Accelerated Exploration of High-Dimensional Parameter Spaces with MCMC Algorithms
When
Markov Chain Monte Carlo methods are widely used to infer probability distributions in model parameter space given a likelihood function (e.g., the likelihood of a model matching observed data). As models increase in dimensionality, however, both MCMC and alternate methods typically have very poor performance scaling. In this informal talk, we discuss the reasons for this performance scaling and introduce three new MCMC techniques that offer improved performance. The last of these new techniques (Searchlight) involves a combination of MCMCs and search algorithms that also solves longstanding issues with exploring probability surfaces that have multiple minima.
https://arizona.zoom.us/j/91826900125 Password: "Locute"