MCMC
Lab
Algorithms
Random Walk Metropolis
Simple local proposals
Indep. Metropolis Hastings
Fixed global proposal
Slice Sampling
Adaptive step size
Elliptical Slice
For Gaussian priors
Hit-and-Run
Random direction lines
Hamiltonian Monte Carlo
Gradient-based physics
About RWM
Proposes a move nearby. Fails in high dimensions or correlated distributions (Random Walk behavior).
Gaussian
Bimodal Mixture
Donut (Ring)
Banana (Rosenbrock)
Start
SAMPLES:
0
Parameters
Step Size (Sigma)
0.5
Speed (Steps/Frame)
1
Pro Tip
Try the
Donut
distribution. Random Walk will get stuck, but
HMC
will glide effortlessly around the ring.