Adaptive Scheduling in MCMC and Probabilistic Programming

We introduce an adaptive output-sensitive inference algorithm for MCMC and probabilistic programming, Adaptive Random Database. The algorithm is based on a single-site updating Metropolis-Hasting sampler, the Random Database (RDB) algorithm. Adaptive RDB (ARDB) differs from the original RDB in that the schedule of selecting variables proposed for modification is adapted based on the output of of the probabilistic program, rather than being fixed and uniform. We show that ARDB still converges to the correct distribution. We compare ARDB to RDB on several test problems highlighting different aspects of the adaptation scheme.