sps newsletter 1 pdfsubject=stochastic ... - mail.airo.org · of the next one by david woodru );...
TRANSCRIPT
SPS NewsletterThe Newsletter of the Stochastic Programming Society
Volume 1 Number 1 March 2020
Contents
Chair’s ColumnGuzin Bayraksan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
ICSP XV Conference ReportAsgeir Tomasgard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
The Dupacova-Prekopa Best Student Paper PrizePhilip Thompson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4Weijun Xie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
In Memoriam:
Marida BertocchiGiorgio Consigli and Francesca Maggioni . . . . . . . 8
Maarten van der VlerkWard Romeijnders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Shabbir AhmedJim Luedtke and Jeff Linderoth . . . . . . . . . . . . . . . . 11
Real World Impact of Stochastic Programming in
Pension Funds: An Example from ChileBernardo Pagnoncelli . . . . . . . . . . . . . . . . . . . . . . . . . 13
Energy: An Example from EuropeWim van Ackooij . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15
SP Events in 2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Announcement of ICSP XVI
David Woodruff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Send your comments to the Editor:
Francesca Maggioni
Chair’s ColumnGuzin Bayraksan
Ohio State University (USA)
Dear Colleagues,
Welcome to the first-ever newsletter of theStochastic Programming Society! Our firstnewsletter comes at a time when many of us areforced to work at home, teach online from home,and limit our physical exposure to each other. Itis a time when decisions under uncertainty andrisk severely impact our daily lives.
One of the goals of the Committee on Stochas-tic Programming is to increase communicationamong our members and to create a strongersense of community. Toward this end, in addi-tion to our existing email list and website, we areinitiating this newsletter, where we share impor-tant announcements, highlight research advancesand real-world impact of our field.
I am also happy to announce that StochasticProgramming Society now has social media pres-ence. Please follow us on Twitter and LinkedIn:
https://twitter.com/stoprogsociety
https://www.linkedin.com/groups/
13799735/
I would like to extend special thanks toFrancesca Maggioni, who has agreed to be the
2 Stochastic Programming Society Newsletter
editor of the SPS Newsletter. Special thanks alsogoes to Phebe Vayanos (Twitter) and BernardoPagnoncelli (LinkedIn) for their efforts. Pleasecontinue sending us your events and announce-ments.
Talking about announcements, we have anearly announcement: The date for thenext ICSP in 2022 is set. You need to makeit to the end of the newsletter to find out. (No,you do not have to solve a stochastic program tofind out.) Please mark your calendars and blockoff that week!
It is exciting to see that our research area hasbeen receiving increased recognition. Here aresome highlights:
• Prof. Alex Shapiro has been elected to theNational Academy of Engineering (NAE),one of the highest honors accorded to anengineer, for his contributions to the theory,computation, and application of stochasticprogramming;
• Prof. Jorge Nocedal has been elected anNAE member, for his contributions to thetheory, design, and implementation of op-timization algorithms and machine learningsoftware, which include both deterministicand stochastic optimization;
• An increasing number of research outputand grants awarded on stochastic program-ming and its interface with other areas ofoptimization and machine learning;
• Many conferences have a growing number oftalks on various aspects of stochastic opti-mization, from theory to modeling to com-putations, with many talks appearing underother streams such as energy, finance, andhealthcare.
The first issue of the SPS Newsletter presentsarticles on the International Conference onStochastic Programming (report on the past con-ference by Asgeir Tomasgard, and announcementof the next one by David Woodruff); researchhighlights from Student Paper Prize Awardees,and In Memoriam of our colleagues who we lost
in the last few years. I would like use this oppor-tunity to thank Asgeir and the organization com-mittee for a great conference in Trondheim andlooking forward to our next big event in 2022.
In addition to research and conference high-lights, we are also presenting a series of articleson Real-World Impact of Stochastic Program-ming. Many thanks to COSP members BernardoPagnoncelli and Wim van Ackooij for initiatingthis series.
I wish all of you healthy, happy, creative days,and I look forward to seeing you at the nextmeeting that we are allowed to have.
ICSP XV ConferenceReport
Asgeir TomasgardNTNU Trondheim (Norway)
The 15th International Conference on Stochas-tic Programming was organized in Trondheimduring July 29–August 2, 2019. The conferencechair was Prof. Asgeir Tomasgard from NTNU,and the program committee chair was Prof. SteinW. Wallace (NHH) with co-chair Prof. Laure-ano F. Escudero (Universidad Rey Juan Car-los). In the weekend berfore the conference, thetraditional pre-conference tutorials were givenby Alejandro Jofre, Andrzej Ruszczynski, DavidWoodruff, Lei Zhao, Trine Krogh Boomsma, andUday V. Shanbhag.
The conference included plenaries, parallel ses-sions, 3 invited sessions, and 23 mini symposia,each featuring a semi-plenary and a stream oftalks. The 6 Plenary talks were given by Clau-dia Sagastizabal, George Lan, Guzin Bayraksan,Jong-Shi Pang, Stein-Erik Fleten, and RudigerSchultz.
With around 300 participants the conferencecovered topics like
• Risk-Averse Stochastic Programming
Vol.1 Number 1, March 2020 3
Figure 1: Insights from the closing plenary byProf. Rudiger Schultz
• Stochastic Programming in Energy
• Decomposition Techniques for Large-ScaleStochastic and Robust Energy System Mod-els
• Equilibrium Models in Energy and Trans-portation
• New Frontiers in Financial Decision Makingunder Uncertainty
• Robust and Distributionally Robust Opti-mization
• Methods and Computations in StochasticProgramming
• Equilibrium and Variational Inequalities
A five-day Ph.D. course was given the week be-fore the conference, organized by Stein W. Wal-lace, Suvrajet Sen, Stein-Erik Fleten, and AsgeirTomasgard. The course had around 50 partici-pants.
The Dupacova-Prekopa Best Student PaperPrize in Stochastic Programming was sponsoredby EWGSO and awarded to Philip Thomp-son (Instituto de Matematica Pura e Aplicada
Figure 2: The Dupacova-Prekopa Best StudentPaper Prize winner Philip Thompson and headof the prize committee Prof. Nilay Noyan
(IMPA), Brazil, advisers: Alfredo N. Iusemand Alejandro Jofre) for the paper by AlfredoN. Iusem, Alejandro Jofre, Roberto I. Oliveira,Philip Thompson, “Variance-Based Extragradi-ent Methods with Line Search For StochasticVariational Inequalities”, SIAM Journal on Op-timization, 29 (1): 175–206, 2019 with runnerup Weijun Xie (Georgia Tech, adviser: Shab-bir Ahmed), and finalists Rui Peng Liu (Geor-gia Tech, adviser: Alex Shapiro) and Junyi Liu(University of Southern California, adviser: Su-vrajeet Sen).
4 Stochastic Programming Society Newsletter
The ICSP XVDupacova-Prekopa
Student Paper Prize
Stochastic line-search forstochastic approximation:
adaptation to Lipschitz constant
Philip ThompsonInstituto de Matematica Pura e Aplicada (IMPA),
Rio de Janeiro (Brazil)
The 2019 Dupacova-Prekopa Best Student Pa-per Prize was awarded to Philip Thompson(IMPA) at ICSP XV for the paper “Variance-Based Extragradient Methods with Line Searchfor Stochastic Variational Inequalities,” coau-thored with Alfredo N. Iusem, Alejandro Jofre,and Roberto I. Oliveira, published in SIAM Jour-nal on Optimization, 29(1): 175–206, 2019. Be-low is a summary of this paper.
Consider the classical stochastic opti-mization problem (SO): solve the problemminx∈Rd{f(x) := E[G(ξ, x)]} given a ran-dom function G : Ξ × Rd → R. Here f hasL-Lipschitz continuous gradients and the ex-pectation is taken with respect to a randomvariable ξ. The Stochastic Approximation (SA)methodology was first proposed by Robbins andMonro: xk+1 := xk − αk∇G(ξk, xk), given ani.i.d. sample sequence {ξk} and positive stepsizesequence {αk}. This is a first instance of thestochastic gradient method (SG).
Another problem of interest is the so calledstochastic variational inequality (SVI): Given aclosed convex set X ⊂ Rd and a random op-erator F : Ξ × X → Rd with L-Lipschitz con-tinuous mean operator T (x) := E [F (ξ, x)] , thestochastic variational inequality problem (SVI),denoted as VI(T,X), is the problem of find-ing a point x∗ ∈ X such that, for all x ∈X, 〈T (x∗), x − x∗〉 ≥ 0. In case of monotone
operators, the SA version of the extragradientmethod (SEG) is given by the subsequent iter-ations zk := Π
[xk − αkF (ξk, xk)
]and xk+1 :=
Π[xk − αkF (ξk, zk)
]. Here, Π denotes the Eu-
clidean projection onto X. The mentioned unre-stricted SO problem is a special instance of theSVI with X := Rd, F (ξ, x) := ∇xG(ξ, x) andT := ∇f . By now the literature on stochasticapproximation is very large spanning the com-munities of statistics and stochastic approxima-tion, stochastic optimization and machine learn-ing. Many variants of SG and SEG are now avail-able and a proper review is out of the scope. Werefer for the recent review [1]. A proper reviewfor SVIs are out the scope. See e.g. references in[2, 3].
As widely known in practice, the performanceof SG and SGE are sensitive to the choice ofstepsize sequence. Small stepsize policies likeαk = θk−1 or αk = θk−1/2 with an additionalfinal iterative average have the advantage of notrequiring precise knowledge of L. Still, they canbe slow and still quite sensitive to θ, motivatingsome sort of adaptation [4, 5]. Fairly recently,variance-reduction methods were proposed ei-ther using gradient aggregation methods (SAG,SAGA, SVRG, Catalyst, etc) or mini-batchingmethods (see e.g. [1] and references there in).These methods aim for a constant stepsize policyαk := O(L−1), improving the iteration complex-ity without serious compromise on the samplecomplexity1. The downside of these methods isthat they require knowledge of L. Different ap-proaches to L-adaptation in SG and SGE havebeen considered: stochastic line-search [6, 7], in-cluding [8] reviewed in this article, adaptationvia probabilistic models or heuristics (see refer-ences in [7] and [9]) and AdaGrad [10] and itsvariants such as Adam. For those with con-vergence guarantees, the interesting works by[6, 9] are based on martingale-difference meth-ods. They are based on a probabilistic modelrequirement which, in a general setting, wouldtypically still require knowledge of some param-eters related to the confidence level of the ap-
1These methods assume that the stochastic oracle canbe queried multiple times per iteration.
Vol.1 Number 1, March 2020 5
proximation. Previous work still imposed stronguniform approximation assumptions that are notobvious from the standard i.i.d. sampling as-sumption. The interesting recent article [7] es-tablishes a simplified line-search analysis withoutdynamic mini-batching under the interpolationassumption: this is satisfied by certain NeuralNetworks but not for a generic stochastic opti-mization problem. Finally, we should mentionthe success and widespread use of AdaGrad andits variants in the machine learning community.The success of AdaGrad does not come withoutlimitations: it does not attain the improved it-eration complexity of variance-reduced methodsand strongly requires an uniform boundednessassumption on the gradient.
In [8], we have proposed a stochastic line-search method which does not require any pa-rameter knowledge, assuming the standard i.i.d.sampling assumption and that mini-batches arepossible. Concerning the tolerance ε > 0, weprovide an efficient dynamic mini-batching pol-icy attaining the optimal O(ε−2) sample com-plexity (up to log factors) and the improved it-eration complexity O(ε−1) for monotone SVIsand smooth convex SOs shared by variance-reduced methods. Our guarantees are valid un-der local variance assumptions. Hence we al-low the less standard setting where the varianceE[‖F (ξ, x)−T (x)‖2] is not bounded over x (mul-tiplicative noise). See also [11] for the convexsetting and [12, 13] for the strongly-convex set-ting. At the core of our analysis is to properlyaddress the fact that simple line search methods2
with efficient mini-batching break the martingaledifference property of stochastic approximation.We handle this fact and mutiplicative noise bymaking use of a novel iterative localization ar-gument based on empirical process theory. Weinvite the reader to see [8] for more details.
REFERENCES
[1] L. Bottou, F. E. Curtis and J. Nocedal. Optimiza-tion methods for large-scale machine learning,
2By this we mean line search methods using cheapstochastic approximation with no a priori confidence levelrequirements.
SIAM Review, 60(2), 223-311, 2018.
[2] A.N. Iusem, A. Jofre, R.I. Oliveira and P. Thomp-son. Extragradient method with variance re-duction for stochastic variational inequalities,SIAM Journal Optimization, 27(2), 686-724,2017.
[3] A.N. Iusem, A. Jofre and P. Thompson. Incremen-tal constraint projection methods for monotonestochastic variational inequalities, Mathematicsof Operations Research, 44(1), 236-263, 2019.
[4] F. Yousefian, A. Nedich, and U. V. Shanbhag. Onstochastic gradient and subgradient methodswith adaptive steplength sequences Automatica,48(1), 56-67, 2012.
[5] R. Pasupathy, H. Honnappa and S.R. Hunter. Openproblem-adaptive constant-step stochastic ap-proximation, Stochastic Systems, 9(3), 307-310,2019.
[6] C. Paquette and K. Scheinberg. A Stochastic linesearch method with convergence rate analysis,https://arxiv.org/abs/1807.07994.
[7] S. Vaswani, A. Mishkin, I. Laradji, M. Schmidt G.Gidel and S. Lacoste-Julien. Painless stochasticgradient: interpolation, line-search, and conver-gence rates, to appear in NeurIPS 2019.
[8] A. Iusem, A. Jofre, R. I. Oliveira and P. Thomp-son. Variance-based extragradient methods withline search for stochastic variational inequali-ties, SIAM Journal Optimization, 29(1), 175-206, 2019.
[9] R. Chen, M. Menickelly and K. Scheinberg. Stochas-tic optimization using a trust-region methodand random models, Mathematical Program-ming, 169(2), 447-487, 2018.
[10] J. Duchi, E. Hazan and Y. Singer. Adaptive subgra-dient methods for online learning and stochasticoptimization, Journal of Machine Learning Re-search, 12, 2121-2159, 2011.
[11] A. Jofre and P. Thompson. On variance reduc-tion for stochastic smooth convex optimizationwith multiplicative noise, Mathematical Pro-gramming, 174 (1-2), 253-292, 2019.
[12] L.M. Nguyen, P.H. Nguyen, M. van Dijk, P.Richtarik, K. Scheinberg and M. Takac. SGDand Hogwild! Convergence without thebounded gradients assumption, ICML 2018.
[13] R.M. Gower, N. Loizou, X. Qian, A. Sailanbayev, E.Shulgin and P. Richtarik. SGD: general analysisand improved rates, ICML 2019.
6 Stochastic Programming Society Newsletter
The ICSP XVDupacova-Prekopa
Student Paper Prize
On deterministic reformulationsof distributionally robust joint
chance constrained optimizationproblems
Weijun XieDepartment of Industrial and Systems Engineering,
Virginia Tech (USA)
The Runner-up of the 2019 Dupacova-PrekopaBest Student Paper Prize was awarded to Wei-jun Xie (Georgia Tech) at ICSP XV for the pa-per “On Deterministic Reformulations of Dis-tributionally Robust Joint Chance ConstrainedOptimization Problems,” coauthored with adviserShabbir Ahmed, published in SIAM Journal onOptimization, 28(2):1151–1182, 2018. Below isa summary of this paper.
Setting: We consider a distributionally robustchance constrained program (DRCCP) of theform:
v∗ = minx
c>x, (1a)
s.t. x ∈ S, (1b)
infP∈P
P[ξ : F (x, ξ) ≥ 0] ≥ 1− ε, (1c)
where x ∈ Rn is a decision vector; the vectorc ∈ Rn denotes the objective coefficients; theset S ⊆ Rn denotes deterministic constraints onx; the random vector ξ supported on Ξ ⊂ Rm
denotes uncertain constraint coefficients with arealization denoted by ξ; the mapping F (x, ξ) :=(f1(x, ξ), . . . , fI(x, ξ))> with fi(x, ξ) : Rn × Ξ→R for all i ∈ [I] := {1, . . . , I} defines a set ofuncertain constraints on x; the ambiguity set Pdenotes a set of probability measures P on thespace Ξ with a sigma algebra F ; and ε ∈ (0, 1)
denotes a risk tolerance. In (1) we seek a deci-sion vector x to minimize a linear objective c>xsubject to a set of deterministic constraints de-fined by S, and a chance constraint (1c) that isrequired to hold for any probability distributionfrom the ambiguity set P with a probability of1− ε. Note that when |I| = 1 the constraint (1c)involves a single chance constraint and if |I| ≥ 2it involves a joint chance constraint.
Figure 1: The Dupacova-Prekopa Best StudentPaper Prize runner up Weijun Xie attendingICSP XV via Skype (photo credit: Kai Pan)
Assumption: The primary difficulty of (1) isdue to the distributionally robust chance con-straint (1c). Let us denote the feasible regioninduced by (1c) as
Z =
{x ∈ Rn : inf
P∈PP[ξ : F (x, ξ) ≥ 0] ≥ 1− ε
}.
In this paper we study deterministic reformula-tions of the set Z and its convexity properties.Our study is restricted to the convex, momentconstrained setting (see, e.g., [1, 2]). That is, weassume that the ambiguity set P is nonemptyand is defined by moment constraints:
P = {P : EP[φt(ξ)] = gt, t ∈ T1,EP[φt(ξ)] ≥ gt, t ∈ T2}
where for each t ∈ T1∪T2, the moment functionφt : Ξ → R is a real valued continuous functionand gt is a scalar. Furthermore, for each t ∈ T1,
Vol.1 Number 1, March 2020 7
the function φt(ξ) is linear, and for each t ∈ T2,the function φt(ξ) is concave.
Main Contributions: The DRCCP set Z isnonconvex in general, making (1) a difficult op-timization problem. In this paper we first pro-vide a deterministic approximation of Z that isnearly tight and then identify a variety of set-tings under which Z is convex. The main resultsof this paper are summarized next.
1. We propose a deterministic conservative ap-proximation of set Z, which is in generalnonconvex and can be formulated as an op-timization problem involving biconvex con-straints.
2. If there is a single uncertain constraint, i.e.|I| = 1, we prove that the proposed deter-ministic approximation is exact and reducesto a tractable convex program. This resultis a generalization of existing works (e.g.,[3, 4, 5]) to arbitrary convex ambiguity setsrather than those involving only first andsecond moment constraints.
3. We prove that if the ambiguity set P con-tains only one moment inequality, i.e. |T1| =0 and |T2| = 1, then Z is a tractable con-vex program; and if the ambiguity set con-tains only one moment linear equality, i.e.|T1| = 1 and |T2| = 0, then Z is equivalentto the disjunction of two tractable convexprograms.
4. We prove that if Ξ = Rm and the momentfunctions {φt(ξ)}t∈T1∪T2 are linear, then Z isequivalent to the feasible region of a robustconvex program.
5. We prove that if Ξ is a closed convex cone,the function fi(x, ξ) for any i ∈ [I] is ofthe (separable) form fi(x, ξ) = wi(x)−hi(ξ)where hi(ξ) is positively homogeneous on Ξ,and the moment functions {φt(ξ)}t∈T2 arepositively homogeneous on Ξ, then set Z isconvex. This result is a generalization of [6],where the authors assumed that wi(·) andhi(·) are affine functions for each i ∈ [I].
6. When the decision variables are pure binary(i.e. S ⊆ {0, 1}n) and uncertain constraintsare linear, we show that the proposed deter-ministic approximation can be reformulatedas a mixed integer convex program. We alsopresent a numerical study to demonstratethat the proposed reformulation can be ef-fectively solved using a standard solver.
REFERENCES
[1] A. Shapiro and S. Ahmed. On a class of mini-max stochastic programs, SIAM Journal on Op-timization, 14, 1237–1249, 2004.
[2] A. Shapiro and A. Kleywegt. Minimax analysis ofstochastic problems, Optimization Methods andSoftware, 17, 523–542, 2002.
[3] G.C. Calafiore and L. El Ghaoui. On distribution-ally robust chance-constrained linear programs,Journal of Optimization Theory and Applica-tions, 130, 1–22, 2006.
[4] W. Yang and H. Xu. Distributionally robust chanceconstraints for non-linear uncertainties, Mathe-matical Programming, 155, 231–265, 2016.
[5] S. Zymler, D. Kuhn and B. Rustem. Distribution-ally robust joint chance constraints with second-order moment information, Mathematical Pro-gramming, 137, 167–198, 2013.
[6] G.A. Hanasusanto, V. Roitch, D. Kuhn andW. Wiesemann. Ambiguous joint chance con-straints under mean and dispersion information,Operations Research, 65(3), ii-iv, 2017.
8 Stochastic Programming Society Newsletter
In Memoriam:Marida Bertocchi
Giorgio Consigli and Francesca MaggioniDepartment of Management, Economics and
Quantitative Methods, University of Bergamo (Italy)
Maria Ida (Marida for friends and colleagues)Bertocchi was born in Bergamo on August 14th,1951. She passed away on November 16th, 2016after a sudden and dramatic illness and a verycomplex surgery. This article is dedicated to hermemory, following a memorial in her honor heldin Bergamo on May 28th, 2017 and a special ses-sion held at the XV International Conference onStochastic Optimization (ICSP XV) in Trond-heim.
Since the 80’s, Marida’s extremely kind, posi-tive and productive personal and professional at-titudes led her to become rapidly a central andinfluential figure of the University of Bergamo inparticular and the Italian mathematical commu-nity at large. Throughout her career Marida fos-tered many international collaborations and sup-ported young scholars on a variety of mathemat-ical fields and application areas with unboundedenergy, good humour, outstanding social skillsand a keen intellect.
Her academic excursus started after graduat-ing at the Faculty of Mathematics of the Univer-sity of Milano in 1974. In 1979, Marida joinedthe Faculty of Economics of the University ofBergamo and in 1985 she was appointed Asso-ciate Professor. In 1989, after being awardedthe Full Professorship, she moved to the Uni-versity of Urbino, before returning in 1992 tothe University of Bergamo, where she has beenever since. Marida was very active at all train-ing levels, from undergraduate to doctoral stud-ies: after being appointed Chair in Mathemat-ical Methods for Economics and Finance, shestarted a very successfull and well-known PhDprogram in Computational Methods for Finan-cial and Economic Forecasting and Decisions,
later named Analytics for Economics and Busi-ness, that she coordinated until 2016. She servedas University Vice-chancellor in the academicyear 1995-1996 before becoming Elective Deanof the Faculty of Economics for two terms, from1996 to 2002 and later on head of the Depart-ment of Mathematics and Statistics from 2009to 2012.
Her early research interests were primarily inoptimization and numerical analysis [1, 2], be-fore being increasingly committed to computa-tional and applied topics of stochastic program-ming. Marida’s interest in stochastic program-ming developed since 1992 thanks to a lastingcooperation and friendship with Jitka Dupacova,when they met at the XVI AMASES Conferencein Treviso (Italy), where Jitka gave a plenarytalk on the topic. They started a fruitful col-laboration on several financial optimization ar-eas [3, 4, 5]. Later on, also thanks to the grow-ing group of colleagues and collaborators intro-duced or already working in this field, includingthe two of us, Marida’s research interests spreadto methodological topics [6, 7] and new applica-tion areas in energy and finance [8, 9, 10, 11, 12]always involving stochastic modeling and ad-vanced computational approaches.
Marida has been a very active member of theStochastic Programming community. Amongthe founding members of the European Work-ing Group in Stochastic Programming (SP) es-tablished in Neringa (LT) in July 2012, sheorganized and was behind several SP confer-ences and streams within International meetings,from the extremely successfull 2007-2009 Inter-national Cariplo PhD program to the XIII In-ternational SP Conference held in Bergamo in2013.
In June 2016, right before getting ill, Maridaattended the 14th International Conference onStochastic Programming in Buzios, Brazil. OnSeptember 15th, 2016, already ill, she gave a ple-nary lecture on Stochastic Programming in Eco-nomics and Finance at the 40th AMASES Con-ference in Catania. Her strength and devotionduring her lecture impressed all the participantsin the meeting who will always remember herlast lecture.
Vol.1 Number 1, March 2020 9
Shortly after, Marida was diagnosed with avery severe form of cancer, that despite a com-plex and challenging surgery, took her away, verysadly in mid November 2016. Since then, notonly us but also her former colleagues, Adri-ana Gnudi, Rosella Giacometti, Sergio Ortobelliand Vittorio Moriggia, as well as younger mem-bers of the research group are committed to fur-ther strenghten and consolidate Marida’s scien-tific and human legacy.
A World Scientific volume including a selectedcollection of Marida’s articles was just publishedfew weeks ago [13], while a special issue in hermemory in Annals of Operations Research enti-tled “Stochastic Optimization: Theory and Ap-plications” with guest editors Giorgio Consigli,Darinka Dentcheva and Francesca Maggioni isforthcoming.
REFERENCES
[1] M. Bertocchi and E. Spedicato. Computational ex-perience with conjugate gradient algorithms,Calcolo, 16(3), 255–269, 1979.
[2] V. Moriggia, M. Bertocchi and J. Dupacova. Highlyparallel computing in simulation on dynamicbond portfolio management, Proceedings of theAPL98 Conference, pp. 215–221, 1998.
[3] M. Bertocchi, V. Moriggia and J. Dupacova. Sensi-tivity of bond portfolio’s behavior with respectto random movements in yield: a simulationstudy, Annals of Operations Research, 99(1),267–286, 2000.
[4] M. Bertocchi, V. Moriggia and J. Dupacova. Bondportfolio management via stochastic program-ming, Handbook of Asset and Liability Manage-ment, 1, 305–336, 2008.
[5] J. Dupacova and M. Bertocchi. From data to modeland back to data: a bond portfolio managementproblem, European Journal of Operational Re-search, 134(2), 261–278, 2001.
[6] F. Maggioni, E. Allevi and M. Bertocchi. Bounds inmultistage linear stochastic programming, Jour-nal of Optimization Theory and Applications,163(1), 200–229, 2014.
[7] F. Maggioni, E. Allevi and M. Bertocchi. Mono-tonic bounds in multistage mixed-integer linearstochastic programming, Computational Man-agement Science, 13(3), 423–457, 2016.
[8] M. Bertocchi, G. Consigli and M.A.H. Demp-ster (Eds.). Handbook on stochastic optimiza-tion methods in finance and energy. New finan-cial products and energy market strategies, FredHillier International Series in Operations Re-search and Management Science, Springer (US),2011.
[9] M. Bertocchi, S. Schwartz and W.T. Ziemba.Optimizing the aging, retirement and pensiondilemma, John Wiley & Sons Inc., Hoboken,New Jersey, 2010.
[10] R. Giacometti, S. Ortobelli and M. Bertocchi. AStochastic model for mortality rate on Italiandata, Journal of Optimization, Theory and Ap-plications, 149, 216–228, 2011.
[11] R. Giacometti, M.T. Vespucci, M. Bertocchi andG.B. Adesi. Hedging electricity portfolio for ahydro-energy producer via stochastic program-ming, International Series in Operations Re-search and Management Science, 163, 163–179,2011.
[12] F. Maggioni, M.T. Vespucci, E. Allevi, M. Bertoc-chi, R. Giacometti and M. Innorta. A stochasticoptimization model for gas retail with temper-ature scenarios and oil price parameters, ImaJournal of Management Mathematics, 21(2),149-–163, 2010.
[13] R.L. D’Ecclesia, S.A. Zenios and W.T. Ziemba(Eds.). Collected works of Marida BertocchiBook Series: World Scientific Handbook in Fi-nancial Economics Series, Vol. 8, 2019.
10 Stochastic Programming Society Newsletter
In Memoriam:Maarten van der Vlerk
Ward RomeijndersDepartment of Operations,
University of Groningen (Netherlands)
Maarten van der Vlerk was born in the year1961 in Assen, The Netherlands. He became adoctoral student at the University of Groningenin the Department of Econometrics under thesupervision of Wim Klein Haneveld and LeenStougie in 1990. Invited and encouraged by hissupervisors, Maarten placed a research focus onthe integer programming side of stochastic opti-mization.
He obtained his PhD at the University ofGroningen in 1995 for his thesis “Stochastic Pro-gramming with Integer Recourse”. After a pe-riod at CORE (Louvain-la-Neuve) he returned toGroningen and in 1999 he received a prestigiousresearch fellowship from the Royal NetherlandsAcademy of Arts and Sciences (KNAW).
He would go on to be named professor ofStochastic Optimization at the Faculty of Eco-nomics and Business, and director of the bache-lor and master programmes Econometrics & Op-erations Research and Econometrics, OperationsResearch & Actuarial Studies. He was also mem-ber of the general board of the LNMB, the DutchNetwork on the Mathematics of Operations Re-search, and from 2004 until 2007 he was Chairof the Stochastic Programming community.
On October 9, 2016, Maarten van der Vlerkpassed away at the age of 54.
From the mid 1990s on, stochastic integerprogramming became the field where Maartenvan der Vlerk’s research earned highest recogni-tion in the stochastic programming communityand beyond. A permanently recurring researchtarget of Maarten, and his various coauthors,has been convex approximation of the notori-ously non-convex, discontinuous, mixed-integerexpected recourse function, together with its al-gorithmic utilization. Papers devoted to these
topics range over a time span of more than 20years, see also [1]. Prominent examples include[2], [3], and [4].
Maarten was not only a brilliant researcher,internationally renowned for his work in Stochas-tic Programming, he was also an outstandingteacher with a gift for imparting knowledge. Hewon the FEB lecturer of the year award in 2014.His courses, Stochastic Programming and theSpecialization Course Applied Operations Re-search, were in the top five of his faculty al-most every year. From information meetingsfor prospective students until their master de-gree ceremonies, his dedication to students wasexemplary.
Maarten was also a very active member of theStochastic Programming community. He devel-oped the Stochastic Programming Bibliographyearly in his career, and served as Secretary andthen Chair of COSP. He was a tutorial speakerat several ICSP conferences, and he organizeda workshop on Stochastic Integer Programmingin Groningen in 2004. His Stochastic Program-ming lecture notes, that he shared with manycolleagues, have recently been published as text-book [5] by Springer.
Apart from beautiful mathematics Maartenenjoyed many other good things in life, which hehappily shared with his friends and colleagues:music, in particular Bach, and tennis, which hemade a fixed point of the social program of any
Vol.1 Number 1, March 2020 11
stochastic programming meeting. We will missMaarten’s boisterous laugh, his warm smile, hissense of humor, and his ability to make everyonearound him a better person.
REFERENCES
[1] D.P. Morton, W. Romeijnders, R. Schultz and L.Stougie. The stochastic programming heritageof Maarten van der Vlerk, Computational Man-agement Science, 15(3-4), 319–323, 2018.
[2] M.H. van der Vlerk. Convex approximations forcomplete integer recourse models, MathematicalProgramming, 99, 297–310, 2004.
[3] W.K. Klein Haneveld, L. Stougie and M.H. van derVlerk. Simple integer recourse models: convex-ity and convex approximations, MathematicalProgramming, 108, 435–473, 2006.
[4] W. Romeijnders, R. Schultz, M.H. van der Vlerkand W.K. Klein Haneveld. A convex approxima-tion for two-stage mixed-integer recourse mod-els with a uniform error bound, SIAM Journalon Optimization, 26, 426–447, 2016.
[5] W.K. Klein Haneveld, M.H. van der Vlerk and W.Romeijnders. Stochastic programming: Model-ing decision problems under uncertainty, Grad-uate texts in Operations Research, Springer,2020.
In Memoriam:Shabbir Ahmed
Jim Luedtke and Jeff LinderothDepartment of Industrial and Systems Engineering,
University of Wisconsin-Madison (USA)
[email protected] [email protected]
Our friend and colleague Shabbir Ahmedpassed away on June 19th, 2019 at the age of49 after a hard-fought battle with cancer. Shab-bir was born in Bangladesh, received a B.Eng.in mechanical engineering from Bangladesh Uni-versity of Engineering and Technology in 1993,and M.S. and Ph.D. degrees in industrial engi-neering from the University of Illinois at Urbana-Champaign in 1997 and 2000. Shabbir joinedthe Industrial and Systems Engineering facultyat Georgia Tech in 2000, and he held the ti-tle of Anderson-Interface Chair and Professor atthe time of his passing. Shabbir leaves behindhis loving wife Rasha and daughters Raeeva andUmana.
At the memorial session held in Shabbir’shonor at the ICSP meeting in Trondheim, hisformer students and colleagues spoke of Shab-bir’s many contributions to the field. But themost resounding message heard from the speak-ers was how deeply Shabbir cared about his stu-dents and in fostering the next generation of re-searchers in stochastic programming. As morethan one speaker recalled, Shabbir would oftensay that students “were the blood” of any goodresearch program. Shabbir graduated 28 Ph.D.students during his 19 years at Georgia Tech.
Shabbir’s selfless focus on students carriedover to his service work. Shabbir served asthe Secretary of Stochastic Programming Soci-ety (SPS) from 2007-2010, and he re-aligned SPSas an official technical section under the Mathe-matical Optimization Society (MOS). However,Shabbir’s proudest accomplishment during histenure was the creation of the Stochastic Pro-gramming Society Student Paper Prize.
Shabbir’s academic contributions are almosttoo numerous to mention, with nearly 100 jour-nal articles and more than 8000 citations to his
12 Stochastic Programming Society Newsletter
credit. Many of his seminal contributions werein stochastic programming. He made signifi-cant contributions to the fields of stochastic in-teger programming [1, 2, 3], chance-constraints[4, 5, 6, 7], and risk-averse methods [8, 9, 10].Shabbir worked vigorously on the application ofstochastic programming in important practicaldomains such as supply chain planning [11], ve-hicle routing [12], and energy systems [13].
For his pioneering work on piecewise-linear op-timization, Shabbir was awarded the INFORMSComputing Society Prize in 2017. In 2018, hewon the Farkas Prize for his outstanding contri-butions to the optimization community from theINFORMS Optimization Society. Shabbir waselected a Fellow of INFORMS in 2017.
Shabbir enjoyed traveling with his family; run-ning; he was a skilled table-tennis player; andhe was an extremely-talented heavy-metal mu-sician. Most of all, Shabbir was a staggeringintellect, an encouraging and energetic force forthe optimization community, a devoted husband,a loving father, and a wonderful colleague andfriend. We all miss him very deeply, but we areconsoled by our memories of him and the lastinglegacy he left behind for our community.
REFERENCES
[1] S. Ahmed, M. Tawarmalani, and N. Sahinidis. A fi-nite branch and bound algorithm for two-stagestochastic integer programs, Mathematical Pro-gramming, 100, 355–377, 2004.
[2] Y. Guan, S. Ahmed, and G. Nemhauser. Cut-ting planes for multistage stochastic integer pro-grams, Operations Research, 57(2), 287–298,2009.
[3] J. Zou, S. Ahmed, and X.A. Sun. Stochastic dualdynamic integer programming, MathematicalProgramming, 175, 461–502, 2019.
[4] J. Luedtke and S. Ahmed. A sample approxima-tion approach for optimization with probabilis-tic constraints, SIAM Journal on Optimization,19, 674–699, 2008.
[5] J. Luedtke, S. Ahmed, and G.L. Nemhauser. Aninteger programming approach for linear pro-grams with probabilistic constraints, Mathe-matical Programming, 122, 247–272, 2010.
[6] W. Xie and S. Ahmed. On quantile cuts andtheir closure for chance constrained optimiza-tion problems, Mathematical Programming,172, 621–646, 2018.
[7] W. Xie and S. Ahmed. Bicriteria approximation ofchance constrained covering problems, Opera-tions Research, 2020.
[8] A. Shapiro and S. Ahmed. On a class of minimaxstochastic programs, SIAM Journal on Opti-mization, 14, 1237–1249, 2004.
[9] S. Ahmed. Convexity and decomposition of mean-risk stochastic programs, Mathematical Pro-gramming, 106, 433–446, 2006.
[10] S. Ahmed and A. Atamturk. Maximizing a classof submodular utility functions, MathematicalProgramming, 128, 149–169, 2011.
[11] T. Santoso, S. Ahmed, M. Goetschalckx, andA. Shapiro. A stochastic programming ap-proach for supply chain network design underuncertainty, European Journal of OperationalResearch, 167, 96–115, 2005.
[12] B. Verweij, S. Ahmed, A. Kleywegt, G. Nemhauser,and A. Shapiro. The sample average approx-imation method applied to stochastic routingproblems: A computational study, Computa-tional Optimization and Applications, 24, 289–333, 2003.
[13] J. Zou, S. Ahmed, and X. A.Sun. Partially adaptivestochastic optimization for electric power gener-ation expansion planning, INFORMS Journalon Computing, 30(2), 217–240, 2018.
Vol.1 Number 1, March 2020 13
Real World Impact of StochasticProgramming in Pension Funds:
An Example from Chile
Bernardo PagnoncelliSchool of Business
Universidad Adolfo Ibanez, Santiago (Chile)
Impact Summary: How to measure risk in adefined contribution pension system? In Chile,the oldest of such systems, the regulator recentlyannounced that risk measures will replace port-folio constraints as a way of defining investmentalternatives. The decision was based on resultsfrom a paper that uses Stochastic Programmingtechniques. The Chilean system manages around80% of the country’s GDP, and it is a blueprintfor many systems around the world.
At some point in our lives, most of us need tostop working, and we all hope to have a goodpension plan to cover our expenses during re-tirement. Governments and private companiesaround the world offer different alternatives toachieve that. In all cases, there is an accumula-tion phase where active workers contribute to-wards retirement (which ends after working acertain number of years, and if workers are pastsome age), and a decumulation period where re-tirees are eligible to receive their benefits.
The most basic retirement schemes are definedbenefit (DB) and defined contribution (DC). InDB systems, the worker is guaranteed to havea fixed benefit after retirement, and the contri-bution during the active phase may vary. InDC systems, the contribution to retirement isa fixed percentage of the salary, but the annu-ity that will be received during retirement de-pends on how well the money was invested andon the returns of the assets in the active phase.Many countries use hybrid schemes, combiningelements from both systems. However, it is un-deniable that in recent decades there has been
a shift to DC systems [1, 2]. The reasons forthis trend include a sharp decrease in fertility ra-tios, longevity risk and a future with extremelylow interest rates. I am going to discuss two is-sues that appear in all DC systems, and I willuse the Chilean case to illustrate the discussionsince it is the oldest (and purest) DC system inthe world. Moreover, deflated data is publiclyavailable, which makes it a natural candidate forexperiments [3].
The first aspect I want to address is how tospecify the financial instruments that will beavailable to workers during the active phase. InChile, a DC system was imposed in 1980, and in2002 it reached its more mature phase with thecreation of five different funds, A, B, C, D and E.Fund A is supposed to be the riskiest, and fund Ethe safest, and each fund is defined by upper andlower bounds on some financial instruments (seeTable 1 for a sample of the constraints imposedby the regulator). We claim that such a systemis based on proxies for risk: it is expected that,by allowing a higher exposure to stocks, fundA will have higher returns—and higher risk—which will decrease gradually until fund E (lowerreturns, lower risk). As an example, a portfo-lio composed of 80% of US stocks and 20% oftreasury bonds does not have the same risk as aportfolio consisting of 80% of stocks from a morevolatile emerging market (and the same 20% oftreasury bonds), and they would both be eligibleas investments for fund A. By analyzing Chileandata, we noticed that the cumulative returns ofthe funds were not properly ordered (i.e. thesupposed higher returns for A, and then B, C, D,and E), even for extended periods of time. Thatis problematic, since the funds were designed tohave such profile, and workers that choose a par-ticular fund (say, fund A) expect to have higherreturns in the long run.
In [4] we proposed to measure risk directly,that is, we suggest to remove all portfolio con-straints (aside from nonnegativity, and that in-vestments must sum to one) and include a riskmeasure for each fund. In our approach, a fundwill be defined by the right-hand side of a riskconstraint on returns: smaller values will corre-spond to fund E (tighter constraint), higher val-
14 Stochastic Programming Society Newsletter
Limits on stocks Limits on foreign instr.
Max % Min% Max % Min%
A 80 40 100 45B 60 25 90 40C 40 15 75 30D 20 5 45 20E 5 0 35 15
Table 1: Limits in % based on the market valueof the portfolio for funds A, B, C, D and E
ues to fund A, and values in between will definefunds B, C, and D. Our out-of-sample resultsindicated that had such strategy been adoptedthe funds would have been correctly ordered interms of risk and return according to the originalintention of the regulator. Moreover, our resultsshowed that higher returns would have been ob-tained in the risk-controlled framework even ifthe investment universe was drastically reducedto include only six index funds instead of thou-sands of individual assets. It is worth mentioningthat the Mexican pension system includes riskmeasures in the fund’s definitions, and they haveexhibited ordered returns since their inception.
I presented, with one of the study’s co-authors(A. Cifuentes), our findings to the technical com-mittee that defines the investment rules for pen-sion funds in Chile. Our conclusions were well-received, and in October 2019, the Pension Su-perintendent (the ultimate authority in pensionsin the country) announced that in 2020 the fundswould not be defined by portfolio constraintsanymore [5]. He mentioned that using risk todefine the different funds will “generate a regu-latory context that facilitates the construction ofefficient portfolios”.
The second topic is related to default strate-gies in DC pension funds. It is a well-known factthat financial illiteracy prevents a large percent-age of workers from adequately managing theirportfolio. Lack of time and even interest (!) havealso been pointed out as factors explaining whyworkers do not manage their pension fund ac-counts. Several polls done in Chile in the last10 years show that 65% of the respondents donot know in which fund they are at the moment,
and only 22% know how much they have accu-mulated in their pension account [6].
Due to those facts, all DC pension fund sys-tems need to have a default strategy for workersthat do not manage their savings. A standardapproach is to follow life-cycle strategies: in thefirst years, the worker should be in a riskier fund(to boost returns), and then move gradually tosafer funds as retirement approaches. The pop-ular target funds follow this strategy, which onlytakes into account the time to retirement. Thedefault strategy in Chile is described in Table 2.A fundamental question is whether the defaultstrategy is a good one and if other strategiesdominate it in some sense. The default strat-egy is like a “Waze-less driver”: no matter thedriving conditions (snow, traffic), I will followthe path that I defined when I left home.
A B C D E
Men ≤ 35 Women ≤ 35
Men 36-55 Women 36-50
Men ≥ 56 Women ≥ 51
Table 2: The default strategy in Chile per genderand age bracket
In a recent work [7], we propose to includewealth as a state variable. In other words, theinvestment strategy will depend on how muchmoney was accumulated so far. We use stochas-tic dynamic programming to obtain an invest-ment strategy that (stochastically) dominatesthe default strategy. In the Chilean case, weshowed that although the default strategy hasan adequate performance overall, our proposedapproach had a higher probability of meeting re-tirement goals, and had a smaller shortfall onaverage. Additionally, the optimal solution in-cludes only funds A and E, calling into questionthe need for five funds.
Constructing, managing, and regulating pen-sion funds are complex problems, and wewill probably continue to see fierce discussionsaround the world in the coming years on howto build a system that offers the best benefitsto workers. DC systems are an excellent op-portunity for optimizers (in particular stochasticones!) because the problems involve meaningful
Vol.1 Number 1, March 2020 15
Figure 1: The challenges of retirement
and clear decisions and the need to measure riskin a context that is related (but different) frompure portfolio allocation problems. I believe theoptimization community should be part of thisdiscussion, working closely with policymakers.
Acknowledgements
The author thanks Arturo Cifuentes, DanielDuque, David Morton, and Wim van Ackooij forproviding helpful comments on an early versionof the article.
REFERENCES
[1] J. Broadbent, M. Palumbo, and E. Woodman. Theshift from defined benefit to defined contribu-tion pension plans-implications for asset alloca-tion and risk management, Reserve Bank of Aus-tralia, Board of Governors of the Federal ReserveSystem and Bank of Canada, 1–54, 2006.
[2] The trouble with pensions: Falling short. TheEconomist, June 12th 2008.
[3] Superintendencia de Pensiones Chile. Monthlyreturns. https://www.spensiones.cl/portal/
institucional/594/w3-propertyvalue-10089.
html.
[4] T. Gutierrez, B. Pagnoncelli, D. Valladao, andA. Cifuentes. Can asset allocation limits de-termine portfolio risk–return profiles in dc pen-sion schemes? Insurance: Mathematics and Eco-nomics, 86, 134–144, 2019.
[5] Oro y mas: anuncian cambios en lasinversiones de los fondos de pen-siones. https://www.elmercurio.com/
Inversiones/Noticias/Analisis/2019/09/10/
Invertir-en-oro-y-mas-anuncian-cambios-\
a-las-inversiones-de-\
los-fondos-de-pensiones.aspx.
[6] La encuesta de proteccion social y la reformaal sistema de pensiones: el caso de chile.https://www.previsionsocial.gob.cl/
sps/download/estudios-previsionales/
encuesta-de-proteccion-social/
estudios-relacionados-eps/2009/
24-la-eps-y-la-reforma-al-sistema-de-\
pensiones_el-caso-de-chile.pdf.
[7] D. Duque, D. Morton, and B. Pagnoncelli. Howgood are default investment policies in definedcontribution pension plans? Available at SSRN3471428, 2019.
Real World Impact of StochasticProgramming in Energy:An Example from Europe
Wim van AckooijEDF R&D, Osiris (France)
Impact Summary: Energy, in particularelectricity, as a fundamental resource, impactssociety at large. The electricity system in West-ern Europe1, for instance, impacts more than150 million people. The trend to increase theshare of wind and solar generation and thus in-termittent generation creates a need to asses andimprove flexibility of the system. Assessing theneed for such flexibility arises as a result of un-certainty. The tools from stochastic program-ming are required for a proper evaluation of thisflexibility. The result of these tools will lead toupdated policy and to changes in the physical sys-tem.
Energy management is about the efficient han-dling and decision making around “assets” thatproduce or consume “energy” in one form oranother. Since “energy” is a vital resource for
1Western Europe is composed of France, Germany,Netherlands, Belgium, Austria, Switzerland, Luxem-bourg, Monaco, Liechtenstein
16 Stochastic Programming Society Newsletter
the economy, and basically any aspect of soci-ety, it is indeed reasonable to desire some effi-ciency in its handling. Although we will mainlydiscuss electrical energy, one can also think ofgas, oil and heat networks. In the electricalsector, assets involve thermal generation units(nuclear, coal, gas,. . . ), hydro-electrical instal-lations (cascading systems, pump storage units,run-of-the river,. . . ), intermittent renewable gen-eration (wind turbines, solar panels,. . . ) and alsodemand-side management programs. Decisionsto consider are investment choices (generationassets, network), strategic planification of re-sources with long cycles (e.g., water), how to op-erate a set of given generation assets, how andwhen to deploy/activate demand-side manage-ment, but also how to handle uncertainty. Un-certainty has always been an important aspectof energy management. Indeed, vital underly-ing aspects are subject to uncertainty, and ac-tually weather related. This is for instance thecase with load, which depends not only on theeconomic cycle, but also on weather through thedemand in heating / cooling. Likewise, the man-agement of hydro-electric installations dependson water, which has unknown inflows (melting ofsnow, rain,. . . ), but may also have unknown out-take for agriculture. Modern systems are furtherdependent on weather as a result of unknowngeneration in wind and solar.
The use of optimization, as a support for deci-sion making, was recognized already as early asin the 1940s, when Pierre Masse laid the foun-dations of many principles of optimization, inparticular to hydroelectric reservoir management[1]. Contemporaneously with Dantzig and Kan-torovitch, Masse realized the importance of dualvariables, and their economic interpretation, andhe also laid the foundations of dynamic program-ming through the recognition of an optimal valuefunction. Optimization has ever since remaineda vital building block of decision making supporttools/software. This is for instance the case inunit-commitment and its great many variants.Optimization contributes to energy applicationsas a tool, but also the practically relevant aspectsof the problem stimulate the development of newtheoretical insights and better algorithms.
Optimization should be understood here in abroader scope. Indeed, uncertainty has alwaysbeen accounted for, in one way or another inall of these problems. First of all, by puttingin place a series of failsafe mechanisms throughorganizational measures, such as spinning re-serve requirements. Second, by carefully study-ing and modelling uncertainty with the tech-niques of statistics. Finally, by incorporatinguncertainty itself within the optimization model:the use of stochastic programming. This too,takes various forms, from the incorporation ofadditional safety margins (derived through con-siderations involving individual probability con-straints or robustness arguments), the incorpo-ration of forecasts, operationally rerunning thesoftware with sliding windows, whenever new in-formation is available, all the way to the consid-eration of more involved and larger models fromstochastic programming itself.
As we have thus seen, these models can belarge in a variety of ways: spatially, temporally,detailed and possibly explicitly accounting foruncertainty. We will illustrate the practical rel-evance of stochastic programming in two “ex-tremal situations”: the context of centrally plan-ning the European energy sector; the context ofan energy management system on a distributiongrid, providing services to the system, dealingwith uncertainty on wind generation and havingaccess to a storage device.
1. Large-Scale System
The EU H2020 funded project2 plan4res fo-cusses on delivering an end-to-end planning toolwhich will help examine the trade-off betweenthe share of integrated renewable energy intothe European Energy System and system reli-ability. The model will account for the Pan-European interconnected electricity system, syn-ergies with other energy systems, emerging tech-nologies, flexibility resources and provide a fullyintegrated modelling environment. As a firstpoint we should emphasize what is meant with“real-world impact” in this particular situation.
2grant 773897, http://plan4res.eu
Vol.1 Number 1, March 2020 17
Indeed, in this case, it is not meant in thesense that some particular new asset will be con-structed in a specific location as a result of theoutcomes of the model, but rather that the latteroutcomes will serve to enlighten policy. This isthe fairly general case, and only rarely will theresults of an optimization problem lead to directdecisions. The policy, enlightened by the resultsof various studies with the optimization tool, willmake better use of resources. One should thussee real-world impact in this light.
Now immediately comes the question of howto properly tackle the study of reliability in con-junction with high intermittent renewable en-ergy insertions. What would be the origin of“unreliability”? It is clear that this results onone hand from the presence of the physical con-straints of the system (e.g., on classic generators)and on the other from the presence of unpre-dictable quantities, i.e., uncertainty. It is thusonly when both features are modelled, that onecan properly address the question. Next, one hasto distinguish that uncertainties impact the sys-tem at different scales of time, and both are rel-evant. Most intermittent generation has a shortterm impact, e.g., within the day, whereas hydroinflows are naturally measured across the fourseasons, i.e., on a yearly basis. Hydro-electricenergy plays an important role in the system,since it is generally a fast, flexible means of gen-eration, that moreover is cheap. The vital issueis thus to properly place this energy with a yearlycycle, thus attributing it a cost of use (a substi-tution cost), since otherwise the resource wouldbe prematurely depleted. While doing so, oneshould evidently account for constraints impliedby other uses of water, such as agriculture ortouristic applications. The thus computed valueof water is typically transmitted, in one way oranother, to shorter time span applications suchas unit-commitment. The vital place that watervaluation has, was recognized a long time ago,e.g., [2], and the typical operationally used ap-proaches are stochastic dynamic programming.These approaches usually involve some form ofaggregation, but are what were/is used in prac-tice. Recently, SDDP (e.g., [3]) has moved to theforeground. This approach will also be used in
plan4res (we also refer to e.g., [4]), where more-over the transition problem will be formed by afull-scale European unit-commitment problem.
2. Small-Scale System
Accounting for uncertainty is not only impor-tant at the larger scale, but also matters at a re-duced scale. This is for instance the case in thedemonstrator built within the EU-Sysflex H2020project 3. Figure 1 below shows the physical in-stallation of the demonstrator.
Figure 1: Facilities of the multi-resources multi-services demonstrator
The purpose of this demonstrator is to ac-quire insights and conduct complex testing cam-paigns that can not be conducted on a real sys-tem (without potentially compromising safety),as a result of handling intermittency and storagesystems. One of the objectives is to examine towhat extent batteries can provide frequency sup-port services to the system. Announcing suchservices (or selling them) is typically somethingdone ahead of real time. Subsequently, the diffi-culty resides in being able to honour the engage-ment as a result of real time deviations. Thelatter being themselves, the result of uncertainty.For this purpose a stochastic programming basedenergy management system is embedded withinthe scheme shown in Figure 2. A stochastic pro-gramming based scheduler is what leads to betterprofits than when considering a scheduler based
3grant 773505, http://https://eu-sysflex.com/
18 Stochastic Programming Society Newsletter
only on forecasts. This results also from hav-ing to pay less penalties in case of discrepanciesbetween engagements and real time capabilities.Contrary to the usual situation, the actual out-
Figure 2: General control layers of the demon-strator
put of the planning tool is this time, taken intoaccount immediately. There are of course somefailsafe mechanisms in so much that the decisionsare actually adapted to account for real timeinformation coming from the actual equipment.Moreover, the planning tool is rerun intradailyto adjust for new forecasts and information suchas the state of charge of the battery. Still, thedecisions mostly get implemented. For furtherdetails on the equipment, we refer to [5].
3. Concluding observations
As we have thus seen in both contexts, whichare illustrative, stochastic programming is a vi-tal building block in the analysis and operationof energy systems and has been so for a signif-icant amount of time. Stochastic programmingis not only vital at larger scales of time / space,as was traditionally the case, but tends also tobecome significantly more important for smallerscales. At these smaller scales, the non-convexityof the underlying physics (e.g., powerflow equa-tions) have a real impact and are hard to neglector simplify. There is thus lots of room for newmodelling, theory and algorithms in stochasticprogramming within these applications – and –most importantly, real impact of one’s research
if the entire path, from application to theory andback, is taken.
REFERENCES
[1] P. Masse and R. Boutteville. Les reserves et laregulation de l’avenir dans la vie economique,Hermann & Cie., 1946.
[2] A. Turgeon. Optimal operation of multi-reservoirpower systems with stochastic inflows, WaterResources Research, 16(2), 275–283, 1980.
[3] M.V.F. Pereira and L.M.V.G. Pinto. Multi-stagestochastic optimization applied to energy plan-ning, Mathematical Programming, 52(2), 359–375, 1991.
[4] D. Beulertz, S. Charousset, D. Most, S. Giannelos,and I. Yueksel-Erguen. Development of a mod-ular framework for future energy system anal-ysis, In 54th International Universities PowerEngineering Conference (UPEC), 2019.
[5] Y. Wang, H. Morais, B. Lenz, V. Gomes,T. Godlewski, and H. Baraffe. The eu-sysflexfrench industrial-scale demonstrator: Coordi-nating distributed resources for multi-servicesprovision, In 25th International Conference onElectricity Distribution, Madrid, 2019.
Vol.1 Number 1, March 2020 19
Stochastic ProgrammingEvents in 2020
Below please find a list of events related toStochastic Programming in 2020. The situa-tion due to COVID-19 is rapidly changing.Therefore, please check the provided linksfor the latest information.
• Workshop title: Optimization under Uncer-taintyWhere/when: Montreal, March 23–27,2020. Postponed to 2021. Please checkthe website for updates: https://lnkd.in/gdHHVgW
• Workshop title: JuMP-dev 2020Where/when: Louvain-la-Neuve, Belgium,June 15–17, 2020More information: https://jump.dev/
meetings/louvain2020/
• School title: 5th DTU CEE SummerSchool: Advanced Optimization, Learning,and Game-Theoretic Models in EnergyWhere/when: Lyngby campus of theTechnical University of Denmark (DTU),Copenhagen, June 21–26 2020More information: https://lnkd.in/
gMsisxC
• Conference title: ECSO-CMS-2020 Confer-enceWhere/when: Ca’ Foscari University ofVenice, July 1–3, 2020, Postponed to2021. Updates will be published in thewebsite: https://lnkd.in/gG987EP
• Conference title: Latin-American Confer-ence on Operations Research (CLAIO2020)Where/when: Madrid, Aug. 31–Sept. 2,2020Stream and sessions devoted to stochasticoptimization (organizers: Laureano Escud-ero and Antonio Alonso Ayuso)More information: www.claio2020.com.
• Conference title: International Conferenceon Optimization and Decision Science(ODS2020)Stream: Optimization under uncertainty(organizers: P. Beraldi and F. Maggioni)Where/when: Rome, September 8–11, 2020More information: http://www.
airoconference.it/ods2020/
• Conference Title: INFORMS Annual Meet-ing 2020Stream: Optimization under uncertainty(organizer: Ruiwei Jiang)Where/When: Washington DC, USA,November 8–11, 2020More information: Contribute orview sessions and talks at https:
//tinyurl.com/OptUncertainty2020
20 Stochastic Programming Society Newsletter
ICSP XVIJuly 25-29, 2022 Davis, California, USA
David Woodruff
UC Davis Graduate School of Management (US)
Mark your calendars! We are going tohold the sixteenth International Conference onStochastic Programming (ICSP) during the lastweek of July 2022: July 25–29, 2022. If youare on the program committee for any other con-ferences, please try to avoid that week.
We are starting work on the website:
https://gsm.ucdavis.edu/
xvi-international-conference
-stochastic-programming-2022
If you want to organize a session or a track,please let us know at [email protected]
Figure 1: UC Davis campus
Committee on StochasticProgramming (COSP)
Members
Chair: Guzin BayraksanIntegrated Systems Engineering, The Ohio State Univer-sity, Columbus, OH (USA)[email protected]
Secretary: Francesca MaggioniDepartment of Management, Economics and QuantitativeMethods, University of Bergamo (Italy)[email protected]
Treasurer: Bernardo PagnoncelliSchool of Business, Universidad Adolfo Ibanez (Chile)[email protected]
Alois PichlerFaculty of Mathematics, University of Technology, Chem-nitz (Germany)[email protected]
Johannes O. RoysetOperations Research Department, Naval PostgraduateSchool, Monterey, CA (USA)[email protected]
Wim van AckooijEDF R&D, Osiris (France)[email protected]
Phebe VayanosViterbi School of Engineering, University of Southern Cal-ifornia, Los Angeles, CA (USA)[email protected]
Webmaster: Wolfram WiesemannBusiness School, Imperial College, London (United King-dom)[email protected]
David Woodruff
UC Davis Graduate School of Management, Davis, CA
(USA)