2008 ISOM/ISE Workshop:
RFID & Supply Chain Information Management
Coordinators:
Amar Sapra
Selwyn Piramuthu
Sponsors:
Center
for Supply Chain Management & DIS Forum(ISOM
Department)
Information
Systems and Operations Management
& Industrial and Systems Engineering
February
15-16, 2008
RFID
& Supply Chain Information Management
2008 ISOM/ISE
Workshop Program
Thursday, February 14, 2008
|
7pm
– 9:30pm |
Kick-off
dinner |
Han’s |
Friday, February 15, 2008
|
Time |
Event |
Location |
|
7:30am
– 8:30am |
Breakfast & welcome |
UF
Hilton |
|
8:30am
– 9:15am |
Adam Mersereau (UNC): Information-Sensitive
Replenishment when Inventory Records are Inaccurate |
|
|
9:15am
– 10am |
Gary M. Gaukler ( |
|
|
10am
– 10:30am |
Break |
|
|
10:30am
– 11:15am |
Benoit Montreuil (Laval): Item-Level
RFID in Retail Facilities: Exploratory Investigation of its Value Creation
Potential for Enabling the Real-Time Retail Demand and Supply Chain |
|
|
11:15am
– 12:00pm |
Manu
Goyal ( |
|
|
12:00pm –
1:30pm |
Lunch |
|
|
1:30pm–
2:15pm |
Metin Cakanyildirim (UT-Dallas): Partially
Observed Inventories: Signals, Sufficient Statistics and Approximations |
|
|
2:15pm
– 3pm |
Jacques Roy (HEC-Montreal):
Cost-Benefit Analysis of a Potential RFID Deployment in a Cruise Ship Supply
Chain Context
|
|
|
3pm
– 3:30pm |
Break |
|
|
3:30pm
– 4:15pm |
Tim
Huh ( |
|
|
7:00 pm– 9:30 pm |
Reception |
Tapas
12 West |
Saturday, February 16,
2008
|
Time |
Event |
Location |
|
8am
– 9am |
Breakfast |
UF
Hilton |
|
9am
– 9:45am |
Diego Klabjan (Northwestern): Next Generation Business Applications for Radio Frequency
Identification |
|
|
9:45
– 10:30am |
|
|
|
10:30
– 10:45am |
Break |
|
|
10:45
– 11:30am |
Sean Marston ( |
|
|
11:30am
– 12:00pm |
Brown-Bag
Lunch |
RFID
& Supply Chain Information Management
2008 ISOM/ISE
Workshop Program
Adam Mersereau
(UNC)
Information-Sensitive
Replenishment when Inventory Records are Inaccurate
The vast majority of inventory management research
assumes that the inventory manager knows with certainty his inventory position.
Recent empirical research, however, calls this assumption into question and
reveals the reality of inventory management in practice: inventory records do
not necessarily match the physical inventory on the shelf. Radio Frequency
Identification (RFID) has been proposed as a solution to the problem of record
inaccuracy. Instead, our interest is in intelligent inventory management tools
that mitigate the costs of record inaccuracy, even without investment in RFID.
We
study an inventory system with imperfect inventory records and unobserved lost
sales. Record inaccuracies in our model are assumed to arise via an “invisible”
demand process that perturbs physical inventory but is unobserved by the
inventory manager. When inventory records are inaccurate, the true inventory
level on the shelf is a random variable from the perspective of the inventory
manager. We propose tracking inventory using a Bayesian Inventory Record (BIR),
a probability distribution that evolves over time to reflect the inventory
manager’s beliefs about the true inventory level, given replenishment and sales
observations.
We
formulate the problem of optimal BIR-based replenishment as a partially
observed Markov decision process (POMDP). We analyze one- and two-period
versions of the problem, isolating and interpreting impacts of record
inaccuracy and invisible demand on the replenishment decision. In our setting,
replenishment decisions in different time periods are coupled for two reasons:
(1) because leftover inventory persists between periods, and (2) because
replenishment decisions impact the shape of the BIR. The latter reason we call
an “information effect,” and we find that it typically incentivizes a
forward-looking inventory manager to stock less than he otherwise would. In this way, our research
connects with known results on demand learning with censored observations,
where an analogous information effect incentivizes an inventory manager to stock
more.
We
examine information-sensitive replenishments over longer horizons using an
approximate POMDP algorithm inspired by the artificial intelligence literature.
In numerical experiments, we find that our approximate POMDP algorithm achieves
lower average cost than the myopic policy by ordering less. The approximate
POMDP algorithm also achieves lower BIR standard deviations on average,
suggesting that an information effect at least partially explains the difference
between the myopic and forward-looking policies.
Gary Gaukler
(
Item-Level RFID in
the Retail Supply Chain: Product Availability and Demand Forecasting
In
this talk we characterize some of the operational benefits of item-level RFID
in a retail environment. We examine a retail operation with backroom and shelf
stock under the assumption of multiple replenishment and sales periods.
Backroom stock is replenished according to a periodic-review order-up to policy
and shelf stock is replenished continually from the backroom.
Replenishment decisions are made
based on demand forecasts that are updated in each sales period based on
previous sales. The influence of item-level RFID is two-fold: first, it
directly affects the amount of products sold. Second, it indirectly affects the
retailer's demand forecast: more products sold mean a higher demand forecast,
which means a higher order-up to level in the backroom. We derive the optimal
order-up to levels for backroom stocking for both the RFID and no-RFID cases,
and we examine the relative magnitude of the direct (i.e., sales) and indirect
(i.e., forecast-driven order-up to levels) effects on expected retailer profit.
A numerical study of the dynamics of this system reveals several insights that
are of managerial interest.
Benoit Montreuil (Laval)
Item-Level RFID in
Retail Facilities: Exploratory Investigation of its Value Creation Potential
for Enabling the Real-Time Retail Demand and Supply Chain
In this paper we focus on
retail demand and supply chains exploiting
RFID enabled retail facilities. Currently, in
retail facilities RFID implementation is
mostly limited to either back store portals
for case identification. Some rare
implementations are geared for item level
identification, such as Gillette’s smart shelves
for its disposable razors. As technology
progresses and costs diminish, there will
be ever more potential for large
scale deployment of item‐level RFID
in retail outlets. Furthermore, as
triangulation capabilities expand, such RFID
implementations will gradually enable real‐time three‐dimensional
positioning of tagged items through retail
facilities. As technology progresses, the
potential for real‐time management
of retail facilities exploiting RFID
generated live positional information. Yet the
adoption of these technologies will depend
strongly on the value generated through
the retail demand and supply chain, from
the consumers to the manufacturers.
Our team has developed the LiveRetail experimental platform for
enabling the experimentation of real‐time management of RFID equipped
retail facilities. It combines a retail
facility configurator, an
agent‐based retail simulator and a web‐connected real‐time retail
management cockpit.
In the paper we first present
the architecture and functionality of the
LiveRetail platform. Second we
then describe the key learnings
from our early experimentation with the
platform relative to value generation through the
retail demand and supply chain. Third
we extrapolate from our early findings so
as to project the potential impact of
item‐level RFID on large retail networks,
large consumer goods manufacturers, and
consumers.
[Joint work with Angel Ruiz
and
Driss Hakimi]
Manu Goyal
(
Strategic Information Management
under Leakage in a Supply Chain
The importance of material flow
management for a profit-maximizing firm has been well-articulated in the supply
chain literature. We demonstrate in our analytical model that a firm must also
actively manage information flows within the supply chain,
which translates to controlling what it knows, as well as what
its competitors and suppliers know.
Our model of a supply chain
consists of two horizontally competing firms sourcing from the same supplier.
One firm (the ‘incumbent’) takes a lead in introducing a new product in the
market, the demand for which is uncertain. The incumbent can invest in
obtaining demand information not directly accessible to its competition. The
second firm (the ‘entrant’) follows the incumbent in the market with the same
or a perfectly substitutable product. Both firms source a component of the
product from the (common) supplier. Now, if the incumbent has acquired
information, his order to the supplier is likely to reflect some of that
information. The supplier in turn could leak the incumbent’s order information to the entrant.
This structure in its barest form captures the essence of numerous examples of supplier-driven
leakage, highlighted as a leading supply chain risk in multiple surveys.
We formally show that the
supplier always leaks the incumbent’s order information to the
entrant. As a result, when the incumbent acquires information, its drive to
control information flows within the supply chain can trigger operational
losses through material flow distortion. Hence the firm may prefer not to
acquire information even when it is costless to do so. However, if acquired,
demand information is always disseminated in the supply chain, aided by leakage.
This result is in stark contrast to the extant literature which argues that
demand information is not shared in similar settings. Thus, in equilibrium,
information asymmetry is dissipated in the supply chain - either all firms are
privy to demand information or none are. Our results underscore the importance
of Strategic Information Management - actively managing the supply chain’s
information flows, and making trade-offs with material flows where appropriate, in
order to maximize profits.
[Joint work with Krishnan S. Anand]
Metin Cakanyildirim
(
Partially
Observed Inventories: Signals, Sufficient Statistics and Approximations
In many
inventory control contexts, inventory levels are only partially (i.e., not
fully) observed. We discuss the recent
developments in the partially observed inventory systems and the associated
models. In these models, the inventory
level or the customer demand is observed via surrogates (signals). The system state turns out to be the
conditional distribution of the inventory/demand given a history of these
signals. In some models, this history
can be summarized by several statistics called sufficient statistics. For
example, the information delay and some censored demand models accept
sufficient statistics. When no
sufficient statistic exists, we are forced to approximate the conditional
distribution.
An option is to approximate the
conditional distribution with its mean and variance. This methodology is applied to the
zero-balance walk model where the demand is not observed, the inventory level
is noticed when it reaches zero, the unmet demand is lost, and replenishment
orders are decided so as to minimize the total discounted costs over an
infinite horizon. This problem has an
infinite-dimensional state space, which makes it difficult to obtain a simple
optimal policy. We compare
approximations that are based on the mean/variance or just the mean of the
inventory level. The mean based
approximation has the customary dynamic programming equation of the fully
observed problem, while the mean/variance based approximation has a novel
equation that resembles a mixture of equations of the fully and partially
observed problems. Value functions of
both the mean/variance based policy and the mean based policy can be used to
obtain lower bounds for the actual cost, but the bound obtained from the former
policy is stronger. Moreover, the former
policy coincides with the latter policy when the variance of the inventory
level is zero. Hence, the mean/variance
based policy generalizes the policy of the fully observed problem.
Another option is to solve the
actual problem by using numerical methods (such as a finite family of
polynomials) to represent the conditional distribution. We report a preliminary comparison of the
mean/variance based policy and the numerical solutions.
Our methodologies can be used to
evaluate the benefit of technologies, such as RFID tags, from the inventory
management point of view. These
technologies provide richer, real-time information to inventory managers in the
form of more accurate measures of inventory or more signals. In a sense, they make a partially observed
problem more of a ``fully observed problem”.
The difference between the optimal cost of the partially observed problem
and that of the fully observed problem is (a bound on) the benefit of the technology. This benefit can be used to make an objective
case against or for the technology. The
objectivity here is critical for companies hesitantly considering new
technology implementations like RFID tags.
[Joint work with Alain Bensoussan, Suresh Sethi]
Jacques Roy (HEC Montréal)
Cost-Benefit
Analysis of a Potential RFID Deployment in a Cruise Ship Supply Chain Context
Despite
simplifying assumptions to the contrary, most project managers understand that
people are both important and different. Sometimes, at the nascence of the
project (i.e., before the final concept or pre-concept) the people can be more
important than the project aspects because the right people might change the
entire direction of a project and vastly improve project outcomes. Moreover,
people are different – they can possess different skills, different
experiences, different levels of creativity and different analytical abilities.
Our
objective is to consider the impact of the people that implement a project on
future project outcomes. To be specific, we seek to determine which
people-metrics (if any) based on past performance will be better at forecasting
future project outcomes and under what conditions. Our findings could be
potentially applicable to many decisions involving people and summarizing their
past performance. In particular, our findings could improve early or
pre-concept forecasting when the people on the project might be the most
important determinants of the project’s future outcome. For example, when
finding a new cure for a disease or inventing a new distribution channel, the
creativity and insights of the people might be far more important than the
initial product concept.
We
study (both theoretically and empirically) six people-metrics based on past
performance - the mean of the past project outcomes for a person, the number of
past project outcomes, the maximum past project outcome, the minimum past
project outcome, the range (i.e., maximum minus minimum) of the past project
outcomes and the last observed past project outcome.
We
find, for example, that when people with a higher number of past project
outcomes have a higher potential (i.e., the probability of more favorable
outcomes), the maximum-metric is always more highly correlated with future
project outcomes (i.e., forecasts better) than the mean-metric. The
minimum-metric is always better than the mean-metric when people with a higher
number of past project outcomes have lower potentials. The range-metric and
number-metric both are better than the mean-metric when people of different
potentials are sufficiently heterogeneous in terms of the number of past
project outcomes. Finally, the mean-metric is always more correlated with
future project outcomes than the last-metric is.
Apart
from pre-concept forecasting, this analysis should be valuable for many
activities including providing information for decisions related to selecting
people for specific projects and demonstrating that individual people do
matter.
[Joint work with Simon Véronneau]
Tim
Huh (
A
Periodic-Review Inventory Model with Unobservable Demand
We
consider a single-product periodic-review inventory system. In each period, we assume that the system
faces two types of uncertain demand, recorded and unrecorded; the recorded
demand refers to the paying customers whose transactions are updated in the
system whereas the unrecorded demand refers to the reduction of
inventory without being updated in the system, either due to information system
incapability (unrecorded sales) or pilferage (loss). Any demand that cannot be satisfied
immediately upon arrival is lost, and incurs a corresponding lost sales penalty
cost. Due to the presence of unrecorded
demand, the actual and the recorded inventory levels may disagree, but the
managerial decisions, such as inventory counting and replenishment, must be
made solely on the recorded inventory level.
In our model, we assume that whenever inventory stocks out, the manager
incurs a fixed penalty cost, and becomes informed of the stock-out event. Furthermore, we assume that inventory
counting is costly, but is necessarily performed as a part of the inventory
replenishment process.
While the actual inventory level
at a given period depends on the entire history of the observed process
(inventory records), we identify that sufficient information is captured by the
pair of (i) the recorded inventory level and (ii) the
number of periods since the last inventory correction. Under mild technical conditions, we obtain
several monotonicity structural results, relating the
actual inventory level and the recorded inventory levels, which are useful
developing the structure of the optimal policy.
If the unrecorded demand consists of unrecorded sales and the inventory
cost is charged based on the maximum storage capacity, then we show the
optimality of a two-parameter policy, called the (l, S) policy. If unrecorded demand is either unrecorded
sales or loss and the inventory cost is charged based on the actual inventory
level in each period, then we identify a sufficient condition for the
optimality of (l, S) policy, which is shown to be optimal in our numerical
experiments.
Diego Klabjan
(Northwestern)
Next
Generation Business Applications for Radio Frequency Identification
RFID is moving from early stages of
slap-and-ship to integration with existing systems and business applications.
It is the latter that will yield a return on investment. In addition to
existing business applications such as promotions execution we discuss in
details two new novel applications.
We
present new models that capture real-time status of shipments and make optimal
inventory control decisions. In addition, we show analytically that RFID
real-time data yield better inventory control policies than the traditional
setting. RFID data can also be explored in expediting replenishment orders. We
introduce so-called sequential systems, which have nicely structured policies.
The regular and expediting orders follow a base stock type policy.
Supply
Disruptions and the Reverse Bullwhip Effect
Hurricane Katrina in 2005 crippled
much of the
We
introduce two analytical models to demonstrate the existence of the RBWE. In the first, we assume that a single buyer
procures product from a single seller that is subject to disruptions in the
form of capacity shocks. A change in
capacity causes a change in the price of the product. If the buyer anticipates further price
changes in the future due to a prolonged disruption, he may purchase a quantity
that differs from the quantity specified by his steady-state demand curve. We provide conditions under which the
variance of (an approximation of) the order quantity exceeds the variance of
the capacity, and therefore that the RBWE occurs. We also prove that the magnitude of the RBWE
increases with either the severity or the duration of the disruption.
Our
second model examines buying patterns when multiple retailers compete for
scarce product from a single supplier.
This model is based on the “rationing game” discussed by Lee, et al.
(1997), who argue that the BWE occurs between the
retailers and their customers (i.e. the retailers’ orders are more volatile
than their customers’ demands). We
examine this claim more closely, verifying it under certain conditions and
questioning it under others.
Furthermore, we argue that the capacity uncertainty causes the RBWE to
occur in the upstream portion of the supply chain; that is, that the retailers’
orders are more volatile than the supplier’s orders. Finally, we consider an alternate pricing
structure in which the retailers pay for every unit ordered, plus a separate
price for units actually received. This
pricing structure discourages retailers from inflating their orders too severely. We demonstrate that this pricing structure
causes a Nash equilibrium of order quantities to exist where it otherwise would
not, and we prove the resulting existence of the (R)BWE.
[Joint work with Zuo-Jun
Max Shen, Ying Rong]
Sean
Marston (
The Impact of
Digital Technologies on Government Cultural Policies
Many countries limit the influence
of foreign cultural products such as music, film, and television programs to
protect their cultural identify.
Commonly observed tools include Quotas, tariffs, and subsidies. However, the advances in digital technology
create new avenues, such as internet, for consumers to access foreign
entertainment programs. This calls a re-examination of the effectiveness of
these traditional tools. We create a
unified analytical framework to study the impact of digital technology on
cultural protection policies. We find that the performances of these tools are
greatly affected by the quality difference between domestic and foreign
entertainment programs (through both traditional channel and Internet), and
quota produces the least social welfare no matter whether there is leakage
through internet.
[Joint
work with Kenny Cheng, Jane Feng, Gary Koehler]