Automated Selection of Optimal Model Abstractions

64 concepts63 papers1 proposal

In engineering and artificial intelligence, researchers often face a "Goldilocks" problem: a model must be detailed enough to be accurate, but simple enough to be computationally efficient. As Iwasaki et al. describe in Automated Model Selection for Simulation, including excessive detail leads to intractable simulations, while omitting critical phenomena results in useless predictions. The research direction of automated model selection seeks to navigate this trade-off by dynamically identifying the "simplest adequate model" for a given task.

The Mechanism of Approximation Reformulations

To automate the choice between models, researchers must first define how different levels of abstraction relate to one another. Daniel Weld proposed a formal framework in Approximation Reformulations that treats simpler models as mathematical limits of more complex ones. By framing the transition between models as a change in parameters within a unified space, systems can switch between abstractions without needing to compare entirely different mathematical structures.

This selection process often relies on Context-Dependent Behaviors (CDBs). As Nayak et al. argue in Automated Model Selection Using Context-Dependent Behaviors, a system can build an adequate model by selecting only the physical phenomena relevant to the device's current function. This prevents the "vicious circle" where one needs a model to understand a system's behavior but needs to understand that behavior to choose the right model.

Domain-Specific Fidelity Trade-offs

The tension between fidelity and complexity appears across diverse fields, each requiring different strategies to manage computational overhead:

  • Robotics and Control: In legged locomotion, Dantec et al. found that while full-body models are the most comprehensive, "kino-dynamics" models offer a superior balance for real-time push recovery and adaptation. Similarly, in Policy Optimization for PDE Control with a Warm Start, Zhang et al. use simplified reduced-order models to "warm start" complex controllers, overcoming the intractability of high-fidelity physical laws.
  • Wireless Communication: Debbah et al. highlight a persistent gap between high-fidelity physical simulators and efficient mathematical frameworks. In Indoor 5G 3GPP-like Channel Models for Office and Shopping Mall Environments, the authors note that incorporating complex phenomena like spherical waves significantly increases computational costs, requiring system-level simulators to balance realism with processing time.
  • Fluid Dynamics and Digital Twins: Creating virtual replicas of human physiology, such as the "digital twins" proposed by Alaa et al., involves choosing between 3D fluid dynamics and simplified "lumped-parameter" models. While the former provides spatial detail, the latter is often necessary to keep the training of differentiable surrogate networks feasible.

Failure-Driven Refinement and Diminishing Returns

Rather than starting with the most complex model, many frameworks adopt a failure-driven approach. Richard Doyle suggests in Constructing and Refining Causal Explanations from an Inconsistent Domain Theory that systems should build explanations from simplified abstractions and only add detail when they encounter real-world inconsistencies. This is echoed by Roehm et al., who use "reachset conformance" to automatically adapt a model's complexity until it matches observed real-world data.

However, increasing complexity does not always yield better results. In the study of thermal impedance, Achiri et al. demonstrated that increasing model order yields diminishing returns in accuracy beyond a certain point. This suggests that the "Goldilocks" model is often found at a moderate level of complexity, where the marginal gain in precision no longer justifies the added computational burden.

Ultimately, the goal of this research is to move away from static, "one-size-fits-all" models toward adaptive systems that can strip away unnecessary detail while preserving the accuracy required for specific tasks, such as real-time diagnosis or large-scale simulation.

Key Foundational Work

Approximation Reformulations (1990) — contributed 1 concept elements to this direction.
Constructing and Refining Causal Explanations from an Inconsistent Domain Theory (1986) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_accuracy_computation_trade_off_model_fid_8409aca7,
  title = {accuracy-computation trade-off (model fidelity-complexity tradeoff)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accuracy-computation-trade-off-model-fidelity-complexity-tradeoff/rdr_8409aca774de403f8348f2345b439483},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_accuracy_computation_trade_off_model_fid_8409aca7_v1,
  title = {accuracy-computation trade-off (model fidelity-complexity tradeoff)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accuracy-computation-trade-off-model-fidelity-complexity-tradeoff/rdr_8409aca774de403f8348f2345b439483/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake