New simulations are drawn from a proposal density based on the current posterior approximation. BUT when the posterior is targeted directly, using a proposal distribution p ̃(θ) different from the prior requires a correction step without it, the posterior under the proposal distribution would be inferred. The so-called proposal posterior:
https://arxiv.org/pdf/1605.06376.pdf, target the proposal posterior p ̃(θ|x) by minimizing the
log likelihood loss −sum log qψ(θn|xn), and then post-hoc solves for p(θ|x).
https://arxiv.org/abs/1711.01861, directly recover p(θ|x) with the weighted loss - sum p(θn)/p ̃(θn) log qψ(θn|xn). However, the importance weights can have high variance during training, leading to inaccurate inference for some tasks.
(from https://arxiv.org/pdf/2101.04653.pdf and https://arxiv.org/pdf/1903.00007.pdf)
Different strategies:
Sequential:
Bayesian optimization. It requires two ingredients: