Bidirectional Learning for Offline Infinite-width Model-based Optimization

Authors: Can Chen, Yingxueff Zhang, Jie Fu, Xue (Steve) Liu, Mark Coates

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various tasks verify the effectiveness of BDI.
Researcher Affiliation Collaboration 1Mc Gill University, 2 Huawei Noah s Ark Lab, 3 Beijing Academy of Artificial Intelligence
Pseudocode No The paper describes the method using mathematical equations and text but does not include a clearly labeled pseudocode or algorithm block.
Open Source Code No The code is available here. (In abstract). We include our code in the supplemental material as a new asset and we will release this new asset upon paper acceptance.
Open Datasets Yes To evaluate the effectiveness of BDI, we adopt the commonly used design-bench, including both continuous and discrete tasks, and the evaluation protocol as in the prior work [8].
Dataset Splits Yes We follow the training settings of [8] for all comparison methods if not specified.
Hardware Specification Yes All Jax experiments are run on multiple CPUs within a cluster and all Pytorch experiments are run on one V100 GPU.
Software Dependencies No We use the NTK library [29] build on Jax [30] to conduct kernel-based experiments and use Pytorch [31] for other experiments.
Experiment Setup Yes We set the regularization β as 1e 6 following [16]. We set the predefined target score yh as a constant 10 across all tasks, which proves to be effective and robust in all cases. We set α as 1e 3 for all continuous tasks and as 0.0 for all discrete tasks, and set the number of iterations T to 200 in all experiments. We adopt a 6-layer MLP (Multi Layer Perceptron) followed by Re LU for all gradient updating methods and the hidden size is set as 2048. We optimize L(Xh) with the Adam optimizer [21] with a 1e 1 learning rate for discrete tasks and a 1e 3 learning rate for continuous tasks.