Principled Preferential Bayesian Optimization
Authors: Wenjie Xu, Wenbin Wang, Yuning Jiang, Bratislav Svetozarevic, Colin Jones
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on sampled instances from Gaussian processes, standard test functions, and a thermal comfort optimization problem all show that our method stably achieves better or competitive performance as compared to the existing state-of-the-art heuristics, which, however, do not have theoretical guarantees on regret bounds or convergence. |
| Researcher Affiliation | Academia | 1Automatic Control Laboratory, EPFL, Lausanne, Switzerland 2Urban Energy Systems Laboratory, Empa, Zurich, Switzerland 3The Institute for Artificial Intelligence Research and Development of Serbia, Serbia. |
| Pseudocode | Yes | Algorithm 1 Principled Optimistic Preferential Bayesian Optimization (POP-BO). |
| Open Source Code | Yes | 2Code link: https://github.com/PREDICT-EPFL/POP-BO |
| Open Datasets | Yes | We compare our method to several well-known global optimization test functions (Dixon, 1978; Molga & Smutnicki, 2005) and To emulate real human thermal sensation, we use the well-known and widely adopted Predicted Mean Vote (PMV) model (Fanger et al., 1970) as the ground truth and generate the preference feedback according to the Bernoulli model as assumed in Assumption 2.5. |
| Dataset Splits | No | The paper describes running its algorithm for a fixed number of steps (e.g., 30 or 100 steps) and across multiple instances or random initial points for optimization, but it does not specify explicit train/validation/test dataset splits with percentages, counts, or predefined partitions. |
| Hardware Specification | Yes | The average update time per step is only 18.0 seconds on a personal computer with one Intel64 Family 6 Model 142 Stepping 12 Genuine Intel 1803 Mhz processor and 16.0 GB RAM. |
| Software Dependencies | No | The paper mentions software like GPy (GPy, since 2012), Cas ADi (Andersson et al., 2019) and Ipopt (W ächter & Biegler, 2006), but it does not provide specific version numbers for these software packages or libraries. |
| Experiment Setup | Yes | We set β = β0 t, where β0 is set to 1.0 by default. For the sampled instances from Gaussian processes, the lengthscale is set to be the ground truth and the norm bound is set to be 1.1 times the ground truth. For the test function examples, we choose the lengthscale by maximizing the likelihood value over a set of randomly sampled data and set the norm bound to be 6 by default (with the test functions all normalized). |