Informs Annual Meeting 2017

TE81

INFORMS Houston – 2017

TE82

4 - Expanding a Bottom-up Energy Model into a General-equilibrium Model: Rationale, Formulation and Implementation for Saudi Arabia Hossa Almutairi, Sr. Research Associate, King Abdullah Petroleum Studies & Research Center, Airport Road, Riyadh, 11672, Saudi Arabia, hossa.mutairi@kapsarc.org, Frederic H. Murphy, Axel Pierru, Shreekar Pradhan We represent the Saudi economy using an applied general equilibrium framework, which combines a bottom-up technology-rich representation of energy-related sectors with a standard top-down CGE representation of the rest of the economy. This work is motivated by the large size of energy and energy- intensive sectors in the Saudi economy and the structural changes that these sectors will experience in the coming years. Our approach can serve as a general template for expanding any bottom-up model into a general equilibrium. To illustrate the capabilities of the model, we applied the energy price reform implemented in 2016 to the static version of the model calibrated to 2013 data. 5 - Osemosys Energy Model: Introducing Elasticity Denis Lavigne, Professor, Royal Military College Saint-Jean, 29, rue Louis-Frechette, Saint-Jean-sur-Richelieu, QC, J2W. 1E9, Canada, denis.lavigne@cmrsj-rmcsj.ca The OSeMOSYS project proposes an open-access energy modeling tool for a wide audience. Its relative simplicity makes it appealing for academic researchers and governmental organizations to study the impacts of policy decisions on an energy system in the context of possible severe greenhouse gases emissions limitations. This paper presents how an elasticity of demand capability can be embedded in an OSeMOSYS analysis, allowing leaders to get insightful results on which end-use demands are sensible to such limitations. A detailed description of the approach will be presented. 382A Multistage Stochastic Programming Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Anthony Downward, University of Auckland, Auckland, 1024, New Zealand, a.downward@auckland.ac.nz 1 - A Deterministic Decomposition Algorithm to Solve Multistage Stochastic Programmes Stochastic dual dynamic programmes are an important class of optimisation problems, especially in energy planning and scheduling. Because of the large scenario trees that these problems induce, current solution methods use random sampling of the tree in order to build a policy. Candidate policies are evaluated using Monte Carlo simulation; convergence of a given policy requires a statistical test at a given level of confidence. In this paper, we present a deterministic algorithm to solve these problems. The main feature of this algorithm is a deterministic path sampling scheme during the forward pass phase of the algorithm. This precludes the use of Monte Carlo simulation, increasing performance. 2 - Multi-period Adaptive Fleet Planning Problem Bruno F. Santos, Assistant Professor, TU. Delft, Kluyverweg 1, Faculty of Aerospace Engineering, Delft, 2629 HS, Netherlands, b.f.santos@tudelft.nl, Laura Requeno Garcia This works presents a stochastic modelling approach to the multi-period airline fleet planning problem. Approximate Dynamic Programming (ADP) is used to model the impact of demand uncertainty on fleet decisions. The proposed ADP algorithm applies local value function approximations resulting from Gaussian kernel regressions to estimate future airline operating profits. A case study is used to show the effectiveness and practicability of the proposed approach. Optimal solutions are achieved for the deterministic case of the problem, while obtained policies excel the most-likely deterministic solution on stochastic versions. 3 - Investments for Disaster Mitigation Given the Uncertain Effects of Climate Change Anthony Downward, University of Auckland, Level 3, 70 Symonds Street, Auckland CBD, Auckland, 1024, New Zealand, a.downward@auckland.ac.nz Climate change is said to be responsible for increases in extreme weather events, however, the exact impact on the frequency or damage of such events is unknown. In this presentation we discuss a multi-criteria multi-stage stochastic programming model for disaster mitigation. In each period, a risk-averse central planner decides on investments that should be made so as to minimize future negative effects on the economy, environment and the society. This is modelled using stochastic dual dynamic programming, with the climate modelled as a Markov chain. TE81 Regan Baucke, University of Auckland, 70 Symonds Street, Auckland, 1010, New Zealand, rbau155@aucklanduni.ac.nz

382B Risk Management in Stochastic Optimization and Machine Learning Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Giorgi Pertaia, gpertaia@ufl.edu 1 - Risky Robust Optimization and Sparsity Matthew Norton, University of Florida, Gainesville, FL, United States, mdnorton@ufl.edu Robust Optimization has proven to be an intuitive and fruitful framework for optimization under uncertainty. Typically, Robust Optimization approaches view uncertainty from a pessimistic, or worst-case, viewpoint and thus can be overly restrictive. We diverge from this tradition, and show that an optimistic, or best- case, viewpoint of uncertainty should not be ignored. We show that this viewpoint can help improve the profitability of robust policies without sacrificing significant robustness properties. In addition, we show that this notion provides new insight into the recently popularized non-convex regularization schemes in machine learning that induce sparsity. 2 - Risk Management with POE, VaR, CVaR, and bPOE Stan Uryasev, University of Florida-Gainesville, AOD, 303 Weilhall, Gainesville, FL, 32611, United States, uryasev@ufl.edu This paper compares four closely related probabilistic measures: Probability of Exceedance (POE), Value-at-Risk (VaR) which is a quantile, Conditional Value-at- Risk which is a Superquantile, and Buffered Probability of Exceedance (bPOE). We discuss basic properties and relations of these functions. We show equivalence of constraints for POE and VaR and for bPOE and CVaR. 3 - Finite Mixtures with CVaR Constraints Giorgi Pertaia, 1221 South West 21st Avenue, Gainesville, FL, 32601, United States, gpertaia@ufl.edu Normal mixtures models are used in financial applications dealing with heavy tail distributions. We have suggested a new CVaR distance between distributions. The CVaR distance is a convex function w.r.t. weights. Weights of normal mixture are found by minimizing CVaR distance between the mixture and the target distribution. We suggested convex constraints on the weights assuring that tail of mixture is as fat as the original distribution. We have conducted a case study demonstrating efficiency of the suggested approach. 382C Optimization, Stochastic Contributed Session Chair: Yanting Wang, Xi’an Jiaotong University, Xi’an, China, wangyt.66@163.com 1 - The Trimmed Lasso: Sparsity and Robustness Martin S. Copenhaver, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, mcopen@mit.edu, Dimitris Bertsimas, Rahul Mazumder Nonconvex penalty methods for variable selection in linear regression have been a topic of fervent interest in recent years. In this talk we introduce a continuous family of nonconvex penalty functions that we call the trimmed Lasso. We show that this family of penalties is an exact method for variable selection with direct control over the desired level of sparsity. In analyzing the properties of the trimmed Lasso, we highlight its connection to existing nonconvex approaches as well as its algorithmic implications. 2 - Generalized Affine Decision Rules for Mixed Integer Linear, Quadratic and Nonlinear Adjustable Robust Optimization Problems by Multiparametric Programming Efstratios N. Pistikopoulos, Texas A&M.Energy Institute, College Station, TX, United States, stratos@tamu.edu, Styliani Avraamidou, Styliani Avraamidou, Chao Ning, Nikolaos A. Diangelakis, Fengqi You We propose a novel method for the derivation of generalized affine decision rules for linear/quadratic/nonlinear and mixed-integer ARO problems through multi- parametric programming. The problem is treated as a multi-level programming problem that is then solved using M-POP, a novel algorithm for the exact and global solution of multi-level mixed-integer programming problems. The main idea behind the proposed approach is to solve the lower optimization level of the problem parametrically. This will result in a set of affine decision rules optimal for the entire feasible space. A set of illustrative numerical examples are provided to demonstrate the potential of the proposed novel approach. TE83

440