Optimally-weighted Estimators of the Maximum Mean Discrepancy for Likelihood-Free Inference

Authors: Ayush Bharti, Masha Naslidnyk, Oscar Key, Samuel Kaski, Francois-Xavier Briol

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Section 5 demonstrates strong empirical performance on a range of simulators and provides extensive simulation studies on benchmark simulators. For example, Table 1 shows 'Average and standard deviation (in parenthesis) of estimated MMD2...computed over 100 runs'.
Researcher Affiliation Academia 1Department of Computer Science, Aalto University, Espoo, Finland 2Department of Statistical Science, University College London, London, United Kingdom 3Department of Computer Science, University of Manchester, Manchester, United Kingdom.
Pseudocode Yes The paper includes "Algorithm 1: Composite goodness-of-fit test" and "Algorithm 2: Random-restart optimiser" in Appendix B.4.
Open Source Code Yes Our code is available at https://github.com/bharti-ayush/optimally-weighted_MMD.
Open Datasets Yes The paper uses various benchmark simulators (e.g., g-and-k, Two moons, MA(2), M/G/1 queue, Lotka-Volterra) and real-world data (e.g., "US dollar to Canadian dollar exchange rate data (Verbeek, 2018) from the Ecdat R package", "multivariate g-and-k distribution introduced in (Drovandi & Pettitt, 2011)", "large scale offshore wind farm model (Niayifar & Porté-Agel, 2016; Kirby et al., 2023)").
Dataset Splits No No explicit mention of specific train/validation/test dataset splits (e.g., percentages or counts) or cross-validation setup for model training/evaluation. The paper discusses 'n' observed data points and 'm' simulated data points for MMD estimation, but not data partitioning for model validation.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are provided for the experimental setup. The paper mentions 'CPU hours' but without specific hardware types.
Software Dependencies Yes For drawing iid or RQMC points, we use the implementation from Sci Py (Virtanen et al., 2020).
Experiment Setup Yes The paper provides specific experimental setup details in Table 4, including 'hyperparameter value alpha 0.05 level of the test B 200 number of bootstrap samples m 100 number of samples from the simulator n 500 number of observations in the data I 50 number of initial parameters sampled R 10 number of initial parameters to optimise S 200 number of gradient steps s 0.04 step size'. Algorithm 2 also details the adam_optimizer with 'S' and 's'.