Thermostat-assisted Continuous-tempered Hamiltonian Monte Carlo for Multimodal Posterior Sampling on Large Datasets

In this paper, we propose a new sampling method named as the thermostat-assisted continuous-tempered Hamiltonian Monte Carlo for multimodal posterior sampling on large datasets. It simulates a noisy system, which is augmented by a coupling tempering variable as well as a set of Nosu0027e-Hoover thermostats. This augmentation is devised to address two main issues of concern: the first is to effectively generate i.i.d. samples from complex multimodal posterior distributions; the second is to adaptively control the system dynamics in the presence of unknown noise that arises from the use of mini-batches. The experiment on synthetic distributions has been performed; the result demonstrates the effectiveness of the proposed method.