Reproduce “Enhancing Gradient-based Discrete Sampling via Parallel Tempering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luxu Liang, Yuhang Jia, Feng Zhou
PT-DMALA runs discrete Langevin chains at different temperatures and exchanges states, allowing exploratory chains to help the target chain move between modes. On a 16-component discretized Gaussian mixture, its MMD was 0.824 ± 0.031 × 10⁻³, versus 1.691 ± 0.042 × 10⁻³ for the strongest listed baseline, indicating better distribution matching in that test, not lower runtime.
Key insights
- Integrating parallel tempering with discrete Langevin proposals effectively enables samplers to escape local minima and explore broader regions of multimodal discrete energy landscapes.
- A specially designed Metropolis swap criterion is essential for ensuring detailed balance and facilitating robust mode jumping within the discrete state space.
- The optimal temperature schedule and number of chains for parallel tempering can be adaptively determined through an automatic tuning scheme, enhancing practical applicability and reducing manual effort.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.