Explore sequential design based on mutual information for computer experiments in this 20-minute talk by Joakim Beck from KAUST. Learn about uncertainty quantification, Gaussian Process emulation, and adaptive design strategies for complex physical systems. Discover the Mutual Information for Computer Experiments (MICE) algorithm and its applications in sensor placement, oscillatory function modeling, piston simulation, and tsunami modeling. Gain insights into practical issues, computational costs, and extensions to GP optimization.
Sequential Design Based on Mutual Information for Computer Experiments - Joakim Beck, KAUST
Alan Turing Institute via YouTube
Overview
Syllabus
Intro
Outline
Problem setting
GP emulation
Sequential adaptive designs
ALM and ALC
Greedy mutual information criterion
MI sequential design algorithm
Designing sensor placements (Krause et al., 2008)
A practical issue with the MI algorithm
Mutual Information for Computer Experiments (MICE)
A nugget parameter for smoothing
The improvement in terms of robustness
A visualisation of the design selection
A comparison of the computational cost
Numerical results: 4-D Oscillatory Function
Numerical results: 7-D Piston Simulation
Case study tsunami modelling
An extension: MICE in GP optimisation (optim-MICE)
Taught by
Alan Turing Institute
