Kernel Adaptive Metropolis-Hastings
Authors: Dino Sejdinovic, Heiko Strathmann, Maria Lomeli Garcia, Christophe Andrieu, Arthur Gretton
ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Kernel Adaptive Metropolis Hastings outperforms competing fixed and adaptive samplers on multivariate, highly nonlinear target distributions, arising in both real-world and synthetic examples. We provide experimental comparisons with other fixed and adaptive samplers in Section 5, where we show superior performance in the context of Pseudo-Marginal MCMC for Bayesian classification, and on synthetic target distributions with highly nonlinear shape. |
| Researcher Affiliation | Academia | Gatsby Unit, CSML, University College London, UK and School of Mathematics, University of Bristol, UK |
| Pseudocode | Yes | MCMC Kameleon Input: unnormalized target π, subsample size n, scaling parameters ν, γ, adaptation probabilities {pt} t=0, kernel k, At iteration t + 1, 1. With probability pt, update a random subsample z = {zi}min(n,t) i=1 of the chain history {xi}t 1 i=0, 2. Sample proposed point x from qz( |xt) = N(xt, γ2I + ν2Mz,xt HM z,xt), where Mz,xtis given in Eq. (3) and H = I 1 n1n n is the centering matrix, 3. Accept/Reject with the Metropolis-Hastings acceptance probability A(xt, x ) in Eq. (4), ( x , w.p. A(xt, x ), xt, w.p. 1 A(xt, x ). |
| Open Source Code | Yes | Python implementation of MCMC Kameleon is available at https://github.com/karlnapf/ kameleon-mcmc. |
| Open Datasets | Yes | We consider the UCI Glass dataset (Bache & Lichman, 2013) |
| Dataset Splits | No | No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning was provided. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions 'Python implementation' but does not provide specific version numbers for Python or any other software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | (γ was fixed to 0.2), and which also stops adapting the proposal after the burn-in of the chain (in all experiments, we use a random subsample z of size n = 1000, and a Gaussian kernel with bandwidth selected according to the median heuristic). Each of these algorithms was run for 100,000 iterations (with a 20,000 burnin) and every 20th sample was kept. |