ORC IAP Seminar - "Quantitative Methods in the Real World"
The annual Independent Activities Period seminar organized by the
Operations Research Center will take place on 01/29/2009, 10am-5pm, in room E51-376. This year's rendition
will be focused on the use of quantitative methods in real-world applications, and will consist
of a series of talks by practitioners from different industries, focusing on the application of analytical techniques (optimization, stochastic processes, statistics) to real-world problems. Each talk will be followed by a Q&A session, and there will be opportunities for networking with the speakers.
Quantitative Methods in the Real World
January 29, 2009. Room E51-376, Tang Center
http://student.mit.edu/iap/nsor.html
Free catered lunch for attendees (RSVP to daniancu@mit.edu)
Speaker: Dr. Benoit Couet
Company: Schlumberger-Doll Research, Uncertainty, Risk and Optimization Group
Title: Practical Methods for Constrained Oilfield Optimization
Speaker: Dr. Michael Prange
Company: Schlumberger-Doll Research, Uncertainty, Risk and Optimization Group
Title: Valuing Future Information Under Uncertainty using Polynomial Chaos
Speaker: Vishal Gupta
Company: Barclays Capital, Commodities Group
Title: Load Following Deal
Speaker: Dr. Alan King
Company: IBM Watson Research Center,
Mathematical Sciences Department
Title: Mathematical Sciences at IBM from a Researcher's Perspective
Speaker: Dr. Bala Chandran
Company: Analytics Operations Engineering
Title: Optimizing operations at a wire and cable manufacturer
- Speaker: Dr. Benoit Couet
Company: Schlumberger-Doll Research, Uncertainty, Risk and Optimization Group
Title: Practical Methods for Constrained Oilfield Optimization
Abstract: Derivative-free optimization methods are mostly designed to solve unconstrained optimization problems.
Constraints are usually accounted for by optimizing an extended objective function obtained by
adding to the desired objective function a penalty term representing a measure of the constraint
violation. A major limitation of these methods is specifying the magnitude of the constraint
penalty term compared to the objective function. Recently, a new class of transformation methods
has been introduced that uses a lexicographic order comparison in which constraint violations
take precedence over objective values. This guarantees that objective values are computed only
for points that satisfy all the constraints. However, in many real-world optimization problems,
both the objective and constraint functions are simulation-based and might be expensive to compute.
In such cases, optimizing using a single lexicographic comparison is not satisfactory, since
expensive nonlinear constraints will be computed for all points, including ones that violate
the simple linear constraints or variable bounds. Not only is this inefficient, but computation
of the nonlinear constraints at points that violate the linear constraints might be unphysical.
We propose two sequential lexicographic methods for solving optimization problems when the
evaluation of both objective and constraint functions may be expensive. To this end, we first
classify all constraints depending on their computational costs. In the first method, the single
lexicographic comparison is replaced by a sequential lexicographic comparison in which the linear
constraints take precedence over nonlinear constraints, and inexpensive nonlinear constraints
take precedence over expensive ones, which in turn take precedence over objective values. In
the second algorithm, points not satisfying the linear constraints are feasibilized prior to
lexicographic comparison of the inexpensive nonlinear constraints, expensive nonlinear constraints
and objective function. We illustrate these two methods using the Nelder-Mead algorithm as the unconstrained
optimizer, although it can also be used with any derivative-free or evolutionary algorithm. An application
to oilfield-production optimization examples shows that both sequential lexicographic methods are well
suited for expensive constrained optimization problems, yielding better oil production rates with fewer
objective and expensive constraint evaluations than when all constraints are treated together.
- Speaker: Dr. Michael Prange
(Schlumberger-Doll Research, Uncertainty, Risk and Optimization Group)
Title: Valuing Future Information Under Uncertainty using Polynomial Chaos
Abstract: We show how to estimate the value of information for highly uncertain projects whose decisions have
long-term impacts. It is a mathematically consistent framework using decision trees, Bayesian updating and
Monte Carlo simulation to value future information today, even when that future information is imperfect.
A polynomial chaos approach suitable for black-box functions is used to reduce the number of Monte Carlo
computations to manageable proportions. In our example it provides a speedup of more than two orders of
magnitude. We demonstrate the approach with an oilfield example involving a future decision on where to place a new
injection well relative to a fault. The example considers the value to the asset holder of a measurement
to be made in the future that reveals the degree of reservoir compartmentalization caused by this fault.
Despite the presence of financial uncertainty on future oil price and private uncertainty on reservoir
variables that are largely unresolved by the measurement, our methodology provides a computationally
efficient valuation framework.
- Speaker: Vishal Gupta
(Barclays Capital, Commodities Group)
Title: Load Following Deal
Abstract: The Load Following Deal is a fundamental derivative on the "serving"
side of US Power markets. In words, the load following deal is an
obligation to provide electricity to a given group of customers for a
predetermined time at a contractual cost. The deal embodies at least
two sources of randomness - that from the changing market price of
electricity (power), and that from the changing quantity demanded by the
customers (load). We outline a pricing model for these deals which goes
beyond "standard" risk-neutral methodology for derivative valuation to
incorporate more subtle tehniques from portfolio allocation and
optimization theory. Particular emphasis is placed on the option
structure embedded in these deals, sometimes known as "gamma trades", as
well as the unhedgeable components of the deals.
We describe several market-based and historical calibration procedures
for load and power price models.
This work was done in cooperation with Mircea Marinescu and Yuri
Zhestkov at Barclays Capital.
- Speaker: Dr. Alan King
(IBM Watson Research Center,
Mathematical Sciences Department)
Title: Mathematical Sciences at IBM from a Researcher's Perspective
Abstract: This talk will give a researcher's overview of the Mathematical Sciences Department at IBM's T. J. Watson Research Center. We'll start with some historical perspectives and then give a snapshot of some current activities. We'll focus on a couple of themes: the interesting process how our research concentrations influence and are influenced by IBM's business concerns over time, and some (possibly irreverent) observations on managing a research career inside IBM.
- Speaker: Dr. Bala Chandran (Analytics Operations Engineering)
Title: Optimizing operations at a wire and cable manufacturer
Abstract: The manufacture of communication cables and wires takes place in an environment containing elements of both a flow shop as well as a job shop. I will describe two independent decision support systems developed for a cable manufacturer--one that optimizes the sequence of machines to use for each order and another that optimizes the cutting of cable reels to desired lengths. In both cases, the underlying problems are modeled as integer programs and solved using CPLEX.
Both manufacturing and cutting-stock-related problems have been extensively studied in the literature. In this talk, I will focus on modeling less frequently-encountered objectives and constraints that are unique to the problems at hand. In addition, I will describe aspects of the implementation and software development process that are are important to the eventual acceptance and use of the decision support tools.
