Edward A. Roualdes [Chico State]

Friday Oct 3, 2025, 11:00 AM → 12:00 PM America/New_York

3rd Floor Classroom (162 Fifth Avenue )

Discussion lead: Edward A. Roualdes [Chico State]

Title: KLHR Sampler

Abstract: Edward will introduce a work in progress MCMC sampling algorithm, which combines ideas from both the Hit and Run Sampler/NURS and Metropolis-Hastings.  Similar to the Hit and Run Sampler, random lines through the latent space of the target distribution are sampled.  Along each random line, a proposal distribution is fit by minimizing the KL-divergence.  A proposal is then sampled and Metropolis-Hastings filtered.  During warmup, the principal component axes are learned, such that the randomly generated lines are more likely to align with the principal component axes.  At the expense of computational efficiency/speed, the stepsize is automatically tuned per step by learning a proposal distribution along the randomly generated lines. Feedback is welcome.

Links: (1) https://roualdes.us/lectures/misc/klhr/main.html#/title-slide

(2) https://github.com/roualdes/klhr