Cost-Normalized Acquisition for Multi-Fidelity Optimization

371 concepts178 papers1 proposal

When optimizing complex systems—such as tuning a bipedal robot's walking controller, designing an analog circuit, or validating the safety of an autonomous vehicle—we face a fundamental bottleneck: evaluating the objective function is incredibly expensive. A physical robot trial might take hours and risk hardware damage, while a post-layout circuit simulation can consume massive computational overhead.

Standard black-box optimization frameworks often treat every evaluation as equally costly. However, practitioners rarely operate in a vacuum devoid of cheaper alternatives. We often have access to a hierarchy of low-fidelity approximations: simplified physics simulators, schematic-level circuit designs, or downsampled datasets. These proxies are fast and cheap but inherently biased or noisy.

As noted by Alonso Marco et al. in Virtual vs. real: Trading off simulations and physical experiments in reinforcement learning with Bayesian optimization, treating these diverse data sources as distinct information assets with different price tags allows us to bridge the "reality gap" efficiently. The core challenge then shifts from simple optimization to multi-fidelity resource allocation: how do we strategically deploy our budget across these varying levels of accuracy to find the global optimum with the lowest cumulative cost?

The Core Mechanism: Information per Dollar

To solve this problem, we must move away from cost-blind acquisition functions. In traditional Bayesian optimization, an acquisition function like Expected Improvement (EI) or Entropy Search evaluates a candidate point $x$ solely based on its potential to improve our knowledge of the optimum.

Multi-fidelity optimization introduces a joint decision space: at each step, the algorithm must choose both the next input parameter $x$ and the fidelity level $m$ at which to evaluate it. This mechanism relies on a cost-normalized acquisition function. Instead of maximizing raw information gain, the algorithm maximizes the ratio of expected information gain to the evaluation cost $C(m)$ of that specific fidelity:

$$\alpha(x, m) = \frac{\text{Information Gain}(x, m)}{C(m)}$$

This simple division fundamentally changes the optimization behavior. As demonstrated by Shion Takeno et al. in Multi-fidelity Bayesian Optimization with Max-value Entropy Search and its Parallelization, cost-weighting forces the optimizer to exhaustively exploit cheap, low-fidelity evaluations during the early stages of search. A low-fidelity query is preferred as long as its information return per unit cost exceeds that of an expensive high-fidelity query.

Henry B. Moss et al. formalize this in MUMBO: MUlti-task Max-value Bayesian Optimization by evaluating information gain with respect to the value of the maximum, ensuring that expensive high-fidelity trials are strictly reserved for moments when low-cost proxies can no longer resolve the remaining uncertainty. Similarly, Aryan Deshwal extends this cost-weighted paradigm to multi-objective landscapes in Multi-Fidelity Multi-Objective Bayesian Optimization: An Output Space Entropy Search Approach, demonstrating how cheap simulations can map out the broad structure of a Pareto front before high-fidelity resources refine it.

Coarse-to-Fine Filtering and Space Pruning

The intuition behind why this cost-normalized approach works lies in treating cheap evaluations as a sieve or a filtering mechanism. Instead of wasting a massive high-fidelity budget exploring unpromising regions of a vast search space, the algorithm uses low-fidelity queries to perform broad exploration. If a cheap simulator indicates that a certain region of the parameter space yields poor performance, that region can be safely pruned or ruled out without ever running a single expensive experiment.

Rajat Sen et al. formalize this hierarchical pruning in Multi-Fidelity Black-Box Optimization with Hierarchical Partitions, where cheap, biased approximations guide a tree-search algorithm to narrow down the search space, triggering high-fidelity evaluations only when the approximation error within a specific partition exceeds its internal uncertainty. This hybrid philosophy is echoed by Ellis Hoag et al. in Bayesian Optimization Meets Search Based Optimization: A Hybrid Approach for Multi-Fidelity Optimization, where search-based optimization rules out sub-optimal regions using cheap approximations, focusing expensive resources solely on high-potential areas.

This filtering mechanism is particularly valuable in safety validation and falsification. For instance, Zahra Shahrooei et al. show in Falsification of Learning-Based Controllers through Multi-Fidelity Bayesian Optimization how cheap, low-accuracy simulations can rapidly scan for potential system failures, allowing high-fidelity resources to be precisely targeted to pinpoint exact failure modes.

Finding the Optimal Allocation Ratio

A natural question arises: how do we determine the exact mathematical balance between these data sources? If we over-invest in cheap, noisy data, our model will be misled by bias; if we over-invest in high-fidelity data, we run out of budget before covering the search space.

The optimal allocation depends heavily on the correlation between the fidelities and their relative costs. Alexey Zaytsev 0002 et al. address this in Minimax Approach to Variable Fidelity Data Interpolation by minimizing the worst-case interpolation error, proving that the optimal ratio of data sources is a direct function of dimensionality, cost ratios, and cross-fidelity correlation.

Furthermore, cheap data can act as a statistical anchor to de-noise expensive observations. In Leveraging Correlation Across Test Platforms for Variance-Reduced Metric Estimation, Rachel Luo et al. demonstrate that treating abundant, low-fidelity simulation data as control variates can drastically reduce the variance of high-fidelity real-world performance estimates, allowing for high-confidence validation with up to 80% fewer physical trials.

This structural trade-off is governed by predictable scaling laws. As Paul Setinek explores in Towards Multi-Fidelity Scaling Laws of Neural Surrogates in CFD, the data scaling axis can be explicitly decomposed into a total computational budget and a composition ratio, revealing the precise mixtures that maximize model accuracy under strict resource constraints.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 205 papers
  2. Version 1 · 2026-10-04 · 295 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_multi_fidelity_resource_allocation_4f767e7e,
  title = {multi-fidelity resource allocation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-fidelity-resource-allocation/rdr_4f767e7ec6d34c6eac6fe7f276c14580},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_multi_fidelity_resource_allocation_4f767e7e_v2,
  title = {multi-fidelity resource allocation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-fidelity-resource-allocation/rdr_4f767e7ec6d34c6eac6fe7f276c14580/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake