Active Pilot Sampling for Multi-fidelity Estimation in Stochastic Optimization

Key points

• Stochastic optimization with high-fidelity nonlinear models is often computationally prohibitive because each objective evaluation requires sampling-based uncertainty quantification inside an outer optimization loop.1
• Multi-fidelity estimators, such as approximate control variates (ACV), can reduce this cost by exploiting cheaper low-fidelity models, but they require model-output covariance information to determine control-variate weights and sample allocations.1
• Active-learning acquisition functions allocate pilot samples to designs where covariance information is most uncertain and most relevant to the outer optimization objective.1
• The results show improved optimization performance relative to single-fidelity Monte Carlo and simpler pilot-sampling strategies.1

Sources (1)

• [1]Active Pilot Sampling for Multi-fidelity Estimation in Stochastic Optimization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 03:24 AM

  We propose an active pilot-sampling framework that learns a probabilistic emulator of the model-output covariance matrix using a positive-definite parameterization based on matrix-logarithm-transformed correlations and log standard deviations.

  Stochastic optimization with high-fidelity nonlinear models is often computationally prohibitive because each objective evaluation requires sampling-based uncertainty quantification inside an outer optimization loop.

Extractive summary: sentences quoted from the sources.