An Automatic Approximate Bayesian Computation Approach Using Metric Learning

摘要

Recent progress in Bayesian inference has allowed for accurate posterior estimations in complex situations with no idea about a likelihood function. Currently, Approximate Bayesian Computation (ABC) techniques have emerged as a widely used set of free-likelihood methods. Although there is a large number of different ABC-based approaches across the literature, all they have in common a hard dependence on free parameters selection, demanding for expensive tuning procedures such as grid search or cross-validation. Here, we introduce an Automatic Metric Learning-based ABC approach, termed AML-ABC. Namely, AML-ABC matches the simulation and observation spaces within an ABC-based framework. Attained results on a synthetic dataset and a real-world ecological system show that our approach is a competitive method compared to other non-automatic state-of-the-art ABC techniques.