DR-ABC: Approximate Bayesian Computation with Kernel-Based Distribution Regression

Authors: Jovana Mitrovic, Dino Sejdinovic, Yee-Whye Teh

ICML 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experimental Results Toy example. The first problem we study is the following Gaussian hierarchical model... In our experiments, we compare full and conditional DR-ABC against SA-ABC and K2-ABC. ... For the performance metric, we calculate the mean square error (MSE) of the parameter of interest on synthetic data. ... Figure 1 describes the performance of our chosen methods across different numbers of particles...
Researcher Affiliation Academia Jovana Mitrovic MITROVIC@STATS.OX.AC.UK Dino Sejdinovic DINO.SEJDINOVIC@STATS.OX.AC.UK Yee Whye Teh Y.W.TEH@STATS.OX.AC.UK Department of Statistics, University of Oxford
Pseudocode Yes Algorithm 1 Distribution Regression... Algorithm 2 Conditional Distribution Regression... Algorithm 3 DR-ABC Algorithm
Open Source Code Yes The code for all presented experiments is available at https://github.com/jovana-mitrovic/dr-abc.
Open Datasets Yes As an example of an ecological system with a dynamic generative process, we examine the problem of inferring the dynamics of the adult blowfly population as introduced in Wood (2010).
Dataset Splits Yes The hyperparameters in the two DRABC methods are set via five-fold cross-validation on appropriately defined grids.
Hardware Specification No The paper discusses computational complexity and efficiency (e.g., 'significantly reduced computational cost'), but it does not specify any hardware details such as GPU or CPU models used for the experiments.
Software Dependencies No The paper refers to code availability on GitHub but does not explicitly list any software dependencies with specific version numbers (e.g., Python version, specific libraries, or frameworks).
Experiment Setup Yes The hyperparameters in the two DRABC methods are set via five-fold cross-validation on appropriately defined grids. For the grids of the different kernel bandwidth parameters, we multiply the respective median heuristics (Reddi et al., 2014) with a set of ten equally spaced points between 10 4 and 1000. For λ and ϵ, we choose the grids by exponentiating 10 to the powers given by ten equally spaced points between 4 and 1. ... We use either 100 or 200 particles in (conditional) distribution regression.