End-to-End Meta-Bayesian Optimisation with Transformer Neural Processes

Authors: Alexandre Maraval, Matthieu Zimmer, Antoine Grosnit, Haitham Bou Ammar

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that our method achieves state-of-the-art regret results against various baselines in experiments on standard hyperparameter optimisation tasks and also outperforms others in the real-world problems of mixed-integer programming tuning, antibody design, and logic synthesis for electronic design automation.
Researcher Affiliation Collaboration Alexandre Maraval Huawei Noah s Ark Lab alexandre.maraval1@huawei.com Matthieu Zimmer Huawei Noah s Ark Lab matthieu.zimmer@huawei.com Antoine Grosnit Huawei Noah s Ark Lab Technische Universität Darmstadt antoine.grosnit2@huawei.com Haitham Bou Ammar Huawei Noah s Ark Lab, University College London haitham.ammar@huawei.com
Pseudocode Yes Algorithm 1 Neural Acquisition Process training.
Open Source Code No The paper does not provide an explicit statement or a link to the open-source code for the methodology described in this paper.
Open Datasets Yes We experiment on the HPO-B benchmark [44], which contains datasets of (classification) model hyperparameters... We use the open-source SCIP solver [45] and the Benchmark suite from the MIPLib2017 [46]... We collected datasets of CDRH3 sequences and their respective binding energies... from the protein database bank [49]... We collected datasets for 43 different circuits. Each dataset consisted of 500 sequences... from Open ABC [50].
Dataset Splits Yes We train our model on data collected from BO traces on 103 MIPs and test on a held-out set of 42 instances... we meta-learn on 109 datasets, validate on 16, and test NAP on 32 new antigens... We train all methods on 30 circuits from Open ABC [50], validate on 4 and test on 9.
Hardware Specification No The paper mentions running experiments 'on the same GPU' but does not specify any particular GPU model, CPU, or other hardware details.
Software Dependencies No The paper mentions using 'open-source SCIP solver [45]' and 'open-source ABC library [51]' but does not provide specific version numbers for these or any other software dependencies crucial for reproduction.
Experiment Setup No The paper mentions '5 initial points' for the experiments (Figure 2 caption) but does not provide specific hyperparameter values (e.g., learning rate, batch size, number of epochs) or detailed system-level training settings for their proposed NAP model.