44 1541-1672/05/$20.00 © 2005 IEEE IEEE INTELLIGENT SYSTEMS
Published by the IEEE Computer Society
T r a n s p o r t a t i o n a n d L o g i s t i c s
Case Study:
An Intelligent Decision-
Support System
Zbigniew Michalewicz, University of Adelaide
Martin Schmidt, Matthew Michalewicz, and Constantin Chiriac, SolveIT Software
Information technology applications that support decision-making processes and prob-lem-solving activities have proliferated and evolved over the past few decades. In the
1970s, these applications were simple and based on spreadsheet software. During the
1980s, decision-support systems incorporated optimization models, which originated
Distributing used cars
to various automobile
auctions is a
complicated problem
with multiple
variables. A software
system that combines
prediction,
optimization, and
adaptation techniques
has generated
impressive profits
for a large auto
manufacturer.
in the operations research and management science
communities. In the 1990s, these systems were fur-
ther enhanced with components from artificial intel-
ligence and statistics.
This evolution led to many different types of deci-
sion-support systems with somewhat confusing
names, including management information systems,
intelligent information systems, expert systems,
management-support systems, and knowledge-based
systems. Because businesses realized that data was
a precious asset, they often based these “intelligent”
systems on data warehousing and online analytical
processing technologies. They gathered and stored
a lot of data, assuming valuable assets were implic-
itly coded in it. Raw data, however, is rarely benefi-
cial. Its value depends on a user’s ability to extract
knowledge that is useful for decision support. Thou-
sands of “business intelligence” companies thus
emerged to provide such services. After analyzing a
corporation’s operational data, for example, these
companies might return intelligence (in the form of
tables, graphs, charts, and so on) stating that, say, 57
percent of the corporation’s customers are between
40 and 50, or product Q sells much better in Florida
than in Georgia.
Many businesses have realized, however, that the
return on investment for pure “business intelligence”
is much smaller than initially thought. The “discov-
ery” that 57 percent of your customers are between
40 and 50 doesn’t directly lead to decisions that
increase profit or market share. Moreover, we live in
a dynamic environment where everything is in flux.
Interest rates change, new fraud patterns emerge,
weather conditions vary, the stock markets rise and
fall, new regulations and policies surface, and so on.
These economic and environmental changes render
some data obsolete and make other data—which
might have been useless just six weeks ago—sud-
denly meaningful.
We developed a software system to address these
complexities and implemented it on a real distribu-
tion problem for a large car manufacturer. The sys-
tem detects data trends in a dynamic environment,
incorporates optimization modules to recommend a
near-optimum decision, and includes self-learning
modules to improve future recommendations. As fig-
ure 1 shows, such a system lets enterprises monitor
business trends, evolve and adapt quickly as situa-
tions change, and make intelligent decisions based
on uncertain and incomplete information.
Problem overview
We developed the system for a US-based car man-
ufacturer that has more than 1 million cars returned
from leases or rentals each year. The manufacturer
owns the cars, and the problem is how to best distrib-
ute these cars among hundreds of auction sites around
the United States. The cars vary by make and model,
mileage, options, wear and tear, and so on. These char-
acteristics, along with others, influence the car’s sale
price at each particular auction. Our central challenge
was to achieve the “best” possible distribution among
these auction sites—that is, the distribution that max-
imizes the net proceeds from all sales.
The process of making optimal recommendations
involves many considerations, ranging from price pre-