Generalized Bayesian Likelihood-Free Inference Using Scoring Rules Estimators

We propose a framework for Bayesian Likelihood-Free Inference (LFI) based on Generalized Bayesian Inference. To define the generalized posterior, we use Scoring Rules (SRs), which evaluate probabilistic models given an observation. As in LFI we can sample from the model (but not evaluate the likelihood), we employ SRs with easy empirical estimators. Our framework includes novel approaches and popular LFI techniques (such as Bayesian Synthetic Likelihood), which benefit from the generalized Bayesian interpretation. Our method enjoys posterior consistency in a well-specified setting when a strictly-proper SR is used (i.e., one whose expectation is uniquely minimized when the model corresponds to the data generating process). Further, we prove a finite sample generalization bound and outlier robustness for the Kernel and Energy Score posteriors, and propose a strategy suitable for the LFI setup for tuning the learning rate in the generalized posterior. We run simulations studies with pseudo-marginal Markov Chain Monte Carlo (MCMC) and compare with related approaches, which we show do not enjoy robustness and consistency.