Structural Kernel Search via Bayesian Optimization and Symbolical Optimal Transport
Authors: Matthias Bitzer, Mona Meister, Christoph Zimmer
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the following section, we show experimental results for our novel meta-GP model and kernel search method. We evaluate our meta-model on a meta-regression task where we predict test kernel-log-evidence pairs based on training pairs. Secondly, we consider kernel search and compare it against the greedy method in [3], the evolutionary algorithm in [4] and against the BO method of [9]. |
| Researcher Affiliation | Industry | Matthias Bitzer Mona Meister Christoph Zimmer Bosch Center for Artificial Intelligence, Renningen, Germany {matthias.bitzer3,mona.meister,christoph.zimmer}@de.bosch.com |
| Pseudocode | Yes | A complete description of the method can be found in Algorithm 1 in Appendix A. While an evolutionary algorithm seems to be a computationally intense procedure that takes place in each BO iteration, we emphasize that the evaluation of the acquisition function is very efficient for our proposed method. The reason for the efficiency is that the evaluations of our kernel-kernel KSOT (k1, k2) at two kernels k1, k2 K is very cheap, compared for example to the method in [9] (see Appendix E for computational time comparision). |
| Open Source Code | Yes | The implementation of our method is available at https://github.com/boschresearch/bosot. |
| Open Datasets | Yes | Furthermore, we consider the following publicly available datasets: Airline, LGBB, Airfoil, Powerplant, Concrete. |
| Dataset Splits | No | The paper specifies 'training set' and 'test set' sizes and generation (e.g., 'ntrain + ntest kernels in Kcomplete', 'divide the set uniformly into Ktrain and Ktest'), but does not explicitly mention a 'validation set' or provide its split details. |
| Hardware Specification | No | The paper analyzes performance over 'CPU-time' but does not specify any particular CPU model, GPU, or other hardware components used for the experiments. |
| Software Dependencies | No | The paper states 'The implementation of both BO methods is based on GPflow [10]', but it does not provide specific version numbers for GPflow or any other software dependencies. |
| Experiment Setup | Yes | We always use the Laplace approximation to calculate the log-model evidence, where we use 10 repeats to do MAP estimation of the kernel parameters. Both BO methods run for 50 iterations and the kernel-kernel hyperparameters are updated in each iteration via marginal likelihood maximization. Our method applied the evolutionary Algorithm 2 (see Appendix A) to optimize its acquisition function, using a population size of 100. |