From saunders at stanford.edu Sat Oct 1 13:43:04 2011 From: saunders at stanford.edu (Michael Saunders) Date: Sat Oct 1 13:44:00 2011 Subject: [BANANA] LA/Opt seminar Thursday (Rob Schreiber) Message-ID: Dear colleagues, The first LA/Opt seminar of the quarter is on Thursday Oct 6. It is about the Netflix Prize. Rob Schreiber is a distinguished researcher. Among other things he helped give us sparse matrices in Matlab. ---------------------------------------------------------------------------------------------- Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 6, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 (Note new room, near last year's room) Dr Rob Schreiber Distinguished Technologist Assistant Director of the Exascale Computing Lab HP Labs, Palo Alto http://www.hpl.hp.com/personal/Robert_Schreiber A Parallel Approach to Collaborative Filtering (joint work with Rong Pan, Dennis Wilkinson, and Yunhong Zhou) Recommendation systems suggest purchases to customers by mining logs of the ratings and purchases of a whole community of buyers, an idea known as collaborative filtering (CF). For example, Netflix actively gathers user ratings of movies, and makes recommendations using such a system. The problem can be viewed as matrix completion -- fill in the missing entries of a large, dense matrix given a sparse subset. Of course, an assumption is necessary -- normally that the matrix can be approximated well by a matrix of moderate rank. CF methods have to contend with scale -- numbers of buyers, products, and ratings, and sparseness -- some buyers and some products don't appear often in the record. We describe a parallel algorithm that we designed for the Netflix Prize, a large-scale collaborative filtering challenge. We used an alternating-least-squares approach to compute the low-rank approximation by minimizing the difference between the known elements of the full ratings matrix and the corresponding elements of the (factored) low-rank approximation. To deal with sparseness, we employed a novel variant of Tikhonov Regularization, which succeeded well in preventing overfitting. We used parallel Matlab on a Linux cluster as the experimental platform, as a means to deal with scale. As a result, we were able to create an approximation of higher rank than the competitive approaches. For the Netflix dataset, we were able to find a rank-1000 approximation, and got an RMSE score of 0.8985, which was one of the best results based on a pure (not boosted) method. Forthcoming talks: Thu Oct 13 William Kahan, UC Berkeley Thu Oct 20 Thu Oct 27 Ming Gu, UC Berkeley Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore From mgu at math.berkeley.edu Mon Oct 3 18:42:02 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Mon Oct 3 18:42:58 2011 Subject: [BANANA] Reminder: LAPACK seminar on Oct. 5, 2011 In-Reply-To: <201109300614.p8U6E3No013461@panda.math.berkeley.edu> References: <201109300614.p8U6E3No013461@panda.math.berkeley.edu> Message-ID: Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 5, 2011 Speaker: Jim Demmel, UC Berkeley Title: More Recent Progress on Communication-Avoiding Algorithms Date: Oct. 12, 2011 Speaker: Michael Mahoney, Stanford U. Title: Linear Algebra and Machine Learning of Large Informatics Graphs From saunders at stanford.edu Thu Oct 6 10:10:29 2011 From: saunders at stanford.edu (Michael Saunders) Date: Thu Oct 6 10:11:21 2011 Subject: [BANANA] LA/Opt seminar TODAY (Rob Schreiber) Message-ID: Reminder: seminar this afternoon Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 6, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 (Note new room, near last year's room) Dr Rob Schreiber Distinguished Technologist Assistant Director of the Exascale Computing Lab HP Labs, Palo Alto http://www.hpl.hp.com/personal/Robert_Schreiber A Parallel Approach to Collaborative Filtering (joint work with Rong Pan, Dennis Wilkinson, and Yunhong Zhou) Recommendation systems suggest purchases to customers by mining logs of the ratings and purchases of a whole community of buyers, an idea known as collaborative filtering (CF). For example, Netflix actively gathers user ratings of movies, and makes recommendations using such a system. The problem can be viewed as matrix completion -- fill in the missing entries of a large, dense matrix given a sparse subset. Of course, an assumption is necessary -- normally that the matrix can be approximated well by a matrix of moderate rank. CF methods have to contend with scale -- numbers of buyers, products, and ratings, and sparseness -- some buyers and some products don't appear often in the record. We describe a parallel algorithm that we designed for the Netflix Prize, a large-scale collaborative filtering challenge. We used an alternating-least-squares approach to compute the low-rank approximation by minimizing the difference between the known elements of the full ratings matrix and the corresponding elements of the (factored) low-rank approximation. To deal with sparseness, we employed a novel variant of Tikhonov Regularization, which succeeded well in preventing overfitting. We used parallel Matlab on a Linux cluster as the experimental platform, as a means to deal with scale. As a result, we were able to create an approximation of higher rank than the competitive approaches. For the Netflix dataset, we were able to find a rank-1000 approximation, and got an RMSE score of 0.8985, which was one of the best results based on a pure (not boosted) method. Forthcoming talks: Thu Oct 13 William Kahan, UC Berkeley Thu Oct 20 Thu Oct 27 Ming Gu, UC Berkeley Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore From mgu at math.berkeley.edu Fri Oct 7 22:03:46 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Fri Oct 7 22:04:50 2011 Subject: [BANANA] LAPACK seminar on Oct. 12, 2011 Message-ID: <201110080503.p9853k74022812@phoenix.math.berkeley.edu> Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 12 Speaker: Michael W. Mahoney, Stanford University Title: Linear Algebra and Machine Learning of Large Informatics Graphs Abstract: A Very large informatics graphs such as large social and information networks typically have properties that render many popular machine learning and data analysis tools largely inappropriate. While this is problematic for these applications, it also suggests that these graphs may be useful as a test case for the development of new algorithmic tools that may then be applicable much more generally. Many of the popular machine learning and data analysis tools rely on linear algebra, and they are typically used by calling traditional numerical linear algebra code as a black box. After briefly reviewing some of the structural properties of large social and information networks that are responsible for the inapplicability of traditional linear algebra and machine learning tools, I will describe several examples of "new linear algebra" and "new machine learning" that arise from the analysis of such informatics graphs. These new directions involve looking "inside" the black box, and they place very different demands on the linear algebra than are traditionally placed by numerical, scientific computing, and small-scale machine learning applications. Bio: Michael Mahoney is at Stanford University. His research interests focus on theoretical and applied aspects of algorithms for large-scale data problems in scientific and Internet applications. Currently, he is working on geometric network analysis; developing approximate computation and regularization methods for large informatics graphs; and applications to community detection, clustering, and information dynamics in large social and information networks. In the past, he has worked on randomized matrix algorithms and applications in genetics and medical imaging. He has been a faculty member at Yale University and a researcher at Yahoo, and his PhD was is computational statistical mechanics at Yale University. Date: Oct. 19 Speaker: Joel Phillips, Cadence Design Systems From saunders at stanford.edu Mon Oct 10 11:10:19 2011 From: saunders at stanford.edu (Michael Saunders) Date: Mon Oct 10 11:12:39 2011 Subject: [BANANA] LA/Opt seminar Thursday Oct 13 (William Kahan) Message-ID: Distinguished address by Professor William Kahan for all concerned with scientific computation Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 13, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 PROFESSOR WILLIAM KAHAN Math Dept and EECS Dept, UC Berkeley http://www.eecs.berkeley.edu/~wkahan Desperately Needed Remedies for the Undebuggability of Large Floating-Point Computations in Science and Engineering If suspicions about the accuracy of a computed result arise, how long does it take to either allay or justify them? Often diagnosis has taken longer than the computing platform's service life. Software tools to speed up diagnosis by at least an order of magnitude could be provided but almost no scientists and engineers know to ask for them, though almost all these tools have existed, albeit not all together in the same place at the same time. These tools would cope with vulnerabilities peculiar to Floating-Point, namely roundoff and arithmetic exceptions. But who would pay to develop the suite of these tools? Nobody, unless he suspects that the incidence of misleadingly anomalous Floating-Point results rather exceeds what is generally believed. Ample evidence supports that suspicion. Bio: Professor Kahan has been widely honored for his dedication to improving scientific computing algorithms and machines. 1989 ACM Turing Award 1993 Honorary Doctor of Mathematics, Chalmers Inst, Sweden 1994 ACM Fellow 1997 SIAM John von Neumann Memorial Lecture 1998 Honorary Doctor of Mathematics, Univ of Waterloo, Canada 2000 IEEE Emanuel R. Piore Award 2003 American Academy of Arts & Sciences 2004 Distinguished Mentor of Undergraduate Research in the College of Letters & Science 2005 National Academy of Engineering, Foreign (Canadian) Associate From mgu at math.berkeley.edu Tue Oct 11 00:03:15 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Tue Oct 11 00:04:18 2011 Subject: [BANANA] Reminder: LAPACK seminar on Oct. 12, 2011 In-Reply-To: <201110080503.p9853k74022812@phoenix.math.berkeley.edu> References: <201110080503.p9853k74022812@phoenix.math.berkeley.edu> Message-ID: Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 12 Speaker: Michael W. Mahoney, Stanford University Title: Linear Algebra and Machine Learning of Large Informatics Graphs Date: Oct. 19 Speaker: Joel Phillips, Cadence Design Systems From saunders at stanford.edu Thu Oct 13 00:51:42 2011 From: saunders at stanford.edu (Michael Saunders) Date: Thu Oct 13 00:52:41 2011 Subject: [BANANA] LA/Opt seminar TODAY (William Kahan, UC Berkeley) Message-ID: Reminder: Distinguished address by Professor William Kahan for all concerned with scientific computation Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 13, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 PROFESSOR WILLIAM KAHAN Math Dept and EECS Dept, UC Berkeley http://www.eecs.berkeley.edu/~wkahan Desperately Needed Remedies for the Undebuggability of Large Floating-Point Computations in Science and Engineering If suspicions about the accuracy of a computed result arise, how long does it take to either allay or justify them? Often diagnosis has taken longer than the computing platform's service life. Software tools to speed up diagnosis by at least an order of magnitude could be provided but almost no scientists and engineers know to ask for them, though almost all these tools have existed, albeit not all together in the same place at the same time. These tools would cope with vulnerabilities peculiar to Floating-Point, namely roundoff and arithmetic exceptions. But who would pay to develop the suite of these tools? Nobody, unless he suspects that the incidence of misleadingly anomalous Floating-Point results rather exceeds what is generally believed. Ample evidence supports that suspicion. Bio: Professor Kahan has been widely honored for his dedication to improving scientific computing algorithms and machines. 1989 ACM Turing Award 1993 Honorary Doctor of Mathematics, Chalmers Inst, Sweden 1994 ACM Fellow 1997 SIAM John von Neumann Memorial Lecture 1998 Honorary Doctor of Mathematics, Univ of Waterloo, Canada 2000 IEEE Emanuel R. Piore Award 2003 American Academy of Arts & Sciences 2004 Distinguished Mentor of Undergraduate Research in the College of Letters & Science 2005 National Academy of Engineering, Foreign (Canadian) Associate Forthcoming talks: Fri Oct 21 David Fong, ICME Thu Oct 27 Ming Gu, UC Berkeley Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore From mgu at math.berkeley.edu Thu Oct 13 14:45:15 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Thu Oct 13 14:46:10 2011 Subject: [BANANA] LAPACK seminar on Oct. 19, 2011 Message-ID: <201110132145.p9DLjFpH019470@phoenix.math.berkeley.edu> Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 19, 2011 Speaker: Dr. Joel R Phillips, Cadence Design Systems Title: Model Order Reduction in Electrical Circuit Analysis Abstract: The engineering of modern electronic systems is critically dependent on electronic design automation tools that enable the synthesis and analysis of very large-scale electrical circuits. One of the key components in contemporary circuit analysis has come from the application of model-order-reduction algorithms that evolved from a union of modern numerical linear algebra techniques and industrial experience in large-scale circuit simulation. Despite some widely-known open problems, little practical progress has been made in the past decade. This talk will begin by discussing the application domain, the types of starting systems it generates, how reduced-order-models interact with the rest of the analysis system, system requirements for algorithmic performance, and how these constraints have impacted algorithmic innovation in the area. Next we will discuss some open problems for linear-system reduction algorithms in an industrial context and outline some work in two directions: constructing models incrementally, and finding algorithms that are suited for systems with underlying sparse network topologies. Finally, we will discuss a problem that may be suitable for application of currently-available technology for reduction of nonlinear systems, discuss verification and validation of the models produced, and show some initial results. Date: Oct. 26, 2011 Speaker: Prof. William Kahan From saunders at stanford.edu Sat Oct 15 23:18:50 2011 From: saunders at stanford.edu (Michael Saunders) Date: Sat Oct 15 23:19:39 2011 Subject: [BANANA] Fwd: CME 500 Colloqium 10/17 (Michael Friedlander) In-Reply-To: <00e101cc8926$c489a7a0$4d9cf6e0$@edu> References: <00e101cc8926$c489a7a0$4d9cf6e0$@edu> Message-ID: Dear Stanford Linear Algebra and Optimization friends, The ICME Colloquium on Monday is of interest to our community. I hope you can come. ---------- Forwarded message ---------- From: Christopher Kabelac Date: Wed, Oct 12, 2011 at 2:34 PM Subject: CME 500 Colloqium 10/17 (Michael Friedlander) To: icme-students@lists.stanford.edu, icme-colloquium@lists.stanford.edu Cc: icme-faculty@lists.stanford.edu -------------------------------------------- ICME Colloquium -------------------------------------------- Michael Friedlander (University of British Columbia) Data fitting and optimization with randomized sampling. Monday, Oct 17, 2011 4:15pm ? 5:15pm Building 380, Rm. 380C For many structured data-fitting applications, incremental gradient methods (both deterministic and randomized) offer inexpensive iterations by sampling only subsets of the data. They make great progress initially, but eventually stall. Full gradient methods, in contrast, often achieve steady convergence, but may be prohibitively expensive for large problems. Applications in machine learning and robust seismic inversion motivate us to develop an inexact gradient method and sampling scheme that exhibit the benefits of both incremental and full gradient methods. -------------------------------------------- Future ICME Colloquium Speakers -------------------------------------------- 10/24/2011 (Monday) - Charles Taylor, Stanford University 10/31/2011 (Monday) - Krishna Gopinathan 11/07/2011 (Monday) - Daniel Niles, AlphaOne Capital 11/14/2011 (Monday) - David Wong, Havok.com 11/28/2011 (Monday) - Joe Perl, Stanford University 12/05/2011 (Monday) - Fumiko Hoeft, Stanford University -------------------------------------------- ICME Colloquium webpage: http://icme/seminars/seminarInfo.php?seminar_id=1 From mgu at math.berkeley.edu Mon Oct 17 14:32:14 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Mon Oct 17 14:33:52 2011 Subject: [BANANA] Reminder: LAPACK seminar on Oct. 19, 2011 In-Reply-To: <201110132145.p9DLjFpH019470@phoenix.math.berkeley.edu> References: <201110132145.p9DLjFpH019470@phoenix.math.berkeley.edu> Message-ID: Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 19, 2011 Speaker: Dr. Joel R Phillips, Cadence Design Systems Title: Model Order Reduction in Electrical Circuit Analysis Date: Oct. 26, 2011 Speaker: Prof. William Kahan From saunders at stanford.edu Wed Oct 19 12:12:48 2011 From: saunders at stanford.edu (Michael Saunders) Date: Wed Oct 19 12:13:45 2011 Subject: [BANANA] LA/Opt this week (David Fong, Friday 10/21) Message-ID: Dear Stanford Linear Algebra and Optimization colleagues, There is no CME 510 seminar tomorrow, but David Fong's PhD defense on Friday is an excellent substitute. Please note different time and place. ---------- Forwarded message ---------- From: Indira Choudhury Date: Fri, Oct 7, 2011 at 11:50 AM Subject: Announcing Oral Examination for David Fong, Friday 10/21 To: icme-faculty@lists.stanford.edu Student: David Chin-lung Fong Advisor: Michael Saunders Day/time: October 21, 2011, from 12:30pm Location: Packard 202 Minimum-Residual Methods for Sparse Least-Squares Using Golub-Kahan Bidiagonalization For 30 years, LSQR has been the standard iterative solver for large rectangular systems Ax~=b. ?It is analytically equivalent to symmetric CG on the normal equations A'Ax=A'b, and it reduces norm(rk) monotonically, where rk=b-A*xk is the k-th residual vector. The techniques pioneered in the development of LSQR allow better algorithms to be developed for a wider range of problems. We first present LSMR, an algorithm that is similar to LSQR but exhibits better convergence properties. ?LSMR is equivalent to applying MINRES to the normal equations, so that norm(A'rk) decreases monotonically. ?In practice we observe that norm(rk) is also monotonic like LSQR and the Stewart backward error norm(A'rk)/norm(rk) is usually very close to optimal. ?LSMR has essentially the same computational cost per iteration as LSQR, while having a more favorable convergence behaviour. ?Thus if iterations need to be terminated early, it is safer to use LSMR. LSQR and LSMR are based on Golub-Kahan bidiagonalization. ?For the second part of the presentation, we leverage the techniques used in deriving LSMR to construct algorithm AMRES for negatively-damped systems (A'A-d^2*I)x=A'b, again using Golub-Kahan bidiagonalization. This problem arises in total least-squares problems, Rayleigh quotient iteration (RQI), and Curtis-Reid scaling for rectangular sparse matrices. ?Our solver AMRES provides a stable method for these problems. ?AMRES also allows caching and reuse of the Golub-Kahan vectors across RQIs, and can be used to compute any of the singular vectors of A, given a reasonable estimate of the singular vector, or an accurate estimate of a singular value. Forthcoming talks: Thu Oct 27 Ming Gu, UC Berkeley Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore From mgu at math.berkeley.edu Fri Oct 21 20:17:49 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Fri Oct 21 20:18:49 2011 Subject: [BANANA] LAPACK seminar on Oct. 26, 2011 Message-ID: <201110220317.p9M3HnBK000627@phoenix.math.berkeley.edu> Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. Note that Prof. Kahan has prepared some reading material for his tutorial on Oct. 26 and Nov. 9 (see below.) Please read it before his lectures if possible. This might help prepare you to ask him questions during the lectures. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 26, 2011 Speaker: Prof. W. Kahan, UC Berkeley Title: A Tutorial Overview of Vector and Matrix Norms Abstract: Intended for new graduate students whose experience as undergraduates may have prepared them inadequately to apply norms to numerical error-analyses and to proofs of convergence, this tutorial surveys norms for finite-dimensional real spaces in a way that may ease a transition to the infinite-dimensional spaces of Functional Analysis. The tutorial?s notation is mostly standard but interpreted in ways not always taught to undergraduates, so attendees may prepare for it by reading just a few of my lecture notes for Math H110 posted at eecs.berkeley.edu/$\sim$wkahan/MathH110/2dspaces.pdf and ... /pts.pdf in that order and afterwards .../geo.pdf skimmed lightly. This tutorial omits proofs; almost all can be found in .../NORMlite.pdf and .../GIlite.pdf, and a few other places cited. This is the third of a series of four lectures on vector and matrix norms. The final lecture will be on Nov. 9. This tutorial?s text is to be posted at eecs.berkeley.edu/$\sim$wkahan/MathH110/NormOvrv.pdf Date: Nov. 2, 2011 Speaker: Oded Schwartz, UC Berkleey From saunders at stanford.edu Mon Oct 24 11:59:58 2011 From: saunders at stanford.edu (Michael Saunders) Date: Mon Oct 24 12:00:47 2011 Subject: [BANANA] ISL Special Seminar: Prof. Dominique Bonvin [10/25]-Modeling for Real-Time Optimization In-Reply-To: <4EA597F7.6060304@stanford.edu> References: <4EA597F7.6060304@stanford.edu> Message-ID: Special optimization seminar at Stanford University tomorrow Title: Modeling for Real-Time Optimization Speaker: Prof Dominique Bonvin, Laboratoire d?Automatique, EPFL, Lausanne, Switzerland (currently on leave at Stanford University) 4:00 - 5:00 pm, Packard 101 Tuesday, October 25, 2011 Refreshments served after the talk Abstract: Optimization in the process industry has received a lot of attention in recent years because, in the face of growing competition, it represents a natural choice for reducing production costs, improving product quality, and meeting safety requirements and environmental regulations. Traditionally, the optimal operating conditions are determined on the basis of a model of the process. However, the resulting process operation can be highly sensitive to uncertainty such as process-model mismatch and process disturbances. This generally gives rise to suboptimal process operation or, worse, infeasible operation, which of course is not acceptable in most industrial applications. Over the last decade, the Laboratoire d'Automatique of EPFL has developed a promising approach that converts a dynamic optimization problem with both path and terminal constraints into a feedback control problem. In this approach, near-optimal process operation can be enforced by tracking appropriate references, namely the necessary conditions of optimality (NCO). The NCO-tracking methodology is appealing in its on-line simplicity and? potential to be robust towards uncertainty. This presentation will start with a tutorial overview of dynamic optimization in the presence of uncertainty. The main discussion will concern ways of dealing with plant model uncertainty via appropriate modeling. Also, a systematic procedure for devising NCO-tracking controllers will be described and illustrated through several case studies. Speaker Bio: PROFESSIONAL EXPERIENCE 1989- Professor of Automatic Control, EPFL Lausanne, Switzerland 1983-89 Senior Lecturer, ETH Zurich, Switzerland 1980-83 Process Automation Engineer, Sandoz Ltd., Basel, Switzerland 1976-80 Ph.D. in Chem. Eng., University of California, Santa Barbara, USA 1971-75 Diploma in Chem. Eng., ETH Zurich, Switzerland PROFESSIONAL ACTIVITIES ? -Dean of Bachelor and Master Studies at EPFL, 2004-2011 ? -Head of the Mechanical Engineering Department, EPFL, 1995-97 ?-Director of the Laboratory of Automatic Control, EPFL, 1993-97, 2003-08 From saunders at stanford.edu Mon Oct 24 12:03:38 2011 From: saunders at stanford.edu (Michael Saunders) Date: Mon Oct 24 12:04:32 2011 Subject: [BANANA] LA/Opt seminar Thursday Oct 27 (Ming Gu) Message-ID: Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 27, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 Professor Ming Gu Dept of Mathematics, UC Berkeley http://www.eecs.berkeley.edu/~mgu Low-Rank Matrix Approximations and Randomized Sampling A classical problem in matrix computations is the efficient and reliable approximation of a given matrix by one of lower rank, with applications throughout wide areas of computational sciences and engineering. The truncated singular value decomposition (SVD) is known to provide the best such approximation for any given fixed rank. However, the SVD is very costly to compute. Recently, a number of randomized algorithms for low-rank matrix approximations have attracted researchers' attention because of their surprising reliability and computational efficiency in different application areas. In this talk, we present a novel analysis based on a connection between randomized algorithms and the traditional subspace iteration methods, allowing us to establish new and better error bounds for both. We present various numerical results that are in support of our analysis and show that different domain applications lead to different accuracy requirements, and therefore different oversampling sizes. Forthcoming: Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore Thu Nov 24 Thanksgiving Thu Dec 01 From mgu at math.berkeley.edu Tue Oct 25 00:32:54 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Tue Oct 25 00:33:52 2011 Subject: [BANANA] Reminder: LAPACK seminar on Oct. 26, 2011 In-Reply-To: <201110220317.p9M3HnBK000627@phoenix.math.berkeley.edu> References: <201110220317.p9M3HnBK000627@phoenix.math.berkeley.edu> Message-ID: Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. Note that Prof. Kahan has prepared some reading material for his tutorial on Oct. 26 and Nov. 9 (see below.) Please read it before his lectures if possible. This might help prepare you to ask him questions during the lectures. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Oct. 26, 2011 Speaker: Prof. W. Kahan, UC Berkeley Title: A Tutorial Overview of Vector and Matrix Norms Reading material can be found at eecs.berkeley.edu/~wkahan/MathH110/2dspaces.pdf and ... /pts.pdf in that order and afterwards .../geo.pdf skimmed lightly. This tutorial omits proofs; almost all can be found in .../NORMlite.pdf and .../GIlite.pdf, and a few other places cited. This tutorial?s text is to be posted at eecs.berkeley.edu/~wkahan/MathH110/NormOvrv.pdf This is the third of a series of four lectures on vector and matrix norms. The final lecture will be on Nov. 9. Date: Nov. 2, 2011 Speaker: Oded Schwartz, UC Berkleey From saunders at stanford.edu Thu Oct 27 11:03:33 2011 From: saunders at stanford.edu (Michael Saunders) Date: Thu Oct 27 11:04:33 2011 Subject: [BANANA] LA/Opt seminar TODAY (Ming Gu) Message-ID: Reminder: seminar this afternoon Linear Algebra and Optimization Seminar (CME 510) http://icme.stanford.edu/seminars/seminars.php 4:15pm Thursday Oct 27, 2011 Y2E2 111 http://campus-map.stanford.edu/index.cfm?ID=04-070 Professor Ming Gu Dept of Mathematics, UC Berkeley http://www.eecs.berkeley.edu/~mgu Low-Rank Matrix Approximations and Randomized Sampling A classical problem in matrix computations is the efficient and reliable approximation of a given matrix by one of lower rank, with applications throughout wide areas of computational sciences and engineering. The truncated singular value decomposition (SVD) is known to provide the best such approximation for any given fixed rank. However, the SVD is very costly to compute. Recently, a number of randomized algorithms for low-rank matrix approximations have attracted researchers' attention because of their surprising reliability and computational efficiency in different application areas. In this talk, we present a novel analysis based on a connection between randomized algorithms and the traditional subspace iteration methods, allowing us to establish new and better error bounds for both. We present various numerical results that are in support of our analysis and show that different domain applications lead to different accuracy requirements, and therefore different oversampling sizes. Forthcoming: Thu Nov 03 Thu Nov 10 Thu Nov 17 Richard Li-Yang Chen, Sandia Livermore Thu Nov 24 Thanksgiving Thu Dec 01 Oren Livne, U of Chicago From mgu at math.berkeley.edu Thu Oct 27 12:39:16 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Thu Oct 27 12:41:20 2011 Subject: [BANANA] LAPACK seminar on Nov. 2, 2011 Message-ID: <201110271939.p9RJdGQd026816@phoenix.math.berkeley.edu> Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Nov. 2, 2011 Speaker: Oded Schwartz, UC Berkeley Title: Minimizing communication costs: new upper and lower bounds using graph expansion considerations Abstract: The communication costs of algorithms are shown to be closely related to the expansion properties of the corresponding computation graphs. We demonstrate this on Strassen?s fast matrix multiplication algorithm, and obtain the first lower bounds on its communication costs, both for sequential and for parallel models. This bound is optimal for the sequential case, as it is attainable by a natural implementation. The technique extends to other classes of algorithms. Motivated by the graph-expansion approach, we then suggest parallel implementations for Strassen's and Strassen-like algorithms, that attain the communication costs lower bounds, and perform better in theory and in practice. Based on joint work with Grey Ballard, James Demmel, Olga Holtz, Ben Lipshitz, and Eran Rom. Date: Nov. 9, 2011 Speaker: W. Kahan, UC Berkeley Title: A Tutorial Overview of Vector and Matrix Norms From saunders at stanford.edu Fri Oct 28 19:18:16 2011 From: saunders at stanford.edu (Michael Saunders) Date: Fri Oct 28 19:19:10 2011 Subject: [BANANA] Special seminar by Prof Rudolf Kalman Message-ID: Special seminar of interest to the Linear Algebra and Optimization community (and others) A NEW APPROACH TO UNCERTAINTY AND STATISTICS Rudolf Kalman, Swiss Federal Institute of Technology, Zurich, Switzerland Monday 31 October 2011, 3:45pm Stanford University, Bldg 370, Room 370 (main quad) Especially in the English-speaking world, the mainstream approach to both problems has crystallized around a single hypothesis: (H) It is a law of Nature that randomness, uncertainty, indeterminacy, fuzziness, ..., and all that, is governed by probability. Rather than engaging in polemics or invoking Newton ("Hypotheses non fingo") we wish to make two simple remarks: 1. (H) is not true in the real world, and consequently is not a law of nature. 2. (H) is not necessary for a scientific analysis of randomness and uncertainty. We shall elaborate on the second point by presenting a mathematical critique (unpublished) of the "method of least squares" to illustrate the fact that dispensing with the a-priori assumptions centered on probability ("prejudices" in the scientific sense) leads to sharp insight and important new results. Bio: Prof Kalman received the IEEE Medal of Honor, the Kyoto Prize, and the National Medal of Science. From mgu at math.berkeley.edu Mon Oct 31 17:44:22 2011 From: mgu at math.berkeley.edu (Ming Gu) Date: Mon Oct 31 17:45:19 2011 Subject: [BANANA] Reminder: LAPACK seminar on Nov. 2, 2011 In-Reply-To: <201110271939.p9RJdGQd026816@phoenix.math.berkeley.edu> References: <201110271939.p9RJdGQd026816@phoenix.math.berkeley.edu> Message-ID: Math 290, Section 25, CS 298, Section 6 Fall 2011 (Matrix Computations and Scientific Computing) We meet WEDNESDAYS 12:10 - 1:00PM in Room 380 Soda Hall, Berkeley campus. The coordinators are Profs. J. Demmel (demmel@cs.berkeley.edu) and M. Gu (mgu@math.berkeley.edu). The program will be a mixture of research talks and tutorials. The tutorials will provide a partial sequel to Math 221. For the schedule and other details about the seminar, please see math.berkeley.edu/~mgu/LAPACKSeminar.htm Date: Nov. 2, 2011 Speaker: Oded Schwartz, UC Berkeley Title: Minimizing communication costs: new upper and lower bounds using graph expansion considerations Date: Nov. 9, 2011 Speaker: W. Kahan, UC Berkeley Title: A Tutorial Overview of Vector and Matrix Norms