An algorithm for sampling from non-log-concave multivariate distributions isproposed, which improves the adaptive rejection Metropolis sampling (ARMS)algorithm by incorporating the hit and run sampling. It is not rare that theARMS is trapped away from some subspace with significant probability in thesupport of the multivariate distribution. While the ARMS updates samples onlyin the directions that are parallel to dimensions, our proposed method, the hitand run ARMS (HARARMS), updates samples in arbitrary directions determined bythe hit and run algorithm, which makes it almost not possible to be trapped inany isolated subspaces. The HARARMS performs the same as ARMS in a singledimension while more reliable in multidimensional spaces. Its performance isillustrated by a Bayesian free-knot spline regression example. We showed thatit overcomes the well-known `lethargy' property and decisively find the globaloptimal number and locations of the knots of the spline function.
展开▼
