Transition Constrained Bayesian Optimization via Markov Decision Processes
Authors: Jose Pablo Folch, Calvin Tsay, Robert Lee, Behrang Shafei, Weronika Ormaniec, Andreas Krause, Mark van der Wilk, Ruth Misener, Mojmir Mutny
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We showcase applications in chemical reactor optimization, informative path planning, machine calibration, and other synthetic examples. We analyze the scheme theoretically and empirically demonstrate its practicality on problems in physical systems, such as electron laser calibration and chemical reactor optimization. |
| Researcher Affiliation | Collaboration | Jose Pablo Folch Imperial College London London, UK; Robert M Lee BASF SE Ludwigshen, Germany; Weronika Ormaniec ETH Zurich Zurich, Switzerland |
| Pseudocode | Yes | Algorithm 1 Transition Constrained Bayes Opt via MDPs |
| Open Source Code | No | We note that the implementation code will be made public after public review. The code will be made public upon acceptance, and an anonymized version is included for the review process. |
| Open Datasets | Yes | Samaniego et al. [20] investigated automatic monitoring of Lake Ypacarai, and Folch et al. [11] and Yang et al. [40] benchmarked different Bayes Opt algorithms for the task of finding the largest contamination source in the lake. We use the simulator from Mutný et al. [67] that optimizes quadrupole magnet orientations for our experiment with varying noise levels. Our chemical reactor benchmark synthetizes Knorr pyrzole in a transient flow reactor. ... The kernel is based on the following ODE model, which is well known in the chemistry literature and given in [27]. |
| Dataset Splits | No | The paper does not explicitly provide training/validation/test dataset splits, specific percentages, or sample counts for the data used in the experiments. |
| Hardware Specification | Yes | The times were obtained in a simple 2015 Mac Book Pro 2.5 GHz Quad-Core Intel Core i7. The bulk of the experiments was ran in parallel on a High Performance Computing cluser, equipped with AMD EPYC 7742 processors and 16GB of RAM. |
| Software Dependencies | No | The paper does not provide specific software dependency versions (e.g., Python, PyTorch, or specific library versions) beyond implicitly using standard machine learning frameworks. |
| Experiment Setup | Yes | For each benchmark, we selected reasonable GP hyper-parameters and fixed them during the optimization. These are summarized in Appendix E.2. Table 2: Benchmark and hyper-parameter information. For the knorr pyrazole synthesis example, we further set αode = 0.6, αrbf = 0.001, k1 = 10, k2 = 874, k3 = 19200, αsig = 5. The number of features for each experiment, M, is set to be M = |X| in the discrete cases and M = min 25+d, 512 where d is the problem dimensionality. In the case of Local Search Region Bayes Opt (LSR) [22] we set the exploration hyper-parameter to be γ = 0.01 in all benchmarks. |