strictly parsimonious.
  • Independence of Corrections: The algorithm assumes that the effects of corrections are largely independent. While this is a reasonable approximation for linear systems, synergistic effects between corrections can occur, potentially influencing the order in which corrections are chosen and their ultimate inclusion.
  • Fitting Approximations: AIM's reliance on "fitting approximations" for corrections restricts the types of corrections that can be made.
  • Single Operating Region: AIM generates models for a single operating region. The paper notes that different operating regions may require different models, and AIM currently does not control transitions between these distinct models.
  • Future Work: The authors suggest that further investigation into the combined effects of corrections is needed to refine the independence assumption. They also hint at the need for more sophisticated methods to handle situations where the insignificant correction assumption breaks down.

Summary updated

BibTeX
@inproceedings{1998new,
  title = {{A New Architecture for Automated Modelling}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1998},
}

Loading related content...

Figure