Proposal Distribution Tuning
Proposal scale trades off acceptance against step size: tiny steps accept nearly always but crawl, huge steps get rejected. For random-walk Metropolis the classic optimum is roughly 23% acceptance in high dimensions (about 44% in 1D), often reached by adapting the proposal covariance during warmup.
This Concept is waiting for its first lesson!
Proposal scale trades off acceptance against step size: tiny steps accept nearly always but crawl, huge steps get rejected. For random-walk Metropolis the classic optimum is roughly 23% acceptance in high dimensions (about 44% in 1D), often reached by adapting the proposal covariance during warmup.
Are you a teacher? Sign in to start contributing.
Sign In