My research highlight for the day was a visit by Eric Vanden-Eijnden (NYU) in Math. He showed me (among other things) some new MCMC methods that can make use of dynamics to sample difficult probability distributions or compute fully marginalized likelihoods (evidences). They involve dissipative dynamics, not fully Hamiltonian dynamics like Hamiltonian methods. We resolved to try them out on the binary-star problem that we solved with The Joker, because these problems are multi-modal but in ways we understand.
No comments:
Post a Comment