Title: Inference with Hamiltonian sequential Monte Carlo
Authors: Martin Burda - University of Toronto (Canada) [presenting]
Remi Daviet - University of Toronto (Canada)
Abstract: A key problem with traditional MCMC methods using Metropolis-Hastings or Gibbs sampling is slow mixing of the chain under multimodality, concentrated mass or a complex shape of the objective function. A new Monte-Carlo method is proposed by combining the advantages of Sequential Monte Carlo simulators and Hamiltonian Monte Carlo simulators. The result is robust to multimodality and complex shapes of the objective function. We further enhance the method with a kernel-based resampling step to enhance robustness and efficiency. We present several challenging simulated examples.