Efficient Gradient-Free Variational Inference using Policy Search
Authors: Oleg Arenz, Gerhard Neumann, Mingjun Zhong
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate VIPS with respect to efficiency of the optimization as well as quality of the learned approximations. For assessing efficiency, we focus on the number of function evaluations, but also include a comparison with respect to the wall-clock time. As the ELBO objective is hard to use for comparisons as it depends on the current sample set, we assess the quality of the approximation by comparing samples drawn from the learned model with groundtruth samples based on their Maximum Mean Discrepancy (MMD) (Gretton et al., 2012). |
| Researcher Affiliation | Academia | 1Computational Learning for Autonomous Systems, TU Darmstadt, Darmstadt, Germany 2Machine Learning Lab, University of Lincoln, Lincoln, UK 3Lincoln Center for Autonomous Systems, University of Lincoln, Lincoln, UK. |
| Pseudocode | Yes | Algorithm 1 Variational Inference by Policy Search |
| Open Source Code | Yes | The full algorithm is outlined in Algorithm 1. An open-source implementation is available online1. 1https://github.com/OlegArenz/VIPS |
| Open Datasets | Yes | We perform Bayesian logistic regression on the German Credit and Breast Cancer datasets (Lichman, 2013) |
| Dataset Splits | No | The paper evaluates on standard datasets and compares to other methods, but it does not explicitly provide specific training/validation/test dataset splits (e.g., percentages, sample counts, or explicit splitting methodology). |
| Hardware Specification | No | Calculations for this research were conducted on the Lichtenberg high performance computer of the TU Darmstadt. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) for its own implementation. |
| Experiment Setup | No | The paper mentions general experimental settings like using the same hyper-parameters for all experiments and heuristics for adding/deleting components, but it does not provide concrete hyperparameter values (e.g., learning rate, batch size, epochs) or detailed optimizer settings in the main text. |