Reinforced Few-Shot Acquisition Function Learning for Bayesian Optimization

Authors: Bing-Jing Hsieh, Ping-Chun Hsieh, Xi Liu

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments, we demonstrate that the FSAF achieves comparable or better regrets than the state-of-the-art benchmarks on a wide variety of synthetic and real-world test functions. and 4 Experimental Results
Researcher Affiliation Collaboration 1Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan 2Applied Machine Learning, Facebook AI, Menlo Park, CA, USA
Pseudocode Yes The pseudo code of the training procedure of FSAF is provided in Appendix A.
Open Source Code Yes The source code for our experiments has been made publicly available4. 4https://github.com/pinghsieh/FSAF.
Open Datasets Yes We proceed to evaluate FSAF on test functions obtained from five open-source real-world tasks in different application domains. and The detailed description of the datasets is in Appendix C. In this setting, we consider 1-shot adaptation for FSAF, a rather sample-efficient scenario of few-shot learning. From Figure 3, we observe that FSAF remains the best or among the best for all the five real-world test functions, despite the salient structural differences of the datasets. and For training, we construct a collection of training tasks, each of which is a class of GP functions with either an RBF, Matern-3/2, or a spectral mixture kernel with different parameters (e.g., lengthscale and periods).
Dataset Splits Yes For any initial model parameters θ and training set Dtr of a task τ, let M(θ, Dtr τ) be an algorithm that outputs the adapted model parameters by applying few-shot fast adaptation to θ based on Dtr τ. The performance of the adapted model is evaluated on Dval τ by a meta-loss function L(M(θ, Dtr τ), Dval τ ).
Hardware Specification Yes Our experiments are conducted on NVIDIA GeForce RTX 2080 Ti GPUs. The average training time for FSAF is around 25 GPU-hours.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions like PyTorch or TensorFlow).
Experiment Setup Yes Configuration of FSAF. For training, we construct a collection of training tasks, each of which is a class of GP functions with either an RBF, Matern-3/2, or a spectral mixture kernel with different parameters (e.g., lengthscale and periods). For the reward design of FSAF, we use g(z) = log z to encourage high-accuracy solutions. We choose N = 5, K = 1, and S = 1 given the limitation of GPU memory. For testing, we use the model with the best average total return during training as our initial model, which is later fine-tuned via few-shot fast adaptation for each task. For a fair comparison, we ensure that FSAF and Meta BO-T use the same amount of meta-data in each experiment.