Important
Surrogate is a library developed in the course of a research project, not as a dedicated tool. As such, it has been tested for a number of example use cases, but not with an exhaustive test suite. Therefore, we currently do not intend to upload this library to a public index.
This repository provides the code base for a greedy, asynchronous surrogate model in the form of an UM-Bridge server. The interface can be used for different surrogate types, but the current applications rely on Gaussian Process Regression as implemented in the scikit-learn library. The basic idea of the implementation is that a surrogate can learn an input-output mapping from another model, from which it fetches new training data on demand. This demand is determined by the variance in the (probabilistic) surrogate prediction for a given input. If the variance is below a user-defined threshold, the surrogate prediction is returned. Otherwise, a new run of the data-generating model is triggered and its output is returned. In addition, the newly generated data is used to retrain the surrogate in an asynchronous background process. The described approach is particularly useful for multilevel-type algorithms. We have successfully employed it for a specialized version of the Multilevel Delayed Acceptance (MLDA) algorithm, a multilevel Markov Chain Monte Carlo sampler. For further details, we refer the reader to the accompanying publication, Scalable Bayesian Inference of Large Simulations via Asynchronous Prefetching Multilevel Delayed Acceptance (to be published).
The library in this repository is a Python package readily installable via pip, simply run
pip install .For development, we recommend using the great uv project management tool, for which we provide a universal lock file. To set up a reproducible environment, run
uv sync --all-groupsThe documentation provides further information regarding usage, technical setup and API. Alternatively, you can check out the runnable examples.
This Software is distributed under the MIT license.
It has been developed as part of the innovation study ScalaMIDA. It has received funding through the Inno4scale project, which is funded by the European High-Performance Computing Joint Un- dertaking (JU) under Grant Agreement No 101118139. The JU receives support from the European Union’s Horizon Europe Programme.