This event will be held virtually.
Speaker: Thanh Nguyen, ECpE Graduate Student
Advisor: Chinmay Hegde
Title: Black-Box Optimization with Deep Surrogate Langevin Dynamics
Abstract: We consider the problem of optimizing by sampling under multiple black-box, expensive constraints. We leverage the posterior regularization framework and show that the constraint satisfaction problem can be formulated as sampling from a Gibbs distribution. The main challenges come from the black-box nature of the constraints, for example the ones obtained by solving complex PDEs. To circumvent these issues, we introduce Surrogate-based Constrained Langevin dynamics for black-box sampling. We devise two approaches for learning surrogate gradients of the black-box functions. The first approach exploits zero-order approximation of the gradients in the Langevin sampling. In practice, this approach can be prohibitive due to the need to evaluate the expensive functions. The second approach approximates the gradients in the Langevin dynamics with deep neural networks, providing us an efficient sampling strategy using the surrogate models. We prove the convergence of both approaches when the target distribution is $\log$-concave and smooth. We also show the effectiveness of our approaches over Bayesian optimization in designing optimal nano-porous material configurations, where the goal is to produce nano-pattern templates with low thermal conductivity and reasonable mechanical stability.