Author Manuscript – Same content as final publication, only the layout differs.
The
final
publication
is
available
at
Wiley
and
Sons
via
http://onlinelibrary.wiley.com/doi/10.1111/tgis.12018/abstract
To cite this paper :
Renard J., Duchêne C. 2014. Urban Structure Generalization in Multi-Agent Process by Use of Reactional
Agents, Transactions in GIS, vol. 18, n°2, April 2014, 201-218, doi:10.1111/tgis.12018.
Urban Structure Generalization in Multi-Agent Process by
Use of Reactional Agents
Jérémy Renard1, Cécile Duchêne1
1
Institut Géographique
[email protected]
National
(IGN
France),
laboratoire
COGIT,
Université
Paris-Est,
This article proposes an improvement of automated cartographic generalization using multi-agent sytems in urban
areas. Indeed the AGENT model, whose robustness has been tested and approved through the European project
AGENT, gives very good results in generalizing dense urban areas by means of enlargement, removal and
displacement of buildings. But this model does not tackle the question of including particular structures like
building alignments in the process, which is a crucial issue. The problem is that integrating such structures does not
fit into the accurate top-down hierarchy of urban agents. In order to face this problem, we propose to partly reengineer the model by introducing the concept of reactional agents whose behavior is very different from
hierarchical agents of the original model as they use bottom-up activation. In this view, urban alignment is
considered to be a reactional agent activated only by its inner buildings, which generalizes the aligned buildings
together into one entire structure. Associating reactional alignment behavior with new generalization actions on
alignments significantly improves the model and gives better results in dense urban areas. Moreover, the idea could
probably be used forother applications.
1. Introduction
Generalization of geographical data is a process which aims at adapting the content of a map
or a geographical database depending on its display scale. It implies simplifying, caricaturing
or displacing all or part of the geographical objects contained in the map in order to ensure its
legibility and its clarity while trying to lower the level of detail. Since it is a complex and
timeconsuming process, its automation has been the topic of many research studies for more
than 20 years, resulting in many research models and effective production processes. Brassel
and Weibel (1988) and McMaster and Shea (1992) laid the foundations of automated digital
generalization as an independent topic by defining the main problems that have to be faced
during map generalization. Subsequently, many operators have been defined to overcome the
problem, and these are usually divided into two themes described by Regnauld and McMaster
(2007) and refined by Foerster et al.’s (2007) classification: model generalization, which
derives the content of a geographical database by using specifications – e.g. class selection,
simplification of geometries, etc. – and cartographic generalization, which tends to solve
legibility problems caused by the symbolization applied to the objects (overlapping,
overcrowding, etc.) – e.g. displacement, typification, etc. In this article, we will focus mainly
on cartographic generalization.
Many different approaches have been tested to tackle the issue of automated cartographic
generalization, resulting in interesting and efficient models. One of them is derived from the
multi-agent system paradigm and takes into consideration all geographical objects on a map
as agents with constraints and autonomous behavior to solve these constraints by using sets of
actions. This purpose was first developed theoretically by Baejis et al. (1996) and Ruas and
Plazanet (1996) and then formalised in the AGENT generalization model (Ruas 1998a, Baejis
2000). The main idea of the model is to describe a map as a fixed hierarchy of agents, going
from the lowest level of detail (micro agents like buildings, road lines) to the highest one
(several groups of agents called meso agents like urban blocks, whole cities). Then, the
TGIS, 18(2): 201-218, 2014
generalization of each agent is governed by its upper level meso agent and depends on
constraints to follow and actions to overcome these constraints. Thus, the model is
particularly well-suited to urban generalization since urban areas have a very hierarchical
structure – buildings are governed by blocks, blocks are governed by districts, and districts
are governed by cities. This approach gives interesting results, but it does not take into
account urban structures like building alignments because such structures do not fit into the
constrained hierarchy of urban agents. However, generalizing these structures is very
important to preserve the general layout of the urban area.
The aim of this article is to propose a way to improve the AGENT model in order to include
urban structures in the process. The next part of the article gives an overview of existing
techniques to generalize urban areas using multi-agent systems. The third part illustrates the
questions raised by integrating a meso level for urban structures into the existing hierarchy of
urban agents, while the fourth part proposes a means of answering these questions by
introducing the principles of reactional agents. Section five focuses on the application of
reactional agents to generalize building alignments, and some results are illustrated in section
six. In the final section of the article, we emphasize flexibility and possible reuse of the main
principles of this model in the future and we suggest examples of further work on the topic.
2 Urban Generalization Using Multi-Agents Systems
2.1 Theoretical Generalization of Towns
Generalization of urban areas is one of the most important problems that a cartographer has to
face while generalizing topographic maps at medium scales. Well generalized cities are
essential to the whole visual rendering of a map, mainly for two reasons:
– Cities are structuring places of the map, so the users expect a precise and legible
representation of these areas; and
– Cities are dense with objects, they have many streets and buildings, so it always poses
problems of overlapping, exaggerated density and general overcrowding, which have to be
corrected through accurate generalization methods.
Urban generalization is a complex issue as it involves different types of geographical objects
with a different level of detail – the core of a town is composed of buildings as well as road
sections aggregated in a complete street network, and this network delineates urban blocks.
Ruas and Mackaness (1997) highlighted the main challenge that has to be faced, which
consists in managing these different levels of detail properly and correctly. Ruas (1997) gave
a classification of all operators that have to be applied in generalizing a whole town:
–
–
–
–
–
–
Individual building generalization: enlargement, shape simplification, squaring, etc.;
Building aggregation for small groups of buildings;
Street selection to ensure map legibility and avoid overcrowded city centres;
Building selection to preserve block density and prevent overlapping;
Building displacement to avoid overlapping with roads or with other buildings; and
Building typification to preserve the structures of building clusters.
To apply these, operators need different levels of detail to be applied: road selection is a
general process affecting the whole town, whereas building selection and displacement are
controlled by inner blocks in a town. According to the practical results of the AGENT project
presented by Lamy et al. (1999) and Barrault et al. (2001), the use of multi-agent systems
(MAS), where geographical objects are described as agents with constraints that have to be
TGIS, 18(2): 201-218, 2014
respected and actions to solve these constraints, appears to be a sensible and efficient method
in overcoming this multi-level issue.
2.2 Current Use of the MAS Approach
Multi-agent systems offer solutions to complex problems involving many entities in
interaction. They describe a world populated by agents where each agent is able to
characterize itself by satisfying its constraints, to propose actions and to communicate with
other agents in the neighborhood. This conceptual approach is suitable for very hierarchical
organizations like the model proposed by Ferber and Gutknecht (1998) which introduces
recursive designs in the organization of MAS. Indeed, cartographic generalization of urban
areas is ruled by a rigid hierarchy between geographic objects, developed in Section 2.1, so
that the concepts of MAS make sense in solving this problem. Ruas (2000) focused on the
interest of using such a multilevel model for urban areas as the first step in using MAS for
cartographic generalization purposes.
The AGENT model, proposed by Ruas (1998a) and whose results are illustrated by Lamy et
al. (1999), describes the geographical agents comprising a town with three different levels of
detail (Figure 1): two successive levels of meso agents respectively composed of towns and
urban blocks, and the last level of micro agents made of individual buildings. The huge
interest in this multi-agent approach is to use each level for particular constraints depending
on which generalization operators (described in Section 1.1) have to be applied in this level.
Figure 1. Urban agent hierarchy (Ruas 2000)
In practice, the initial data is exclusively composed of individual buildings, considered as
micro agents, then town and block agents are created by enriching this initial data. The
enrichment consists in calculating concave hulls around buildings, using pure geometrical
methods like alpha shapes (Galton and Duckham 2006) or methods inspired from an accurate
analysis of urban spaces and shapes (Boffet 2000, 2002). These concave hulls represent the
geometrical shape of towns, then they are partitioned using structuring networks – mainly
streets and rivers – to create the town’s inner urban blocks. As a consequence, the hierarchy
of urban agents is composed of:
– Towns, which are composed of blocks and of a street network, ruling the generalization of
their blocks;
– Blocks, which are part of a town and are composed of buildings and street sections
surrounding themselves, ruling the common generalization of their buildings; and
– Buildings, which are part of a block, ruling their own generalization.
This hierarchy relies on a conceptual model illustrated in Figure 2, and each level is adapted
to a few processes described in Section 2.1 with the classification of Ruas (1997).
TGIS, 18(2): 201-218, 2014
Generalization operators are triggered depending on the degree to which cartographic
constraints are met.
Figure 2. Conceptual model (UML) of the AGENT model hierarchy for urban areas (Ruas 2000)
The constraints applied on buildings (mainly size and legibility) do not consider the spatial
neighborhood and the agent hierarchy of buildings; only micro agents are concerned, and only
the lowest level of the AGENT hierarchy is taken into account. It results in individual
building generalization using enlargement, squaring, shape simplification, etc.
On the contrary, the constraint of not overcrowding city centres, mainly due to high street
density, involves all parts of the town. Road selection is needed to solve this constraint, ruled
by the town meso agent that will echo the results of block meso agents after the process. To
reach this aim, Jiang and Claramunt (2002) and Jiang and Harrie (2004) divided street
network selection into two steps: elimination of insignificant dead ends, and aggregation of
urban blocks overcrowded by road symbolization.
The role of the block meso level is to control intermediate constraints: non-overlapping
between buildings and streets, optimizing map legibility within the blocks, preserving the
initial urban density while decreasing the number of buildings. The operators involved in
overcoming these constraints are mainly building selection and displacement, launched and
ruled by the urban blocks. There are many algorithms using agents for these two operators, for
instance those based on the AGENT model developed by Ruas (1998b) or those based on the
GAEL model developed by Gaffuri (2007).
2.3 Actual Results
The practical implementation of the theoretical AGENT model already gives good automated
results. The process has even been used on production lines (Lecordix et al. 2007), as
illustrated in Figure 3, with fully automated urban generalization. Nevertheless, satisfactory as
these results may seem to be, some imperfections still emerge. Focusing on the right part of
the generalized area in Figure 3, urban alignments structuring the town in the initial data are
not taken into account throughout the generalization process. As a result, some of these
alignments disappear and other ones are distorted. To avoid such problems, the model should
be partly re-engineered.
TGIS, 18(2): 201-218, 2014
Figure 3. Actual results of the AGENT process on a city centre at IGN France (Lecordix et al. 2007). It is
significant to notice that building alignments are not well represented (e.g. on the right of the area)
In fact, the overview of constraints and actions used in AGENT and described in Section 2.2
leads us to the first conclusion: among all urban generalization operators, there is still one
which has not been processed, and that is preserving the shape of particular building
structures like alignments by use of typification or caricature. The problem is that we do not
know which meso agent should cover such structures during the whole AGENT process, as
block level is too high and building level is too low to control intermediate structures.
Integrating a new meso level for urban structures seems to be necessary, but it raises some
questions that are quite tough to answer due to the characteristics of these structures. It is
more of a question of the control of urban structures and on their hierarchy level in the urban
AGENT model than on their own generalization, as some interesting ideas on typification of
urban alignments already exist like those presented by Regnauld (2001) or Hangouët (1998).
To face this major setback of the AGENT model, we propose a means of introducing urban
structure agents to the AGENT hierarchy, to ensure the shape of these structures are preserved
during the generalization process. Indeed, the question of building alignments is a study case
leading to a more generic model.
3 Urban Structures Integration in the AGENT Hierarchy
3.1 Which Agent Level for Urban Structures?
To represent special urban structures like building alignments in the urban agents hierarchy,
the integration of a new intermediate agent level is needed. In practice, the agent level for
urban structures should be integrated below the block level and above the building level – a
structure is part of a block and comprises buildings. But this solution poses a problem as
buildings, depending on their context, could be part of an urban structure or not, in which case
their controlling meso agent would be different depending on whether they are part of a
structure or not. So this idea is less straightforward than it appears at first.
Indeed, the original AGENT hierarchy is very systematic: towns are exclusively composed of
blocks, blocks are exclusively composed of buildings, and no direct interaction is possible
between a town and its buildings. Considering the question of building alignments, the
behavior of an alignment meso level leads to a huge problem (explained above and illustrated
in Figure 4) since alignments are not systematic structures and a block could now be
composed of:
– Buildings exclusively, directly ruled by the block, without any alignment;
– Alignments exclusively, each of them composed of buildings; or
TGIS, 18(2): 201-218, 2014
– Both individual buildings and alignments composed of buildings.
Figure 4. Naive integration of the meso level “building alignment” in the urban agents hierarchy
In such an architecture, building micro agent generalization may be ruled either by an
alignment meso agent or directly by a block meso agent. The problem is that the behavior of a
block towards an alignment or towards a single building is very different. If we take a look at
the original model through the UML diagram presented in Figure 2, the complexity of the
issue is even clearer as it leads to the the problem of preserving direct access from a block to a
building if there is an intermediate alignment level between them, as this link is necessary to
perform displacement or contextual elimination.
According to these observations, it appears that urban structures such as alignments just do
not fit into the AGENT hierarchy since they are not systematic. That is why re-engineering
the urban AGENT model is needed to allow the integration of this new particular level.
3.2 First Possible Ideas
To integrate a meso level for building alignments into the hierarchy of urban agents, there are
two opposite conceptual models:
– A block is partly composed of alignments, and each alignment activates and rules the
generalization of its buildings, which is similar to the theoretical proposal by Gaffuri
andTrevisan (2004); or
– A block is exclusively composed of buildings, but some of these buildings are part of an
alignment and can activate the generalization of the whole alignment.
In the first solution, the block agent considers its inner alignments and its individual buildings
like equal components. To perform building displacement, the block agent decides which
alignment agent or individual building agent has to be displaced according to how they
overlap. If an alignment is involved in this displacement, then each of its components has to
be displaced. Such behavior is common for building selection, except that if an alignment
generates too high a density it will launch typification on its micro components.
In the second solution, the block agent keeps a direct link with all of its buildings, deciding
whether one of them should be removed or displaced. The difference is that if the considered
TGIS, 18(2): 201-218, 2014
building is part of an alignment, it will refer to its alignment meso agent instead of acting as a
micro agent. Then the alignment agent activates itself and decides if global displacement or
typification should be launched.
As we talk about theory and model, both solutions have advantages and drawbacks, but both
suppose an update of the urban AGENT model. Nevertheless, none of them could be
considered a perfect method. In fact, the first appears to partially contradict the AGENT
model as it can break the direct link between a block and its buildings. This severe breakdown
would probably result in problems in creating a generic model. So, the second idea seems to
be more sensible but still raises the question of considering the alignments as meso agents
while they do not have the main characteristics of a meso agent which would perfectly fit into
the generic agent hierarchy. In fact, the idea is to go beyond the second solution in order to
properly handle a new agent level called “urban structures”. The next part of the article
explains how to reach this goal through a new concept: reactional agents.
4 Principles of Reactional Agents in a Top-Down Organization
4.1 Definition of Reactional Agents
Generalization using MAS usually holds two types of agents. First the agents of the AGENT
model described in Section 2 and which could be called “hierarchical agents” with a top-down
activation, then the agents of the CartaCom model (Duchêne 2003) which could be called
“conversational agents” with a transversal activation. Unfortunately, none of these agents can
tackle the issue of activating a structure through the action of one of its components – like a
building activating its urban structure – which is like bottom-up activation, i.e. a component at
the lowest level should be able to activate a structure at a higher level. The behavior of
contextual agents is entirely top-down, and the behavior of conservational agents is
transversal, both being unadapted to this particular problem. In fact, the activation of a
structure is a reactional consequence of the activation of one of its components. That is why
we propose to call structures “reactional agents”.
The idea is close to the early proposal by Ferber and Müller (1996) who stressed the
reciprocity between influences and reactions in an agent world. Several practical studies
proved how sensible this proposal is, the experiments by Michel (2005) with his IRM4S
model could be cited as an example. To make a parallel with our concrete problem of urban
structures, alignments influence the generalization of buildings – a building is not generalized
the same way if it is part of an alignment or not – but activated buildings trigger the reaction
of alignments. Hayes-Roth (1995) also focused on the difference between deliberative agents
and reactive agents: the first have intelligence capacity to find out when and how they should
be activated, the second only react to external requirements. Reactional structure agents could
be described as reactive agents in Hayes-Roth’s model: an agent whose activation can only be
triggered by one of its components, not by other agents at the same level of hierarchy (other
structures in the neighborhood) or by agents at a higher level (meso agents like blocks or
towns), and that are able to make decisions that will be applied by their components.
4.2 Integration in a Top-Down Organization
The behavior of so-called reactional agents relies on a bottom-up principle – agents at a low
level trigger the activation of reactional structures at a higher level. As soon as this concept is
considered, it leads to the problem of its correct integration in a top-down hierarchical model
like AGENT. Such paradoxical integration is tackled by Picard et al. (2009) in a very
theoretical and generic way, but it needs to be improved and adapted to really fit the present
TGIS, 18(2): 201-218, 2014
issue. Consequently, it is clearly not possible to integrate an independent level for reactional
agents in a very hierarchical and systematic organization like that presented in Figure 1
because the two types of behavior are contradictory.
Figure 5. Integration of reactional structures agents in a top-down agent system
In fact, because of their bottom-up behavior, reactional agents must take place next to the
existing hierarchy. They are activated by lower level agents, and they have absolutely no
activation link towards a meso agent, so they just need to be linked to their triggering agents.
Figure 5 illustrates a generic top-down organization similar to the AGENT structure – two
levels of meso agents, and one level of micro agents. The new level of reactional structure
agents has to be linked with lower levels of micro agents – structure agents just react to the
activation of micro agents, and afterwards micro agents refer to their meso agents to continue
the process if necessary. With such organization, both types of agent activations can co-exist
without any conflict – the hierarchical top-down activation using the classical model, and the
reactional bottom-up activation in parallel. Thus, when a micro agent is activated by its meso
agent, two things may occur:
– If the micro agent is not part of a more complex structure agent, the process remains the
same – the micro agent generalizes itself and then other micro agents are activated by the
meso agent; or
– If the micro agent is part of a reactional structure agent, it refers to this structure that reacts
and controls the activation of each of its components, then the structure stops its work and
the process ruled by the meso agent continues on other micro agents.
This model of reactional agents is built to be as generic as possible. Since reactional structure
agents are not constrained by the hierarchy of agents but live next to it, they are very
permissive in the way they are constructed and they act. A micro agent can be linked to
several reactional structures, like a building being part of two alignments. A reactional
structure can also have micro components controlled by different meso agents, e.g. a building
alignment passing through two neighboring city blocks. Reactional agents could also be
composed of meso agents, not only of micro agents, in order to open up new process
possibilities. All these possibilities highlight the structure agent independence from the
classical hierarchical structure of agents.
4.3 Use in the AGENT Model
To ensure the possible reuse of the model and make it as generic as possible, reactional
structure agents are integrated at a high level of abstraction. In the particular study case of
building alignments, the generic model described in Section 4.2 has to be fully integrated in
the contextual pattern of towns, blocks and buildings described in Figure 2, by integrating
alignment agents in direct link with the lower level of building agents.
TGIS, 18(2): 201-218, 2014
This leads to an update of the original generic data pattern which is presented in Figure 6.
Reactional structure agents are composed of agents which could be either meso or micro.
Therefore an agent can now have a meso agent controlling its contextual activation and also a
structure agent it can activate in reaction. A light direct link is maintained between structure
agents and meso agents to compute the satisfaction of meso agents, but this link is not used
for activation.
Figure 6. Conceptual model (UML) of the upgraded AGENT model integrating urban alignments and structure
agents at a higher level
When implementing this generic organization in our study case, alignment agents are linked
to building agents since an alignment is composed of buildings. So buildings keep direct link
with their immediate meso agents – urban blocks – but have the possibility of being part of a
reactional alignment so that this alignment is taken into account during generalization. The
hierarchical top-down activation is guided by the link meso block – micro building, and the
reactional bottom-up activation is guided by the link micro building – reactional alignment.
Both behavior schemas co-exist and are triggered separately, thus allowing inclusion of
building alignments without any loss on already existing processes like building removal or
displacement.
5 Building Alignments Processing
5.1 Automated Detection
The first step towards complete generalization using reactional alignment agents is to be able
to create them through automated detection of urban alignments in a city. Derived from the
Gestalt theory of Wertheimer (1980), alignment is defined as a general relationship between
several buildings located on the same line, which could be straight or curvilinear.
Christophe and Ruas (2002) proposed a method to detect straight building alignments based
on centroid projection along lines, considering that clusters of narrow projected points
represent buildings that are aligned straight. This process seems to be efficient, especially for
complex blocks with many buildings, but it is reserved for straight alignments.
TGIS, 18(2): 201-218, 2014
Besides, the approach to detect curvilinear alignments is to use minimal spanning trees
(MST), as described by Regnauld (2001). The computation of a MST creates a skeleton
joining up all buildings inside a block. Then the MST has to be split and filtered according to
homogeneity criteria (size, shape, distance, etc.) and the different isolated parts of the MST
are considered as curvilinear building alignments. This method is a little more complex to
parameterise and heavier to process, but it goes further than only detecting straight
alignments.
However, the road network can be used as a means to detect building alignments along the
streets. Such a method was developed by Zhang et al. (2010, 2011) and offers some
advantages among which is good computation time and a way to discriminate alignments
depending on their type – along a street or not. It is sensible to consider this detection step
before using other methods, as buildings are most of the time very close to a street.
Considering these three approaches, they offer different possibilities with various advantages.
Ideally, the best solution is to merge them in a general process which would be able to benefit
from the best results of each component. Such a process tries to combine the three methods
one after the other in three steps: first detection of building alignments along the streets using
buffers on the road network, then detection of straight alignments using centroid projection
and clustering and finally detection of curvilinear alignments using MST computation and
segmentation. At each step, attempts are made to reconnect the new detected alignments to
those already detected if possible, in order to treat particular cases like a straight alignment
being part of a bigger curvilinear alignment. Figure 7 illustrates in detail how the three
situations are addressed by the whole detection process, and some final results are illustrated
in Figure 8.
Figure 7. Illustration of the three steps of the whole detection process
5.2 Reactional Behavior of Building Alignment Agents
The first role of an alignment is to react to an action applied to one of its inner buildings, so
the generalization of each building is ruled by its surrounding alignment. This action is
imposed by the urban block (meso agent) and can be simple activation, removal or
displacement of the current building. In response to this action, the reactional behavior of the
alignment is based on the pattern explained in Section 4.3 and it relies on three principles:
TGIS, 18(2): 201-218, 2014
– If one of its buildings is activated, the alignment is activated instead (no single activation of
the building);
– If one of its buildings is removed, the removal is validated then the alignment is activated;
and
– If one of its buildings is displaced, the displacement is cancelled then the alignment is
activated.
Figure 8. Automated detection of building alignments in a city centre
The activation of an alignment will afterwards correct the unsatisfying constraints that led to
the activation of one of its buildings, for instance by correcting the spatial heterogeneity
introduced by building removal or by performing coordinated displacement of all alignment
buildings. So, urban alignments first act as reactional structures activated by buildings, and
then as independent agents with their own constraints and actions to control the generalization
of their inner buildings.
5.3 Alignment Generalization
Usual methods to generalize building alignment are based on the general structure of the
whole alignment. When an alignment is activated, it will act as a classical agent which
generalizes itself in order to satisfy its main constraints:
– Homogeneity in the building shape, size and orientation;
– Regular distance between each building;
– Preserving the starting and ending points of the alignment, avoiding outcomes that become
larger or shorter than in reality;
– Preserving the proximity to a street if needed; and
– Preserving the general density of buildings.
On the other hand, there is no constraint on the number of buildings to be preserved, i.e. it can
remain the same if there is enough space for all buildings to co-exist, but it can be decreased if
building removal is needed. The only issue is to perform coordinated generalization by
keeping the original layout of the alignment. To solve these constraints, processes have
already been developed based on building removal to avoid overlapping and building
TGIS, 18(2): 201-218, 2014
displacement to preserve homogeneity, for instance by Regnauld (2001) or Hangouët (1998).
Such a sequential process is a good working base but it needs to be transformed into a real
AGENT process with actions driven by constraints.
The final goal of alignment generalization is to preserve the general shape of an alignment
while satisfying density and proximity constraints. To describe this shape, an alignment is
modeled through two characteristics in addition to its inner buildings:
– Its first and last components, to take into account the extremities of the alignment duringthe
generalization process in order to preserve its length; and
– Its general shape which can be described with a skeleton guiding the relative positions of
buildings in the whole alignment, to make sure that the generalized alignment will followits
original layout.
On the basis of these characteristics, there are several constraints applied to an alignment and
several actions to solve these constraints, the goal being to respect the main characteristics of
the alignment during generalization: non overlapping, shape preservation, density
preservation, homogeneous distribution and orientation, etc. The main constraints and actions
controlling AGENT generalization of an alignment are detailed in Table 1, with different
levels of priorities that can be defined by the user. When an alignment agent computes these
constraints and actions, it results in a generalized alignment taking into account each criterion
that leads the theoretical generalization of such a structure.
Table 1. Constraints and actions applied on an alignment agent to perform its generalization
TGIS, 18(2): 201-218, 2014
6 Results
6.1 Implementation
The theoretical model described in this article has been implemented on the agent-based
generalization platform CartAGen. The architecture of this platform, discussed by Renard et
al. (2010), allows to process automated generalization on geographical data using multi-agent
systems, some generalization results obtained on this platform are illustrated by Renard et al.
(2011). The AGENT model is completely integrated in CartAGen and this implementation
has been used as the basis of the present work. The tests that have been carried out focused
mainly on scales like 1 to 25K, 40K or 50K, as the work presented in this article makes sense
at such scales because they demand to take into account urban alignments during
generalization.
Figure 9. Automated detection of alignments on original data at 1:25K(left) – AGENT generalization at 1:50K
without alignments (middle) – AGENT generalization at 1:50K with alignments using reactional structure agents
(right)
Figure 10. Automated detection of alignments on original data at 1:25K (left) – AGENT generalization at 1:25K
without alignments (middle) – AGENT generalization at 1:25K with alignments using reactional structure agents
(right)
Figures 9 and 10 show some examples directly captured from the CartAGen GUI illustrating
the differences between the original AGENT process and the improvement presented in this
article to include urban alignments. Alignments are automatically detected and represented
with a green halo around them representing their maximum theoretical extent at the display
scale by taking into account future theoretical enlargement of buildings. It appears that in the
original AGENT process (in the middle) the spatial distribution of buildings within urban
blocks is partly lost during the generalization process because no coordinated treatment is
triggered on building alignments. On the contrary, with the integration of alignments as
reactional structure agents (on the right), the spatial distribution of blocks is well preserved.
TGIS, 18(2): 201-218, 2014
As a consequence, the generalized map is more precise and accurate, giving a far better idea
of the real situation on the ground. Figure 10 is very interesting as it proves the quality of the
generalization process even when there is overlapping between two alignments. This situation
can often be seen in block corners where one building can be the extremity of two alignments,
so it is crucial that the model be able to tackle this case properly.
Considering the results, the whole process is actually able to perform either typification
(Figure 9) by replacing a larger number of objects with a smaller number while preserving
spatial structure as defined by McMaster and Shea (1992), or caricature (Figure 10) by
exaggerating some characteristics of the alignment like similar orientation and size. The final
result obviously depends on the choice of constraints and actions, and on their priorities.
6.2 Evaluation of Results
The results have been tested through visual inspection by cartographers and generalization
experts at IGN France. These tests show that urban AGENT generalization including building
alignments as reactional agents gives better results at scales larger than 1:40K. Most of the
urban blocks processed at this scale are very well generalized, even those with a high density
of buildings in city centres, and their spatial distribution is completely preserved. This is more
complicated at the scale of 1:50K – if blocks of medium density do not result in any
problems, those in city centres are sometimes unable to give a better solution than the original
AGENT process due to the high density of buildings and alignments. Anyway, during a
generalization process, urban blocks with too high a density of buildings are often combined
as uniform polygons without inner buildings, like central blocks in Figure 3, so in this case
the question of alignments does not really make sense.
Defining a more precise evaluation method than this visual inspection is a complex issue
which involves the creation of special metrics on urban areas. One possibility is to destroy the
relationships “alignment” with generalized data and to re-detect the alignments afterwards,
then to compare the result to the original detected alignment. At a scale of 1:25K in a dense
city centre, no more than 60% of the original alignments are preserved with the original
AGENT process, other ones being destroyed or distorted during generalization. With the
introduction of the reactional mechanism on urban structures, around 95% of alignments are
preserved, resulting in a better description of the city layout. More accurate metrics could also
be used to numerically estimate the quality of the results, like the visual similarity model
proposed by Mao et al. (2010), with the advantage of giving an external and objective
estimation of the quality of the generalization process, but such methods have not yet been
tested on our results. Anyway, we could say that the quality of the results is always guided
and ensured by the choice of constraints applied on the alignments, as was already the case in
the original AGENT model, overall if constraints monitoring are used like presented by
Touya (2012).
The creation of alignment agents adds 10% to 15% more agents on the map in dense urban
areas, resulting in a much longer computation time at the initialization of the process due to
the automated detection of alignments, around five times longer than it used to be.
Nevertheless, considering the process itself the computation time remains substantially the
same as taking into account the alignments often offers optimization while finding common
solutions for building groups instead of considering each building individually.
TGIS, 18(2): 201-218, 2014
7 Perspectives and Further Work
In this article, we proposed a model that seems to give the correct answer to the problem of
the preservation of building alignment distribution during the agent generalization process.
The implementation of the model confirms this conclusion with very good results.
Nevertheless, it still needs some improvement before it is considered as a fully operational
solution, as the tests that have been carried out still point out several problems in dense areas.
The reactional mechanism is not involved since it is the most sensible way to perform
coordinated generalization of buildings being part of an alignment, whereas improvements are
possible for a finer definition of constraints and actions used by the alignment agents. Indeed,
the major contribution of this work is to propose a way to update the AGENT model through
the concept of reactional structures and further work is in progress to better handle the
characteristics of alignments during their generalization.
Moreover, the analysis of the results highlights the fact that in most of the cases where
alignment generalization using reactional agents causes problems, it is mainly because of a
lack of quality in the automated detection of alignments. With a better detection process, most
problems would probably be solved. One way to improve this detection could be to use expert
knowledge to estimate the quality of alignments as proposed by Ruas and Holzapfel (2003), in
order to clean up the detected alignments and remove those which are not significant. Being
able to merge accurate research on automated detection of urban structures with the present
model of reactional structure agents is probably the main challenge that has to be faced to
generalize urban structures perfectly.
As building alignments are also controlled by the block meso level, a block agent is
theoretically able to generalize two alignments that are too close to each other through the
aggregation of the two alignments. This process illustrated by Figure 11 is a classical
generalization operator, but it was impossible in the original AGENT process to deal with the
aggregation of two structures – only micro agents aggregation was triggered, e.g. aggregation
of two neighboring buildings by Regnauld (2001). Aggregating two urban structures is now
possible with the integration of the alignment agent level, which is another important step to
perfect improve urban generalization. This particular operator gives first results that are
encouraging but still needs to be improved and precisely formalized.
Figure 11. Aggregation of two close alignments at 25K if there is not enough space for both of them
Besides, if the model has been tested through the study case of urban alignments, it has been
created as a generic model which could be used in many other cases. The main idea is that the
model can be applied in any multi-agent hierarchical system where intermediate structures
emerge at particular points. Another study case could consider for example the generalization
of two adjacent buildings that have been generalized separately but should have been kept
TGIS, 18(2): 201-218, 2014
adjacent; the creation of a reactional structure “adjacent buildings” should theoretically solve
this issue, i.e. both buildings are still separately generalized but the reactional structure
ensures their adjacency. The model is generic enough to allow such an application. Urban
generalization is obviously not the only application of the model and we could imagine even
more complex applications, not only for generalization purposes but perhaps for other
domains that use multi-agent systems with multi-level hierarchies and intermediate structures.
Such possibilities are at the moment purely prospective, and we are looking for other MAS
domains which could reuse our proposal to test its robustness and its perspectives.
References
Baejis C 2000 Quand les agents jouent aux cartes... systèmes multi-agents pour la généralisation cartographique.
In Proceedings of the 8èmes Journées Francophones d’Intelligence Artificielle Distribuée et Systèmes
Multi-Agents (JFIADSMA’00), Hermès, France: 291–6
Baejis C, Demazeau Y, and Alvares L 1996 Sigma: Application of multi-agent systems to cartographic
generalization. In Meyer J-J C and Tambe M (eds) Intelligent Agents: VIII, Agent Theories,
Architectures, and Languages. Berlin, Springer Verlag Lecture Notes in Artificial Intelligence Vol.
1038: 163–76
Barrault M, Regnauld N, Duchêne C, Haire K, Baejis C, Demazeau Y, Hardy P, Mackaness W, Ruas A, and
Weibel R 2001 Integrating multi-agent, object-oriented, and algorithmic techniques for improved
automated map generalisation. In Proceedings of the Twentieth International Cartographic Conference
(ICC’01), Beijing, China: 2110–16
Boffet A 2000 Creating urban information for cartographic generalization. In Proceedings of the Ninth
International Symposium on Spatial Data Handling (SDH’00), Beijing, China: 4–16
Boffet A 2002 Analyse multiniveau des espaces urbains. Revue Internationale de Géomatique 12: 215–23
Brassel K E and Weibel R 1988 A review and conceptual framework of automated map generalization.
International Journal of Geographical Information Systems 2: 229–44
Christophe S and Ruas A 2002 Detecting building structures for generalisation purposes. In Proceedings of the
Tenth International Symposium on Spatial Data Handling (SDH’02), Ottawa, Ontario: 419–32
Duchêne C C 2003 Automated map generalisation using communicating agents. In Proceedings of the
Twentyfirst International Cartographic Conference (ICC’03), Durban, South Africa: 105
Ferber J and Müller J-P 1996 Influences and reaction: A model of situated multiagent systems. In Proceedings of
the Second International Conference on Multi-Agents Systems (ICMAS’96), Kyoto, Japan: 72–9
Ferber J and Gutknecht O 1998 Aalaadin: A meta-model for the analysis and design of organizations in
multiagent systems. In Proceedings of the Third International Conference on Multi-Agent Systems
(ICMAS’98), Paris, France: 128–35
Foerster T, Stöter J, and Kobben B 2007 Towards a formal classification of generalization operators. In
Proceedings of the Twenty-third International Cartographic Conference (ICC’07), Moscow, Russia
Gaffuri J and Trévisan J 2004 Role of urban patterns for building generalisation: An application of AGENT. In
Proceedings of the Seventh ICA Workshop on Generalisation and Multiple Representation, Leicester,
United Kingdom
Gaffuri J 2007 Field deformation in an agent-based generalisation model: The GAEL model. In Proceedings of
the Fifth Geographic Information Days (GIDAYS’07) Young Researcher’ Forum, Münster, Germany:
1–24
Galton A and Duckham M 2006 What is the region occupied by a set of points? In Raubal M, Miller H J, Frank
A U, and Goodchild M F (eds) GIScience 2006: Proceedings of the Fourth International Conference on
Geographic Information Science. Berlin, Springer-Verlag Lecture Notes in Computer Science Vol.
4197: 81–98
Hangouët J-F 1998 Approches et méthodes pour l’automatisation de la generalization cartographique:
Application en bord de ville. Unpublished Ph.D. Dissertation, University of Marne-La-Vallée-France
Hayes-Roth B 1995 An architecture for adaptive intelligent systems. Artificial Intelligence 72: 329–65
Jiang B and Claramunt C 2002 A structural approach to model generalisation of an urban street network.
GeoInformatica 8: 157–71
Jiang B and Harrie L 2004 Selection of streets from a network using self-organizing maps. Transactions in GIS
8: 335–50
TGIS, 18(2): 201-218, 2014
Lamy S, Ruas A, Demazeau Y, Jackson M, Mackaness W A, and Weibel R 1999 The application of agents in
automated map generalization. In Proceedings of the Nineteenth International Cartographic Conference
(ICC’99), Ottawa, Ontario: 1225–34
Lecordix F, Le Gallic J M, Gondol L, and Braun A 2007 Development of a new generalisation flowline for
topographic maps. In Proceedings of the Tenth ICA Workshop on Generalisation and Multiple
Representation, Moscow, Russia
Mao B, Fan H, Harrie L, Ban Y, and Meng L 2010 City model generalization quality assessment using nested
structure of earth mover’s distance. In Proceedings of the Thirteenth ICA Workshop on Generalisation
and Multiple Representation, Zurich, Switzerland
McMaster R B and Shea S K 1992 Generalization in Digital Cartography. Washington, D.C., Association of
American Geographers
Michel F 2005 Le modèle IRM4S: le principe influence/réaction pour la simulation de systèmes multi-agents. In
Proceedings of the 14èmes Journées Francophones sur les Systèmes Multi-Agents (JFSMA’06),
Annecy, France: 9–22
Picard G, Hübner J-F, Boissier O, and Gleizes M-P 2009 Réorganisation et auto-organisation dans les systèmes
multiagents. In Proceedings of the 17èmes Journées Francophones sur les Systèmes Multi-Agents
(JFSMA’09), Lyon, France: 89–108
Regnauld N 2001 Contextual building typification in automated map generalization. Algorithmica 30: 312–33
Regnauld N and McMaster R B 2007 A synoptic view of generalisation operators. In Mackaness W, Ruas A, and
Sarjakoski T (eds) Generalisation of Geographic Information: Cartographic Modelling and
Applications. Oxford, Elsevier: 37–66
Renard J, Gaffuri J, and Duchêne C 2010 Capitalisation problem in research: Example of a new platform for
generalisation, CartAGen. In Proceedings of the Thirteenth ICA Workshop on Generalisation and
Multiple Representation, Zurich, Switzerand
Renard J, Gaffuri J, Duchêne C, and Touya G 2011 Automated generalisation results using the agent-based
platform CartAGen. In Proceedings of the Twenty-fifth International Cartographic Conference
(ICC’11), Paris, France: 115
Ruas A and Plazanet C 1996 Strategies for automated map generalisation. In Proceedings of the Seventh
International Symposium on Spatial Data Handling (SDH’96), Delft, The Netherlands: 319–36
Ruas A 1997 Generalisation of an urban situation by me ans of road selection and building aggregation.
typification and displacement. In Proceedings of the First ICA Workshop on Generalisation and
Multiple Representation, Stockholm, Sweden
Ruas A 1998a OO-constraint modelling to automate urban generalisation process. In Proceedings of the Eighth
International Symposium of Spatial Data Handling (SDH’98), Vancouver, British Columbia
Ruas A 1998b A method for building displacement in automated map generalisation. International Journal of
Geographical Information Systems 12: 789–803
Ruas A 2000 The role of meso objects for generalisation. In Proceedings of the Ninth International Symposium
of Spatial Data Handling (SDH’00), Beijing, China
Ruas A and Holzapfel F 2003 Automatic characterisation of building alignments by means of expert knowledge.
In Proceedings of the Twenty-first International Cartographic Conference (ICC’03), Durban, South
Africa: 1604–15
Ruas A and Mackaness W A 1997 Strategies for urban map generalisation. In Proceedings of the Eighteenth
International Cartographic Conference (ICC’97), Stockholm, Sweden: 1387–94
Touya G 2012 Social welfare to assess the global legibility of a generalized map. In Xiao N, Kwan M-P,
Goodchild M F, and Shekhar S (eds) GIScience 2012: Proceedings of the Seventh International
Conference on Geographic Information Science. Berlin, Springer-Verlag Lecture Notes in Computer
Science Vol. 7478: 198–211
Wertheimer M 1980 Max Wertheimer, Gestalt Prophet. Gestalt Theory 2: 3–17
Zhang X, Ai T, and Stoter J 2010 Characterization and detection of building patterns in cartographic data: Two
algorithms. In Proceedings of the Fourteenth International Symposium on Spatial Data Handling
(SDH’10), Hong Kong: 267–72
Zhang X, Ai T, Stoter J, Kraak M-J, and Molenaar M 2012 Building pattern recognition in topographic data:
Examples on collinear and curvilinear alignements. GeoInformatica 16: in press