Andrew
R. Barron (Department of Statistics, Yale University)
andrew.barron@yale.edu
Probability
Theory of Compounding Wealth and Universal Portfolio Estimation
We
review roles of probability theory, statistics, and information
theory in examining compounding wealth in gambling and stock
market settings. Whereas maximizing conditional expected log
return produces the highest long-run growth rate of wealth,
we show that in contrast maximizing other common utilities
expressed by power laws optimize large deviation probabilities
(maximize the slim chance of unusually large returns). Universal
portfolios, their relationship to Bayes strategies, and their
minimax characteristics are reviewed. Building on past work
by Tom Cover and his colleagues, universal portfolios are
determined that have nearly minimal maximum regret compared
to the best with hindsight (in customary classes of portfolio
strategies) uniformly over all possible stock price sequences.
Strategies for estimation and computation of universal portfolios
are discussed.
Joe
H. Chow
(Department of Electrical & Computer Systems Engineering,
Rensselaer Polytechnic Institute (RPI))
Optimization
and Risk Management in Open-Access Electric Energy Markets
Slides: pdf
In
an open-access electric energy market, electricity suppliers
and load serving entities are allowed to trade energy and
provide bids into a daily energy market. The power system
is managed by an independent system operator, who usually
has the role of administrating the energy market as well.
Suppliers and loads need to optimize their expected energy
position as well as any real-time variation, under a limited
information structure imposed by the independent system operator.
Thus the optimization has to be based on past data predicting
future market conditions. The talk will provide an overview
of open-access electric energy markets as well as some research
on optimal bidding strategies by Ning Lu, a PhD candidate
at RPI.
George
Cybenko
(Dartmouth College) gvc@dartmouth.edu
http://www.dartmouth.edu/~gvc
Dynamic
Dynamical Systems Slides:
html
pdf
ppt
A
new class of control problems are emerging in which the state
space of the system changes dynamically. This presents two
novel challenges: how to dynamically define these changes
and; how to develop effective controls for dealing with systems
that dynamically change. This talk will present examples and
ongoing work to address both challenges.
Joao
P. Hespanha (Center for Control Engineering and
Computation, University of California, Santa Barbara) hespanha@ece.ucsb.edu
Complexity
Issues in Probabilistic Mapping Slides:
pdf
This
talk addresses the issue of estimating the positions of a
group of objects using a stream of noisy sensor measurements.
This is often called probabilistic mapping. From a formal
point of view, probabilistic maps are just the probability
densities of object positions, conditioned to the available
sensor measurements. In this talk we will explore issues related
to the computational complexity of constructing probabilistic
maps and also utilizing them in the context of path planning.
Arthur
Kordon
(Research Leader, Corporate R&D, The Dow Chemical Company)
AKKordon@dow.com
Hybrid
Intelligent Systems for Data-Driven Monitoring and Optimization
Slides: pdf
A
novel approach for data-driven modeling based on integration
of four key computational intelligence approaches (genetic
programming, analytical neural networks, support vector machines,
and particle swarm optimizers) is proposed. The integrated
methodology amplifies the advantages of the individual techniques,
significantly reduces the development time, and delivers robust
empirical models with low maintenance cost. The advantages
of the proposed methodology for data-driven monitoring and
optimization will be illustrated with several successful applications
in The Dow Chemical Company.
Rudolf
Kulhavy (Honeywell ACS Advanced Technology) kulhavy@htc.honeywell.cz
Data-driven
Decision-making: The Good, the Bad, and the Ugly Slides:
html
pdf
pps
The overwhelming amount of data stored in databases gives
sometimes rise to exaggerated expectations. One of the popular
myths is that a large amount of data carries necessarily a
large amount of information. It is clearly not so^×data stored
in databases is often redundant or showing just a couple of
patterns from the multitude of all possible patterns of the
process behavior. Very rarely the data collected is the result
of a planned experiment, rather it is a series of snapshots
of routine operation. What is so exciting then about the massive
data sets available to us today? It is not that a huge amount
of data can replace the domain knowledge and the art of modeling.
It is that for the first time we have the whole process history
at disposal to make decisions affecting the future behavior.
This makes database-centric decision-making an exciting alternative
to the current paradigms. The presentation discusses opportunities
and challenges presented by the new paradigm. Special attention
is paid to selection of a "data cube" capturing multi-dimensional
data, definition of "similar" historical data points, and
similarity search in high-dimensional spaces, while sharing
experience from real-life applications of data-centric decision
support systems.
Steffen
L. Lauritzen
(Department of Mathematical Sciences, Aalborg University)
steffen@math.auc.dk
http://www.math.auc.dk/~steffen
LIMIDs
- Representing and Solving Decision Problems with Limited
Information Slides:
pdf
The
notion of a Limited Memory Influence Diagram (LIMID) is introduced
as a Bayesian network augmented with nodes representing decisions
and utility functions. For each decision it is specified what
information is available at the time when the decision is
to be made. In contrast with traditional influence diagrams,
the assumption of no forgetting is relaxed, and there is no
additional constraints on the order in which decisions are
to be taken. This allows for multiple decision makers and
decision makers with limited memory, and reduces complexity
of strategies. We give a local computation algorithm for finding
locally optimal policies, conditions for the policies to be
globally optimal, and indicate how this can be exploited to
obtain bounds for the loss of utility, for example in partially
observed Markov decision processes (POMDPs). The lecture is
largely based upon:
Lauritzen,
S.L. and Nilsson, D. (2001). Representing and Solving Decision
Problems with Limited Information, Management Science, 47,
1238-1251.
Can
be obtained from http://www.math.auc.dk/~steffen/papers/limids.pdf
Jay
H. Lee
( School of Chemical Engineering, Georgia Institute of Technology)
jay.lee@che.gatech.edu
Simulation
Based Approximation of Value Function for Process Control
Slides:
pdf
Although
model predictive control (MPC) has firmly etched itself in
process control practice, its large on-line computational
demand and inability to rigorously consider information feedback
under uncertainty limits its usage in complex systems, which
are characterized by multi-scale, nonlinear, hybrid dynamics
and significant uncertainties. In this talk, we propose an
alternative approach based on the infinite horizon cost-to-go
(the 'value function'). The key issue lies in obtaining an
accurate approximation of the value function for the relevant
regions of state space. We propose to build an approximation
using simulation data and improve it iteratively through policy
or value iteration and additional simulation. We demonstrate
the efficacy of the approach on two different bioreactor optimal
control problems. Along the way, we also point out some critical
issues and outstanding theoretical problems.
Susan
A. Murphy
(Department of Statistics, Quantitative Methodology Program,
University of Michigan) samurphy@umich.edu
http://www.stat.lsa.umich.edu/~samurphy/
Dynamic
Treatment Regimes for Chronic, Relapsing Disorders
The management of chronic, relapsing disorders can be viewed
as a control problem in that multi-stage treatment decisions
are made with the goal of optimizing mean response. For example,
in the prevention of relapse by recovering alcoholics, the
response might be percent days abstinent and the treatment
decisions might be, which preventative treatment should be
used initially, how long should we wait to declare the initial
treatment ineffective and switch to a secondary treatment,
which secondary treatment should be used, when should treatment
be stopped, etc. These treatment decisions would be made on
the basis of time varying covariates such as number of days
heavy drinking, measures of craving, measures of stress, patient
preference and results of urinalyses.
A
important open problem in this area is how we might use a
batch of data, i.e., a longitudinal sample of individuals
for whom both response, covariates and treatment decisions
are recorded for each time period, so as to estimate the optimal
decision rules. This challenging area is characterized by
delayed effects of treatment, an unknown model relating past
treatment and covariates to future covariates and a high noise
to signal ratio.
Gregory
Piatetsky-Shapiro (KDnuggets) gps@kdnuggets.com
Knowledge
Discovery in Microarray Gene Expression Data
Slides: html
pdf
ppt
DNA
Microarrays are revolutionizing molecular biology, allowing
simultaneous analysis of many thousands of genes. Microarray
hold the promise of important applications, including creating
novel, genetic-based diagnostic tests, finding new molecular
targets for therapy, and developing personalized treatments.
Microarrays
allow analysis of dynamic processes and deeper insight into
biological pathways.
However,
the large number of genes and a typically small number of
samples, present unique challenges for DNA microarray data
analysis. We discuss issues in normalization of microarray
data, selecting the best set of genes for classification and
clustering, randomization techniques, and building classification
and clustering models.
We
illustrate these processes using a number of software tools
and show new results with potential biological significance.
Maria
Prandini (Department of Electronics for Automation,
University of Brescia, Italy) prandini@ing.unibs.it
Cautious
Hierarchical Switching Control of Stochastic Linear Systems
(Poster Session)
We
address the problem of controlling an unknown stochastic linear
system and propose a new methodology that incorporates the
advantages of cautious stochastic control and switching control
in a hierarchical scheme. The design of cautious switching
controllers is based on the following two-step procedure:
i) a probability measure describing the likelihood of different
models is updated on-line based on observations; and ii) at
each switching time, the controller in the candidate controller
set that optimizes a certain average control cost with respect
to the updated probability measure is selected. If a certain
structured set of candidate controllers is used in the above
cautious switching scheme, then a controller is automatically
chosen that suitably compromises performance against robustness.
Randomized algorithms are used to make the controller selection
computationally tractable.
This
is a joint work with M.C. Campi
and J.P. Hespanha.
Daniel
E. Rivera
(Department of Chemical and Materials Engineering, Arizona
State University) daniel.rivera@asu.edu
http://www.eas.asu.edu/~csel/rivera.html
Model-on-Demand
Estimation for Improved Identification and Control of Process
Systems Slides:
pdf
In
recent years we have been pursuing the concept of nonlinear
identification and control through a data-driven framework
named Model-on-Demand (MoD). The MoD approach enhances traditional
local modeling and provides the potential for performance
rivaling global methods (such as NARX models and neural networks)
while involving substantially less detailed knowledge of model
structure from the user and much more reliable numerical computations.
Research
in our laboratory (performed in collaboration with the Division
of Automatic Control at Linkoping University, Sweden) has
focused on demonstrating the MoD estimation framework as an
effective, practical means for modeling and controlling nonlinear
process systems. Research topics have included such diverse
problems as MoD-based automated smoothing of empirical transfer
function estimates (ETFEs), systematic design of databases
for MoD estimation using multi-level pseudo-random and minimum
crest factor multisine input signals, and the development
of a comprehensive MoD-based Predictive Control methodology.
A Matlab-based tool for MoD estimation and control, developed
in our laboratory in collaboration with Linkoping researchers,
is available in the public domain.
The
presentation will describe our general experiences with MoD
estimation in each of these topical areas. Some pressing challenges
and open issues in the application of MoD estimation will
be discussed. The talk will conclude with a summary of current
activities, among these the application of MoD-based estimation
and control to inventory management in supply chains.
Tariq
Samad (Honeywell Automation and Control Solutions)
tariq.samad@honeywell.com
http://www.htc.honeywell.com/people/tariq_samad
High-Confidence
Control: Ensuring Reliability in High-Performance Real-Time
Systems Slides:
html
pdf
ppt
Technology
transfer is an especially difficult proposition for real-time
control. To facilitate it, we need to complement the "high
performance" orientation of control research with an emphasis
on demonstrating "high confidence" in real-time implementation.
We focus on a particular problem in this context: Complex
algorithms have unpredictable computational characteristics
that nevertheless need to be modeled. Statistical verification
is suggested as a possible approach and we are exploring the
application of statistical learning theory. A synthesis of
control engineering and computer science is required if effective
solutions are to be devised.
Concluding
Remarks Slides:
html
pdf
ppt
Steve
Smale (Department of Mathematics, University of
California, Berkeley) smale@math.berkeley.edu
Fast Algorithms for Dealing with Data and Understanding
Them
Recent
developments in learning theory help to broaden and deepen
methods for analysing data. Non-linear algorithms get replaced
by linear ones in high dimensional spaces.
Bruce
F. Wollenberg
(Department of Electrical and Computer Engineering, University
of Minnesota) wollenbe@ece.umn.edu
http://www.ece.umn.edu/faculty/wollenberg.html
Solving
the ISO "Seams" Problem for Uniform Boundary LMP's Slides:
pdf
The
US Department of Energy, Federal Energy Regulators Commission,
has released a Standard Market Design (SMD) which introduces
the problem of enabling two Independent System Operators (ISO's)
which independently calculate the market clearing prices for
their respective markets to reach consistent Locational Marginal
Prices (LMP's) along a shared boundary. Without consistent
LMP's trading across a boundary (seam) can be difficult or
impossible. This presentation will focus on the issue of enabling
two ISO's to reach a common set of LMP's on the boundary.
William Hogan has presented some preliminary work toward a
solution wherein multiple ISO's solutions are iterated until
a common solution is reached. In Hogan's work, only transmission
limit constraints were imposed on the solution. We have extended
this to include first contingency constraints as well. The
market clearing calculations are done with an Optimal Power
Flow (OPF) based on a full Alternating Current model of the
power system. The LMP's are the bus power constraint Lagrange
multipliers from the solution. Both Hogan's and our own work
so far have been with linear networks not with full AC OPF
solutions and full AC contingency analysis. The talk will
explore many of the difficulties of achieving a common boundary
bus LMP when each ISO is using an AC OPF and AC contingency
analysis to calculate the LMP's and what research directions
we see as promising. The aim of the work we are conducting
is to achieve tools for ISO's to enable them to continue to
operate independently yet to have uniform LMP's along the
boundaries with other ISO's.
