Table of Contents
Innovation Criterion for Evaluating the Organizational
Effectiveness of a Retail Chain using a Complex Adaptive
System Model and the SWARM Simulation Environment
G. P. Papaioannou, John Vlahos and A. N. Yannacopoulos
Abstract
In this paper we examine the application of a Complex Adaptive System
(CAS) on studying the relative organizational effectiveness of Centralized and
Decentralized Retail Chains. The criterion for the evaluation of this
effectiveness is the rate of creation of new (innovative) ideas, related to
management policies and practices, by the Shops of the Chain, which are the
agents of the CAS Model.
The diffusion of these innovations throughout the Organizational structure and
finally their adoption as a new standard practice or policy by the whole chain,
is also an important dimension of the evaluation criterion. We provide here
some basics for the CAS modeling and their connection with the traditional
Systems Dynamics modeling, justifying as well our decision to use the CAS
formalism. We examine a generic Retail Chain organization, the
corresponding structure of the CAS model and finally we give details on the
simulation of the model with the SWARM software system. We produce
various scenaria by changing the chain’s main structural parameters and
finally we discuss the results obtained and draw conclusions on the relative
organizational effectiveness of centralized and decentralized Retail Chains.
KA/NAN/NEW YORK 1
1. Introduction
Our main objective is to study the relative effectiveness of centralized
and decentralized organizational structures of a retail chain using
innovation as a criterion. By innovation we mean the continuous creation
of information that improves organization’s ability to adapt to a constantly
changing competitive environment. Since we view the organization as a
collection of agents, each of whom is capable of creating new ideas, the
use of aggregated models such as the usual stock and flow
presentations, is not recommended here. Stock and flow models are good
for capturing the time evolution of the average behaviour of the underlying
original system but they are leaving out the details related to fluctuations.
The missing part of the dynamics attributed to these ignored fluctuations
are often extremely important for the understanding of the underlying
dynamics and may be responsible for unpredictable emerging collective
properties, which can be revealed only through a multi-agent system
formalism.
This work is an extension of the work of Chang and Harrington model
(afterwards referred as CH) of a retail chain (Chang and Harrington,
2000). The main modification we made on the C-H model is the
introduction of the notion of the firm topology, which is related to how
different subunits are connected and how they interact. The topology in
the C-H model is extremely simple and this is the reason why the creators
of the model surprisingly claim that it is the centralized and not the
KA/NAN/NEW YORK 2
decentralized organization, with coordinated search for a global optimum
in the competitive landscape, that is more effective in a constantly
changing environment. In contrary, our study shows that when we make
the firm topology a little more complicated, then decentralization
outperforms centralization, almost in every case, when the markets are
not stable (i.e. in a state of volatility).
The performance of a firm is related to its structure. The same firm, under
the same conditions may have different performance in the same
environment, depending on how it is organized and structured.
At this point we need to define the term structure. A working definition
could be that structure is how the different sub-units constituting the firm
are connected, how they interact and how they exchange information.
Loosely speaking, we may understand the importance of structure in the
performance of a firm, as the structure of a firm is related to how flexible
it is, and how well it may interact with its (possibly changing) environment
and adapt to external conditions. Borrowing from biology, it is related to
how well an organism may take evolutionary steps and move in the
fitness landscape towards more desirable states, (Kauffman, 1993).
Thus, a crucial question in management science, is the determination of
the optimal structure of a firm, depending on the nature of the firm and the
market in which it functions.
The aim of the present research is to try and answer some questions
related to this matter, using the methodology of complex adaptive
systems (CAS) or Multi-Agent Adaptive Systems. To this end, we will
model a firm by a CAS that will reproduce the basic functions and
KA/NAN/NEW YORK 3
characteristics of a given firm. Then, by the use of computer simulations
we may “run” different scenarios for the evolution of a firm, and check the
performance of different structures to provide solutions. The present
model is a rather “generic” model of a firm that has the basic structure of a
retail chain. As a result of that we may be able to draw general
conclusions about the effect of structure on a firm’s performance. The
simulations were performed in the computer simulation package SWARM
which is very well suited for simulations of complex adaptive systems (see
section 2 for more details on SWARM). Although this work is of academic
nature, the computer software developed using the simulation package
may be developed into the form of a micro-world (Casti, J.D., 1997,
Morecroft J., et.al, 2000) that may be used for the simulation of the
function of a given firm. The micro-world may be used by managing
directors as a tool for planning and strategic decision-making.
The present work may be considered as one of the many recent efforts
of systems dynamics to encapsulate Complex Adaptive Systems (CAS) or
Agent-based models into the area of the traditional Systems dynamic
modeling, within the frames of new “challenges for the future” as they
described in Chapter 22 of the Sterman’s book (Sterman J.D., 2000) and
his related paper (Sterman, J.D., 1994).
2. CAS and the SWARM computational environment
In this section we provide some basic information on CAS, their relation
with the traditional Systems Dynamics modeling and the SWARM
simulation environment.
KA/NAN/NEW YORK 4
2.1. CAS basics
We take complexity to mean the intricate inter-relationships that arise from
the interaction of agents, which are able to adapt in and evolve with a
changing environment. The theoretical framework being developed is
based on work in the natural sciences (in physics, chemistry, biology,
mathematics and computer simulation) studying complex adaptive
systems (CAS).
In an organizational context, complexity provides an explanatory
framework of how organizations behave. How individuals and
organizations interact, relate and evolve within a larger social ecosystem.
Complexity also explains why interventions may have un-anticipated
consequences. The intricate inter-relationships of elements within a
complex system give rise to multiple chains of dependencies. Change
happens in the context of this intricate intertwining at all scales. We
become aware of change only when a different pattern becomes
discernible. But before change at a macro level can be seen, it is taking
place at many micro-levels simultaneously. Hence micro-agent change
leads to macro system evolution.
Complex Adaptive System (CAS) is found in everyday life. A crucial
distinguishing characteristic of such systems is that their component
elements are living “agents” capable of autonomous behaviour, which can
be adapted to changing circumstances. This contracts with complex
systems in chemistry, physics and engineering founded on established
KA/NAN/NEW YORK 5
theories explaining observable phenomena encompassing interactions
between non-living elements.
An overwhelming spectrum of living systems falls within the scope of
current research, from stock markets to supermarkets urban. traffic
networks, national economies to global ecosystems and_ business
organizations. However, although many effective models have been
created (Casti 1997), there is still no real science to provide theoretical
foundations for building these kinds of systems. One of the main reasons
for this slow progress is that researchers into social and behavioural
phenomena have not had the ability to conduct the controlled, repeatable
experiments that are an integral part of the methods employed in natural
sciences to test hypotheses and establish new theories.
Until the advent of widespread and usable computer power, it was
generally impractical to perform experiments on everyday social and
behavioural systems. For example, Wall Street cannot be asked to
change its rules to allow an economist to check a new theory of financial
markets. And even if such an unlikely event happened, a genuinely
repeatable experiment could not be conducted because too many
variables would have to be considered. In other cases, experiments would
take too long to be of practical value or may pose too much danger to the
real world, say by trying to evaluate a theory about biological diversity by
introducing a new species to an environment.
The power and versatility of computer technology has now reached
the point where we can create realistic “silicon surrogates” (Casti, 1997),
KA/NAN/NEW YORK 6
encapsulating inside a computer the full scope and richness of interactive
patterns of the social systems we want to experiment with.
Research at SFI (Casti 1997) has indicated that the existence of a
medium-sized number of intelligent adaptive agents making decisions on
the basis of local information can be regarded as the “fingerprint”
indicating that a system being studied can be classified as a CAS.
However, these features do not constitute a full definition of complex
adaptive systems.
The three distinguishing characteristics of a CAS fingerprint involve:
1. A medium-sized number of individual agents. An agent is the
basic element in a CAS. A customer in a shop, or a shop in a retail
chain of an industry, are examples of potential agents in industry
microworlds. The number of agents must be neither so small that all
their interactions could be worked out “on the back of the envelope”,
nor so large that statistical aggregation methods could tell you
everything you want to know about the system. In the type of CAS we
are concerned with, here the actual number of agents can be
considered “low grained”, in the range of a few hundred to a few
thousand.
2. Intelligent agents with the ability to adapt. Agents need to be
“intelligent” in the sense that they can use in-built rules to decide what
actions to take at any given moment. If they find a current rule isn’t
working well, agents should be “adaptive” in their ability to discover
and change to new or different rules.
KA/NAN/NEW YORK 9
3. Local information only. All agents invoke their rules to make
decisions on the basis of only partial, or “local”, information. This
means there is no agent within the system, which knows what every
other agent is doing. The “localness” can relate to physical or
informational dimensions.
2.2. CAS and “Traditional” System Dynamics Models
So far, traditional system dynamics and complex adaptive systems
have been treated as two completely separated aspects of reality,
whether physical or social. However, nowadays with complexity and
nonlinearity coming of age, it is high time to reconsider and view things
under a new perspective. We should therefore try and reconcile the two
approaches into one unified view of reality (Sterman, J.D., 1994). Our
philosophy is that complex adaptive systems and system dynamics are
just two different glimpses of the same phenomena but in different
scales. Modeling uses one and unique methodology but depending on
the coarse graining and the detail with which we wish to study a system,
we may end up with a system dynamics or a complex adaptive system
model. As a matter of fact, we wish to stress that a complex adaptive
system model, being more detailed, may under appropriate averaging,
or coarse graining be reduced to an appropriate system dynamics
model. The degree of coarse graining used in a system, may be
equivalent to different averaging procedures, which in turn reduce the
original fully detailed complex adaptive system to a series of system
KA/NAN/NEW YORK 8
dynamics models of increasing level of complexity. This is the analogue
of interacting particle systems models and mean filed approximation
models used in the physical sciences.
2.3. The SWARM Simulation System
The Swarm simulation system’s objective is to provide researchers
with a standardized, flexible, reliable set of software tools for
experimenting with complex adaptive system of the type we will discuss in
this paper. SWARM is a set of libraries that facilitate implementation of
agent — based models. Artificial life, which tries to explain biological
phenomena, is the inspiration of SWARM.
At the time of Swarm’s inception, researchers in the field of CAS were
finding that ad-hoc programming was not a sufficiently powerful, reliable,
or economical way to ask the kinds of questions that needed to be asked.
Chris Langton of Santa Fe Institute (Langton, C., et al 1995) having
seen this problem decided to initiate the SWARM project in 1994 at the
Santa Fe Institute.
Virtual machine is the primary feature of SWARM. The virtual machine
allows the researcher to describe agent behaviors one by one, agent by
agent, context by context, all while keeping an exact notion of time and
currency in the world. Swarm also makes it possible to compose or
decompose hierarchies of agents. This is the composability attribute.
This notion of composability is useful because it often isn’t clear where to
begin a modeling effort. For example, in modeling a large organization, it
KA/NAN/NEW YORK 9
3.
may be the case that the subjective understandings of individuals’ or
departments’ roles and responsibilities differ widely, and that this variance
includes poor performance in some cases and outstanding performance in
other cases. Rather than seeking denotation on how the organization
should work and looking for deviations, one can built independent model
components from many perspectives and then combine them (mirroring
abstractions of people for real people). This bottom-up approach has the
advantage of documenting the both unexpected bad and good things in
the organization, as well as contextual sensitivities (Casti, J., 1997).
A Schedule is an agent’s to-do list. There are different kinds of to-do
lists, and different attributes that Action items on the to-do list can have.
An Action is something that happens in the world. In Swarm,
Schedules and Actions are typically closely associated with an agent or
model component. Agents may have their own Schedules (perhaps
several) and a repertoire of Actions they know how to perform.
A model for a retail chain
We will briefly describe a model of a retail chain that may easily bring
into the general framework of complex adaptive systems. The model is
very broad and versatile, and can be used in the modeling of a wide rage
of firms but here to be precise we will present a brief description of how
this may be used to model a retail chain. The model is a generalization of
a model for a retail chain that was first proposed by Chang and Harrington
(CH) (Chang M. -H. et al, 2000).
KA/NAN/NEW YORK 10
3.1 The original model of Chang and Harrington
In the original model of a retail chain, first proposed by Chang and
Harrington, there is a headquarters (HQ) controlling the activity of M
stores (sub-units). The stores interact with each other only through the
headquarters. Each store has N dimensions in its policy. A dimension is
related to some activity of a store e.g. pricing policy or customer support
policy. Every dimension consists of R practices. In this abstract model
each store at any time may be described by an N dimensional vector
Z=(a1,a2,.....an) where each of the aj may take R distinct values. The
vector z will hereafter be called the store policy.
The market of each store is assumed to consist of a collection of
economic units-consumers (agents). Each agent has some preference
towards the policy of each store, and is considered to be a rational entity,
acting to maximize some utility function. This utility function depends on
the quantity that an agent will buy from a store, on the price and on the
“distance” of the agents preferred store practice w and the store’s actual
practice z. The price is considered as given by the store and the agent
only has control over the quantity that will consume. This is chosen
optimally, and is a function of the price, the preference and the policy.
At the next level, the store itself may be modeled as a collection of
agents, having a distribution of preferences. The probability distribution
of agent’s preferences F(w) is considered to characterize the market in
which the store acts. For each store we may define a profit function Pj,
which is simply the total demand for some product (weighted average over
KA/NAN/NEW YORK ll
all agents) multiplied by the net profit per unit. This profit function depends
on the price (over which the store has control) and the deviation of the
stores actual practice from the preferred practice by the agents. The aim
of each store is to maximize its profit. This is done by choosing
optimally the price p of the product in question.
Finally, the whole unit is modeled by a profit function, which is simply
the sum of the profit function of each store. A good example of how the
structure of a firm may affect its behavior is in its reaction towards
innovation. We give here some details on what we mean by this term.
3.2 Innovation as explanation of competitive landscape and as a
process of information creation and communication.
Traditional methods usually ignore an organization’s capacity to learn
and change and to maintain diverse and varied strategies, assuring that a
single “optimum” strategy is both possible and desirable. For an
organization, such our firm, to survive and thrive it needs to explore its
space of possibilities and to encourage variety. When markets were stable
and growth was a constant, single optimum strategies based on
extrapolation from historical data, were thought to be feasible. But as
Ashby has shown (Ashby, 1964, 1969), unstable environments and rapidly
changing markets require flexible approaches based on variety.
In our work we adopt the view of Radner (Radner, 1993) and Van Zandt
(Van Zandt, 1998) of the innovation as the process of information creation
and communication, as well as the views of Chang and Harrington (Chang
KA/NAN/NEW YORK 12
and Harrington, 1998, Chang and Harrington, 2000) according to which
innovation is regarded also as a random search carried out in a fixed
space of ideas.
An innovation (idea) may be though of as a process that alters the
policy of a store. In the model at hand, an idea may be though of as some
process altering some of the entries of the policy vector z. The complexity
of an idea is related to the number of entries of the policy vector, which are
altered. An idea may originate either at store level or a headquarter level.
An idea originating at headquarter level, may change some of the entries
of the policy vector of some store. This number is related to the degree of
centralization that a firm has. The smaller it is the grater the freedom that
the headquarter allows the store managers towards innovation.
The modeling of innovation adoption will be as follows:
e An idea is randomly generated either at store level or at
headquarter level.
e If an idea is generated at HQ level it is tested on whether it
increases the potential profit of the chain and if it does it is adopted
at a global level.
. If an idea is generated at store level, it is tested on whether it
increases the profit of the store and then it is communicated to
HQ where it is tested on whether it may increase profit at chain
level. If the idea increases the profit of y store it is adopted,
otherwise it is dropped. How easy the communication of an idea
KA/NAN/NEW YORK 13
to HQ is and how easy it is to adopt some idea or not is related to
the structure of a firm and to the degree of centralization.
The markets may be stable (that is the agents preference distribution is
independent of time) or fluctuating (the agents preference distribution may
change stochastically in time).
3.3 A generalization of the Chang and Harrington model
The Chang and Harrington model is interesting and versatile enough to
model decision making in a wide range of firms, even though it was
originally proposed in the context of retail chains. However, we propose
here some extra features that we feel may lead to a generalization of the
model which may be used to shed even more light on the problem of
interaction of the structure of a firm on decision making and its
performance.
We begin by briefly describing the change we feel are major regarding
the structural characteristics of the model and towards the end of this
section mention some less important changes which will as well lead to
more realistic features.
The major point in our generalization of the model of Chang and
Harrington is the introduction of the notion of firm topology in the model.
This notion is related to how different sub-units are connected and how
they interact (e.g. how they exchange information). In the original CH
model, the topology of the firm is extremely simple and essentially is the
KA/NAN/NEW YORK 14
topology of a simple tree where we have a headquarters (HQ) and all
stores (S) directly connected to it. While this may be a plausible structure,
real life firms may display more complicated connections between their
sub-units. For instance we may generalize this structure by introducing
various levels of sub-headquarters (S-HQ) that will be responsible for
decision-making and will only be responsible for the function of certain
groups of stores. This will lead to a different firm topology, different
connectivity properties etc. The idea of local management, as is clear
intuitively may lead to effective management of local units so as the firm as
a whole may be able to interact more efficiently to an inhomogeneous
market. Furthermore, this structure may not be static. The topology and
the structure of the firm may be made to vary depending on the
performance and the long-time scale properties of the fluctuations of the
market. This feature of the model may be used to model the potential
strategic restructuring of a firm in the course of its function. This
dynamic feature is built in our generalized model and it turns out that the
simulation package we employ is well suited to deal with that. The
structure proposed here is just one possibility. It is interesting to try and
test different connection topologies (in this task ideas from neural network
theory may be very useful) with regards to their performance and try to find
the optical connection topology.
We now describe some less major, but all the same interesting changes
in the original model.
KA/NAN/NEW YORK 15
. In the model of Chang and Harrington, ideas are randomly
generated. To make the model realistic, we have to introduce at
store level a search policy to increase profit at store level.
. We have to take into account interaction of agents and store, in
what the agents’ preferred practice may not be considered as
given and unchanged in time, but will change according to what
the store offers. In other words there will be some sort of feedback
between the store and the consumers that will affect the
consumers taste.
e¢ The policy of HQ may not be uniform towards the whole chain
but may change from store to store. For instance the HQ may put
more weight in certain markets neglecting others, or may allow
more freedom in certain stores depending on store managers
abilities etc.
3.4 Description of the implementation
For the implementation of the simulation software we used Swarm (see
section 2.3). The basic architecture of Swarm is the simulation of
collections of concurrently interacting agents: with this architecture, we can
implement a large variety of agent-based models. Swarm is a collection of
object oriented software libraries, which provide support for simulation
programming (see Langton, C., 1995). We build simulations by
incorporating Swarm library objects in our programs. Figure 1 shows the
KA/NAN/NEW YORK 16
main objects that Swarm provides and their interactions in a simulation
application.
One can define agents (either independent or belonging to a group) and
their behavior (actions and schedule). This is the model of the simulation.
The observer object monitors the model execution and provides methods
to output the results (to a GUI or to files on the hard disk) (Benedic S., et
al, 2000, Daniels M., 2000).
BASIC OBJECTS OF SWARM
Figure 1a Main Objects of Swarm environment
KA/NAN/NEW YORK 17
If
<cond>
then
<action1>
else
<action2>
©
+ Tlapampytiis
animated agents
<=
Organization of
agents
Figure 1b
3.4.1. Definition of categories of Agents
For the implementation of a simulation with Swarm, the first step is to
define the agents of the simulation. In the case of our simulation of a retail
chain, we defined the agents, as shown in figure 2.
KA/NAN/NEW YORK 18
Sub-Headquarter
Sub-Headquarter
Figure 2 : The agents of the simulation
‘Sub-Headquarter
Customers:
The main agent is the Store. Each store has a number of Customer
agents and it belongs to a Headquarter or a Sub-Headquarter agent,
depending on the structure of the Retail chain. As one can see, stores and
headquarters are modeled as a whole, without being analyzed to sub-
agents (for example, employees of the store).
3.4.2. Agent characteristics
Customers: Each customer belongs to one store for the whole simulation
period. Each one has a set of preferences, which is described by a
vector of arithmetic values. Depending on the simulation parameters,
the preferences of a customer may remain the same or change during
the simulation period.
KA/NAN/NEW YORK
19
Stores: Each store has a set of practices, which is described by a vector
of arithmetic values. The store’s market consists of a number of
customer agents. The profit of the store is maximized when its
practices are closer to its customers’ preferences. The markets can
be the same for all the stores, or may vary from store to store. This is
accomplished with differently distribution of the customers (actually
the customer preferences). The stores generate new ideas at every
repetition of the simulation. These ideas are evaluated and are
accepted if they result in raising the profit of the store and the retail
chain.
Sub-Headquarters: The Sub-Headquarters agents introduce another level
of complexity to the simulation. Each Sub-Headquarter has a number
of stores under its authority. Depending on the mode of simulation
selected by the user, the Sub-Headquarter either takes part in the
evaluation of new ideas generated by their stores, or they just act as a
carrier, in order to transfer information among the stores.
Headquarter: It is the central point of the model. It gives the total profit of
the Retail chain. Depending on the structure of the chain, which was
selected by the user, it has a number of stores or sub-headquarters
under its authority.
KA/NAN/NEW YORK 20
3.4.3. Model flow
The simulation starts by building the different agents. The creation of
the agents starts from the bottom. First, we create the customer agents.
Then the stores are created and the customers are assigned to them. The
next step is the creation of the sub-headquarters, if the user has selected a
three-level simulation. The stores are then assigned to the sub-
headquarters, in an equal manner. That means that if we have 4 stores
and 2 sub-headquarters, the first two stores are assigned to the first sub-
headquarter and the other two to the second. Finally, the headquarter
agent is constructed. Depending on the structure of the Retail chain (two
or three levels), the Headquarter agent is connected either to the sub-
headquarter agents or to the store agents.
When the creation of the agents is completed, the simulation starts
executing. In each iteration of the simulation, we have the following steps:
Each store has a new idea (innovation). The new idea is a change in one
of the practices of the store. The store then calculates its profit with the
new idea and compares it with its previous profit (before the emergence of
the innovation). If the new profit is lower, then the idea is discarded.
Otherwise, it is passed to the next level of the retail chain structure for
further evaluation (headquarter or sub-headquarter). The Headquarter or
sub-headquarter then passes the idea to the stores that are under its
authority. If the mode of operation of the retail chain is the decentralization,
then the idea is considered independently by each store. If it raises its
profit it is realized, otherwise it is discarded. If the mode of operation of the
KA/NAN/NEW YORK 21
retail chain is the centralization, the idea is mandated to all the stores, if
more than y stores (y can be determined by the user) profit from it.
3.4.4. User Interface
The user interface of the application was designed with the principle to
facilitate the use of the software. We have combined GUI elements that
are provided by the Swarm with custom-made frames.
Number of Stores:
Number of Sub Headquarters:
Dimensions of Operation:
Max Value of Pract
Number of Periods tn t
Number of Customers
Comtratization Mode:
Market is Changing:
‘Show Practices Frame:
Show Structure,
Use Normal Distribution:
Profit function paramet
Figure 3: The screen where the user enters the simulation parameters.
KA/NAN/NEW YORK 22
Figure 4: The graphs that present the profit of the stores and the
headquarters
KA/NAN/NEW YORK 23
‘Store Practices and Profit:
‘SubHO 0 - Stor.../SubHO 0- Stora..| SubHO 1 - Stor.../SubHO 1 - Store. /SubHa 1 - Stor...
a “91 45 Fi
19 oR 79)
38 55) 76 a
59
92 92 53)
67) En 83
40
$4
rofit= 221382... Profit= 263372...
Profit 247881...
284: 76:91:72: 78:19:13
99:94°61:6:43:92°19
99:67. 40:77:65, 89:1
SubHO 1-Store 1; 45:43: 76:6:94.9.15:4:91 28
(SubHO 1-Store 2: 77:79: 42:53:83: 4321487 -64217
SubHO O-Store 0: 21 24: 76:91:72: 78:19: 19-> Profit: 221380 2502051
Figure 5: Custom made screens for showing the practices of the stores
and the structure of the Retail chain.
KAIMANINEW YORK 24
4. Results
4a. Description of the simulation procedure
The software developed permits to conduct simulation runs with
different parameters. Some of them are mentioned here:
e The structure of the Retail chain may vary. The number of stores
and the existence (or not) of sub-headquarters can be defined by
the user.
e The number of customers that are the market of a store may vary.
The preferences of the customers are distributed over the type
space either according to an uniform distribution, or according to a
normal distribution with given mean and standard deviation. The
distribution of customers may be either the same for all the stores
(homogeneous markets) or different for each store (heterogeneous
markets).
e Many parameters of the profit function may be changed by the
user.
e The duration of the simulation runs can also change, in order to
simulate over short or long horizons.
In order to perform the evaluation and compare the different models of
operation, we have to keep some parameters fixed:
KA/NAN/NEW YORK 25
e¢ The number of customers for each store was set to 1000.
e The number of preferences for each customer is set to 10, each
having a value from 1 to 100. So, each store also has a set of 10
store practices, each having a value from 1 to 100.
e The parameters of the profit function for each store have remained
unchanged for all the simulation runs.
e The maximum duration of the simulations was set to 500
repetitions. Results were gathered for 100 repetitions (short
horizon) and 500 repetitions (long horizon).
The simulation runs were conducted on the following retail chain
structures:
e 2,4,6 and 8 stores without Sub-Headquarters.
e 5 stores with 2 Sub-Headquarters, where the first Sub-
Headquarter has 2 stores and the second 3.
e 8 stores with 2 Sub-Headquarters, each having 4 stores.
For each of the above configurations, runs were performed to compare
the results among centralization and decentralization with different market
conditions (heterogeneous vs. homogeneous markets, markets that are
stable vs. markets that change during time etc). A total of 1000 runs were
performed.
KA/NAN/NEW YORK 26
4b. Description of results
The
results show that generally decentralization outperforms
centralization in the majority of cases. Partial centralization gives better
results when the retail chain has a three-level organization (it includes sub-
headquarters). The best solution in that case seems to be a combination of
centralization (sub-headquarter over stores) and decentralization
(headquarter over sub-headquarters).
In more details, we can say the following:
o
Centralization is more likely to outperform decentralization when
the stores have similar markets, while decentralization is more
likely to outperform centralization when markets are different.
Since centralization imposes uniformity of practices, these results
are not surprising. When the markets are different, it is better to let
each store change its practices according to the needs of its local
market. On the other hand, when markets are similar, common
practices can be imposes and can give better results.
Centralization gives better results in this case, especially when we
have a large number of stores and long horizons of simulation.
The presence of sub-headquarters does not have an impact when
markets are similar. However, if markets are different, the
presence of sub-headquarters can flavor centralization, if markets
are grouped by similarity (sub-headquarters have under their
authority stores with sufficiently similar markets).
KA/NAN/NEW YORK 7
¢ Decentralization outperforms centralization, almost in every case,
when the markets are not stable (customer preferences change
over time). Centralization does not seem to be able to follow the
changes of the market of every store. On the other hand,
decentralization gives the ability to each store to better adjust its
practices, according to its customers needs. The presence of sub-
headquarters does not change the situation very much in this
case.
KA/NAN/NEW YORK 28
‘Number ofMowidose
‘Store Frofit
Figure 6: Store profit evolution example for 4 stores with different markets, for
decentralization (up) and centralization (down).
KA/NAN/NEW YORK 29
Decentralization gives better results, especially for two of the 4 stores. The reason
is that the similar practices enforced by centralization keep the two stores
away from their optimum practices.
We have of course to comment that, in our simulation, the communication
and transfer of ideas among stores is always perfect, in both modes of
operation, centralization and decentralization. In the real world, that may not
happen every time, so decentralization would probably have worse results,
since the information about a new idea may never (or very late) reach all the
stores of the retail chain.
5. Future improvements — Conclusion
One important improvement that will be made to the initial model is the
introduction of the ability of customers to move. In that way, a customer that is
not satisfied with one store’s practices will be able to move to another store in
his vicinity.
Another important step in enriching the model would be the introduction of
rival businesses. In that scenario, a customer will also be able not only to move
from store to store of the same retail chain, but also, if he is not satisfied by the
retail chain in general, to move to a rival's store in the region.
Finally, improvements can be made in the user interface of the software. In
future versions, the user will be able to design the structure of the retail chain
in a graph. The software will then read the graph and construct the simulation
KA/NAN/NEW YORK 30
for the specified structure. As a result, even more complicated structures will
be supported.
> The model of Chang and Harrington, does not take into account
competition at all. We intend to introduce into the model a competing
firm B that will complete for the same market trying to adopt policies to
approximate better the agents preferences. The stores of the rival firm,
may observe the firm’s A practice and learn from its experience.
> Finally, a detailed study of the agents’ utility function has to be made.
This will give us the characteristics of the market. We also need to
introduce inhomogeneity of the markets (i.e. different preference
distribution from market to market).
References
e Ashby, W., R., (1969), Self-regulation and Requisite Variety, in Emery
(1969)
e Ashby, W.,R., (1964), An introduction to Cybernetics, London: Methuen
e Benedikt Stefannson’s and Franceso Luna, Economic Simulations in
SWARM: Agent-Based Modeling and Object Oriented Programming,
Kluwer, 2000
e Casti, John, L., “Would-be Worlds” Chichester: John Willey & Sons, 1997
e Chang, M. H. and J. E. Harrington, Jr (1998), “Organizational Structure
and Firm Innovation in a Retail Chain” Computational and Mathematical
Organization Theory, 3(4), 267-288
KA/NAN/NEW YORK 31
Chang, M. H. and J. E. Harrington, Jr (2000), “Centralization vs.
Decentralization in a Multi-Unit Organization: A Computational Model of a
Retail Chain as a Multi-Agent Adaptive System” Santa Fe Paper 00-02-010,
(2000)
Daniels Marcus, An open framework for agent-based modeling, slides for a
talk delivered at “Applications of Multi-Agent Systems in Defense Analysis”,
a workshop held at Los Alamos National Labs in April 2000 (see
http:/Awww.swarm.org/intro-papers.html)
Kauffman, S., Origins of Order. New York: oxford University Press, 1993
Langton, C., N., Minimar, and R. Burkhart. “The SWARM Simulation
System: A tool for studying Complex Systems”. Santa Fe Institute preprint,
March 1995. Also, for free download of the SWARM:
http://www.santafe.edu/projects/swarm or www.swarm.org
Morecroft, J., Sterman, J., Modeling for learning Organizations, Productivity
Press, 2000.
Sterman, John D., "Learning in and about Complex Systems”, System
Dynamics Review 10 (2-3), 291-330, 1994.
Srerman J. D., 2000. Business Dynamics: System Thinking and Modeling
for a Complex World. Inwin McGraw — Hill: Boston, MA
Radner, R., (1993) “The Organization of Decentralized Information
Processing”, Econometrica, 61, 1109-1146
Van Zandt, T., (1998), “Real-Time Hierarchical Allocation”, Northwestern
University, Center for Mathematical Studies in Economics and Management
Science, Discussion Paper No 1231
KA/NAN/NEW YORK 32
Back to the T
