likelihood free estimation vs test-quantities driven MCMC

[hyunjimoon]

Likelihood-free estimation and test-quantities driven MCMC share spirit in that we don't believe the structure of the model too much (e.g. whether the parameter it self should perfectly be recovered).

I think these attempts are exactly equivalent to parameterization attempts then using MCMC which passes test-diagnostics based SBC.

Brehmer20_Mining gold from implicit models to improve likelihood-free inference.pdf
Cranmer16_Approximating Likelihood Ratios with Calibrated Discriminative Classifiers.pdf