Local Bayesian optimization via maximizing probability of descent

Authors: Quan Nguyen, Kaiwen Wu, Jacob Gardner, Roman Garnett

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

Reproducibility Variable Result LLM Response
Research Type Experimental We now present results from extensive experiments that evaluate our method MPD against three baselines: (1) GIBO [18], which performs local BO by minimizing the trace of the posterior covariance matrix of the gradient and uses the expected gradient in the update step; (2) ARS [16], which estimates the gradient of the objective via finite difference with random perturbations; and (3) Tu RBO [5], a trust region-based Bayesian optimization method.
Researcher Affiliation Academia Quan Nguyen 1 Kaiwen Wu 2 Jacob R. Gardner2 Roman Garnett1 1Washington University in St. Louis 2University of Pennsylvania {quan,garnett}@wustl.edu {kaiwenwu,jacobrg}@seas.upenn.edu
Pseudocode Yes Algorithm 1 Local BO via MPD
Open Source Code Yes Our implementation is available at https://github.com/kayween/local-bo-mpd.
Open Datasets Yes The second set of experiments are reinforcement learning Mu Jo Co locomotion tasks [23], where each task involves learning a linear policy that maps states to actions to maximize the reward received from the learning environment. We use the same three environments in Müller et al. [18], Cart Pole-v1 with 4 parameters, Swimmer-v1 with 16, and Hopper-v1 with 33, to evaluate the methods and show the results in Fig. 3. ... The second is a problem from cosmology [20] ... Our third objective function uses the rover trajectory planning problem [27].
Dataset Splits No The paper describes the budget of function evaluations for each run (e.g., 'Each run has a budget of 500 function evaluations.') and how starting points are sampled, but does not provide explicit train/validation/test splits for the objective functions themselves, which are black-box.
Hardware Specification Yes Experiments were performed on a small cluster built from commodity hardware comprising approximately 200 Intel Xeon CPU cores (no GPUs), with approximately 10 GB of RAM available to each core.
Software Dependencies No The paper mentions extending an existing codebase for implementation but does not specify software dependencies with version numbers (e.g., Python, specific libraries like PyTorch or TensorFlow versions).
Experiment Setup Yes In our experiments, we set the step size to δ = 0.001 and the descent probability threshold to p = 65%, which we find to work well empirically. (Sect. 3.2) and For each objective function tested, we run each algorithm ten times from the same set of starting points sampled from a Sobol sequence over the (box-bounded) domain. In each of the following plots, we show the progressive mean objective values as a function of the number of queries with error bars indicating (plus or minus) one standard error. For the other objectives (Sect. 4.3), we set the number of samples to learn about the gradient per iteration M = 1. (Sect. 4)