DOCSLIB.ORG

A CMA-ES with Multiplicative Covariance Matrix Updates

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
A CMA-ES with Multiplicative Covariance Matrix Updates A CMA-ES with Multiplicative Covariance Matrix Updates Oswin Krause Tobias Glasmachers Department of Computer Science Institut für Neuroinformatik University of Copenhagen Ruhr-Universität Bochum Copenhagen,Denmark Bochum, Germany [email protected] [email protected] ABSTRACT N (m; σ2C), namely the global step size σ and the covari- Covariance matrix adaptation (CMA) mechanisms are core ance matrix C. Adaptation of the step size enables linear building blocks of modern evolution strategies. Despite shar- convergence on scale invariant functions [4], while covari- ing a common principle, the exact implementation of CMA ance matrix adaptation (CMA) [10] renders the asymptotic varies considerably between different algorithms. In this pa- convergence rate independent of the conditioning number of per, we investigate the benefits of an exponential parametriza- the Hessian in the optimum of a twice continuously differ- tion of the covariance matrix in the CMA-ES. This technique entiable function. was first proposed for the xNES algorithm. It results in a The most prominent algorithm implementing the above multiplicative update formula for the covariance matrix. We principles is CMA-ES [10, 8, 12]. Nowadays there exists a show that the exponential parameterization and the mul- plethora of variants and extensions of the basic algorithm. tiplicative update are compatible with all mechanisms of A generic principle for the online update of parameters CMA-ES. The resulting algorithm, xCMA-ES, performs at (including the search distribution) is to maximize the ex- least on par with plain CMA-ES. Its advantages show in pected progress. This goal can be approximated by adapting particular with updates that actively decrease the sampling the search distribution so that the probability of the pertur- variance in specific directions, i.e., for active constraint han- bations that generated successful offspring in the past are dling. increased. This is likely to foster the generation of better points in the near future.1 The application of the above principle to CMA means to Categories and Subject Descriptors change the covariance matrix towards the maximum likeli- [Continuous Optimization] hood estimator of successful steps. To this end let N (m; C) denote the search distribution, and let x1; : : : ; xµ denote suc- General Terms cessful offspring. The maximum likelihood estimator of the covariance matrix generating step δi = xi − m is the rank- T Algorithms one matrix δiδi . All CMA updates of CMA-ES are of the generic form C (1 − c) · C + c · δδT , which is a realization Keywords of this technique keeping an exponentially fading record of previous successful steps δ. CMA-ES variants differ in which evolution strategies, covariance matrix adaptation, CMA- step vectors enter the covariance matrix update. Early vari- ES, multiplicative update, exponential coordinates ants were based on cumulation of directions in an evolution path vector pc and a single rank-one update of C per gener- 1. INTRODUCTION ation [10]. Later versions added a rank-µ update based on Evolution Strategies (ES) are randomized direct search immediate information from the current population [8]. algorithms suitable for solving black box problems in the A different perspective on CMA techniques is provided d within the framework of information geometric optimization continuous domain, i.e., minimization problems f : R ! R defined on a d-dimensional real vector space. Most of these (IGO) [14], in particular by the natural evolution strategy algorithms generate a number of normally distributed off- (NES) approach [17, 16]. It turns out that the rank-µ update spring in each generation. The efficiency of this scheme, equation can be derived from stochastic natural gradient de- at least for unimodal problems, crucially depends on online scent of a stochastically relaxed problem on the statistical adaptation of parameters of the Gaussian search distribution manifold of search distributions. This more general perspec- tive opens up new possibilities for CMA mechanisms, e.g., reparameterizing the covariance matrix in exponential form as done in the xNES algorithm [7]. This results in an update Permission to make digital or hard copies of all or part of this work for equation with the following properties: a) the update is mul- personal or classroom use is granted without fee provided that copies are tiplicative, in contrast to the standard additive update, b) it not made or distributed for profit or commercial advantage and that copies is possible to leave the variance in directions orthogonal to bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific 1 permission and/or a fee. This statement holds only under (mild) assumptions on the GECCO’15, July 11-15, 2015, Madrid, Spain. regularity of the fitness landscape, which remain implicit, Copyright 2015 ACM TBA ...$15.00. but may be violated, e.g., in the presence of constraints. all observed steps unchanged, and c) even when performing Algorithm 1: CMA-ES \active" (negative) updates the covariance matrix is guaran- Input: m, σ teed to remain positive definite. C I In this paper we incorporate the exponential parameteri- while stopping condition not met do zation of the covariance matrix into CMA-ES. We derive all // sample and evaluate offspring mechanisms found in the standard CMA-ES algorithm in for i 2 f1; : : : ; λg do this framework, demonstrating the compatibility of cumula- 2 xi N (m; σ C) tive step size adaptation and evolution paths (two features end missing in xNES) with exponential coordinates. The new sort fxig with respect to f(xi) algorithm is called xCMA-ES. Its performance on standard // internal update (paths and parameters) (unimodal) benchmarks coincides with that of CMA-ES, 0 Pµ m wixi however, in addition it benefits from neat properties of the i=1 p (1 − c ) · p + pc (2 − c )µ · C−1=2(m0 − m) exponential parameters, which show up prominently when s s s s s eff p m0−m performing active CMA updates with negative weights, e.g., pc (1 − cc) · pc + cc(2 − cc)µeff · σ T for constraint handling. C (1 − c1 − cµ) · C + c1 · pcpc T In the next section we recap CMA-ES and xNES. Based Pµ xi−m xi−m +cµ · i=1 wi σ σ thereon we present our new xCMA-ES algorithm. In sec- σ σ · exp cs · kpsk − 1 tion 3 its performance is evaluated empirically on standard Dσ χd 0 benchmarks. We demonstrate the superiority for special m m tasks involving active CMA updates. end 2. ALGORITHMS In this section we provide the necessary background for it is assumed that its computational cost is substantial, our new algorithm before introducing the xCMA-ES. We will usually exceeding the internal computational complexity cover the well-known CMA-ES algorithm as well as xNES, of the algorithm. both with a focus on the components required for our own 3. Sort the offspring by fitness. Post-condition: f(x ) ≤ contribution. In the following algorithms, a few implemen- 1 f(x ) ≤ · · · ≤ f(x ). tation details are not shown, e.g., the eigen decomposition 2 λ 4. Perform environmental selection: keep fx ; : : : ; x g, dis- of C required for sampling as well as for the computation of 1 µ card fx ; : : : ; x g, usually the worse half. the inverse matrix square root C−1=2. µ+1 λ 5. Update the evolution path for cumulative step size adap- 2.1 CMA-ES tation: the path ps is an exponentially fading record of steps of the mean vector, back-transformed by multipli- The Covariance Matrix Adaptation Evolution Strategy −1=2 (CMA-ES) [10, 12] is the most prominent evolution strategy cation with the inverse matrix square root C into a in existence. It comes in many variants, e.g., with extensions coordinate system where the sampling distribution is a for handling of fitness noise [9] and multi-modality [5]. Here standard normal distribution. This path is supposed to we describe what can be considered a baseline version, fea- follow a standard normal distribution. turing non-elitist (µ, λ) selection, cumulative step size con- 6. Update the evolution path for covariance matrix adap- trol (CSA), and two different types of covariance matrix up- tation: the path pc is an exponentially fading record of dates, namely a rank-1 update based on an evolution path, steps of the mean vector divided (corrected) by the step and the so-called rank-µ update based on the survivors of size. This path models the movement direction of the environmental truncation selection. distribution center over time. d The state of CMA-ES is given by the parameters m 2 R , 7. Update the covariance matrix: a new matrix is obtained d×d σ > 0, and C 2 R of its multi-variate normal search dis- by additive blending of the old matrix with a rank-one tribution N (m; σ2C), as well as by the two evolution paths matrix formed by the path pc and a rank-µ matrix formed d by successful steps of the current population. ps; pc 2 R . Pseudo-code of a basic CMA-ES is provided in algorithm 1. The algorithm has a number of tuning con- 8. Update the step size: the step size is changed if the norm stants, e.g., the sizes of parent and offspring population µ of ps indicates a systematic deviation from the standard and λ, the various leaning rates, and the rank-based weights normal distribution. Note (for later reference) that this w1; : : : ; wµ. For robust default settings for the different pa- update has a multiplicative form, involving the exponen- rameters we refer to [12]. tial function. In each generation, CMA-ES executes the following steps: 9.
Load more

Recommended publications
  • Conjugate Gradient Method (Part 4)  Pre-Conditioning  Nonlinear Conjugate Gradient Method
    18-660: Numerical Methods for Engineering Design and Optimization Xin Li Department of ECE Carnegie Mellon University Pittsburgh, PA 15213 Slide 1 Overview Conjugate Gradient Method (Part 4) Pre-conditioning Nonlinear conjugate gradient method Slide 2 Conjugate Gradient Method Step 1: start from an initial guess X(0), and set k = 0 Step 2: calculate ( ) ( ) ( ) D 0 = R 0 = B − AX 0 Step 3: update solution D(k )T R(k ) X (k +1) = X (k ) + µ (k )D(k ) where µ (k ) = D(k )T AD(k ) Step 4: calculate residual ( + ) ( ) ( ) ( ) R k 1 = R k − µ k AD k Step 5: determine search direction R(k +1)T R(k +1) (k +1) = (k +1) + β (k ) β = D R k +1,k D where k +1,k (k )T (k ) D R Step 6: set k = k + 1 and go to Step 3 Slide 3 Convergence Rate k ( + ) κ(A) −1 ( ) X k 1 − X ≤ ⋅ X 0 − X κ(A) +1 Conjugate gradient method has slow convergence if κ(A) is large I.e., AX = B is ill-conditioned In this case, we want to improve convergence rate by pre- conditioning Slide 4 Pre-Conditioning Key idea Convert AX = B to another equivalent equation ÃX̃ = B̃ Solve ÃX̃ = B̃ by conjugate gradient method Important constraints to construct ÃX̃ = B̃ à is symmetric and positive definite – so that we can solve it by conjugate gradient method à has a small condition number – so that we can achieve fast convergence Slide 5 Pre-Conditioning AX = B −1 −1 L A⋅ X = L B L−1 AL−T ⋅ LT X = L−1B A X B ̃ ̃ ̃ L−1AL−T is symmetric and positive definite, if A is symmetric and positive definite T (L−1 AL−T ) = L−1 AL−T T X T L−1 AL−T X = (L−T X ) ⋅ A⋅(L−T X )> 0 Slide
    [Show full text]
  • Hybrid Algorithms for Efficient Cholesky Decomposition And
    Hybrid Algorithms for Efficient Cholesky Decomposition and Matrix Inverse using Multicore CPUs with GPU Accelerators Gary Macindoe A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy of UCL. 2013 ii I, Gary Macindoe, confirm that the work presented in this thesis is my own. Where information has been derived from other sources, I confirm that this has been indicated in the thesis. Signature : Abstract The use of linear algebra routines is fundamental to many areas of computational science, yet their implementation in software still forms the main computational bottleneck in many widely used algorithms. In machine learning and computational statistics, for example, the use of Gaussian distributions is ubiquitous, and routines for calculating the Cholesky decomposition, matrix inverse and matrix determinant must often be called many thousands of times for com- mon algorithms, such as Markov chain Monte Carlo. These linear algebra routines consume most of the total computational time of a wide range of statistical methods, and any improve- ments in this area will therefore greatly increase the overall efficiency of algorithms used in many scientific application areas. The importance of linear algebra algorithms is clear from the substantial effort that has been invested over the last 25 years in producing low-level software libraries such as LAPACK, which generally optimise these linear algebra routines by breaking up a large problem into smaller problems that may be computed independently. The performance of such libraries is however strongly dependent on the specific hardware available. LAPACK was originally de- veloped for single core processors with a memory hierarchy, whereas modern day computers often consist of mixed architectures, with large numbers of parallel cores and graphics process- ing units (GPU) being used alongside traditional CPUs.
    [Show full text]
  • Matrix Inversion Using Cholesky Decomposition
    Matrix Inversion Using Cholesky Decomposition Aravindh Krishnamoorthy, Deepak Menon ST-Ericsson India Private Limited, Bangalore [email protected], [email protected] Abstract—In this paper we present a method for matrix inversion Upper triangular elements, i.e. : based on Cholesky decomposition with reduced number of operations by avoiding computation of intermediate results; further, we use fixed point simulations to compare the numerical ( ∑ ) accuracy of the method. … (4) Keywords-matrix, inversion, Cholesky, LDL. Note that since older values of aii aren’t required for computing newer elements, they may be overwritten by the I. INTRODUCTION value of rii, hence, the algorithm may be performed in-place Matrix inversion techniques based on Cholesky using the same memory for matrices A and R. decomposition and the related LDL decomposition are efficient Cholesky decomposition is of order and requires techniques widely used for inversion of positive- operations. Matrix inversion based on Cholesky definite/symmetric matrices across multiple fields. decomposition is numerically stable for well conditioned Existing matrix inversion algorithms based on Cholesky matrices. decomposition use either equation solving [3] or triangular matrix operations [4] with most efficient implementation If , with is the linear system with requiring operations. variables, and satisfies the requirement for Cholesky decomposition, we can rewrite the linear system as In this paper we propose an inversion algorithm which … (5) reduces the number of operations by 16-17% compared to the existing algorithms by avoiding computation of some known By letting , we have intermediate results. … (6) In section 2 of this paper we review the Cholesky and LDL decomposition techniques, and discuss solutions to linear and systems based on them.
    [Show full text]
  • AMS526: Numerical Analysis I (Numerical Linear Algebra) Lecture 3: Positive-Definite Systems; Cholesky Factorization
    AMS526: Numerical Analysis I (Numerical Linear Algebra) Lecture 3: Positive-Definite Systems; Cholesky Factorization Xiangmin Jiao Stony Brook University Xiangmin Jiao Numerical Analysis I 1 / 11 Symmetric Positive-Definite Matrices n×n Symmetric matrix A 2 R is symmetric positive definite (SPD) if T n x Ax > 0 for x 2 R nf0g n×n Hermitian matrix A 2 C is Hermitian positive definite (HPD) if ∗ n x Ax > 0 for x 2 C nf0g SPD matrices have positive real eigenvalues and orthogonal eigenvectors Note: Most textbooks only talk about SPD or HPD matrices, but a positive-definite matrix does not need to be symmetric or Hermitian! A real matrix A is positive definite iff A + AT is SPD T n If x Ax ≥ 0 for x 2 R nf0g, then A is said to be positive semidefinite Xiangmin Jiao Numerical Analysis I 2 / 11 Properties of Symmetric Positive-Definite Matrices SPD matrix often arises as Hessian matrix of some convex functional I E.g., least squares problems; partial differential equations If A is SPD, then A is nonsingular Let X be any n × m matrix with full rank and n ≥ m. Then T I X X is symmetric positive definite, and T I XX is symmetric positive semidefinite n×m If A is n × n SPD and X 2 R has full rank and n ≥ m, then X T AX is SPD Any principal submatrix (picking some rows and corresponding columns) of A is SPD and aii > 0 Xiangmin Jiao Numerical Analysis I 3 / 11 Cholesky Factorization If A is symmetric positive definite, then there is factorization of A A = RT R where R is upper triangular, and all its diagonal entries are positive Note: Textbook calls is “Cholesky
    [Show full text]
  • 1. Positive Definite Matrices a Matrix a Is Positive Definite If X>Ax > 0 for All Nonzero X
    CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 8 GENE H. GOLUB 1. Positive Definite Matrices A matrix A is positive definite if x>Ax > 0 for all nonzero x. A positive definite matrix has real and positive eigenvalues, and its leading principal submatrices all have positive determinants. From the definition, it is easy to see that all diagonal elements are positive. To solve the system Ax = b where A is positive definite, we can compute the Cholesky decom- position A = F >F where F is upper triangular. This decomposition exists if and only if A is symmetric and positive definite. In fact, attempting to compute the Cholesky decomposition of A is an efficient method for checking whether A is symmetric positive definite. It is important to distinguish the Cholesky decomposition from the square root factorization.A square root of a matrix A is defined as a matrix S such that S2 = SS = A. Note that the matrix F in A = F >F is not the square root of A, since it does not hold that F 2 = A unless A is a diagonal matrix. The square root of a symmetric positive definite A can be computed by using the fact that A has an eigendecomposition A = UΛU > where Λ is a diagonal matrix whose diagonal elements are the positive eigenvalues of A and U is an orthogonal matrix whose columns are the eigenvectors of A. It follows that A = UΛU > = (UΛ1/2U >)(UΛ1/2U >) = SS and so S = UΛ1/2U > is a square root of A. 2. The Cholesky Decomposition The Cholesky decomposition can be computed directly from the matrix equation A = F >F .
    [Show full text]
  • Cholesky Decomposition 1 Cholesky Decomposition
    Cholesky decomposition 1 Cholesky decomposition In linear algebra, the Cholesky decomposition or Cholesky triangle is a decomposition of a symmetric, positive-definite matrix into the product of a lower triangular matrix and its conjugate transpose. It was discovered by André-Louis Cholesky for real matrices and is an example of a square root of a matrix. When it is applicable, the Cholesky decomposition is roughly twice as efficient as the LU decomposition for solving systems of linear equations.[1] Statement If A has real entries and is symmetric (or more generally, is Hermitian) and positive definite, then A can be decomposed as where L is a lower triangular matrix with strictly positive diagonal entries, and L* denotes the conjugate transpose of L. This is the Cholesky decomposition. The Cholesky decomposition is unique: given a Hermitian, positive-definite matrix A, there is only one lower triangular matrix L with strictly positive diagonal entries such that A = LL*. The converse holds trivially: if A can be written as LL* for some invertible L, lower triangular or otherwise, then A is Hermitian and positive definite. The requirement that L have strictly positive diagonal entries can be dropped to extend the factorization to the positive-semidefinite case. The statement then reads: a square matrix A has a Cholesky decomposition if and only if A is Hermitian and positive semi-definite. Cholesky factorizations for positive-semidefinite matrices are not unique in general. In the special case that A is a symmetric positive-definite matrix with real entries, L can be assumed to have real entries as well.
    [Show full text]
  • The CMA Evolution Strategy: a Tutorial
    The CMA Evolution Strategy: A Tutorial Nikolaus Hansen June 28, 2011 Contents Nomenclature3 0 Preliminaries4 0.1 Eigendecomposition of a Positive Definite Matrix...............5 0.2 The Multivariate Normal Distribution.....................6 0.3 Randomized Black Box Optimization.....................7 0.4 Hessian and Covariance Matrices........................8 1 Basic Equation: Sampling8 2 Selection and Recombination: Moving the Mean9 3 Adapting the Covariance Matrix 10 3.1 Estimating the Covariance Matrix From Scratch................ 10 3.2 Rank-µ-Update................................. 12 3.3 Rank-One-Update................................ 14 3.3.1 A Different Viewpoint......................... 14 3.3.2 Cumulation: Utilizing the Evolution Path............... 16 3.4 Combining Rank-µ-Update and Cumulation.................. 17 4 Step-Size Control 17 5 Discussion 21 A Algorithm Summary: The CMA-ES 25 B Implementational Concerns 27 B.1 Multivariate normal distribution........................ 28 B.2 Strategy internal numerical effort........................ 28 B.3 Termination criteria............................... 28 B.4 Flat fitness.................................... 29 B.5 Boundaries and Constraints........................... 29 C MATLAB Source Code 31 1 D Reformulation of Learning Parameter ccov 33 2 Nomenclature We adopt the usual vector notation, where bold letters, v, are column vectors, capital bold letters, A, are matrices, and a transpose is denoted by vT. A list of used abbreviations and symbols is given in alphabetical order. Abbreviations CMA Covariance Matrix Adaptation EMNA Estimation of Multivariate Normal Algorithm ES Evolution Strategy (µ/µfI;Wg; λ)-ES, Evolution Strategy with µ parents, with recombination of all µ parents, either Intermediate or Weighted, and λ offspring. RHS Right Hand Side. Greek symbols λ ≥ 2, population size, sample size, number of offspring, see (5).
    [Show full text]
  • Reducing Dimensionality in Text Mining Using Conjugate Gradients and Hybrid Cholesky Decomposition
    (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 8, No. 7, 2017 Reducing Dimensionality in Text Mining using Conjugate Gradients and Hybrid Cholesky Decomposition Jasem M. Alostad The Public Authority for Applied Education and Training (PAAET), College of Basic Education P.O.Box 23167, Safat 13092, Kuwait Abstract—Generally, data mining in larger datasets consists The data reduction is the process of reducing the size or of certain limitations in identifying the relevant datasets for the dimensionality of the data, however, the representation of the given queries. The limitations include: lack of interaction in the data should be retained. Selection of instance is one better way required objective space, inability to handle the data sets or to reduce the data by reducing the total number of instances. In discrete variables in datasets, especially in the presence of spite of many efforts to deal with such instances, data mining missing variables and inability to classify the records as per the algorithm, however undergoes severe challenges due to non- given query, and finally poor generation of explicit knowledge for applicability of datasets with large instances. Hence, the a query increases the dimensionality of the data. Hence, this computational complexity of the system increases with larger paper aims at resolving the problems with increasing data instances [3], [4] and leads to problems in scaling, increased dimensionality in datasets using modified non-integer matrix storage requirements and clustering accuracy. The other factorization (NMF). Further, the increased dimensionality arising due to non-orthogonally of NMF is resolved with problems associated with larger data instances include, Cholesky decomposition (cdNMF).
    [Show full text]
  • Stable Computations of Generalized Inverses of Positive Semidefinite
    Stable computations of generalized inverses of positive semidefinite matrices A. Markopoulos, Z. Dostál, T. Kozubek, P. Kovár,ˇ T. Brzobohatý, and R. Kuceraˇ Centre of Excellence IT4I, VŠB-Technical University of Ostrava, Czech Republic [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] Summary. Effective implementation of some efficient FETI methods assumes application of a direct method for solving a system of linear equations with a symmetric positive semidefinite matrix A. The latter usually comprises a triangular decomposition of a nonsingular diagonal block AJJ of the stiffness matrix A of a subdomain and an effective evaluation of the action of a generalized inverse of the corresponding Schur complement. The other posibility consists in a regularization of A resulting in a non-singular matrix Ar whose (standard) inverse is the generalized inverse to A. It avoids the necessity to identify zero pivots with negligible fill-in. We review both these techniques. 1 Introduction Due to the rounding errors, effective elimination of the displacements of “floating” subdomains is a nontrivial ingredient of implementation of FETI methods, as it can be difficult to recognize the positions of zero pivots when the nonsingular diagonal block of A is ill-conditioned. Moreover, even if the zero pivots are recognized prop- erly, it turns out that the ill-conditioning of the nonsingular submatrix defined by the nonzero pivots can have a devastating effect on the precision of the solution. Most of the results are related to the first problem, i.e., to identify reliably the zero pivots.
    [Show full text]
  • Hierarchical Sparse Cholesky Decomposition with Applications to High-Dimensional Spatio-Temporal filtering
    Hierarchical sparse Cholesky decomposition with applications to high-dimensional spatio-temporal filtering Marcin Jurek∗ Matthias Katzfuss∗y Abstract Spatial statistics often involves Cholesky decomposition of covariance matrices. To ensure scalability to high dimensions, several recent approximations have assumed a sparse Cholesky factor of the precision matrix. We propose a hierarchical Vec- chia approximation, whose conditional-independence assumptions imply sparsity in the Cholesky factors of both the precision and the covariance matrix. This remarkable property is crucial for applications to high-dimensional spatio-temporal filtering. We present a fast and simple algorithm to compute our hierarchical Vecchia approxima- tion, and we provide extensions to non-linear data assimilation with non-Gaussian data based on the Laplace approximation. In several numerical comparisons, our methods strongly outperformed alternative approaches. Keywords: state-space model, spatio-temporal statistics, data assimilation, Vecchia approxima- tion, hierarchical matrix, incomplete Cholesky decomposition 1 Introduction Symmetric positive-definite matrices arise in spatial statistics, Gaussian-process inference, and spatio-temporal filtering, with a wealth of application areas, including geoscience (e.g., Cressie, 1993; Banerjee et al., 2004), machine learning (e.g., Rasmussen and Williams, 2006), data assimilation (e.g., Nychka and Anderson, 2010; Katzfuss et al., 2016), and the analysis arXiv:2006.16901v1 [stat.CO] 30 Jun 2020 of computer experiments (e.g., Sacks et al., 1989; Kennedy and O'Hagan, 2001). Inference in these areas typically relies on Cholesky decomposition of the positive-definite matrices. However, this operation scales cubically in the dimension of the matrix, and it is thus com- putationally infeasible for many modern problems and applications, which are increasingly high-dimensional.
    [Show full text]
  • LAPACK Working Note
    ISSN 1360-1725 UMIST LAPACK-Style Codes for Level 2 and 3 Pivoted Cholesky Factorizations C. Lucas Numerical Analysis Report No. 442 February 2004 Manchester Centre for Computational Mathematics Numerical Analysis Reports DEPARTMENTS OF MATHEMATICS Reports available from: And over the World-Wide Web from URLs Department of Mathematics http://www.ma.man.ac.uk/nareports University of Manchester ftp://ftp.ma.man.ac.uk/pub/narep Manchester M13 9PL England LAPACK-Style Codes for Level 2 and 3 Pivoted Cholesky Factorizations Craig Lucas¤ February 5, 2004 Abstract Fortran 77 codes exist in LAPACK for computing the Cholesky factor- ization of a symmetric positive de¯nite matrix using Level 2 and 3 BLAS. In LINPACK there is a Level 1 BLAS routine for computing the Cholesky factorization with complete pivoting of a symmetric positive semi-de¯nite matrix. We present two new algorithms and Fortran 77 LAPACK-style codes for computing this pivoted factorization: one using Level 2 BLAS and one using Level 3 BLAS. We show that on modern machines the new codes can be many times faster than the LINPACK code. Also, with a new stopping criterion they provide more reliable rank detection and can have a smaller normwise backward error. 1 Introduction The Cholesky factorization of a symmetric positive de¯nite matrix A 2 Rn£n has the form A = LLT ; where L 2 Rn£n is a lower triangular matrix with positive diagonal elements. If A is positive semi-de¯nite, of rank r, there exists a Cholesky factorization with complete pivoting ([6, Thm. 10.9], for example).
    [Show full text]
  • A Scalable High Performant Cholesky Factorization for Multicore with GPU Accelerators LAPACK Working Note #223
    A Scalable High Performant Cholesky Factorization for Multicore with GPU Accelerators LAPACK Working Note #223 Hatem Ltaief1, Stanimire Tomov1, Rajib Nath1, Peng Du1, and Jack Dongarra1,2,3? 1 Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville 2 Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee 3 School of Mathematics & School of Computer Science, University of Manchester {ltaief, tomov, rnath1, du, dongarra}@eecs.utk.edu Abstract. We present a Cholesky factorization for multicore with GPU accelerators. The challenges in developing scalable high performance al- gorithms for these emerging systems stem from their heterogeneity, mas- sive parallelism, and the huge gap between the GPUs’ compute power vs the CPU-GPU communication speed. We show an approach that is largely based on software infrastructures that have already been devel- oped for homogeneous multicores and hybrid GPU-based computing. The algorithm features two levels of nested parallelism. A coarse-grained parallelism is provided by splitting the computation into tiles for con- current execution between GPUs. A fine-grained parallelism is further provided by splitting the work-load within a tile for high efficiency com- puting on GPUs but also, in certain cases, to benefit from hybrid com- putations by using both GPUs and CPUs. Our resulting computational kernels are highly optimized. An efficient task scheduling mechanism ensures a load balanced execution over the entire multicore with GPU accelerators system. Furthermore, the communication overhead is min- imized by trading off the amount of memory allocated on GPUs. This results in a scalable hybrid Cholesky factorization of unprecedented per- formance.
    [Show full text]