Applied Physics Seminar | Bayesian Optimal Experimental Design
Xun Huan, Ph.D., Associate Professor of Mechanical Engineering, College of Engineering, University of Michigan
Wednesday, September 18, 2024
12:00-1:00 PM
Virtual
Abstract:
Experiments are crucial for developing and refining models in engineering and science. When experiments are expensive, a careful design of these limited data-acquisition opportunities can be immensely beneficial. Optimal experimental design (OED) leverages the predictive power of simulation models to systematically quantify and maximize the value of experiments.
We first introduce OED for a batch of experiment under a Bayesian setting using the expected information gain (EIG) (uncertainty reduction) objective. We then formulate OED for a sequence of experiments via a Markov decision process (MDP), where an optimal design rule (policy) can (a) adapt to newly collected data along the way (feedback) and (b) anticipate future consequences (lookahead). We solve the sequential OED problem with policy gradient techniques from reinforcement learning together with an efficient lower bound EIG estimator. We demonstrate our method on several examples, including an optimal sensor movement application for source inversion in a convection-diffusion field.
Experiments are crucial for developing and refining models in engineering and science. When experiments are expensive, a careful design of these limited data-acquisition opportunities can be immensely beneficial. Optimal experimental design (OED) leverages the predictive power of simulation models to systematically quantify and maximize the value of experiments.
We first introduce OED for a batch of experiment under a Bayesian setting using the expected information gain (EIG) (uncertainty reduction) objective. We then formulate OED for a sequence of experiments via a Markov decision process (MDP), where an optimal design rule (policy) can (a) adapt to newly collected data along the way (feedback) and (b) anticipate future consequences (lookahead). We solve the sequential OED problem with policy gradient techniques from reinforcement learning together with an efficient lower bound EIG estimator. We demonstrate our method on several examples, including an optimal sensor movement application for source inversion in a convection-diffusion field.
| Building: | West Hall |
|---|---|
| Event Link: | |
| Event Password: | Passcode: 898441 |
| Event Type: | Lecture / Discussion |
| Tags: | Engineering, Mechanical Engineering, Physics |
| Source: | Happening @ Michigan from Applied Physics |
