You are on page 1of 23

This article was downloaded by: [Universiteit Twente]

On: 25 September 2014, At: 06:33


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Geographical


Information Science
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tgis20

Modeling-in-the-middle: bridging the


gap between agent-based modeling and
multi-objective decision-making for
land use change
a b c
Christopher Bone , Suzana Dragicevic & Roger White
a
Department of Biology , University of Alaska Anchorage ,
Anchorage, AK, USA
b
Department of Geography , Simon Fraser University , Burnaby,
British Columbia, Canada
c
Department of Geography , Memorial University of
Newfoundland , St. John's, Newfoundland, Canada
Published online: 04 Nov 2010.

To cite this article: Christopher Bone , Suzana Dragicevic & Roger White (2011) Modeling-in-the-
middle: bridging the gap between agent-based modeling and multi-objective decision-making for
land use change, International Journal of Geographical Information Science, 25:5, 717-737, DOI:
10.1080/13658816.2010.495076

To link to this article: http://dx.doi.org/10.1080/13658816.2010.495076

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
Content) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at http://www.tandfonline.com/page/terms-
and-conditions
Downloaded by [Universiteit Twente] at 06:33 25 September 2014
International Journal of Geographical Information Science
Vol. 25, No. 5, May 2011, 717737

Modeling-in-the-middle: bridging the gap between agent-based


modeling and multi-objective decision-making for land use change
Christopher Bonea*, Suzana Dragicevicb and Roger Whitec
a
Department of Biology, University of Alaska Anchorage, Anchorage, AK, USA; bDepartment of
Geography, Simon Fraser University, Burnaby, British Columbia, Canada; cDepartment of
Geography, Memorial University of Newfoundland, St. Johns, Newfoundland, Canada
(Received 26 February 2010; final version received 16 April 2010)
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

A spectrum of methods exists for investigating and providing solutions for land use
change. These methods can be broadly categorized as either top-down or bottom-up
approaches according to how land use change is modeled and analyzed. Although there
has been much research in recent years advancing the use of these techniques for both
theoretical and practical applications, integrating top-down and bottom-up approaches
for enhancing land use change modeling has received minimal attention.
The objective of this study is to address this gap in the literature by bridging the
bottom-up simulation of agent-based modeling and the top-down analytical capabilities
of multi-objective decision-making by means of a heuristic modeling approach called
reinforcement learning (RL). A model is developed in which computer agents represent-
ing households and commercial enterprises select locations to inhabit based on popula-
tion densities and attractivity preferences. The land use change resulting from these
dynamics is evaluated by a set of agents representing different stakeholders who are
embedded with RL algorithms that allow them to influence the land use change process
so that their objectives are addressed. The results demonstrate that bridging bottom-up
and top-down models leads to negotiated land use patterns in which the desires and
objectives of all individuals are constrained by behaviors of others. This study suggests
that a movement toward a modeling-in-the-middle approach is desirable to incorporate
the real yet conflicting forces that shape land use change and that are rarely considered in
unison.
Keywords: agent-based modeling; multi-objective decision-making; reinforcement
learning; land use change

1. Introduction
Land use change takes place at the confluence of bottom-up processes and top-down
constraints that lead to the emergence of landscape patterns over time (Johnson 2002).
With regard to the growth of cities, bottom-up processes are driven by the desires of current
and potential residents to inhabit specific locations, the desires of commercial, industrial, and
institutional interests to be situated in strategic locations, as well as the dynamics that such
desires entail (Jacobs 2002). Meanwhile, top-down constraints are manifest in policies that
are intended to direct landscape dynamics in a fashion that satisfies the objectives of an often
diverse set of stakeholders. As such, the ability to provide effective land use planning

*Corresponding author. Now affiliated with The Resilience and Adaptive Management Group,
University of Alaska Anchorage. Email: afceb@uaa.alaska.edu

ISSN 1365-8816 print/ISSN 1362-3087 online


# 2011 Taylor & Francis
DOI: 10.1080/13658816.2010.495076
http://www.informaworld.com
718 C. Bone et al.

requires methods computational models in particular that can incorporate the interplay
involved in the bottom-up/top-down relationship.
The literature surrounding land use science, urban planning, and regional science
provides distinct classifications for the types of models that can be implemented for assisting
in the understanding and planning of land use change. Malczewskis (2004) classification,
for example, places computational methods under the headings of computer-assisted map
overlay, multi-criteria analysis (MCA), and artificial intelligence (AI) methods. More
detailed classifications have been suggested within these groupings to further distinguish
between the methods belonging to MCA (Jankowski 1995, Herwijnen and Rietveld 1999,
Malczewski 1999, 2006, Chakhar and Martel 2003) and AI (Openshaw and Openshaw
1997) that can be applied for land use planning and other geographical inquiries. However, it
can be argued that a much broader classification exists: those methods that devise land use
patterns for satisfying a host of global objectives (top-down models) and those that derive
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

patterns from the simulation of local land use change dynamics (bottom-up models).
Generally speaking, top-down models operate by algorithmically combining different
criteria or objectives to produce an overarching plan of how an area should be spatially
structured. Such methods are strictly concerned with the pattern of land use change and
include a variety of methods that integrate GIS with some form of analysis of multiple
criteria, objectives, or attributes (Carver 1991, Church et al. 1992, Eastman et al. 1998,
Chakhar and Martel 2003, Feick and Hall 2004, Malczewski 2006). In general, these
methods are advantageous because of their ability to incorporate a wide range of spatial
objectives, to weight objectives based on their importance, and to provide results that
intuitively relate to model inputs (Heywood et al. 1998). More recently, top-down
approaches have improved with the use of generative modeling techniques that are able to
produce hundreds or thousands of potential but different solutions to a given land use
problem. These include the use of generative modeling for attempting to achieve different
land use growth objectives (Ligmann-Zielinska et al. 2008) and for determining optimal site
selection (Xiao et al. 2002). These methods contribute an additional benefit by providing a
spectrum of potential solutions from which the trade-offs between all objectives can be
observed (Xiao et al. 2007). However, the limitation with the top-down approach is that
feedback processes are absent from the estimation of suitable land use patterns, thus ignoring
the processes by which land use patterns arise as a result of the attraction of housing and
commercial, industrial, and institutional activity to each other and to areas with specific
densities. For example, an increase in the number of residents in a specific area of a city will
attract commercial enterprises, which in turn may make the location more desirable as a
place to live, thus attracting still more residents.
Unlike their top-down counterparts, bottom-up models are developed to simulate the
interactions among individual decision-makers (e.g., residents, commercial/institutional
enterprises, developers, etc.) or, at a slightly higher level of abstraction, interactions
among individual land use parcels to simulate the emergence of land use patterns over
time. The absence of a central governing equation changes the focus from the pattern to the
process, or more specifically, to understanding how the process generates the resulting
patterns. Although several modeling approaches may be classified as bottom-up, cellular
automata (CA) and agent-based modeling (ABM) have by far dominated the land use
modeling literature. Both CA (Batty 1997, Clarke et al. 1997, White and Engelen 1997,
Liu et al. 2008, Jantz et al. 2009) and ABM (Benenson 1998, Parker et al. 2003, Xie et al.
2007, Zellner 2007, Filatova et al. 2009, Robinson and Brown 2009) have their roots in other
disciplines, but their use for modeling land use change has received much attention as
landscapes can be discretized into appropriate spatial units which can then experience
International Journal of Geographical Information Science 719

changes in state because of the states of neighboring units (in the case of CA) or because of
the decisions of computer agents that represent the behaviors of different individuals (as is
the case with ABM).
The great strength of bottom-up approaches is that they capture the effects of the
dominant processes involved in generating urban land use patterns, and consequently they
can give very good predictions of land use (Benenson 1998, Portugali 2000, Engelen and
White 2007). However, all such models that aim for any degree of realism include a
stochastic element to represent the variability of the agents and their needs and goals. This
is the case both for explicit ABM and for CA (technically, CA are ABM, with cells being
agents that are spatially fixed; but in land use applications they are better thought of as
models using small aggregates of agents representing people or businesses looking for a
suitable location). The stochastic element ensures realism, but one aspect of realism is that
the land use system has an open future that is, the dynamic process can generate many
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

possible future land use patterns. Most of these will be almost indistinguishable from each
other, but there will usually be several areas that have two or more possible types of future
land use patterns that are qualitatively quite different, and yet each type of pattern has a
significant probability of being realized. These areas represent spatial bifurcations in the
dynamics of the system. Because of their difference, the subsequent evolution of the system
will depend on which of the two types is chosen when the bifurcation point is reached.
Because of the multiple possible outcomes inherent in good bottom-up land use models, they
are often thought to be of limited use for policy and planning: they provide no optimal land
use patterns that can be implemented as a policy goal (Batty and Torrens 2005). But in fact
they are the only reasonable basis for planning because they provide an understanding of the
reality the planner must deal with. So a more reasonable assessment might be that the model
provides a picture of the range of possible land use futures, and the planners job is to assess
the possibilities in terms of their desirability and then, through the implementation of
appropriate plans and policies, guide the system toward one of the more desirable possibi-
lities, or at least away from the less desirable futures.
But the planner is not outside the system. The real system includes numerous higher level
agents like governments and their agencies that are active along with the bottom-level agents
represented explicitly in the model. Some bottom-up models include, in a limited way, some
of the more important effects of these higher level agents on the bottom-up dynamics without
including the agents themselves in the model. For example, some CA land use models
include land use zoning and changes in the transportation infrastructure products of top-
down agents (TDAs) as influences on the land use dynamics (Engelen et al. 2007).
However, in the real system the higher level agents are continually interacting with the
system both with each other, and, especially, with the bottom-up agents (BUAs) and the
effects of those agents actions. Thus, for a better understanding of the system and its
dynamics, the TDAs should be included in the land use model.
The dichotomy between bottom-up and top-down approaches for modeling land use
change has polarized our means of understanding land use processes and our ability to
achieve desirable land use patterns. This research proposes a move toward a modeling-
in-the-middle approach that bridges the gap between bottom-up and top-down modeling for
land use change by simulating how individual behaviors influence the morphology of
landscapes while simultaneously evaluating and improving land use patterns to meet
different objectives. This is achieved by integrating a bottom-up model with BUAs that
simulate the behaviors of individuals, enterprises, and institutions with regard to their desires
for residing in specific locations, and a top-down model with TDAs representing planning
stakeholders who are interested in achieving multiple land use objectives.
720 C. Bone et al.

Previous research has demonstrated the potential for bridging top-down and bottom-up
models, specifically in Castella et al.s (2007) integration of a multi-agent model with a
multiple-goal linear programming approach for modeling sustainable natural resource man-
agement. Although the model presented here shares a similar theoretical approach, its novelty
resides in the nature of top-down and bottom-up integration. The bottom-up model follows the
conventional tendencies of agent-based models as described by Parker et al. (2003). BUAs are
embedded with behaviors that focus their decisions on local scale patterns, such as which land
use parcel is most desirable as a residential site with regard to proximity to retail, or where to
establish retail space to serve the largest consumer catchment. BUAs behave purely in self-
interest and as such are responsible for driving both positive and negative feedbacks within the
land use change process. Conversely, the top-down model represents a multi-objective
decision-making mechanism as outlined by Cohon (1978). However, instead of using a single
equation to guide the multi-objective decision-making process, the objectives are evaluated by
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

TDAs representing different stakeholders. TDAs are external to the bottom-up process as they
evaluate the global outcome of BUA-driven land use change. At the final time step of a given
land use change scenario, each TDA examines the resulting pattern and calculates the degree to
which its objectives are achieved. As a result, it acquires knowledge about how the bottom-up
process of land use dynamics impacts its ability to achieve its own objectives. The scenario is
simulated a number of times, each time providing the TDAs with some additional knowledge
about land use dynamics. After a given number of scenario simulations, the TDAs start to
influence the behaviors of the BUAs by imposing their acquired knowledge on BUA decision-
making. The final result of the model is a set of land use change outcomes that represent a
balance between local interactions and global objectives.
Because the TDAs must learn the behavior of the land use model by observing many
simulation runs, the top-down model does not directly represent the actual behavior of real
TDAs who have only one history to learn from. This lack of correspondence is due to the
limitations of current learning algorithms, which are far less efficient than those used by the
human brain. But the final learned behavior of the TDAs should be a reasonable approxima-
tion to what a real agent would accomplish during an actual period of time.
For TDAs to acquire knowledge about the land use change processes, each one is guided
by a set of reinforcement learning (RL) algorithms. RL, derived from the work of Barto et al.
(1981) and Sutton (1988), is a subset of heuristic modeling methods that are used for self-
educating a model to improve solutions to a given problem (Russell and Norvig 2003). In
this research, RL operates by evaluating the spatial patterns resulting from bottom-up-driven
land use change and then assigning numeric rewards as a function of how successful the
pattern is at achieving the objectives of a TDA. The rewards are used as a means for
reinforcing the actions of the BUAs that lead toward patterns that the TDA considers to be
more advantageous. The learning of each TDA is guided by an independent set of RL
algorithms that are specific to the achievement of its own objectives. As a result, the
inclusion of multiple TDAs will have a diverse impact on the behaviors of the BUAs.
Although different heuristic approaches could potentially meet the needs of bottom-up and
top-down integration, RL was specifically chosen for its ability to incorporate individual-
level learning mechanisms (Tang 2008), which is necessary for the TDAs to independently
observe and evaluate the land use change process.

2. Methods
The methods of this study are explained in three subsections to provide independent
descriptions of the bottom-up model with BUAs, the top-down model with TDAs, and the
International Journal of Geographical Information Science 721
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 1. Schematic of agent types, their behaviors, and their influence on the land use change
process.

integration of the two. Figure 1 provides a depiction of agent types, their behaviors, and how
they influence changes to the landscape from either emerging bottom-up patterns or top-
down decision-making. A description of model implementation is then provided using land
use and demographic data from the City of Chilliwack, British Columbia.

2.1. Bottom-up model


The general theoretical structure of the bottom-up model with BUAs in this study is designed
to capture the feedbacks involved in the evolving spatial distribution of population and
commercial activity within city limits (Brown et al. 2005). As shown in Figure 1, the bottom-
up model is composed of two BUAs: (1) household agents that represent individuals or
families that reside each in a single residential unit and (2) commercial agents that purchase
land and develop commercial space in areas with large consumer catchments (i.e., with high
population densities). Although the model can be applied to both raster and vector data
types, this study focuses on the use of vector polygons for representing the irregular structure
of land use parcels that constitute the landscape. Land use parcels are represented by those
722 C. Bone et al.

land uses present in a given region, such as residential, commercial, forest, and agriculture,
which have the potential to change based on agent behaviors.
In addition to land use type, each parcel contains an Attractivity value that provides a signal
to household agents depicting how desirable the location is as a place to live. The Attractivity
value is composed of variables such as distance to desirable land use (e.g., commercial sites or
nature areas) and distances to undesirable land uses (e.g., airports or landfills). An equation
defining the Attractivity value A is calculated by
      
1 1 1 1 1 1
A  1 1  1  (1)
DV1 DV2 DVn UV1 UV2 UVn

where DV and UV represent desirable and undesirable land use variables, respectively, and
n is the total number of variables. As a result, the Attractivity of each parcel is a real value in
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

the interval [0, 1]. Each parcel also contains two population parameters: the first represents
the population of households within the parcel and the second represents the household
population density within a defined neighborhood of the parcel.
The bottom-up model commences by adding a defined number of household agents to
the area. Each household agent is assigned a neighborhood density preference from a
distribution of neighborhood density values. The shape of this distribution is dependent on
the estimated density preference of household agents and will likely be skewed toward
higher densities to represent a standard desire to reside in areas with higher population
densities (Brown et al. 2005). A household agent selects the neighborhood that has a density
most similar to its preference, and then selects a random set of land use parcels from that
neighborhood and decides to reside in the parcel that has the highest Attractivity value and
has vacant space. Selecting from a random set of parcels ensures that household agents have
bounded knowledge that represents individual preferences, personal relationships, limited
search capacities, and timing (Brown et al. 2005).
After all household agents have located a parcel in which to reside, a commercial agent
assesses all neighborhoods and determines if household densities are high enough in any
particular neighborhood to warrant the development of a commercial space. If so, the
commercial agent occupies a vacant parcel within a desired neighborhood, and then addi-
tional commercial agents evaluate the same prospect until it is decided that further commer-
cial space is not warranted. At this point, the Attractivity values for all parcels are
recalculated to incorporate changes to the landscape, and the initial time step of the agent-
based model is complete. The model is performed for a given number of time steps that
represents a timeframe for a land use change process to complete.

2.2. Top-down model


Each TDA in the top-down model represents a separate stakeholder. These agents, unlike the
BUAs, exist as a means to evaluate land use change outcomes. Each TDA is assigned a
separate set of RL algorithms that are implemented for evaluating patterns and assigning
rewards to landscape patterns that are considered successful for achieving their objectives.
The rewards essentially provide a signal representing a type of numeric Quality, depicting
which non-residential parcels should be converted to residential or commercial.
The objectives of the TDAs can be classified into one of three categories: Category 1
the objective is to minimize the loss of a particular land use type to development, where
development is the change in state from non-residential or non-commercial to either
International Journal of Geographical Information Science 723

residential or commercial; Category 2 the objective is to minimize the distance of


developed parcels to a specific land use; and Category 3 the objective is to maximize the
distance of developed parcels to a specific land use. A single TDA can have multiple
objectives belonging to the same or to different categories which can be weighted to
represent the relative importance that a TDA places on one objective over the others. The
RL algorithms combine these objectives to estimate a single quality value for each TDA,
which can be mapped to depict where development should occur so that their objectives are
addressed. The quality values for the different TDAs are then combined to form a collective
Quality value to represent the relative benefit of developing specific parcels so that the
objectives of all TDAs are considered.
The model commences by generating a solution depicting where a defined number of
households and commercial enterprises could be located. The first solution is completely
random as there is no prior information available to direct the model, and as such the
households and commercial enterprises will be randomly distributed in a manner that is
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

far from satisfying any objectives. Each TDA evaluates the solution and, based on the RL
algorithms, assigns a numeric reward to each parcel that has experienced a change in its land
use. Furthermore, a threshold can be defined which specifies the number of new household
agents required for a non-residential parcel to be converted to residential. This ensures that
agriculture or forest land use parcels that are populated with only a small number of
household agents are treated differently by the TDAs during evaluation than those parcels
that are populated with significantly larger numbers of household agents. For example, a
forest parcel with a single household agent exerts a different impact on local water bodies
than it would if it were populated with say 100 household agents. The rewards are calculated
differently for each objective category based on the schematic described below, where a
reward is a real number in the interval [0, 1].

2.2.1. Category 1: Minimize the loss of a particular land use type x to development
Let Px represent a parcel P that has changed from a specific non-residential state x to a
developed state. Then the reward rPx for Px is
2 P  3
min aP x  
tt1;t2;... 7 minaPx
Px
6 P
rP x 4 5 (2)
Px aPx aPx

where aPx represents the area of parcel PPx, minaPx is the minimum area of a parcel from the
set of Px for the current solution, and Px aPx and aPx are calculated from the current solution,
P 
but min Px a Px
is the minimum summed area of Px from the set of land use
tt1,t2,...
solutions from all previous simulations in the learning set (note that the set of parcels Px will in
general be different in each simulation). Equation (2) ensures that all parcels of a particular
land use that change to residential will receive a relatively high reward if the total area changed
to residential is minimized and if the area of a specific parcel is relatively small.

2.2.2. Category 2: Minimize the distance of developed parcels to a specific land use
Let P represent a parcel that has changed from any non-residential state to a developed state.
Then the reward rPy for P is
724 C. Bone et al.
" P #
min P dP;Pj tt1;t2;... mindP;Pj 
rPy P (3)
dP;Pj dP;Pj

where dP;P j is the shortest distance between parcel P and a parcel Pj with specific land use j,
max
P dP;Pj is the minimum distance between P and Pj from the current P solution,
and
d
P P;Pj and d P;Pj are calculated from the current solution, but min d
P P;Pj tt1;t2;...
is the minimum sum of distances between P and Pj from all previous solutions.

2.2.3. Category 3: Maximize the distance of developed parcels to a specific land use
Let P represent a parcel that has changed from any non-residential state to a developed state.
Then the reward rpz for P is
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

" #" #
X d P;Pj t t  1; t  2 . . . dP;Pj
rPz P  (4)
P max P dP;Pj max dP;Pj

where  dP;P j is the shortest distance between parcel P and a parcel Pj with specific land use j,
P dP;Pj is the maximum distance between P and Pj from the current
max P solution,
and
d
P P;Pj and d P;Pj are calculated from the current solution, but max d
P P;Pj tt1;t2;...
is the maximum sum of distances between P and Pj from all previous solutions.
As mentioned above, the calculated rewards are used to estimate the quality of each
parcel that was developed in the present solution. The quality, q, is calculated for
h  i
Category 1: qPxt1 max 0; qPxt a rPx  qPxt (5)

h  i
Category 2: qPyt1 max 0; qPyt a rPy  qPyt (6)

h  i
Category 3: qPzt1 max 0; qPzt a rPz  qPzt (7)

where the initial values for qPxt1 , qPyt1 , and qPzt1 are set to 0.0, t equals the time step of
the current solution, t + 1 represents time step of the new value to be used when developing
the next solution, and a represents a real number in the interval [0, 1] that diminishes with the
generation of each new solution to decrease changes to estimates in the quality of parcels. A
maximum condition is present in Equations (57) to ensure that qPxt1  0:0, qPyt1  0:0,
and qPzt1  0:0. Each estimate of quality can be weighted based on its importance using a
weighting factor, w, and then combined to formulate a quality value for each TDA, g, based
on the equation

qgPt1 wx qPxt1 wy qPyt1 wz qPzt1 (8)

Thus, each TDA possesses a map of the landscape represented by quality values
indicating where development should occur to achieve its own objective(s). It is important
to note that Equation (8) contains a single instance from each of the three categories, but can
be altered to represent scenarios in which, for TDA g, quality is represented by a single
International Journal of Geographical Information Science 725

category, or, conversely, by multiple instances of one or more categories. The quality values
of the TDAs are then combined to form the final Quality, Q, of parcel P using the equation

QPt1 qg1 Pt1 qg2 Pt1    qgn Pt1 (9)

where n is the total number of TDAs.


After the Quality of each land use parcel has been estimated, the model generates another
solution that is again based on the random assignment of households and commercial
enterprises to specific land use parcels, and the rewards and subsequent Quality are esti-
mated. As the number of generated solutions increases, the model gradually makes a
transition from randomly assigning location to utilizing the estimated Quality of each parcel.
The probability of using the Quality values to determine household and commercial
enterprise location increases toward 1.0 as the model draws closer toward generating its
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

final solution. The number of solutions required to generate a solution that is acceptable for
achieving all TDAs is dependent on the number of objectives and the number of land use
parcels in the data set that the model must search. If a suitable number of solutions are used
and the probability of using the Quality values to determine household and commercial
enterprise locations diminishes in an appropriate fashion, those solutions generated toward
the end will be considered acceptable for meeting the objectives of the different stakeholders.

2.3. Integrating bottom-up and top-down models


The bottom-up model from Section 2.1 and the top-down model from Section 2.2 are
integrated to formulate the notion of modeling-in-the-middle. This allows for land use
patterns to emerge from the interactions among household agents and commercial agents
while simultaneously allowing TDAs to evaluate the resulting patterns and estimate the
Quality of parcels for development. Modeling-in-the-middle is accomplished by simulating
the land use change processes for a given number of time steps during which period
household and commercial agents select parcels in which to reside. At the completion of
the final time step, the evaluative process contained in the top-down model is activated to
estimate how well the resulting land use pattern achieves the objectives of the various TDAs.
The estimated Quality values of each TDA are combined with the Attractivity values of each
parcel to form a combined TDABUA notion of Suitability, S, of parcel P using the equation

SPt1 APt1 QPt1 (10)

The land use change process is simulated a number of times, each time updating the
Suitability of each parcel in the landscape. The two types of models become fully integrated
as the household and commercial agents select parcels in which to reside based on the logic
of the top-down model; that is, their decisions during the initial simulations of the process are
random, but slowly they convert their behavior by utilizing the Suitability values and density
preferences to determine where to reside. The initial random behavior followed by the switch
to selective behavior allows the model to explore a wide variety of land use outcomes, gain
knowledge on how well such outcomes are able to achieve various objectives, and gradually
improve the outcomes to increasingly satisfy the TDAs. Furthermore, because the household
and commercial agents are responsible for ensuing land use change based on their prefer-
ences, the integrated model ensures that individual decision-making behavior and the
resulting feedbacks are manifest in the outcome. Formulating the model as such does not
726 C. Bone et al.

represent a direct interaction between the TDAs who are primarily concerned with policies
and the BUAs who are interested in preferential locations because it is unrealistic to assume
that TDAs would have repetitive opportunities to observe and evaluate the emerging land-
scape patterns. However, the intention with the modeling-in-the-middle approach is not to
impose a direct link between the decision-making behavior of both types of agents, but
instead to evaluate how emerging landscape patterns can be improved to achieve global land
use objectives.

3. Model implementation
The models were developed in Agent Analyst 1.0b (Argonne National Laboratory 2006) in
which Python programming language (Python Software Foundation 2007) was used for
designing agent behavior and constructing RL algorithms. The bottom-up and top-down
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

models and the modeling-in-the-middle approach are implemented in the context of land use
change in the City of Chilliwack, located in southwestern British Columbia, Canada. A
significant portion of Chilliwacks land base is dedicated to agriculture, most of which is
protected under the Agricultural Land Reserve as a means to prevent agricultural land from
being transformed due to growth. The city also contains several areas of sensitive wetlands
as well as a portion of the Fraser River, a significant waterway for fishing and transport. In
addition, the city contains forested land, some of which is considered old growth and
provides habitat for threatened species. Conflicting with these land uses is the need to
develop part of the city to facilitate a projected increase in 16,000 households and a
substantial amount of new commercial space over a 16-year period between 2005 and
2021. Although some of the urbanized areas in Chilliwack have been slated for redevelop-
ment, there is a public reluctance to have a high-density town center. As a result, there is
much pressure on agriculture and forests, as well as lands close to wetlands and waterways,
for conversion to residential and commercial land uses. The situation that Chilliwack faces,
common in many cities and regions undergoing population growth, leads to the establish-
ment of both bottom-up dynamics and top-down constraints, a situation that is well-suited
for modeling with the approach developed in this research. The scenario presented here
motivates the definition of the TDAs, which, along with their objectives and categories as
defined in Section 2.2, is presented in Table 1.
The data used for implementing the model were derived from the City of Chilliwacks
land use data (Figure 2a) and cadastral data (Figure 2b) in addition to census data from the
Government of Canada (Figure 2c). The land use data set is composed of polygons
representing specific land uses over broad regions within the city, whereas the cadastral
data set contains polygons that represent division of property. The rural land class in the land

Table 1. TDA types and their objectives.


Top-down agent type Objective Category
Agriculture agent Minimize conversion of agriculture to residential 1
and commercial land use
Forest conservation agent Minimize conversion of forest to residential 1
and commercial land use
Planner agent Minimize distance of new residential land 2
use parcels to existing residential land use parcels
Environmental Maximize distance of new residential and commercial 3
conservation agent land use parcels to rivers and wetlands
International Journal of Geographical Information Science 727
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 2. City of Chilliwacks land use distribution (a) and population density based on census
tracts (b).

use data set represents undeveloped areas that do not contain forest or agriculture. Examples
of rural land use are open fields or meadows, land cleared for future development, or areas of
extremely low density such as a single house on land adjacent to agricultural fields. The
cadastral data were overlaid onto the land use data to provide a finer scale land use data set.
The polygons were then aggregated to a scale deemed appropriate for simulating land use
change with the modeling-in-the-middle approach (Figure 2d). The cadastral data, while
appropriate for simulating bottom-up processes, exist at too fine a scale for evaluating land
use change patterns with a top-down approach because the number of potential solutions is
innumerable due to the large number of cadastral units that can potentially experience
change. This presents a significant challenge when implementing learning algorithms that
are intended to evaluate and improve upon land use solutions. Conversely, the land use data
exist at far too coarse a scale for adequately representing land use change that unfolds
because of the interactions of household agents and commercial agents.
The census data provide population values for each census tract in the City of Chilliwack
in 2006. A census tract represents a neighborhood in a large urban area that generally has a
population between 2500 and 8000 people. The information from the census data was
utilized for defining neighborhoods that household agents analyze to make their decisions,
as is described below.
The parameters used to define the household and commercial agents were developed
from information contained in the City of Chilliwacks Official Community Plan (City of
Chilliwack 1998). Based on the citys planning initiatives, the model operates representing a
728 C. Bone et al.

single year with one time step. At each time step, 1000 new households are added to the city
for a total of 16 time steps to represent the time period between 2005 and 2021. This
population growth trajectory is based on the citys projection for 16,000 new residents
between 2005 and 2021 (City of Chilliwack 1998), and assumes an equal number of
residents are added to the city each year.
Census data from the Government of Canada are used to identify the neighborhood to
which each parcel belongs. Household agents first examine the density of the census tracts to
determine which one is closest to its preferred density, and then looks at a subset of parcels
within the census tract and searches for the one with the highest Attractivity. The Attractivity
value is based on three criteria: (1) to be located at minimum distance to commercial space;
(2) to be located at minimum distance to nature reserves; and (3) to be located at a maximum
distance to airport. Future implementation of this model can incorporate additional variables
that influence location preferences, such as employment, land price and transportation. The
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

City of Chilliwacks Official Community Plan was used in coordination with the spatial
dimensions of the data sets to determine that a non-residential land use parcel will convert to
residential once 20 new households have been added to the parcel, and restricting the density
of each parcel to 20 households per acre.
The citys Official Community Plan states that 21.74 acres of new commercial space will
need to be added by 2021. This area of commercial space and the total land needed to support
it translates, within the context of the model, into converting 1 hectare of undeveloped land
to commercial land use for each additional 600 households. Therefore, after every 600
households are added, a commercial agent looks for the census tract with the highest density
of households and converts an undeveloped parcel to commercial. Although this is a
simplification of how commercial spaces are purchased and converted, it provides a
mechanism for ensuring that feedbacks are present in the model as households desire to
live near commercial land uses and commercial enterprises strive to establish a large
consumer catchment.

4. Results
The results are presented separately for the bottom-up model, the top-down model, and the
integrated bottom-up and top-down model.

4.1. Bottom-up model


The results from the bottom-up model are presented in Figure 3. The map depicts the location of
household and commercial agents among the different land use types at the final time step
(Figure 3a), and, for visualization purposes, the location of household and commercial agents in
the land use parcels without land use types displayed (Figure 3b). The results demonstrate that
household agents tend to cluster in areas close to existing residential land use parcels and that
commercial agents are located in areas adjacent or close to newly added households. Those
parcels that contain household agents but have not changed to residential land use exist in their
initial state because the number of new household agents is below the threshold for change.

4.2. Top-down model


The top-down model was calibrated to determine the number of solutions required for the
model to generate acceptable results. Generating too few solutions leads to results that do not
International Journal of Geographical Information Science 729
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 3. Simulation results for the bottom-up model displayed with land uses (a) and without land
uses (b).
730 C. Bone et al.

adequately address the TDAs objectives, while generating too many solutions is unneces-
sary and leads to larger than desired processing times. Calibration was performed by running
the model and observing the point at which TDAs gain sufficient knowledge for achieving
their objectives. Sufficient knowledge is gained when the Quality values estimated by each
TDA are representative of their objectives. The calibration procedure resulted in the selec-
tion of 1000 solutions, meaning that the model is able to sufficiently achieve TDAs
objectives no later than the 1000th generated solution. Figure 4 presents the Quality values
estimated by the agriculture agent (a), the forest conservation agent (b), the planning agent
(c) and the environmental conservation agent (d). Each map depicts the Quality values as of
the 1000th solution. The spectrum of low to high Quality values represents the degree to
which it is disadvantageous or advantageous to develop a parcel for the specific TDA to
achieve its objective.
The final land use solution of the top-down model is presented in Figure 5. The equal
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

weighting of the four objectives can be observed in the fact that the majority of households
are added in rural areas (i.e., non-agricultural and non-forested areas) that are close to
existing residential land use parcels. The only objective that does not appear to be fully
satisfied is the environmental conservation agents objective, as developing at the furthest
point from the rivers and wetlands would entail sacrificing the other objectives.
Furthermore, the land use solution also demonstrates that clustering of household agents
is more prominent than with the bottom-up model, but commercial space is more spatially
distributed.
The top-down model is further analyzed by generating solutions with unequal weighting
distributions. Figure 6 represents the final solution from four different modeling scenarios in
which a proportionately higher weight is assigned to different TDAs. The results reveal the
locations of household agent clusters when different objectives are emphasized and how
certain objectives are sacrificed at the expense of achieving others.

4.3. The integrated bottom-up and top-down model


The simulation results for the integrated model at the final time step of the 1000th solution is
presented in Figure 7a with land uses and without land uses in Figure 7b. The most notable
observation from this simulation is that the inclusion of household and commercial agents
detracts from the ability of the TDAs to achieve their objectives. This should be somewhat
expected as including further parameters in the model will lead to more diverse results, but
the results indicate that the bottom-up process has a disproportionate influence on urban land
use change. This is due to the fact that household agents still control their desire to live
within a neighborhood with a certain density, which dictates the general location where they
live. The specific location is then dictated by the Quality and Attractivity values within the
neighborhoods. Furthermore, the locations of commercial agents are more similar to the
results from the bottom-up model than the top-down model: commercial agents occupy
larger areas that are adjacent to densely populated parcels. This is due to the fact that the
location of commercial agents in the integrated model is dictated by the preference to be in
densely populated areas, which will force the commercial agents to cluster in areas where the
most new household agents have been added.
For the modeling approach presented in this study to be useful for land use planning, it is
important to understand how different TDA objectives are achieved under the different
models. Figure 8 presents the results from all the performed modeling scenarios with regard
to outcomes of the four objectives. The bar graphs demonstrate how much forest land and
agricultural land is lost due to development and how far the developed parcels are from
International Journal of Geographical Information Science 731
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 4. Distribution of Quality values for the final solution of the top-down model shown
independently for the (a) agriculture agent, (b) forest conservation agent, (c) planning agent, and
(d) water conservation agent.

existing residential and river/wetland parcels. The different bars represent these land use
change characteristics for the bottom-up model (A); the top-down model with objectives
equally weighted (B); the top-down model with higher weight apportioned to the agriculture
agent (C); the forest conservation agent (D); the planning agent (E); and the environmental
conservation agent (F); and the integrated bottom-up and top-down model with all objec-
tives equally weighted (G).
The bar graphs reveal that, as expected, the top-down model is most successful at
achieving the specific objectives of the TDAs when the objectives are weighted in their
favor. That is, the simulation with the minimal amount of forest area lost was biased toward
the forest conservation agents objective (bar D in Figure 8a); the least amount of agriculture
land is lost when the simulation is biased toward the agriculture agents objective (bar C in
Figure 8b); the cumulative distance to residential is reduced when focusing on the planner
agents objective (bar E in Figure 8c); the cumulative distance to water is maximized when
biased toward the environmental conservation agents objective.
Figure 8 also reveals important information regarding each specific objective. Specifically,
emphasizing the environmental conservation agents objective results in the highest loss of
forest land parcels (bar F in Figure 8a), which is most likely due to the forested parcels existing
at the furthest distance from bodies of water; the agriculture land use receives a dispropor-
tionate negative impact when land use change solutions integrate the bottom-up behavior of
the household and commercial agents (bars A and G in Figure 8b); and the planner agents
732 C. Bone et al.
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 5. Simulation results for the final solution of the top-down model displayed with land uses (a)
and without land uses (b). All objectives are equally weighted.
International Journal of Geographical Information Science 733
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 6. Simulation result for the final solution of the top-down model when a higher weight is
apportioned to the (a) agriculture agent, (b) forest conservation agent, (c) planning agent, and
(d) environmental conservation agent.

objective is moderately sacrificed when the bottom-up behavior of the household and com-
mercial agents is included (bars A and G in Figure 8c).
As mentioned above, the most significant observation that can be made from these
results is that achieving top-down objectives becomes challenging when the bottom-up
processes led by the desire of individuals to locate in specific areas influence land use
change. However, Figure 8 demonstrates that the land use patterns resulting from the
integrated model are more similar to the bottom-up model, which is counter to the expecta-
tion that the integrated model would result in an even compromise between the bottom-up
and top-down models because of the fact that the Attractivity and Quality values were
equally treated. The reason for the disproportionate similarity to the bottom-up model can
be attributed not only to the desires to live in specific neighborhoods, but also to the
feedbacks that result from the bottom-up interactions between agents and the urban
landscape.

5. Conclusion
Bridging the gap between bottom-up and top-down models through the modeling-in-the-
middle approach is an important direction for land use modeling and planning. The bottom-
up component of this study was a generalized representation of land use dynamics because
734 C. Bone et al.
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 7. Simulation result for the final solution of the integrated bottom-up and top-down model
displayed with land uses (a) and without land uses (b). All objectives are equally weighted.
International Journal of Geographical Information Science 735
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Figure 8. Land use measurement relating to the objective of each TDA. The bars represent the results
from the bottom-up model (A); the top-down model with objectives equally weighted (B); the top-
down model with higher weight apportioned to the agriculture agent (C); the forest conservation agent
(D); the planning agent (E); and the water conservation agent (F); and the integrated bottom-up and
top-down model with all objectives equally weighted (G).

of the fact that household agent heterogeneity and competition between commercial agents
were not explicitly considered. However, for the purpose of this study, the ABM adequately
provides a mechanism for simulating the emerging land use patterns resulting from indivi-
dual behaviors, and how such behaviors influence the ability to achieve various top-down
objectives.
The top-down multi-objective decision-making component of this study was
implemented by way of heuristic modeling algorithms for TDAs to improve their estimation
of parcel Suitability so that they can achieve their objective. One may view this method as
similar to other multi-criteria analytical procedures, such as multi-criteria evaluation, espe-
cially considering the results that depict parcel Suitability, the agents appear to be merely
learning through repetitive simulation what simpler methods can achieve. However, the
main advantage of implementing the heuristic component is that the RL algorithms are
useful for showing how to achieve stakeholder objectives among the bottom-up dynamics
generated by individual households and commercial enterprises. The final solution from the
integrated model is thus not the optimal solution for addressing all objectives, but instead is a
representation of where it is most suitable for land use change to occur given the location
preferences, and the ensuing feedbacks will play a large role in dictating the dynamics of the
urban landscape.
In closing, the literature on land use science and modeling has progressed in recent years
by incorporating a host of methods from a diverse set of disciplines to explore, predict, and
generate urban land use change scenarios. Bringing together bottom-up and top-down
approaches will assist in this progress by representing two conflicting yet very real forces
in the land use change process.
736 C. Bone et al.

Acknowledgments
The authors thank the Natural Sciences and Engineering Research Council of Canada (NSERC) and
the Social Science and Humanities Research Council of Canada (SSHRC) for their financial support of
this study. Acknowledgment is also given to the City of Chilliwack for providing the land use data and
cadastral data for use in this study.

References
Argonne National Laboratory, 2006. Agent analyst. Chicago, IL: Center for Complex Adaptive Agent
Systems Simulation: Argonne National Laboratory.
Barto, A.G., Sutton, R.S., and Brouwer, P.S., 1981. Associate search network: a reinforcement learning
associative memory. Biological Cybernetics, 40, 201211.
Batty, M., 1997. Urban systems as cellular automata. Environment and Planning B: Planning and
Design, 24 (2), 159305.
Batty, M. and Torrens, P.M., 2005. Modeling and prediction in a complex world. Futures, 37 (7),
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

745766.
Benenson, I., 1998. Multi-agent simulations of residential dynamics in the city. Computers,
Environment and Urban Systems, 22 (1), 2542.
Brown, D.G., et al., 2005. Path dependence and the validation of agent-based spatial models of land
use. International Journal of Geographical Information Science, 19 (2), 153174.
Carver, S.J., 1991. Integrating multi-criteria evaluation with geographical information systems.
International Journal of Geographical Information Systems, 5 (3), 321339.
Castella, J.-C., et al., 2007. Combining top-down and bottom-up modelling approaches of land use/
cover change to support public policies: application to sustainable management of natural
resources in northern Vietnam. Land Use Policy, 24, 531545.
Chakhar, S. and Martel, J.-M., 2003. Enhancing geographical information systems capabilities with
multi-criteria evaluation functions. Journal of Geographic Information and Decision Analysis, 7,
4771.
Church, R.L., Loban, S.R., and Lombard, K., 1992. An interface for exploring spatial alternatives for a
corridor location problem. Computers and Geosciences, 8, 10951105.
City of Chilliwack, 1998. City of Chilliwack official community plan. Chilliwack, British Columbia:
City of Chilliwack.
Clarke, K.C., Hoppen, S., and Gaydos, L., 1997. A self-modifying cellular automation model of
historical urbanization in the San Francisco Bay area. Environment and Planning B: Planning and
Design, 24, 247261.
Cohon, J.L., 1978. Multiobjective programming and planning. New York: Academic Press.
Eastman, J.R., Jiang, H., and Toledano, J., 1998. Multi-criteria and multi-objective decision making for
land allocation using GIS. In: E. Beinat and P. Nijkamp, eds. Multicriteria analysis for land-use
management. Dordrecht, The Netherlands: Kluwer Academic Publishers, 227251.
Engelen, G., et al., 2007. The MOLAND modeling framework for urban and regional land use
dynamics. In: E. Koomen, et al., eds. Modelling land-use change: progress and applications.
Springer: New York, 297320.
Engelen, G. and White, R., 2007. Validating and calibrating cellular automata based models of land use
change. In: S. Albeverio, et al., eds. The dynamics of complex urban systems: an interdisciplinary
approach. New York: Springer, 297320.
Feick, R.D. and Hall, B.G., 2004. A method for examining the spatial dimensions of multi-criteria
weighting sensitivity. International Journal of Geographical Information Science, 18, 815840.
Filatova, T., Parker, D.C., and van der Veen, A., 2009. Agent-based urban land markets: agents pricing
behavior land prices and urban land use change. The Journal of Artificial Societies and Social
Simulation, 12 (1): Article Number 3.
Herwijnen, V.M. and Rietveld, P., 1999. Spatial dimensions in multicriteria analysis. Berlin: Springer
Verlag.
Heywood, I., Cornelius, S., and Carver, S., 1998. An introduction to geographical information systems.
New York: Prentice Hall.
Jacobs, J., 2002. The death and life of great American cities. New York: Random House.
Jankowski, P., 1995. Integrating GIS and multiple criteria decision making methods. International
Journal of Geographical Information Systems, 9 (3), 252273.
International Journal of Geographical Information Science 737

Jantz, C.A., et al., 2009. Designing and implementing regional urban modeling system using the
SLEUTH cellular urban model. Computers, Environment and Urban Systems, 34, 116.
Johnson, S., 2002. Emergence: the connected lives of ants, brains, cities and software, New York:
Touchstone.
Ligmann-Zielinska, A., Church, R.L., and Jankowski, P., 2008. Spatial optimization as a generative
technique for sustainable multiobjective landuse allocation. International Journal of
Geographical Information Science, 22 (6): 601622.
Liu, X.P., et al., 2008. A bottom-up approach to discover transition results of cellular automata using
ant intelligence. International Journal of Geographical Information Science, 22 (1112),
12471269.
Malczewski, J., 1999. GIS and multicriteria decision analysis. New York: John Wiley and Sons.
Malczewski, J., 2004. GIS-based land-use suitability analysis: a critical overview. Progress in
Planning, 62, 365.
Malczewski, J., 2006. GIS-based multicriteria decision analysis: a survey of the literature.
International Journal of Geographical Information Science, 20 (7), 703726.
Openshaw, S. and Openshaw, C., 1997. Artificial intelligence in geography. New York: Wiley.
Downloaded by [Universiteit Twente] at 06:33 25 September 2014

Parker, D.C., et al., 2003. Multi-agent systems for the simulation of land-use and land-cover change: a
review. Annals of the Association of American Geographers, 93 (2), 314337.
Portugali, J., 2000. Self-organization and the city. New York: Springer.
Python Software Foundation, 2007. Python. New Hampshire: Python Software Foundation.
Robinson, D.T. and Brown, D.G., 2009. Evaluating the effects of land-use development policies on ex-
urban forest cover: an integrated agent-based approach. International Journal of Geographical
Information Science, 23 (9), 12111232.
Russell, S. and Norvig, P., 2003. Artificial intelligence: a modern approach. Upper Saddle River, NJ:
Prentice Hall.
Sutton, R.S., 1988. Learning to predict by the methods of temporal differences. Machine Learning, 3, 944.
Tang, W., 2008. Simulating complex adaptive geographic systems: a geographically aware intelligent
agent approach. Cartography and Geographic Information Science, 35 (4), 239263.
White, R. and Engelen, G., 1997. Cellular automata as the basis of integrated dynamic regional
modeling. Environment and Planning B: Planning and Design, 24, 235246.
Xiao, N., Bennett, D.A., and Armstrong, M.P., 2002. Using evolutionary algorithms to generate
alternatives for multiobjective site search problems. Environment and Planning A, 34 (4),
639656.
Xiao, N., Bennett, D.A., and Armstrong, M.P., 2007. Interactive evolutionary approaches to multi-
objective spatial decision making: a synthetic review. Computers, Environment and Urban
Systems, 31, 232252.
Xie, Y.C., Batty, M., and Zhao, K., 2007. Simulating emergent urban form using agent-based model-
ing: Desakota in the Suzhou-Wuxian region in China. Annals of the Association of American
Geographers, 97 (3), 477495.
Zellner, M., 2007. Generating policies for sustainable water use in complex systems: an integrated
land-use and water-use model of Monroe County, Michigan. Environment and Planning
B: Planning and Design, 34 (4), 664686.

You might also like