Minimizing UCB: a Better Local Search Strategy in Local Bayesian Optimization
Authors: ZHEYI FAN, Wenyu Wang, Szu Hui Ng, Qingpei Hu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our algorithms on different synthetic and real-world functions, and the results show the effectiveness of our method. Our algorithms also illustrate improvements on local search strategies from an upper bound perspective in Bayesian optimization, and provides a new direction for future algorithm design. [...] In this section we apply varies experimental settings to assess the efficacy of our proposed two algorithms, Min-UCB and LA-Min UCB, relative to two established methodologies. |
| Researcher Affiliation | Academia | Zheyi Fan1,2, Wenyu Wang3, Szu Hui Ng3 , Qingpei Hu1,2 1Academy of Mathematics and Systems Science, Chinese Academy of Sciences, China 2School of Mathematical Sciences, University of Chinese Academy of Sciences, China 3Department of Industrial Systems Engineering & Management, National University of Singapore, Singapore |
| Pseudocode | Yes | Algorithm 1: Min UCB: Local Bayesian Optimization through Minimizing UCB [...] Algorithm 2: LA-Min UCB: Look Ahead Bayesian Optimization through Minimizing UCB |
| Open Source Code | Yes | Our code is based on Nguyen et al. [23], where they provide the program of GIBO, MPD, and various of objective functions. Our experimental settings on synthetic (Sect. 4.1) and reinforcement learning (Sect. 4.2) objectives are same as their papers. Each algorithm is executed a total of ten times for every objective function that we examine, initiating from an identical set of starting points sampled across the bounded domain via a Sobol sequence. We illustrate the results in Fig. (2) (synthetic objective) and (3) (reinforcement learning objective). The figures shows the mean of current best reward on the number of queries with an error bar (defined through the standard derivation). Our experimental framework was executed on a workstation of 20 Intel Xeon CPU cores, with a 32GB of memory. The codes can be viewed on https://github.com/chinafzy1/Minimizing-UCB. |
| Open Datasets | Yes | In our first experiments, we focused on optimizing synthetic objective functions within the ddimensional unit hypercube [0, 1]d. These functions were generated by sampling from a Gaussian Process (GP) with a Radial Basis Function (RBF) kernel. [...] In this experiment, we turned to reinforcement learning, specifically Mu Jo Co-based locomotion tasks [30]. [...] The second function addresses a cosmological issue[24], where the objective is to fine-tune a cosmological model or physical simulator to match observational data from the cosmos. [...] The third function concerns rover trajectory planning [33]. |
| Dataset Splits | No | Each algorithm is executed a total of ten times for every objective function that we examine, initiating from an identical set of starting points sampled across the bounded domain via a Sobol sequence. [...] Each experiment was allocated a budget of 500 function evaluations, under the dimension {25, 50, 100} separately. The paper does not specify explicit training/validation/test dataset splits. |
| Hardware Specification | Yes | Our experimental framework was executed on a workstation of 20 Intel Xeon CPU cores, with a 32GB of memory. |
| Software Dependencies | No | To handle this task we use the idea of one-shot optimization in Bo Torch [2], that transform the original problem into a deterministic optimization problem with fixed sampling. This method is first proposed to solve the optimization of Knowledge Gradient. However, considering that our method is very similar to Knowledge Gradient, their optimization method is also applicable to our problem. For more information about this one-shot optimization method, please refer to the Section 4.2 in Balandat et al. [2]. The paper mentions software tools like BoTorch but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Each algorithm is executed a total of ten times for every objective function that we examine, initiating from an identical set of starting points sampled across the bounded domain via a Sobol sequence. [...] Each experiment was allocated a budget of 500 function evaluations, under the dimension {25, 50, 100} separately. [...] As for our Min UCB and LA-Min UCB, we use a fixed batch size in each experiment, i.e. b(1) t , b(2) t and bt are unchanged in each experiment. The UCB coeffieient βt in Min UCB and LA-Min UCB is fixed as 3, which has been shown to be able to balance the convergence speed and accuracy of the algorithm. |