Embedding sustainable development strategies in agent

International Journal of Geographical Information Science
Vol. 22, No. 1, January 2008, 21–45

Research Article

Embedding sustainable development strategies in agent-based models for
use as a planning tool

XIA LI* and XIAOPING LIU
School of Geography and Planning, Sun Yat-sen University, Guangzhou, 510275, PR
China

(Received 22 June 2006; in final form 11 January 2007 )

Rapid land development in rapidly growing countries has created a series of land-
use problems. The implementation of sustainable land use can alleviate some of
these problems. It needs a set of tools for the exploration, design, modification,
illustration, and evaluation of alternative planning scenarios. This paper
demonstrates that the integration of cellular automata and agent-based
modelling can provide a spatial exploratory tool for generating alternative
development patterns. Sustainable development strategies are embedded in the
modelling to regulate agents’ behaviours. The use of agents can help to represent
human–environment interactions in solving complex land-use problems. It is able
to examine the effects of different stakeholders in influencing the process of land
development. The proposed model has been applied to the simulation of
planning scenarios for residential development in a rapidly expanding city in the
Pearl River Delta.

Keywords: Agent-based modelling; Cellular automata; GIS; Urban planning

1. Introduction
Rapid urban expansion in rapidly growing countries has created a major concern for
sustainable land use in these regions (Li and Yeh 2001). Massive conversion of non-
urban land into urban land has created a series of land-use problems, such as a
decrease in food production, destruction of sensitive ecosystems, water and air
pollution, and deprivation of future land supply (Yeh and Li 1999, Jantz et al. 2005).
Sustainable land use, which should also coordinate the land-use demands from
multiple aspects and different interest groups, can provide a useful tool to alleviate
some of these land-use problems. The implementation of sustainable land use is
quite complex because it involves social, economic, and environmental factors.
Modelling systems can be developed to provide assistance in implementing the
initiatives of sustainable land use (Zandera and Kachele 1999). These models are
¨
useful for carrying out a scenario analysis which is a promising and interesting
planning tool for investigating future possibilities in a changing environment
(Nijkamp et al. 1997).
Cellular automata (CA), a type of bottom-up approach, have been used to
investigate the ‘business-as-usual’ scenario, that is, further development of present
conditions. These models have been widely used to simulate complex geographical
phenomena which have nonlinear and emergent features (White and Engelen 1993,

*Corresponding author. Email: lixia@mail.sysu.edu.cn
International Journal of Geographical Information Science
ISSN 1365-8816 print/ISSN 1362-3087 online # 2008 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/13658810701228686

22 X. Li and X. Liu

Batty and Xie 1994, Li and Yeh 2000). However, CA have limitations reflecting the
decision behaviours of individuals, such as governments and investors, in shaping
urban growth. The influence of human factors is difficult to include in traditional
CA (Torrens and Benenson 2005).
Agent-based modelling (ABM) can be used as a tool for analysing complex
natural systems (Courdier et al. 2002). A major feature and advantage of ABM is
the ability to produce nonlinear and emergent phenomena based on behaviour of
individuals. In the past, ABM have been used mostly in purely social contexts
(Gilbert and Conte 1995). They were used to validate or illustrate social theories
(e.g. biological, economic, and political theories) or to predict the behaviour of
interacting social entities (e.g. actors in financial markets and consumer behaviour)
(Basu and Pryor 1997). However, this type of model does not make use of spatial
information central to geographical analyses.
Both CA and ABM are limited in their geographic functionality when considered
in isolation (Torrens and Benenson 2005). However, increasingly researchers are
turning to the integration of CA with ABM to produce better simulation results.
Research has indicated that not only the neighbourhoods (or cell states of CA) but
interactions between local actors and their environment must be considered in order
to forecast landscape transition with a higher accuracy (Loibl and Toetzer 2003).
The integration of CA with ABM promises to provide a powerful spatial approach
to the modelling of complex geographic systems that are affected by physical factors
(e.g. land use and accessibility) and individuals (e.g. organizations and human
objects) (Torrens and Benenson 2005).
There are as yet no published studies on the integration of both techniques as a
planning tool for implementing the initiative of sustainable land use. Sustainable
land-use planning generally requires the analysis of a vast array of spatial data. It
needs a set of tools for the exploration, design, modification, illustration, and
evaluation of alternative planning scenarios (Henton and Studwell 2000).
This paper will examine the integration of CA and ABM as a planning tool for
managing residential development. The crucial part of this model is to define agents’
behaviours based on sustainable development strategies. The efficiency criteria in
using land resources are adopted to alleviate land-use conflicts in rapidly growing
cities. This bottom-up approach is well adapted to the simulation of the interactions
and negotiations of different stakeholders. It can provide a useful spatial
exploratory tool for comparing various development options and evaluating the
potential impacts of implementing certain land-use policies.

2. Study area and data
The study area is situated in the Haizhu district of Guangzhou, a rapidly growing
city in the Pearl River Delta, China. Unprecedented land-use changes have been
witnessed in the region in the last two decades (Li and Yeh 2004). The land-use
changes are associated with many environmental problems, such as agricultural land
loss, urban sprawl, and soil erosion (Yeh and Li 1999). In particular, urban
expansion has triggered the loss of a large amount of agricultural land in the Pearl
River Delta.
The spatial information for the proposed integrated CA and ABM model is
obtained using remote sensing and GIS data. The common land-use types in this
study area include urban land, farmland, forest, orchard, and water. GIS are used to
provide the spatial information related to land-use changes. This type of spatial

Sustainable development strategies 23

information includes the maps of planning schemes, land price, land use, and public
facilities (e.g. hospitals, schools, and parks) (figure 1). Additional information (e.g.
age and income) is also obtained from the statistical yearbooks of Guangzhou and
the Fifth National Censu. The above information is used as the inputs to the
modelling and the basis to define agents’ properties.

3. Integrated CA and ABM planning model
The proposed model consists of three components, GIS, CA, and ABM, for
simulating planning options related to residential development (figure 2). The GIS
component is used to provide the inputs to simulation and model calibration. The
CA component is to reflect neighbourhood influences of physical factors. The ABM
component provides a flexible tool to address the interactions between various
stakeholders that affect residential development. The following sections describe the
detailed procedures in implementing this planning model.

3.1 Retrieving physical factors using GIS
3.1.1 Land use. Land use is one of the important factors in urban simulation.
Agents have different decision behaviours with regard to land-use types. For

Figure 1. Spatial information as the inputs to the simulation.

24 X. Li and X. Liu

Figure 2. Planning model by the integration of cellular automata, agent-based modelling
and GIS.

example, a resident agent has a preference to live in the sites surrounded by a large
area of green land (e.g. forest and orchard) and water, instead of densely developed
land.
3.1.2 Land price. Land price plays a key role in affecting urban development,
especially residential development. Land price is correlated to housing price, which
is a major concern for a potential home buyer. Residents’ financial status determines
their location preferences in buying a home. High-income residents choose locations
of high housing prices to live, while low-income residents choose places of low
housing prices.
3.1.3 Surrounding environment. The attraction of a site for urban development is
related to its surrounding living environment. The surrounding environment is
measured using two indicators, the percentage of green land and the percentage of
water in the neighbourhood. These are calculated using a moving 969 window in


classified satellite images. Finally, the utility (attraction) of a site related to this
amenity is obtained using the following equation:
Benv ðiÞ~ 1 Gpercent ðiÞz 1 Wpercent ðiÞ 0ƒGpercent ðiÞzWpercent ðiÞƒ1
2 2 ð1Þ
where Benv(i) is the utility of the surrounding environment, and Gpercent(i) and
Wpercent(i) are the percentages of green land and water at location i, respectively.
These two variables are treated with equal importance, since there is no prior
knowledge.

3.1.4 Accessibility. Accessibility is related to its geographical location (e.g.
distance to roads and town centres) and the conditions of road networks. A site
will be more likely to develop if it is easily accessed. The utility (benefits) of a site
related to the accessibility is represented as follows:
1 : 1 : 1 :
Baccess ðiÞ~ e{b1 Droad ðiÞ z e{b2 Dexpress ðiÞ z e{b3 Dcentre ðiÞ ð2Þ
3 3 3
where Baccess(i) is the utility related to accessibility at location i; the variables
Droad(i), Dexpress(i), and Dcentre(i) are the Euclidean distances to roads, expressways,
and urban centres, respectively; and b1, b2, and b3 are the decay coefficients for
these variables. The same weight (1/3) is also applied to all these variables for
simplicity.

3.1.5 General public facilities. A site will be more likely to develop if it is closer to
facilities, such as hospitals, gardens, commercial centres, and entertainment centers.
Therefore, the utility of a site in terms of facility provision can be represented as
follows:
1 : 1 : 1 : 1 :
Bfacil ðiÞ~ e{b1 Dhospital ðiÞ z e{b1 Dgarden ðiÞ z e{b1 Dcommercial ðiÞ z e{b1 Dentertainment ðiÞ ð3Þ
4 4 4 4
where Bfacil(i) is the utility related to the provision of public facilities at location i,
such as hospitals, gardens, commercial centres and entertainment; and the variables
Dhospital(i), Dgarden(i), Dcommercial(i), and Dentertainment(i) are the Euclidean distances
to these facilities, respectively. The same decay coefficient of b1 in equation (2) is
used, since these facilities are mainly accessed by roads. All these variables are
treated with the same weight (1/4) in the calculation.

3.1.6 Education benefits. Education is an important attraction factor to home
buying. A Euclidean distance function can also be used to represent the accessibility
of a location to education facilities (e.g. schools and libraries). More education
benefits can be achieved if the location is closer to these facilities. This utility is
estimated as follows:
1 : 1 :
Bedu ðiÞ~ e{b1 Dschool ðiÞ z e{b1 Dlibrary ðiÞ ð4Þ
2 2
where Bedu(i) is the utility related to the provision of educational facilities in terms of
schools and public libraries at location i; and the variables Dschool(i) and Dlibrary(i)
are the Euclidean distances to these facilities, respectively. The same decay
coefficient of b1 in equation (2) is used, since these facilities are mainly accessed
by roads. The same weight (1/2) is also used for these two variables.

26 X. Li and X. Liu

3.2 ABM component
This study assumes that land-development patterns are affected by three types of
agents—government agents, developer agents, and resident agents. Government
agents have no location attributes, since their influences are uniform for the whole
region. It is also difficult to define the exact locations for developer agents. The main
objective of developer agents is to make the profit as high as possible. Resident
agents are movable, and their decisions to reside in a place can influence land-
development patterns. The resident agents are randomly located in the initial stage.
They can move into a place for residency according to their financial status and the
site attributes. However, they do not actually move around the landscape with every
time step for reducing computation time.
3.2.1 Implementing the initiatives of sustainable development by government
agents. The strategies of sustainable development can help to develop methods on
how to grow with harmony with the environment (Markandya and Richardson
1992). Some principles related to sustainable development can be incorporated in
formulating land-development plans. In this model, these principles are defined as
follows:
N Land demand is a factor for promoting regional economic development.
However, a mechanism is required to ensure the proper distribution of land
consumption at different planning stages.
N Land development should avoid the use of good-quality agricultural land as
much as possible. This can be realized by incorporating the criterion of spatial
efficiency.
N Negotiations are necessary to achieve practical solutions to land-use conflicts.
In this model, government agents will consider spatial and temporal efficiencies in
using land resources. The first step is to incorporate the criterion of spatial efficiency
for government agents. Government agents will decide if an application for land
development is successful or not, according to a number of factors. Existing land use
is a major factor in determining land-use conversion. Different land uses will have
different values of approval probability for land development. For example, land
development is not allowed in ecological sensitive areas. The probability for land
development in wetland areas or mountainous areas is much lower. The approval
probability is also related to existing plan schemes. It is more likely that an
application can be approved if there are no conflicts with existing land-use plans. In
this study, the approval probability for government agents is defined to represent
various planning objectives.
The second step is to implement the equity of using land resources in a temporal
dimension by government agents. The temporal efficiency criterion is to produce the
maximum benefits from the use of land resources across generations. Tietenberg
(1992) proposes a method to realize efficient allocation of depletable resources and
maintain the equity between generations in a time dimension. It assumes that the
demand curve for a depletable resource is linear and stable over time (figure 3).
Thus, the inverse demand curve in year t can be written as follows:
Dt ~aÀbqt ð5Þ
where a and b are the intersect and slope of the curve of the marginal benefit,
respectively, and qt is the proposed amount of resource consumed in each period t.


Figure 3. Maximizing the total net benefit derived from the use of land resources.

Then, the total benefit BT from extracting an amount qt in year t is the integral of
equation (5):
BT ~I ða{bqt Þ dqt
ð6Þ
~aqt {bq2 =2:
t

The marginal cost of extracting that resource is further assumed to be a constant
c. The total cost CT of extracting the amount qt is:
CT ~cqt ð7Þ
where c is a constant.
Then, the efficient allocation of a resource over n years should satisfy the
following maximization condition (Tietenberg 1992):
" #
XÀn Á. t{1
Xn
Max aqt {bq2 2{cqt ð1zrÞ zl Q{
t qt ð8Þ
qt
t~1 t~1

where Q is the total available amount of the resource supplied, and r is the discount
rate.
When the factor of population growth is considered, the maximization is revised
by solving the following equations (Yeh and Li 1998):
.
ða{bqt =Pta {cÞ ð1zrÞt{1 {l~0 t~1, Á Á Á , n
X
n ð9Þ
Q{ qt ~0
T~1

28 X. Li and X. Liu

where Q is the total available amount of land resource supplied. Pta is the projected
additional population in period t.
3.2.2 Making profits for developer agents. The main objective of property
developers is to achieve a certain amount of profit above expectations. The
following equation is used for the assessment of development potentials:
Dt t t t
profit ði Þ~Hprice ðiÞ{Lprice ðiÞ{Dcost ðiÞ ð10Þ

where Dt t
profit ðiÞ represents the investment profit at location i, Hprice ðiÞ is the housing
t t
price, Lprice ðiÞ is the land price, and Dcost ðiÞ is the development cost.
The development probability related to developer agents can thus be represented
as follows:
Dt
profit ðiÞ{Dtprofit
Pt
developer ðk,iÞ~ ð11Þ
Dmprofit {Dtprofit
where Pt developer ðk,iÞ is the development probability related to developer agents,
Dtprofit is a threshold value, and Dmprofit is the maximum value of the investment
profit.
3.2.3 Location choice by resident agents. The behaviours of resident agents are
determined by two types of factors: the location factors and agents’ status factors
(e.g. income and family size). These factors are reflected in a combined utility
function, which is defined to assess the value of residency of each site for a resident
agent. The main objective of resident agents is to maximize the following utility
function as much as possible in site selection. This combined utility function of
location (i) for agent k can be represented as follows:
U ðk,iÞ~wprice :Bprice ðiÞzwenv :Benv ðiÞ
ð12Þ
zwaccess :Baccess ðiÞzwfacil :Bfacil ðiÞzwedu :Bedu ðiÞzetij
where wprice + wenv + waccess + wfacil + wedu51; the variables of Bprice(i), Benv(i),
Baccess(i), Bfacil(i), and Bedu(i) are the utilities (benefits) related to land price,
surrounding environment, accessibility, general facilities, and education for the
development of location (i); the parameters of wprice, wenv, waccess, and wedu are the
preferences (weights) for these variables, respectively; and the term of etij is a
stochastic variable which accounts for unexplained factors in site selection.
These weights are dependent on agents’ status, such as income and family size. In
this study, resident agents are classified into a number of categories according to
their attributes, such as income and family size. The weights are then determined for
each group of agents according to Saaty’s pairwise comparison procedure (Eastman
1999). The heterogeneity of resident agents is reflected by the weights in this
combined utility function.
The probability of selecting a site is estimated according to the utility function.
For resident k, the probability of location (i) to be selected is equal to the utility
probability that the utility value at that location is greater than or equal to those at
other locations (McFadden 1978):

0 expðU ðk,iÞÞ
Pt
resident ðk,iÞ~PðU ðk,iÞ§U ðk,i ÞÞ~ P ð13Þ
expðU ðk,iÞÞ
k


3.2.4 Interactions between government agents, developer agents and resident
agents. Although the initial approval probability is determined by governments, it
is subject to changes with the influences from residents and property developers. The
following equation can be used to represent this type of interaction between
government agents, developer agents, and resident agents in affecting the
development probability of a cell (i):

Pt ðiÞ~Pt{1 ðiÞzg:DP1 zzh:DP2 if Pt ðiÞw1, then Pt ðiÞ~1
gov gov gov gov ð14Þ

where the initial value of Pt ðiÞ is P0 ðiÞ, which is related to land-use types; the
gov gov
coefficients g and h are the total numbers applied for development at cell (i) by
resident agents and developer agents, respectively; and DP1 and DP2 are the
incremental probability for each application by developer agents and resident
agents, respectively.

3.3 Integrating CA with ABM
CA are an important component in this integrated model. In this study, the
development probability related to the local interactions of physical factors is
estimated using a logistic-CA model (Wu 2002):
1
Pt ðiÞ~
ca
! :cont ðiÞ:Vt ðiÞ ð15Þ
P
1zexp { D0 z Dh : xh ð i Þ
h

where Pt ðiÞ is the development probability of location (i), determined by the
ca
neighbourhood function, xh is the hth spatial variable, D0 is a constant, and Dh is the
weight of the hth variable. The function of con(i) is a combined physical constraint,
and V(i) is the percentage of developed cells in the neighbourhood.
The final decision is made according to a joint development probability, which
reflects the combined effects of human factors (government agents, resident agents,
and developer agents) and environmental factors. The joint probability is
represented as follows:
Pt ~A:Pt
i
: t : t : t
resident ðk,iÞ Pdeveloper ðk,iÞ Pgov ðiÞ Pca ðiÞ ð16Þ

where A is an adjusted coefficient.
The Monte Carlo method is used to determine the final selection of a location for
development (Wu and Webster 1998). The final land-use conversion is determined
by comparing the development probability with a random variable:

Development, Pt wRandðÞ
i
Stz1 ðiÞ~ ð17Þ
Non À development, Others
where Rand() is a random variable ranging from 0 to 1.
This simulation is to determine which sites will be developed based on the
combined assessment from various individuals. The final decision is based on the
joint probability calculated by equation (16). This equation consists of four
components of interactions for determining land-use conversion. The first three
components are obtained by the ABM method, and the last component is obtained
by the CA method.

30 X. Li and X. Liu

4. Model implementation and results
4.1 Programming
The prototype of this integrated model is developed using the Visual Basic and
ArcObjects component of ARCGIS. The use of ArcObjects can allow this model to
access the spatial data in a GIS database directly. The computation will be too
intensive if the model is implemented using common pure agent-modelling shells
(e.g. the swarm package). They have difficulties in coupling with GIS and CA
directly. Figure 4 shows the interface of this proposed prototype. Agents are only
implied through model results for simplicity.

4.2 Preparing spatial variables and determining model coefficients
The original layers of land-use types, land price, living environment, accessibility,
general public facilities, and education were transformed into a raster format with
the resolution of 1006100 m for the programming. The utility (benefit) of each
spatial variable for land development was estimated using GIS.
Some coefficients should be estimated before calculating the utilities in equations
(2)–(4). The values of b1, b2, and b3 in equation (2), which are related to transport
conditions, can be estimated according to empirical traffic data. It is assumed that a
transport tool with a larger traffic density will have a larger area of influence (a
smaller value for the decay coefficient). For example, expressways which have larger
traffic densities will be assigned smaller values for the coefficient. The following
equation can be used to represent this relationship:

b1 =b2 ~fexpress froad ð18Þ
where froad and fexpress are the average traffic densities for roads and expressways,
respectively.
The same method can be applied to the estimation of b3. If there are z1 number of
roads and z2 number of expressways connected to urban centres, the equation
becomes:
À Á
b1 =b3 ~ z1 :froad zz2 :fexpress froad ð19Þ

Figure 4. Estimating population growth of Guangzhou in 1990–2010.


Table 1 provides estimates of froad and fexpress according to statistical data. When
b1 is set to 0.00100, b2 and b3 become 0.00023 and 0.000125, respectively, according
to the above equations. The same value of b1 is also used in equations (3) and (4).
The original spatial variables were normalized into the range of [0, 1] before they
were used for the calculation. The values of the incremental probabilities, DP1 and
DP2, were decided by experiments. In this study, DP1 was set to 0.005 and DP2 was
set to 0.1. The coefficients of the CA component were calibrated according to
logistic regression (Wu 2002). Landsat TM images dated on 30 December 1995 and
13 June 2004 were used to obtain training data about actual land-use conversion.
Table 2 lists the coefficients of the logistic-CA model in equation (15) based on the
regression analysis.
CA and ABM are based on discrete time steps in simulating urban dynamics. Too
few time steps will neglect local interactions and cannot allow spatial details to
emerge (Yeh and Li 2006). An increase in the number of time steps can help to
generate more accurate simulation results. In many applications, 200–300 time steps
are required to guarantee sufficient temporal accuracy for simulation (Yeh and Li
2006). In this study, 1500 time steps are adopted to simulate land development in the
period of 1995–2010. Therefore, 100 time steps correspond to 1 year in the
simulation.

4.3 Implementing sustainable use of land resources by government agents
The initiatives for sustainable use of land resources should be implemented by
government agents, who determine the proper distribution of land consumption
across different planning periods. The first step is to estimate the population growth
before the appropriate land consumption can be obtained for each planning period.
A regression model was established for estimating the population growth using
empirical data (table 3):

Y ðTz1Þ~673300:46e0:0166T ð20Þ
where Y(T + 1) is the predicted population in year T + 1 based on the initial
population in 1990.

Table 1. Traffic densities of roads and expressways according to statistical dataa.

Length Traffic Average Traffic Density
Transport types (km) (1000 persons) (1000 persons km21)
Roads 4637.2 147 330 31.77
Expressways 382.8 52 310 136.7
a
Sources: Guangdong Statistical Yearbook (2000).

Table 2. Coefficients of the logistic-CA model.

D1 D2 D3 D4 D5
Distance to Distance to Distance to Distance to Distance to
D0 main centres sub-centres main roads roads expressways
0.625 20.002 0.005 20.009 20.006 0.002

32 X. Li and X. Liu

Table 3. Empirical data about the population growth in the Haizhu district of Guangzhou in
1990–1999.

Year 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
Population 684 887 691 557 702 004 710 153 727 045 738 910 751 486 759 256 763 959 778 984

Figure 5 indicates that the regression model can predict the population growth
satisfactorily. The predicted population was used to calculate the optimized land
consumption for each period according to the modified Tietenberg model. The
available land for development was estimated according to land-use information.
The whole study area was 101.40 km2, of which the urban area made up 34.09 km2 in
1995. There was only 66.38% left for future land supply since 33.62% of the area had
been urbanized. The detailed provision of land consumption at each period in 1995–
2010 was obtained by using equation (9), assuming that 50% of the area could be

Figure 5. Interface of the proposed model.


Table 4. Optimal land consumption for different periods with various discount rates
according to the Tietenberg’s model.

Land consumption (km2)
Year Population growth r50 r50.02 r50.1
1995–2000 63 880 5.08 6.04 7.66
2000–2005 69 420 5.52 5.39 5.08
2005–2010 75 439 6.01 5.18 3.87

urbanized in 2010 (table 4). The number of new urban cells was then determined for
the simulation.

4.4 Defining resident agents’ properties using empirical data
The decision behaviours of resident agents are defined using aggregated census data
because of the lack of detailed information. Some simplification procedures have to
be carried out for obtaining the attributes of resident agents. First, resident agents
should be classified into a few categories so that their properties can be heuristically
defined. The attributes for the aggregated agents are obtained using social and
economic data. This study considers two major attributes, income and household
size, which are obtained from the statistical yearbook of Guangzhou in 2004, and
the Fifth National Census, respectively.
Residents can be classified into three groups by their income—low-income class
(income,9600 RMB year21), middle-income class (9600 RMB year21,income,
60 000 RMB year21), and high-income class (income.60 000 RMB year21). (1 US$
is roughly equivalent to 7.6951 RMB as of 8 May 2007.) They can also be classified
into two groups by household size: without children and with children. Six classes of
residents were obtained using these two attributes. The actual percentages for these
six groups were calculated according to the statistical yearbook of Guangzhou in
2004, and the Fifth National Census (table 5). These percentages were used to create
the actual numbers for various groups of resident agents in the simulation.
Each group of resident agents has distinct behaviours or preferences in the
location choice of residency. In this model, their preferences are reflected by the
weights in the utility function as described in equation (12). The weights were
obtained by using Saaty’s pairwise comparison procedure (Eastman 1999). The
comparison was mainly based on experts’ knowledge and preferences. A higher
value of the weight means that the variable will be treated more importantly. A
matrix can be constructed to indicate the relative importance based on the
comparison. Saaty (1990) proposes a consistency ratio (CR) to examine the
consistency of the matrix. He suggests that the matrix should be re-evaluated if

Table 5. Proportion of each group of resident agents.

Types of resident agent
Household size Without children With children
Income Low Middle High Low Middle High
income income income income income income
Proportion (%) 9 39 9 6 31 6

34 X. Li and X. Liu

Table 6. Weights for different groups of resident agents obtained using Saaty’s method.

Weights
Types of Land Surrounding Public
residents price environment Accessibility facilities Education Total CR
Low income 0.443 0.093 0.206 0.155 0.103 1 0.042
without children
Low income 0.401 0.081 0.154 0.081 0.283 1 0.087
with children
Middle income 0.175 0.379 0.165 0.194 0.087 1 0.057
without children
Middle income 0.220 0.276 0.142 0.140 0.222 1 0.094
with children
High income 0.048 0.526 0.194 0.141 0.091 1 0.072
without children
High income 0.084 0.434 0.171 0.076 0.235 1 0.064
with children

the ratio value is greater than 0.10. Table 6 shows the results of the weights derived
from Saaty’s method.

4.5 Generating planning scenarios
This simulation assumes that each new urbanized cell can accommodate one
resident agent. The total number of resident agents was determined according to the
allowed amount of land consumption. The discount rate r was set to 0.1 for
calculating the optimal distribution of land consumption. The detailed procedures
for generating development alternatives are as follows:
1. Determining the total number of resident agents according to the allowed
amount of land consumption for each planning period.
2. Using the Monto Carlo method to create a resident agent according to the
proportion of various types of residents based on census data (table 3).
3. The development probability related to the local interactions of physical
factors is estimated by the logistic-CA model, which is calibrated using the
classified satellite images.
4. Using equation (12) and table 6 to compute the utility function for this
resident agent.
5. Selecting the locations with the highest utility values and estimating
development probability for these places according to the interactions
described in equation (16).
6. Determining whether the locations of the highest combined probability values
will be developed using the Monto Carlo method. If yes, the location will be
marked and go to step 2 to create a new resident agent. If no, the next site of
the second highest utility value will be evaluated until this existing agent has
been accommodated.
7. This procedure continues until all the required resident agents have been
accommodated.
This model was used to simulate both baseline development scenarios and
planning development scenarios. The baseline scenarios were generated according to
the development trend. Planning scenarios were produced by incorporating the


initiatives of sustainable development in the modelling. The intervention from
government agents is crucial for producing planning scenarios instead of baseline
scenarios. In this study, the intervention was first represented by using the initial
pre-defined approval probability (P0 ) for government agents. There are five
gov
regimes of land development for the simulation, as follows.
4.5.1 Baseline scenario. This simulation is based on the trajectory of past
development. This regime assumes that no sustainable development strategies are
adopted to regulate existing development trends. No spatial and temporal
efficiencies are implemented in this simulation. Table 7 lists the initial pre-defined
approval probability (P0 ) from government agents for this regime. This simulation
gov
can generate the scenario provided that the city continues to develop without any
constraints. Urban planners can compare this baseline scenario with the following
planning scenarios.
4.5.2 Planning scenario 1: compact development. Some government intervention is
implemented by controlling land consumption in the spatio-temporal dimension.
The equity of using land resources is emphasized by properly arranging land-use
conversion at each planning stage according to equation (9). This planning scenario
also adopts a high priority on implementing the spatial efficiency in terms of
compact development by using the initial pre-defined approval probability (P0 )gov
(table 8). Land-use types will not impose restrictions on land development so that
compact patterns can be formulated in the simulation.

Table 7. Initial pre-defined approval probability (P0 ) for simulating baseline
gov
development.
Planning
Existing Urban land Water Farmland Forest Orchard Other
Urban 1.00 0.00 0.00 0.00 0.00 0.00
land
Water 0.02 0.00 0.00 0.00 0.00 0.00
Farmland 0.60 0.00 0.15 0.20 0.20 0.30
Forest 0.65 0.00 0.20 0.25 0.25 0.35
Orchard 0.68 0.00 0.20 0.30 0.20 0.40
Other 0.90 0.00 0.40 0.40 0.45 0.60

Table 8. Initial pre-defined approval probability (P0 ) for simulating planning scenario
gov
1: compact development.
Planning
Urban 1.00 0.00 0.00 0.00 0.00 0.00
land
Water 1.00 0.00 0.00 0.00 0.00 0.00
Farmland 1.00 0.00 0.00 0.00 0.00 0.00
Forest 1.00 0.00 0.00 0.00 0.00 0.00
Orchard 1.00 0.00 0.00 0.00 0.00 0.00
Other 1.00 0.00 0.00 0.00 0.00 0.00

36 X. Li and X. Liu

4.5.3 Planning scenario 2: farmland protecting development. This regime is to
ensure the equity of using land resources across generations and avoid the
encroachment on agricultural land as well. The former is realized by arranging the
proper land consumption at different planning stages, and the latter is implemented
by using the initial pre-defined approval probability (P0 ) for government agents
gov
(table 9). A very small probability will be given to the development of agricultural
land. This planning scenario can allow land resources to be used more efficiently
than the baseline pattern.
4.5.4 Planning scenario 3: green-land protecting development. This regime pays
special attention to implementing the concept of ‘garden cities’ while land
consumption is also constrained by the equity criterion. It addresses the growing
concern for a better living environment after residents have secured their basic
housing demand. This is a further development stage compared with planning
scenario 2. Table 10 shows the initial pre-defined approval probability (P0 ) which
gov
imposes extreme restrictions on converting green land and orchard land into
residential use.
4.5.5 Planning scenario 4: housing-demand development. This scenario just
completely satisfies housing demand from resident agents at each location without
government controls. Land-use types do not impose any restrictions on land
development. Therefore, all the initial probability values are set to 1 for this regime.
Figure 6 is the outcome from the simulation of baseline patterns in the study area
in 2000–2010 according to historical growth. Planning scenarios can be simulated by

gov
2: farmland protecting development.
Planning
Urban 0.00 0.00 0.00 0.00 0.00 0.00
land
Water 0.00 0.00 0.00 0.00 0.00 0.00
Farmland 0.10 0.00 0.00 0.05 0.05 0.10
Forest 0.75 0.00 0.25 0.30 0.30 0.35
Orchard 0.80 0.00 0.05 0.35 0.25 0.40
Other 0.95 0.00 0.15 0.40 0.45 0.70

gov
3: green-land protecting development.
Planning
Urban 0.00 0.00 0.00 0.00 0.00 0.00
land
Water 0.00 0.00 0.00 0.00 0.00 0.00
Farmland 0.65 0.00 0.35 0.20 0.20 0.50
Forest 0.10 0.00 0.05 0.00 0.00 0.10
Orchard 0.15 0.00 0.05 0.00 0.00 0.15
Other 0.95 0.00 0.50 0.30 0.35 0.70


Figure 6. Simulation of baseline development patterns of Guangzhou based on historical
trends.

incorporating the criteria of sustainable development and properly modifying the
parameters of this agent-based model. Figure 7(a) is to simulate compact
development, which can reduce the energy consumption in transportation. It is
also able to reduce the encroachment on agricultural land by introducing
government intervention (figure 7(b)). However, this scenario may result in some
fragmented patterns. Planning scenario 3 (green-land protecting development)
emphasizes the preservation of green land and orchard land (figure 7(c)). Planning
scenario 4 (housing-demand development) is associated with significant dispersed
development patterns, since it just satisfies housing demand from resident agents
(figure 7(d)).

38 X. Li and X. Liu

Figure 7. Simulation of planning development patterns of Guangzhou in 2010. (a) Compact
development. (b) Farmland protecting development. (c) Greenland protecting development.
(d) Housing-demand development.


4.6 Metrics for the comparisons among the simulated scenarios
Statistical comparisons among the simulated scenarios were carried out for
providing planning implications according to a number of metrics. These metrics,
which will indicate the gain and loss of land development, include the indicators of
compactness, development suitability gain, agricultural suitability loss, green-land
loss, and farmland loss. The first two indicators are related to the gain, and the last
three are related to the loss of land development.
The compactness of land development can be calculated according to the average
comparison between the perimeter of a developed cluster and the standard perimeter
of the circle which has the same area (Li and Yeh 2004). The index can be
represented using the following equation:
sffiffiffiffiffiffiffiffiffiffiffiffi,X
X
CI~ Sj Pj ð21Þ
j j

where CI is the value of the compactness index, and Sj and Pi are the area and
perimeter of the developed cluster (polygon) jj. It is obvious that land development
with average narrow shapes or dispersed development patterns will have low values
for the index.
The developed sites should have higher values of development suitability and
lower values of agricultural suitability for spatial efficiency (Li 2005). Therefore, the
development suitability gain can be obtained by summing up the urban development
suitability for all the developed cells:
X
Dgain ~ Sur ðiÞ ð22Þ
i

where Dgain is the development suitability gain, and Sur(i) is the urban development
suitability at cell i where land development takes place.
The agricultural suitability loss can be calculated using this similar method:
X
Aloss ~ Sag ðiÞ ð23Þ
i

where Aloss is the agricultural suitability loss, and Sag(i) is the agricultural suitability
at cell i where land development takes place.
The last two indicators are to sum up the total amounts of green-land loss and
farmland loss for each scenario. Table 11 is the analysis results from these five
metrics. Figure 8 shows the gain of land development for these simulated scenarios.
The baseline scenario and planning scenario 4 (housing-demand development) have
lower values of the compactness, although they have higher values of development
suitability gain.
Figure 9 further displays the loss related to land development for these scenarios.
The baseline scenario and planning scenario 4 (housing-demand development) have
larger values for the indicators of agricultural suitability loss, green-land loss, and
farmland loss. Planning scenario 1 (compact development) and planning scenario 3
(green-land protecting development) have lower values for the green-land loss.
Planning scenario 2 (farmland protecting development) has the lowest value for
farmland loss.
The final assessment of these scenarios is based on a linear combination of these
five indicators. The values from these five indicators should be normalized into the

40 X. Li and X. Liu

Table 11. Comparisons among the simulated scenarios using various metrics.

Development Agricultural
suitability suitability Green-land Farmland
Development Compactness gain loss loss loss
patterns Year (61023) (6103) (6103) (6106m2) (6106m2)
Baseline scenario 2000 19.1 46.3 72.5 1.4 1.9
2005 19.5 73.7 109.2 4.9 3.1
2010 19.7 91.2 135.8 6.5 4.9
Planning scenario 2000 20.1 44.3 60.1 1.0 1.2
1: compact 2005 22.3 71.9 95.7 2.2 2.5
development 2010 23.4 89.6 106.5 3.0 3.6
2: farmland 2005 21.2 68.4 89.8 4.7 0.8
protecting 2010 21.0 85.2 99.6 6.2 1.1
development
3: green-land 2005 22.4 66.4 103.5 2.5 1.5
protecting 2010 23.8 82.4 115.8 3.4 2.2
development
Planning 2000 18.7 54.9 79.0 3.0 3.1
scenario 4: 2005 18.9 81.4 112.0 7.2 4.3
housing-demand 2010 19.3 102.4 148.6 10.5 5.2
development

Figure 8. Gain of the simulation scenarios.


Figure 9. Loss of the simulation scenarios.

42 X. Li and X. Liu

range of 0–1 before the use of the linear combination. The normalization is different
for these two types of factors: gain (higher scores are better; e.g. compactness and
development suitability gain) and loss (lower scores are better, e.g. agricultural
suitability loss, green-land loss, farmland loss).
The gain factors are as follows:
x{Min
x0 ~ ð24Þ
Max{Min

The loss factors are as follows:
Max{x
x0 ~ ð25Þ
Max{Min
where x is the original data, Max and Min are the maximum and minimum values,
x9 is the normalized value.
Figure 10 shows the final result from the linear combination of these five
normalized metrics. The same weight is applied to each indicator, since there is no
prior knowledge. It is also clear that the baseline scenario and planning scenario 4
(housing-demand development) have a poorer performance from the assessment.
Planning scenario 1 (compact development) and planning scenario 3 (green-land
protecting development) have a better performance according to the combined
indicator.

Figure 10. Final assessment of the simulated scenarios using a linear combination of five d
metrics.


5. Conclusion
This paper has demonstrated that agent-based modelling techniques can be further
extended to the simulation of development alternatives. The strategies for
sustainable development are incorporated in the modelling by properly defining
agents’ behaviours. Spatial efficiency of using land resources is implemented by
selecting suitable sites for development according to planning objectives. The
efficient allocation of land resources over the temporal dimension is realized by
defining decision behaviours of government agents.
Sustainable land development is a complex issue which involves negotiations and
compromises of various stakeholders. Local interactions from this integrated model
are essential for dealing with these complex situations. In this study, the
heterogeneity of agents is reflected by using different sets of weights according to
GIS data. Development plans can be generated to implement sustainable
development initiatives through the interactions between government agents,
developer agents, and resident agents. Since several compromises have been
adopted in the simulation, the simulated alternatives should be more realistic and
practical for planning practice.
Five scenarios of land development have been simulated by using this proposed
model. The effects of these scenarios are compared according to a number of
metrics, such as compactness, development suitability gain, agricultural suitability
loss, green-land loss, and farmland loss. These metrics are used to quantify the gain
and loss of land development by providing planning implications. The comparison
can identify the best scenario for a certain planning objective. For example, the
baseline scenario and planning scenario 4 (housing-demand development) have
lower values for gain, and larger values for loss. However, planning scenario 2
(farmland protecting development) has the lowest values of farmland loss.
A combined index can be devised to take account of all these five metrics. This
index also indicates that the baseline scenario and planning scenario 4 (housing-
demand development) have a poorer performance, whereas planning scenario 1
(compact development) and planning scenario 3 (green-land protecting develop-
ment) have a better performance for land development.
Like other agent-based models, this model also involves many parameters that
have critical effects on the simulation results. Although some calibration procedures
have been carried out, a finer tuning of this model still requires considerable effort.
Further studies should be carried out in developing the methods of calibrating
agents’ properties in a more consistent way. There is still a general lack of detailed
spatial information that can be used to define the decision behaviours of agents for
Chinese cities. More detailed resident agents can be defined to improve simulation
performance when such spatial information is available.

Acknowledgements
This study was supported by the National Outstanding Youth Foundation of China
(Grant No. 40525002), the National Natural Science Foundation of China (Grant
No. 40471105), and the PhD development program from the Ministry of Education
of China (Project No. 20040558023).

References
BASU, N. and PRYOR, R.J., 1997, Growing a Market Economy, Technical Report SAND-97-
2093 (Albuquerque, NM: Sandia National Laboratories).

44 X. Li and X. Liu

BATTY, M. and XIE, Y., 1994, From cells to cities. Environment and Planning B: Planning and
Design, 21, pp. 531–548.
COURDIER, R., GUERRIN, F., ANDRIAMASINORO, F.H. and PAILLAT, J.M., 2002, Agent-based
simulation of complex systems: application to collective management of animal
wastes. Journal of Artificial Societies and Social Simulation, 5. Available online at:
http://jasss.soc.surrey.ac.uk/5/3/courdier.html (accessed 12 January 2006).
EASTMAN, J.R., 1999, Multi-criteria evaluation and GIS. In Geographical Information
Systems, P.A. Longley, M.F. Goodchild, D.J. Maguire and D.W. Rhind (Eds), pp.
493–502 (New York: Wiley).
GILBERT, N. and CONTE, R., 1995, Artificial Societies: the computer simulation of social life.
UCL Press.
HENTON, D. and STUDWELL, K., 2000, Informed Regional Choices: How California’s Regional
Organizations are Applying Planning and Decision Tools (Oakland, CA: California
Center for Regional Leadership).
JANTZ, P., GOETZ, S. and JANTZ, C., 2005, Urbanization and the loss of resource lands in the
Chesapeake bay watershed. Environmental Management, 36, pp. 808–825.
LI, X. and YEH, A.G.O., 2000, Modelling sustainable urban development by the integration
of constrained cellular automata and GIS. International Journal of Geographical
Information Science, 14, pp. 131–152.
LI, X. and YEH, A.G.O., 2001, Zoning land for agricultural protection by the integration of
remote sensing, GIS and cellular automata. Photogrammetric Engineering Remote
Sensing, 67, pp. 471–477.
LI, X. and YEH, A.G.O., 2004, Analyzing spatial restructuring of land use patterns in a fast
growing region using remote sensing and GIS. Landscape and Urban Planning, 69, pp.
335–354.
LI, X., 2005, A four-component efficiency index for assessing land development using
remote sensing and GIS. Photogrammetric Engineering Remote Sensing, 71, pp.
47–57.
LOIBL, W. and TOETZER, T., 2003, Modelling growth and densification processes in suburban
regions—simulation of landscape transition with spatial agents. Environmental
Modelling Software, 18, pp. 553–563.
MARKANDYA, A. and RICHARDSON, J., 1992, The economics of the environment: an
introduction. In Environmental Economics, A. Markandya and J. Richardson (Eds),
pp. 7–25 (London: Earthscan).
MCFADDEN, D., 1978, Modelling the choice of residential location. In Spatial Interaction
Theory Planning Models, A. Karlqvist, L. Lundqvist, F. Snickars and J. Weibull
(Eds), pp. 75–96 (Amsterdam: North Holland).
NIJKAMP, P., OUWERSLOOT, H. and RIENSTRA, S.A., 1997, Sustainable urban transport
systems: an expert-based strategic scenario approach. Urban Studies, 34, pp. 693–712.
SAATY, T.L., 1990, The Analytic Hierarchy Process: Planning, Priority Setting, Resource
Allocation (Pittsburgh, PA: University of Pittsburgh).
STATISTICAL BUREAU OF GUANGZHOU 2000, Guangzhou Statistical Yearbook (China
Statistics Press), pp. 187–190.
TIETENBERG, T., 1992, Environmental and Natural Resource Economics (New York:
HarperCollins).
TORRENS, P.M. and BENENSON, I., 2005, Geographic automata systems. International Journal
of Geographical Information Science, 19, pp. 385–412.
WHITE, R. and ENGELEN, G., 1993, Cellular automata and fractal urban form: a cellular
modelling approach to the evolution of urban land use patterns. Environment and
Planning A, 25, pp. 1175–1199.
WU, F., 2002, Calibration of stochastic cellular automata: the application to rural–urban land
conversions. International Journal of Geographical Information Science, 16, pp.
795–818.


WU, F. and WEBSTER, C.J., 1998, Simulation of land development through the integration of
cellular automata and multicriteria evaluation. Environment and Planning B, 25, pp.
103–126.
YEH, A.G.O. and LI, X., 1998, Sustainable land development model for rapid growth areas
using GIS. International Journal of Geographical Information Science, 12, pp. 169–189.
YEH, A.G.O. and LI, X., 1999, Economic development and agricultural land loss in the Pearl
River Delta, China. Habitat International, 23, pp. 373–390.
YEH, A.G.O. and LI, X., 2006, Errors and uncertainties in urban cellular automata.
Computers, Environment and Urban Systems, 30, pp. 10–28.
ZANDERA, P. and KACHELE, H., 1999, Modelling multiple objectives of land use for
¨
sustainable development Agricultural Systems, 59, pp. 311–325.

Embedding sustainable development strategies in agent

Embedding sustainable development strategies in agent

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Embedding sustainable development strategies in agent (20)

Recently uploaded (20)

Embedding sustainable development strategies in agent