Batch Multi-Fidelity Active Learning with Budget Constraints
Authors: Shibo Li, Jeff M Phillips, Xin Yu, Robert Kirby, Shandian Zhe
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For evaluation, we examined BMFAL-BC in five real-world applications, including three benchmark tasks in physical simulation (solving Poisson s, Heat and viscous Burger s equations), a topology structure design problem, and a computational fluid dynamics (CFD) task to predict the velocity field of boundary-driven flows. We compared with the budget-aware version of DMFAL, single multi-fideity querying with our acquisition function, and several random querying strategies. Under the same budget constraint, our method consistently outperforms the competing methods throughout the learning process, often by a large margin. |
| Researcher Affiliation | Academia | Shibo Li , Jeff M. Phillips , Xin Yu, Robert M. Kirby, and Shandian Zhe School of Computing, University of Utah Salt Lake City, UT 84112 {shibo, jeffp, xiny, kirby, zhe}@cs.utah.edu |
| Pseudocode | Yes | Algorithm 1 Weighted-Greedy( {λm}, budget B) |
| Open Source Code | No | The paper does not provide a direct link to source code or an explicit statement in the main body that the code for their methodology is released. While the checklist indicates |
| Open Datasets | No | The paper describes generating its own training and test data using numerical solvers and meshes, rather than using publicly available datasets with specific access information. For example, |
| Dataset Splits | No | The paper describes how initial training data was collected and how evaluation was performed, but it does not specify explicit training/validation/test dataset splits (e.g., percentages or sample counts for each split) or refer to predefined splits from standard benchmarks. |
| Hardware Specification | No | The paper mentions that |
| Software Dependencies | No | The paper mentions |
| Experiment Setup | Yes | All the methods were implemented by Pytorch (Paszke et al., 2019). We followed the same setting as in Li et al. (2022) to train the deep multi-fidelity model (see Sec. 2.2), which employed a two-layer NN at each fidelity, tanh activation, and the layer width was selected from {20, 40, 60, 80, 100} from the initial training data. The dimension of the latent output was 20. The learning rate was tuned from {10 4, 5 10 4, 10 3, 5 10 3, 10 2}. We set the budget for acquiring each batch to 20 (normalized seconds), and ran each method to acquire 25 batches of training examples. |