Mixed-Variable Black-Box Optimisation Using Value Proposal Trees
Authors: Yan Zuo, Vu Nguyen, Amir Dezfouli, David Alexander, Benjamin Ward Muir, Iadine Chades
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that this approach significantly outperforms existing mixed-variable optimisation approaches across several mixed-variable black-box optimisation tasks. and We evaluated the optimisation performance of our proposed VPT method against existing methods. |
| Researcher Affiliation | Collaboration | 1 Amazon 2 CSIRO |
| Pseudocode | Yes | Algorithm 1: Value Proposal Trees |
| Open Source Code | No | The paper mentions that 'All baseline methods are implemented according to their publicly available Python repositories' but does not state that the code for the proposed VPT method is open-source or provide a link. |
| Open Datasets | Yes | For the SVM-Boston, XG-MNIST and NASBench-101 datasets, we performed 10 random trials. and Func2C, Func3C and Reizman-Suzuki datasets. |
| Dataset Splits | No | The paper states 'We performed 20 random trials for the Func2C, Func3C and Reizman-Suzuki datasets' and 'For the SVM-Boston, XG-MNIST and NASBench-101 datasets, we performed 10 random trials', but does not provide specific train/validation/test dataset splits or cross-validation details for reproduction. |
| Hardware Specification | Yes | all experiments were conducted on an 8-core 3.4Ghz Intel Xeon processor with 64GB RAM. |
| Software Dependencies | No | The paper mentions using Python for baselines, DBSCAN for clustering, and Extremely Randomised Trees (ERT) for the regression model, but does not specify version numbers for any software dependencies. |
| Experiment Setup | Yes | VPT Settings We utilise the same mixed-kernel GP surrogate as (Wan et al. 2021), which uses a fixed kernel mixture hyperparameter of λ = 0.5. We follow the optimisation approach of (Wan et al. 2021), optimising GP hyperparameters by maximising the log marginal likelihood using variational inference (VI) (Ranganath, Gerrish, and Blei 2014), with a learning rate of α = 0.03. At each optimisation iteration, we generate N = 1000 randomly selected candidates, sampling from a local trust region of the current incumbent. For the tree-based regression model, we use the Extremely Randomised Trees (ERT) approach (Geurts, Ernst, and Wehenkel 2006), constructing an ensemble of 100 trees, each with a maximum depth equal to the number of categorical variables in the black-box function of interest. Clustering is performed using DBSCAN (Ester et al. 1996). We use the proximity matrix generated from the ERT model to specify ϵ, where we use the 20th percentile of distances for the value of ϵ. We set the minimum number of points clustered around a region for it to be considered dense as 5. |