SQUIDS Seminar: Sampling as Bandits: Evaluation-Efficient Design for Black-Box Densities
| Dates: | 21 January 2026 |
|---|
| Times: | 14:00 - 15:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Department of Mathematics |
|---|
| Who is it for: | University staff, Current University students |
|---|
| Speaker: | Takuo Matsubara |
|---|
|
|
Speaker: Takuo Matsubara (University of Edinburgh)
Abstract: We introduce Bandit Importance Sampling (BIS), a novel class of importance sampling methods designed for scenarios where the target density is computationally expensive to evaluate. In contrast to adaptive importance sampling, which optimises a proposal distribution, BIS directly optimises samples through a sequential selection process combined with multi-armed bandits. BIS serves as a general framework accommodating bespoke bandit strategies. Crucially, we establish guarantees of weak convergence for the weighted samples, which hold regardless of the choice of bandit strategy. We present a practical strategy that leverages Gaussian process surrogates to guide sample selection, enabling efficient exploration of the parameter space with a minimal number of target evaluations. BIS yields accurate approximations with fewer target evaluations, demonstrating superior performance across multimodal, heavy-tailed distributions, and real-world Bayesian inference tasks involving computationally intensive models.
Speaker
Takuo Matsubara
Organisation: University of Edinburgh
Travel and Contact Information
Find event
Frank Adams Room 2
Alan Turing Building
Manchester
