A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models

Similar documents
Hierarchical Parallel Solution of Stochastic Systems

Greedy algorithms for high-dimensional non-symmetric problems

Parametric Problems, Stochastics, and Identification

NON-LINEAR APPROXIMATION OF BAYESIAN UPDATE

Non-Intrusive Solution of Stochastic and Parametric Equations

Sampling and Low-Rank Tensor Approximations

Schwarz Preconditioner for the Stochastic Finite Element Method

Fast Numerical Methods for Stochastic Computations

Quantifying Uncertainty: Modern Computational Representation of Probability and Applications

Sampling and low-rank tensor approximation of the response surface

Multilevel stochastic collocations with dimensionality reduction

Lecture 1: Center for Uncertainty Quantification. Alexander Litvinenko. Computation of Karhunen-Loeve Expansion:

Spectral Representation of Random Processes

Poisson Equation in 2D

Optimisation under Uncertainty with Stochastic PDEs for the History Matching Problem in Reservoir Engineering

Tensor Sparsity and Near-Minimal Rank Approximation for High-Dimensional PDEs

Dinesh Kumar, Mehrdad Raisee and Chris Lacor

A Spectral Approach to Linear Bayesian Updating

Numerical Analysis of Differential Equations Numerical Solution of Elliptic Boundary Value

Actor-critic methods. Dialogue Systems Group, Cambridge University Engineering Department. February 21, 2017

9. Iterative Methods for Large Linear Systems

Stochastic methods for solving partial differential equations in high dimension

Proper Generalized Decomposition for Linear and Non-Linear Stochastic Models

On semilinear elliptic equations with measure data

Polynomial chaos expansions for sensitivity analysis

On a Data Assimilation Method coupling Kalman Filtering, MCRE Concept and PGD Model Reduction for Real-Time Updating of Structural Mechanics Model

Uncertainty Quantification and related areas - An overview

Solving the Stochastic Steady-State Diffusion Problem Using Multigrid

Sequential Monte Carlo Samplers for Applications in High Dimensions

Interval model updating: method and application

Introduction to Uncertainty Quantification in Computational Science Handout #3

Principal Component Analysis

Introduction to Iterative Solvers of Linear Systems

On the numerical solution of the chemical master equation with sums of rank one tensors

Sparse grids and optimisation

Numerical Approximation of Stochastic Elliptic Partial Differential Equations

Breaking Computational Barriers: Multi-GPU High-Order RBF Kernel Problems with Millions of Points

Matrix assembly by low rank tensor approximation

PCA with random noise. Van Ha Vu. Department of Mathematics Yale University

Scattering of internal gravity waves

Krylov subspace methods for linear systems with tensor product structure

An Efficient Computational Solution Scheme of the Random Eigenvalue Problems

Scientific Computing II

Hyperbolic Polynomial Chaos Expansion (HPCE) and its Application to Statistical Analysis of Nonlinear Circuits

Exploiting off-diagonal rank structures in the solution of linear matrix equations

On the adaptive finite element analysis of the Kohn-Sham equations

Stochastic Elastic-Plastic Finite Element Method for Performance Risk Simulations

Three remarks on anisotropic finite elements

Uncertainty Quantification in Computational Science

Numerical Solutions to Partial Differential Equations

Proper Orthogonal Decomposition for Optimal Control Problems with Mixed Control-State Constraints

Inverse Problems in a Bayesian Setting

. D CR Nomenclature D 1

Regularity Theory a Fourth Order PDE with Delta Right Hand Side

Numerical Solution I

A very short introduction to the Finite Element Method

Two-parameter regularization method for determining the heat source

DISCRETE MAXIMUM PRINCIPLES in THE FINITE ELEMENT SIMULATIONS

A regularized least-squares method for sparse low-rank approximation of multivariate functions

Multifidelity Approaches to Approximate Bayesian Computation

Diamond Fractal. March 19, University of Connecticut REU Magnetic Spectral Decimation on the. Diamond Fractal

Dynamic System Identification using HDMR-Bayesian Technique

THE HIDDEN CONVEXITY OF SPECTRAL CLUSTERING

Sparse Tensor Galerkin Discretizations for First Order Transport Problems

Introduction to PDEs and Numerical Methods: Exam 1

Magnetic Spectral Decimation on the. Diamond Fractal. Aubrey Coffey, Madeline Hansalik, Stephen Loew. Diamond Fractal.

LECTURE 2 LINEAR REGRESSION MODEL AND OLS

Scientific Computing WS 2017/2018. Lecture 18. Jürgen Fuhrmann Lecture 18 Slide 1

Partitioned Methods for Multifield Problems

Numerical Methods - Numerical Linear Algebra

LECTURE 3: DISCRETE GRADIENT FLOWS

An example of Bayesian reasoning Consider the one-dimensional deconvolution problem with various degrees of prior information.

Non-Newtonian Fluids and Finite Elements

Kasetsart University Workshop. Multigrid methods: An introduction

Tracking using CONDENSATION: Conditional Density Propagation

Multigrid and Iterative Strategies for Optimal Control Problems

Clustering. Professor Ameet Talwalkar. Professor Ameet Talwalkar CS260 Machine Learning Algorithms March 8, / 26

Nucleon Spectroscopy with Multi-Particle Operators

Math 671: Tensor Train decomposition methods

Fundamentals of Data Assimila1on

The Particle Filter. PD Dr. Rudolph Triebel Computer Vision Group. Machine Learning for Computer Vision

Bayesian Inference and MCMC

Econometric Methods and Applications II Chapter 2: Simultaneous equations. Econometric Methods and Applications II, Chapter 2, Slide 1

Minimum Message Length Analysis of the Behrens Fisher Problem

Learning Gaussian Process Models from Uncertain Data

Covariance Matrix Simplification For Efficient Uncertainty Management

A Stochastic Collocation based. for Data Assimilation

APPROXIMATE SOLUTION OF A SYSTEM OF LINEAR EQUATIONS WITH RANDOM PERTURBATIONS

Autonomous Navigation for Flying Robots

Computational Homogenization and Multiscale Modeling

extended Stochastic Finite Element Method for the numerical simulation of heterogenous materials with random material interfaces

Basic Concepts of Adaptive Finite Element Methods for Elliptic Boundary Value Problems

Chance-constrained optimization with tight confidence bounds

Time domain sparsity promoting LSRTM with source estimation

Linear Factor Models. Sargur N. Srihari

Part I. Discrete Models. Part I: Discrete Models. Scientific Computing I. Motivation: Heat Transfer. A Wiremesh Model (2) A Wiremesh Model

Statistical Mechanics of Active Matter

1330. Comparative study of model updating methods using frequency response function data

Spectral methods for fuzzy structural dynamics: modal vs direct approach

8 Basics of Hypothesis Testing

Transcription:

A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models Rainer Niekamp and Martin Krosche. Institute for Scientific Computing TU Braunschweig ILAS: 22.08.2011 A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 1/11

Motivation (1) Diffusion equation with uncertain parameter: ( x κ(x, ω) x u(x, ω) ) = f (x, ω) κ, u and f are stochastic fields spatial dimension 2 or 3, stochastic dimension m Discretisation by Spectral Stochastic Finite Element Method (SSFEM) ansatz functions: N = N x N s DoFs Linear system to be solved: ( l i K i )u = f i=0 l K i U i = F A(U) = F i=0 u, f R Nx Ns Nx Ns U, F R A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 2/11

Motivation (2) Numerical complexity in SSFEM: U is dense and may become quite large e.g. N x N s = 10.000 1.000 = 10.000.000 classical construction of stochastic solution space suffers from the Curse of Dimension Recent approaches: Low-Rank approximation of U [Nouy 2007, Zander/Matthies 2008/2009] a priori solution space adaption [Doostan/Ghanem/Red-Horse 2007, Bieri/Schwab 2009] Low-Rank and a posteriori solution space adaption [Moselhy 2010, Krosche/Niekamp 2010] A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 3/11

Basic Low-Rank SSFEM (1) Minimisation of the expectation of the total potential energy: A(U) = F E(U) := 1 A(U) : U F : U min 2 with A : B := i,j A ij B ij, A, B R n 1 n 2 Ansatz: U = U + gh T (optimal rank-1 update) geometrical part g stochastic part h A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 4/11

Basic Low-Rank SSFEM (2) Differentiating E(U) with respect to U and... perturbing U by varying g leads to H g (h) g = b g (h) perturbing U by varying h leads to H h (g) h = b h (g) alternating iterative process Output: U r = r i=1 g i h T i = G H T A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 5/11

Optimisation in Low-Rank SSFEM Optimisation: obtain rank-1 updates: G = G + gv T and H = H + hw T perform the tensor products and add rank-1 updates stand-alone optimiser A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 6/11

Adaption in Low-Rank SSFEM Adaptive construction of the solution space: add new stochastic basis functions: ( i i := i i + i ) computation of the extended residual with current solution U rating of new basis functions maintain best rated basis functions A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 7/11

Basic Low-Rank SSFEM =0 geometrical dimension 2 stochastic dimension 5 rank-1 updates till to full rank u = 0 100 50 u = 1 1 0.1 0.01 expected value 2nd moment 3rd moment variance =0 relative error 0.001 0.0001 1e-05 1e-06 1e-07 0 10 20 30 40 50 60 70 80 90 100 110 120 130 rank A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 8/11

Optimisation Comparison: basic vs. optimised Low-Rank approach on the fly optimisation: during each rank-1 update maximum number of optimisation steps per rank: 10 0.1 0.01 basic optimised 0.001 relative error 0.0001 1e-05 1e-06 1e-07 1e-08 0 10 20 30 40 50 60 70 80 90 100 110 120 130 rank A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 9/11

Adaption Comparison: basic vs. adaptive Low-Rank approach basic approach: till order 3 adaptive approach: possible till order 5; chosen till order 4 the same number of stochastic DoFs at final rank post optimisation: at the final rank 0.001 optimised: order 3 adaptive, optimised: order 4 relative error 0.0001 2000 3000 4000 5000 number of iterations A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 10/11

Conclusion Problem: large, dense solution U Curse of Dimension in classical solution space construction Proposal: Low-Rank SSFEM to tackle problem of large, dense U stand-alone optimiser to tackle problem of sub-optimal approximations residual-based, a posteriori solution space adaption to tackle Curse of Dimension A Posteriori Adaptive Low-Rank Approximation of Probabilistic Models ILAS: 22.08.2011 11/11