Automation in Construction 20 (2011) 299–310

Contents lists available at ScienceDirect

Automation in Construction

j ourna l homepage: www.e lsev ie r.com/ locate /autcon

Empiric design evaluation in urban planning

Gideon D.P.A. Aschwanden a,⁎, Simon Haegler b,c, Frédéric Bosché b, Luc Van Gool b, Gerhard Schmitt a

a Information Architecture, ETH Zurich, Switzerlandb Computer Vision Laboratory, ETH Zurich, Switzerlandc Procedural Inc., Zurich, Switzerland

⁎ Corresponding author.E-mail addresses: aschwanden@arch.ethz.ch (G.D.P.A

shaegler@vision.ee.ethz.ch (S. Haegler), bosche@vision.vangool@vision.ee.ethz.ch (L. Van Gool), schmitt@arch.e

0926-5805/$ – see front matter © 2010 Elsevier B.V. Aldoi:10.1016/j.autcon.2010.10.007

a b s t r a c t

a r t i c l e i n f o

Article history:Accepted 24 September 2010Available online 8 December 2010

Keywords:Artificial intelligenceProcedural modelingAgent-based simulationSpace syntaxUrban planningDesign evaluationOccupant movement

We propose a system to simulate, analyze and visualize occupant behavior in urban environments bycombining parametric modeling and agent-based simulation. A procedurally generated 3D city model, withsemantic information about the functions and behaviors of buildings, is automatically populated with artificialagents (i.e. pedestrians, cars, and public transport vehicles). In a simulation the built environment and theagents interact with each other. The system identifies empiric correlations between properties such as:functions of buildings and other urban elements, population density, utilization and capacity of the publictransport network, and congestion effect on the street network. Practical applications include the assessmentof a) bottlenecks, b) public transit efficiency, c) accessibility of amenities, d) quality of service of publictransport and the traffic network, as well as e) the stress level and exhaustion of pedestrians. All these aspectsultimately relate to the quality of life within the given urban areas.

. Aschwanden),ee.ethz.ch (F. Bosché),thz.ch (G. Schmitt).

l rights reserved.

© 2010 Elsevier B.V. All rights reserved.

1. Introduction

The majority of the world population already lives and works incities [30]. This influx of new residence puts a lot of pressure on theexisting infrastructure, and on the planning of new and the upgrade ofexisting areas of cities. Cities, like Laos, cannot keep up with buildingthe necessary infrastructure, with the consequence that the qualityof life remains insufficient in general. This is partially confirmed by thelevel of stress experienced by citizens when compared to ruralinhabitants. Even though inhabitants of cities are consuming lessenergy than rural dwellers [40], cities still consume too much energy,produce too much waste and emit too much CO2 for a sustainableway of living. As a result, we are facing the unprecedented challengeof simultaneously improving the quality of life and sustainabilityof cities.

Sustainability and quality of life are both complex matters thatdepend on numerous other, sometimes conflicting, aspects. In thelast century, urban planning patterns put emphasis on path andnetwork optimization for motorized traffic and made drastic changesto the structure of the city. These changes not only impact traffic,they also change the allocation of amenities, land price etc. Adjustingone aspect of the city has an influence on different other equi-libriums within the city. It has become clear that optimizing the

urban layout for pedestrians has a positive effect on the sustainabilityof the urban environment and the quality of life of its citizens. A shiftin the mindset has thus been taking place, with the humanperspective shifting into the focus of attention. Taking the interestsof pedestrians at heart, we present a robust and efficient methodfor simulating and visualizing the related performance of differenturban environment alternatives. This method combines crowdsimulation with procedural city modeling techniques, therebyenabling: a) assessment of the impact of a given built environmenton pedestrians, and b) efficient iterative analysis of different builtenvironments. Such tools empower planners with the means toefficiently investigate subtle ways to adjust the urban fabric. Ourautomated method also offers added value for the entertainmentindustry. It delivers high quality output imagery through standardproduction pipelines and decreases the workload to generate theurban layouts. These are simulated as realistic urban environmentsinhabited by virtual occupants. Traditionally, costs and time neededto produce populated digital urban sets for movies or games areenormous.

The rest of the paper is organized as follows. Section 2 reviewsrelated work in the field of city modeling and urban simulation,Section 3 presents the proposed system for the simulation ofpedestrians within a city environment. We also assess the impact ofthe built environment on pedestrians, and vice versa. We introduceour city model in Section 4 and the semantic data in Section 5. Howthe city model and the semantic data are affecting the agents isdescribed in Section 6 and the in depth study of their agents is inSection 7. In Section 8 the performance of the proposed system isanalyzed through three examples.

300 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

2. Related works

We organized this part in Crowd simulation (Section 2.1), Urbanplanning (Section 2.2) and Pedestrian movement in urban planning(Section 2.3), and Functional urban models (Section 2.4).

2.1. Crowd simulation overview

Models for crowd behavior have been an active research field sincethe late 19th century (e.g. see [4]). Today's computer simulationmodels have a relatively young history. Most relevant approacheshave been realized within the last 20 years and are specialized todifferent fields. A good example is Reynolds [20] flocking method,which uses particle systems and represents one of the most commonapproaches for simulating group movements.

For multi-agent and large crowd simulation, several techniqueshave been proposed to animate or simulate large groups ofautonomous agents and crowds. These methods can be classified infive groups [22]:

• Potential-based methods: Pedestrian-agents are modeled as particleswith potentials and forces [9].

• Flock-like methods: These approaches were introduced by Reynolds,creating simple rules like separation, alignment, and cohesion forcomputing the velocities [21].

• Geometric method: The aim of these approaches is to computecollision free paths. They either integrate the velocity space or useoptimization methods [6,31].

• Field based methods: These algorithms generate fields for agents tofollow, or generate navigation fields for different agents based oncontinuum theories of flows or fluid modes [15,27,41].

• Least effort crowds: These algorithms compute the paths of crowdagents using Zipf's [39] principle of Least Effort [39]. Recently thesehave been combined with collision avoidance algorithms andemergency behaviors for a large number of agents.

Brooks [42] provides a comprehensive foundation on whichmany of the recent agent models and theories are based. Hedescribes many failing artificial intelligence approaches to set upintelligent agents. Musse and Thalmann [14] introduced a moreflexible model with hierarchical behavior. Physics and body effectshad been described by [9] to simulate escape behavior and panicseffectively. In other fields, like robotics [32], safety science [26], andsociology [11], similar approaches have created simulations involv-ing groups of individual, intelligent units. For a more comprehensivedescription of agent-based pedestrian movement, we refer toThalmann and Magnenat-Thalmann [12]. Furthermore, Hillier [10]introduced the idea that a city and spaces in general can be dividedinto components, that can be analyzed as a network of choices andthat can be represented as maps and graphs. Penn and Turner [18,19]described urban agents within their space syntax system. Parallelsystems evolved in computer graphics to populate urban environ-ments in real-time [18,19,29].

2.2. Urban planning

Current planning methods are based on the experience of theplanner and endwhen the architect finishes the building. This methodis insufficient for future needs, changes, and increased pressure on theperformance of urban configurations. There is a demand for simula-tions that incorporate awider range of scales and interdependencies—from building codes to the level of the entire city. They should helpto evaluate and predict the impacts of planning efforts and the changesover time. Single layer optimizations, such as emergency evacuationor traffic simulations, exist in many forms and are widely availablecommercially. These tools incorporate a very limited number of

parameters and either focus on the building performance, neglectingthe interdependencies of buildings, or work at the scale of ametropolitan region, using a grid with a scale compiling whole areasof a city.

Recently, there have been approaches to combine behavioral andstructural modeling [33,34]. These approaches treat the behavioralstate as static input. For example, they try to find the equilibriumbetween street access and the housingmarket, an iteration of this loopis still far from being an integrated, interactive system, or dynamic.A closer integration of function, behavior, and state will reducecalculation time and allows a practical integration into the earlydesign phase [28].

2.3. Pedestrian movement in urban planning

There is an effort in transportation planning to include pedestrianmovements and to move from car- to pedestrian-oriented trans-portation planning. However, this development is currently stillhampered by a lack of empiric simulation methods. Current methodsfor pedestrian movement analysis are targeted towards early designstages and thus treat pedestrians more like a statistical input:common parameters are land use and modes of transportation.

According to the classification of crowd simulation methods(Section 2.1), commercial software tools for pedestrian flow analysistend to either use a particle system with a social force model,neglecting spontaneous decision processes, or incorporate simulatedmovements of crowds, but this requires knowledge of the positionof every source and sink, which has to be manually incorporated.

2.4. Functional urban models

Wegener [37] proposes urban models, in which he definesinteractions between different entities such as land use, networks,population, house, employment, and transportation of goods, andthen uses the model to compare existing operational models. Wu andWebster [38] developed a model integrating multi-criteria evaluationwith cellular automata to simulate land use and its changes. Cellularautomata include the special features of traditional urban models andcapture the spatial features of the urban fabric. Artificial intelligence(AI) has been adopted to use a knowledge-basedmethodology, so thatmodels can be constructed in a compositional way, and to providethe ability to simulate decision-making processes of a single agent or agroup of agents [3].

Waddell and Ulfarsson broadly define urban simulation as“operational models that attempt to represent dynamic processesand interactions of urban development and transportation” [36].Their introduction gives a good overview of specific techniques (suchas cellular automata or multi-agent systems) that have been suc-cessfully implemented in a state-of-the-art urban simulation system,UrbanSim [35]. Traditionally, the tools operate on the level of regulargrids considerably larger than individual building lots. More recentapproaches overcome the rigid grid-cell model by taking morerefined scales into account, typically starting from parcels to zones[43].

3. Overview of the proposed approach and contribution

In Aschwanden et al. [2], an occupant simulation method wasintroduced, which is extended by the present work. The proceduralmodeling technique, on which our urban model is based, was initiallypresented by Parish and Müller [16,17] for the modeling of cities.Muller et al. [13] extended it for the modeling of buildings, and it hasbeen applied to the context of urban planning by Ulmer et al. [25] andHalatsch et al. [7,8]. Generative city models not only allow adjustingand updating the physical representation of the city and its streetlayout or a single building, but they also incorporate metadata about

301G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

the usage, construction details and capacity of most of the cityelements. With the knowledge of floor space, function, and location(e.g. city, distance to the centre) of every building, we are thus able topredict the amount of traffic (e.g. cars, pedestrians, etc.) everybuilding generates and absorbs over the course of a day. For example,the number of pedestrian-agents generated by a given building-agentrelates to its size, function and location. The adaption of methods fromartificial intelligence enables to also incorporate behavior: building-agents are used to represent not only the buildings according to themaster plan but also their adaptive functions. Every building-agenttherefore has some meaning to other agents. Every pedestrian- andcar-agent has an individual set of preferences towards specific placesand amenities and has individual abilities, like strength, endurance,speed etc., to reach them. On the other hand, each point-of-interest isadaptive. For example, a parking lot loses its attraction for car-agentsas soon as it is occupied — empty parking lots will regain theirattraction again. Moving agent has a major goal and visits other placesalong the way. These agents find their way through the city andconstantly interact with moving and immobile agents. Their experi-ence is color-coded in an individual line, analyzing their path towardsinternal and objective attributes, ranging from empiric, e.g. total timeof travel, to personal, e.g. effort to achieve all goals. Each agent thendraws an individual path on the ground, color-coded according toa set of attributes specific for the user, e.g. exhaustion, amount ofpersonal space etc. It is important to detect patterns in the problemsexperienced by different agents. If some problems occur repeatedly,they may very well be intrinsic to the design [24]. We can measure‘the level of exhaustion’ of agents, which then enables us to put publictransport stations or resting areas accordingly. To avoid unnecessarilylong traveling distances for an agent, changes to the traffic guidancesystem or the street network itself may be investigated.

3.1. Overview of the proposed approach

The proposed approach is based on agent-based crowd simulation(the ‘Massive’ software package). At the centre of the simulations arepedestrian-agents that are created in relevant locations in the city andattached with goals (final and intermediary) and behavior models.The generation locations and the goals of the agents are automaticallyestablished based on the functions of the buildings, which constitutethe city model. The city model is augmented with a public transportnetwork with mobile agents (e.g. buses). These agents are alsoprovided with goals and behavior models. They impact the simulationthrough their predefined services and their limited capacity foragents. Finally, the system simulates car traffic to take into account itsimpact on both the pedestrian and the public transportation trafficflow.

In order to make the intended simulation feasible and realistic, adetailed and semantically rich city model is necessary. For instance,the city model should clearly specify building functions (used todefine the building capacity to generate and absorb pedestrian-agents) and street characteristics (i.e. sidewalk size, number of roadlanes, crossroads characteristics). Furthermore, as the ultimate goal ofthis work is to support urban planning and thus to assess the impact of

Fig. 1. Workflow for design iteration

the above characteristics on the overall traffic, it is important thatthese characteristics can be easily changed. This motivated the use ofgenerativemodeling. In our approach, we use the CityEngine softwarepackage to parametrically create city models. The digital city modelis not limited to the 3d representation, but includes semantic infor-mation about the buildings and the street network.

3.2. Contribution

There are multiple contributions in this work. First, the relation-ships between all urban agents are more empirically modeled. Thepedestrian-agents are attributed with dynamic (i.e. adaptive) behav-ior models with preferences for intermediate and ultimate goals andthe individual abilities (such as strength, endurance, speed etc.) toreach them. The goals themselves are given adaptive behavior modelsthat depend on internal and external dynamic parameters. Theseresults show a dynamic interaction of the urban environment and thepedestrian-agents.

The second contribution of this work is a method to create citymodels, containing physical representations of the city and detailedfunctions of the interaction that can be used in a simulation softwarepackage. This significantly impacts the integration and the level ofautomation of the overall simulation process. The city models can beeffectively created from existing information such as master plans,street networks, and elevation maps.

The third contribution is a method for automatically defining thelocations where agents must be generated from the semanticinformation contained in the city model imported into the simulationplatform. This further boosts the efficiency and robustness of theoverall simulation tool.

Finally, we provide a means to visually assess the performance ofthe simulated urban environment, enabling urban planners to identifyareas where improvements could be made and to test alternativesolutions. The generative parametric city model and the level ofautomation achieved with the proposed system are critical to enablethe planners to effectively and efficiently conduct such an iterativeprocess.

4. Urban model

We organize the built environment of the city in 2 layers: I) thephysically built environment and II) the semantic distribution offunctions. Both layers are dynamic but have different time scales, e.g. abuilding can change its function several times throughout its life cycle,the same function of a building is used with different intensitiesduring the course of the day etc. Therefore, we interpret the term‘city’ as a complex, distributed, interconnected and rapidly changingsystem that can be understood as a fabric of space depending ona) people, b) function, c) space, and d) physical environment. Eachof those aspects is typically influenced by a number of variablesand forces. As urban planners, our aim is to try to control both thephysical layout and the function contained in the urban system.

In the case of a complex city space, the population may eitheraccept the designers' work or not, but only the realization and use of

s (input–output based systems).

302 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

the designed urban space will actually show whether the urban planwas successful. In urban planning there exist many importantcorrelations. These call for a balance, which must be reassessedconstantly. In order to start the simulation we are required a) tounderstand the given situation and the aim of the planer, b) to identifythe necessary semantic metadata and c) to incorporate the insightsinto the procedural 3D model and the agents' model. The goal is toestimate as accurately as possible the acceptance of the givenconfiguration of a city by the inhabitants.

4.1. Density

The classification of ‘urban density’ is still disputed. The factors bywhich to classify density are not agreed upon in the first place. Inaddition, the density value does not have a clear unit. Some urbanplanners consider the number of occupants per square kilometer auseful figure. Another useful figure may be the number of carsrecorded between urban centres, but these two figures have alreadyvery little correlation. Therefore, a major achievement is a correlationbetween the built environment, number of people and traffic. Ourcontribution is a first empiric explanation of this correlation. We usethe existing parcels, together with those being developed, to predictthe number of people likely to be occupying that area. Othercalculation methods, like Arnott and Small [1], use functions of thedistance from a virtual centre or a grid overlaid on the city. The ‘grid-cell’ and ‘virtual-centre’ methods are crude and only produce usefulresults in large metropolitan areas.

In our approach (Fig. 2), we want to know how many people arecoming out of a building and arewalking in the public space. The threemost important aspects are: a) the floor size of the building, b) thefunction, and c) the location of the building. These factors enable us topredict how many people are coming out of a building at a certainmoment in time. Hence, the availability of statistics about such factorsis the key to derive comprehensive results. The ‘Bundesamt fürStatistik’ of Switzerland provides us with the statistical input data ofevery municipality in Switzerland.

We used the Swiss census data provided by the government todefine the available floor space per person in a building living areaas well as the space requirements for offices and other workspaces.The overall number of occupants is not decisive enough, only incombination with a time-dependent fluctuation, we are able to knowhow many people are moving in and out of the building at a specifictime.We automated the calculation and specified it according to everycommune in Switzerland. This was necessary due to the high disparityof the input data. For residential areas we used the floor space of theapartments and divided it by the number of residents to get theaverage space available to every resident. The combination of net floorspace, available space, location, and the fluctuation over time allowed

Fig. 2. Human locator (input–output based calculation method for th

us to predict the traffic produced by every building. The availableaverage floor space per resident varies in Swiss cities from 35.5 m2

(Zürich) to 55.7 m2 (Maisprach). Both numbers are higher than thespace provided for office workstations from aminimum of 4.46 m2 forsecretarial workstations to 27.89 m2 for a vice-president office. Thisshows that the traffic generated by office buildings is much higherthan the one generated by residential buildings in general. It isalso more volatile over time. We note that other places, such asrestaurants, cinemas, as well as hospitals, produce even highernumbers at their peak hours, but it is difficult to define them. Thelimits of such automated calculation are reached in cases where thegenerated traffic is not a stable function, such as a stadium emptyingitself within 30 min twice a week or when the number of amount ofbuildings is too small.

4.2. Urban environment

Most recently, research in architecture (and subsequently com-puter graphics) has produced a number of production systemsfor architectural models, such as Semi-Thue processes, Chomskygrammars, graph grammars, shape grammars, attributed grammars,L-systems or set grammars [33,34]. All of these methods exposedifferent application possibilities and levels of efficiency to the user.The shape grammar concept has recently been made more applicableto computer graphics and daily usage [13] and is now commerciallyavailable in the software package “CityEngine”, which is used in thispaper [23]. The key tools of the context-sensitive shape grammaras implemented in the CityEngine consist of: shape operations formass modeling, component splits for the transition between massand facade, split operations for building facades, and spatial queryoperations.

The workflow to create a virtual urbanmodel for crowd simulationis similar to the workflow used by architects in urban planning.Based on a number of maps (color-coded images, raster data or vectormaps), parametric rules are triggered to create street networks andvolumetric models of the buildings. These volumes are used todistribute urban functions and building density and to guide the agentnavigation as described in the previous section. Based on these roughbuilding volumes, we create the final detailed geometry needed forvisualization and the low-polygon geometry used in simulation. Theoriginal aspect of our workflow is the use of procedural modeling toautomate the creation of street networks, building geometries, andsemantic data. A big advantage of this is the adaptability — the urbanenvironment is able to change for each simulation without redrawingit from scratch. We automatically export the data set for thesimulation and the visualization, which differ radically from oneanother. They both are stored inside the same procedural scene

e prediction of the number of people coming out of a building).

Fig. 3. Color-coded attribute maps are used to control the generation of street networks and building geometries. a) Topology b) obstacles c) height map and d) building heights(skyline).

303G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

description and will therefore always be in sync. The followingsections describe our workflow in more detail.

4.2.1. Creating the urban layoutWe use a fictional scenario in the area of Zurich, Switzerland to

describe ourworkflow. In this example,weuse differentmaps (Fig. 3) toencode aspects of the urban layout (e.g. topography, elevation,obstacles, skyline, and land usage). To create the street network, wehave two options: (1) generate a generic, procedural (i.e. rule-based)network which follows the attribute maps (e.g. along the colorgradients) or (2) import a network from a vector data source (e.g.from openstreetmap.org). For simplicity, we use a generic streetnetwork. The algorithm is an extension of the L-system based streetnetwork generator described in Parish and Müller [16,17]. The buildingparcels can be created in a similar way: (1) by a generic subdivisionalgorithm or (2) by importing a vector data source. We use genericparcels, also created with the algorithm described by Parish and Müller[16,17]. Once thebuildingparcels have beengenerated, procedural rulescan be applied to create the building and the street geometry.

4.2.2. Procedural modeling using shape grammarsWe now show how the shape grammar, as implemented in the

CityEngine, is used tomodel some simple building volumes for the Zurichscene. Each building lot (parcel) is assigned a shape grammar rule set. Arule set consists of production statements in the form (Fig. 4):

With this syntax the successors can be composed of several shapeoperations and query statements. Listing 1 contains a small rule set

Predecessor→ [case Condition1:] Successors 1 [case

example and Fig. 4 shows a possible result. Fig. 5 shows the result ofthe application of the rule set example to the complete scene.

5. Generation of the control data for crowd simulation

Our crowd simulation setup needs four types of input from the citymodel to initiate the simulation:

1. Simplified building and street geometry to visually guide theagents in order to prevent collisions with the built environment.

2. Locators for initial agent placement and points-of-interest toimplement building functions.

3. Directed (color-coded) street lanes to guide cars and buses.4. Terrain geometry with a texturemapwhere additional color-coded

control data is stored. For example, we are using color intensity inthe door areas to encode the capacity and function of a building.

Next we describe how these channels are derived from theprocedural urbanmodel. For this we need to extend our grammar ruleset with an optional set of rules, which trigger the generation ofcontrol data for the simulation.

Upon export, CityEngine scans the names of the terminal shapes ofthe grammar generation process (in the U-shaped building exampleabove these were called ‘Facades’) and triggers the creation of one ofthe input data types for the crowd simulation based on a set of namepatterns. For a typical agent simulation scenario the relevant terminalshapes are usually called ‘door’, ‘window’, ‘sidewalk’, ‘street lanes’, etc.Fig. 6 depicts the data-flow between the CityEngine and the crowd

Condition2:] Successors 2 [else:] Successor N.

Fig. 4. Left: A simple U-shaped mass model consisting of three volumes (shapes). Right:listing 1. This “CGA shape” source code produces simple U-shaped mass modelsconsisting of three shapes. Some parts have been omitted for simplicity.

304 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

simulation tool (expansion of Fig. 1). The necessary modifications toan existing grammar rule are:

• Based on the simulation scenario and the simulation tool (in ourcase “Massive”), choose the appropriate control features (vision,sound, colors, vector fields, etc.) that trigger certain actions of theagent (walk, stand, wait, etc.).

Fig. 5. On the left, the resulting scene is shown with the same simple set of shape grammar rarea in the centre.

Fig. 6. Illustration of the data-flow between the CityEngine (left) and the crowd simulation tautomatically from a single grammar rule set.

• Identify the parts of the urban environment with which the crowdagents interact (e.g. doors, sidewalks, etc.).

• Introduce rules, which trigger certain simulation features notnecessarily visible when rendering the scene (e.g. a locator on thebottom-face of a building entrance). These rules are dependent onattributes to comfortably switch between model generation forvisualization and model generation for simulation.

• Introduce conditional rules to deactivate all geometry not needed inthe crowd simulation.

For our final results, the previously mentioned modificationsresulted in the following control data for the simulation readable bythe agents (see Fig. 7):

• Spline segments for street lanes and turnings on crossings. We usethe underlying street graph connectivity to connect different streetlanes on crossings.

• Locators for agent sources: e.g. all terminal shapes called ‘door’trigger the generation of a locator and a colored area on the bottom-face of the doors. The color is used to encode the amount of agentsemitted from this door.

• Locators for points-of-interest (POI): the doors trigger a soundsource to attract agents. We distributed n different frequenciesfor the n different POI according to the master plan. If necessary,the distribution of the POI can also be controlled by grammarrules.

• Locators for “background” agent sources: the shape grammar splitsthe sidewalks into small stripes and assigns agent sources to themaccording to a certain probability.

ules applied to all building lots. The figure on the right shows a close-up of the high-rise

ool (right). Note that all the necessary input data to the simulation have been generated

Fig. 7. From left to right: (1) detailed geometry used for visualization. (2) The rough volumes used for agent vision in the simulation. (3) The blue/red marked areas show the ‘door’terminal shapes used to generate agent source locators. (4) Volumes (gray), color-code (red), and locators (yellow) imported into the simulation tool (“massive”).

305G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

6. Agents

In order to test the usability of the semantic 3D city models, thesimulation is conducted with agents representing pedestrians,buildings, cars, and mass transit. Every agent also acts as a measuringdevice of the city, using its perceptive channels to learn about theenvironment and report back its simulation history. The agentsconsist of perceptual, behavioral, and cognitive components incombination with a physical body to evaluate the design and itscorrelating semantic 3D city model. The agents have multiple,perceptive channels. They are able to see and hear their environmentand read information encoded on the terrain of the 3D environment(usually using color-codes). These abilities are exploited by eachindividual agent at a frequency of 25 times per simulated second.According to their evaluation, they emit sound, tag the ground, andchange their color. This enables agents to interact with theirsurroundings, their built environment, as well as with other agents.This interaction is dynamic. For instance, building-agents lose theirattraction for cars and pedestrians temporarily if they reach theircapacity. A secondary level of interaction is needed in some cases,like between buses and pedestrians where a bus stop indicateswhether a pedestrian-agent is waiting for the bus.

6.1. Sensory channels for perception

We are using multiple sensors as receptors for each agent tounderstand its environment. Like humans, the agents have a limitedset of senses through which to experience the environment. Thefollowing steps of filtering, selection, rating, and simplification havebeen incorporated into each agent's brain to enable it to operatewithin and interact with their surroundings.

6.1.1. VisionVision measures the distance, orientation, and color of the agent

(e.g. head) towards all geometries around it. Some agents are usingvision for collision detection. The definition of distance (vision.z) inrelation to the agent's direction (vision.x) and height (vision.y), incombination with the color (vision.h) of an object or other agentallows the agent to define if an object is or is going to be in the planedpath.

6.1.2. SoundSound always carries a frequency and is limited to a distance. Each

agent who perceives the sound is able to determine the direction(sound.x), distance (sound.d), frequency (sound.f), and emissionradius and amplitude (sound.a) of the sound source.

6.1.3. GroundThe ground consists of two parts, geometry and texture map.

The geometry is used to determine the global height of each part

of the agent (ground.z). The ground color can be read absolutely(ground.r/.g/.b) as well as its change (ground.dr/.dg/.db).

6.1.4. LaneFor streets we use lanes, consisting of a splinewith width and color

(lane.h). The agent can determine its absolute position on the lane inrelation to the centre (lane.x), the deviation of its own angle towardsthe angle of the spline (lane.ox), and the rate of change (lane.ax).

Every channel has specific implications for the agent. Reading theground allows the pedestrian-agents to avoid collisions or to changetheir perceptive model, e.g. while crossing the street the agent ispaying more attention laterally, where cars are likely to appear.

Vision is computationally very expensive but offers numerousadvantages, for instance by putting information also on walls foragents to read. Sound, on the other hand, is computationally efficientand therefore used as the common interaction channel between allagents.

6.2. Set of goals/points-of-interest (POI)

Several analytical tools have been published which describe thecorrelation between the built environment and the movement ofpeople inside it, and use the urban environment [18,19]. In ourapproach this urban environment is not static. It is a dynamicenvironment reacting to its surrounding, adapting to changes in theflow of occupants. Each pedestrian, car, and mass transit agent has itsindividual set of preferences in relation to the specific places it wantsto visit. These places are usually workplaces, home, and also the placesvisited in between. By compiling such information we reducebuildings to symbols of their functions. As presented before, weincorporate the allocation of these points-of-interest (POI) into therules of the procedurally generated city. These rules can be triggeredby a color map representing the master plan of an area consistingof different functional layers (e.g. living, working, industry). Eachbuilding is defined as an agent that emits a specific sound to representits function, and thus attracting other agents. The capacity of abuilding-agent is determined by its size, position within the city, andfunction. If the capacity is reached, the building-agent changes itssound frequency until at least one agent leaves the building.

6.3. Decision process

Pedestrians and car-agents are equipped with preferences and adecision process to determinewhichpoint-of interest to visit first. Usingthe sound channel the agent is able to define the distance to the point-of-interest. The overlay of distance and preferences gives the agent thepoint-of-interest he wants to visit first. If the point-of-interest is losingits attraction, the decision is recalculated, possibly leading the agent to adifferent point-of-interest.

Fig. 9. Input–output details of building-agents.

306 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

7. Specific agents

7.1. Pedestrians

A unique set of personal characteristics and abilities for each agentallows the crowd to realistically represent a standard population. Bodyand brain of each agent are entities enabling the agent to interactwith thebuilt environment and other agents in a distinct way according to itsfeatures and abilities. Virtual agents evaluate their path through the citybased on empiric indicators (e.g. Level of Service (LoS), time of travel etc.)aswell as on secondary individual indicators (e.g. exhaustion, clarityof thecity layout, available personal space, stress etc.), which are otherwise onlyavailable with time intense and expensive surveys.

Agents are designed to prefer walking on the sidewalk, i.e. to avoidtraffic lanes. The agents read the ground and understand the color-codes(Fig. 8). As simplified replicas of humans, the agents are primarilycontrolled by their visual input for collision detection, which is anoverruling factor [5]. A change in color of an agent is used to communicateits state to other agents, allowing typical emotions to be conveyed insituations such as when a collision occurs. For instance, red representsanger and blue indifference. Apart from this visual input, agents use thefrequency and the amplitude of sound to guide them towards distantlocations or points-of-interest, which are out of sight.

7.1.1. Tracked outputThe artificial intelligence conveys information about the effort it took

the agent to reach the designatedpoints-of-interest. The initial strength ofan agent is reduced faster by walking upwards than downwards. Theagent is also equipped to measure the time and determine if his journeywas too time consuming according to the distance it had to walk. Thevision and distance measurements define howmuch personal space wasavailable to each agent during the journey.

7.2. Buildings/POI

Building-agents areprovidedwithaposition, function, andcapacity, allencoded in the parametric model. The building-agent is emitting a soundfrequencywithin a defined radius. This frequency encodes its function forother agents.When it has reached its capacity, the sound radius is dampedand the frequency adjusted. The building-agent receives the soundfrequency from all other surrounding agents (e.g. pedestrians). Bycomputing the distance to other agents, the building-agent is able toknow whether these agents have crossed the door and are inside.Building-agents are also sources of pedestrian-agents, according to thetime of the day each building-agent spawns other agents (e.g.pedestrians), which then start to interact with the environment (Fig. 9).

7.3. Individual transport/cars

Cars are bound to the streets, which are represented by color-codedlanes. At its creation, each car-agent decides on a goal. Guided by thesound emissions, a car-agent is able to define its distance to each goal.Based on its own preferences, it then decides where to go first (Fig. 10).

Encoded in the lane is also the kind of street, limiting the maximumspeed for the car-agents. But the agent is also aware of all theother agentsand is measuring the distance towards each of them. Pedestrian-agentsare recognized when crossing streets, which leads to braking and

Fig. 8. Input–output details for pedestrian-agents.

stopping. Other car- and public transport-agents are also monitored.Based on their locations and speed, each car-agent adjusts its speed inorder tomaintain a safe distance to these other vehicles. This also leads totraffic density fluctuations and ultimately to congestions.

The path finding process is a ‘greedy function’, where the agent istrying tominimize its distance towards the goal. Because each agent isbound to the street, it can only change its path at a crossing, where itdecides upon arrival to turn in the direction of its goal. For ease ofinteraction between pedestrian-agents and car-agents, traffic lightswere introduced, restricting one direction for the car-agents and theother for the pedestrian-agents.

7.3.1. Tracked outputWe used the output channels to gain information about I) their

absolute deviation from the maximum speed II) their accelerationsand III) their deviations from a direct line to their goal. This enables usto identify streets with a high congestion risk and the demand foralternative routes or more lanes.

7.4. Public transport-agent/bus

Similar to the car-agents, public transport-agents are bound to thestreet, but they also have a strict route to follow with stops. Thesestops are only taken if pedestrian-agents are actually present at thebus stop when the agent is arriving. Also the speed and accelerationare different from those of cars, to take the different motion patterninto account.

Each public transport-agent is aware of its total and currenttransport capacity. Additionally, higher numbers of pedestrian-agentsgetting off and on at a bus stop lead to longer waiting times at thestops (Fig. 11).

7.4.1. Tracked outputWewere interested in the capacity level at each stop and the actual

time each public transport-agent needs to go from one stop to thenext or to perform the whole route.

8. Results

8.1. Output by agent

During the simulation, every agent draws an individual color-coded line on the ground. The color change, triggered by the agentbrain, is representative of both its emotions and its empiricalevaluation of the environment. Every agent is unique.

We have to note that a single agent is not representative for realhuman behavior, but if a significant number of agents appear to beexhausted at the very same spot, we know that we have to adjust.

Fig. 10. Input–output details for car-agents.

Fig. 11. Input–output details of bus-agents.

307G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

Also, if the net travel time appears to be too long, we assume that theagents would rather take public or private transport than walk. Withthis information we can adjust the stops of the public transportationsystem to increase the time efficiency for the majority of occupants.

8.2. Dübendorf example

Dübendorf, a suburban centre of Zürich, currently houses amilitary airport, which will be phased out. The area in question isthe biggest to be developed in Switzerland in the near future. With anincreasing trend of living in suburban areas, there is increasedpressure to develop the area into a compact new centre for Dübendorfand surrounding areas. We divided the results in two parts: apedestrian based evaluation (Sections 8.2.1 and 8.2.2) and a streetagent interaction. Both have a similar underlying structure andrepresent the result of our workflow.

8.2.1. Primary design evolutionIn this case we can show the density differences between the area

of high-rise buildings in the centre and the low-density living quarterson the outskirts. There is no absolute rule about density or traffic— forsome functions an increased flow of occupants is favored, while forothers this is not the case. We show in Fig. 12 the layer of personalspace. The problem we identified here was that the entrances/exits ofthe buildings were opposite to each other. They would be betterplaced in an asymmetric way to avoid collisions.

To handle the amount of agents in this case, we group agents intothree major groups [11]. The group which is coming from home, thegroup which comes from work, and the group which has no specificgoal to reach but occupies the streets as background noise.

Fig. 12. Left: Quantified visualization of 6800 agents within the given area. Middle and

8.2.2. Secondary design evolutionThe setup consists of a pedestrian crossing with typical activities

and a street that represents an obstacle to the agents. This time theagents find their way from their source to their final destination bymaking a selection from a list of possible secondary goals andmarking their way within their individual path. As a result we cansee increased traffic around the high-rise buildings as well as thecrossing of the streets in the north (Fig. 13). As seen in the collisionmap, there is no stress problem at the street crossing. However,even with the same amount of pedestrians, there is a problem in themiddle of the walkway. This is due to several reasons. One reason isthe agents' flocking behavior, making the agents align themselvesthere, after increasing the amount of iterations. Additionally, if anagent is making decisions in that particular area, it is likely that itwill change its goals. The agents' evaluation shows more collisionand stress with two asymmetric pedestrian crossings. In orderto find the next crossing, they can easily block each other's path.The total-time-spent map reveals the need for another publictraffic stop at the south end to even out the time needed to reach allgoals.

8.2.3. Third design evolutionIn this case we focus on the evaluation of the street bound agents.

We extracted only the maps drawn by the cars or buses. In this casewe also would like to show the difference over the day. The caseconsists of two areas, to the west, there are bigger volumes with workfunctions and corresponding secondary functions, such as cafeteriaetc.; in the east area, there are residential buildings with lower heightand capacity.

In this case we show a single bus line and the utilization of itscapacity (Fig. 14). The bus is turning clockwise and absorbs agentstraveling further than typical walking distance. In general themorning (07:00–08:00) and evening (17:00–18:00) simulationsshow a higher usage over the whole loop.

In this case we show that the morning traffic is dissolving over 2 h(Fig. 15). These images are part of a dynamicmap showing the densityfluctuations. In Fig. 16 the same agents are also showing where theirpath deviates from the optimum.

right: Occupant stress analysis; dark dots represent not enough individual space.

Fig. 13. High-density crossing from left to right: density, collision, and time map.

308 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

9. Impact on planning decisions

Our simulation method enables the urban planner to adjust thedesign on an arbitrary scale. Each evaluation has different planningimplications (e.g. allocation of public transport, expanding sidewalk,and allocation of functions).

9.1. Pedestrian-agent

The density maps have implications on the functional layer of thecity — commercial function welcomes a higher rate of pedestriantraffic but residential function might be better placed in a differentarea of the city. The collision map shows where the stress level forpedestrians is increased and where the sidewalks have to get wider.The time map has an impact on another system, which defines theposition of the public transports stops.

9.2. Bus-agent

The evaluation of the bus-agents has to be analyzed at differenttimes of the day. The congestion effect within the bus-system

Fig. 14. Time-depende

Fig. 15. Congestion effects of cars in

becomes visible. The efficiency is varying for each individual busand has implications for providing additional buses for a specific timeand area of the bus-system.

9.3. Car-agent

The most profoundly influential part of a city is its street network.Therefore an initial setting of the streets has to be done with a lot ofcare. But the knowledge of congestion effects also allows redirectingtraffic in an existing configuration prior to a planned change, e.g. aconstruction site.

With a limited amount of resources, this allows the urban plannerto choose the smallest possible intervention to gain the best result.Small obvious changes are more likely to be accepted by the occu-pants and therefore have a higher chance of creating a permanenteffect on the city. These evaluations might lead to contradicting needsallowing stakeholders to make a decision reflecting their opinions andto communicate their decisions.

The high-resolution pictures and movies (Fig. 17) are then amethod of communicating to all stakeholders, supporting a discussionabout different design ideas and ways to address problems.

nt use of the bus.

the morning traffic over time.

Fig. 16. Deviation from optimal path in morning traffic over time.

Fig. 17. Detailed graphical output for communication.

309G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

10. Discussion

The multi-faceted results allow the different stakeholders (likeurban/traffic planners or politicians) to discuss the implications ofpossible changes. Even with useful results that are relatively easy tounderstand, we have to research possible deviations between ourprediction and reality. Despite reliable statistics, it is virtuallyimpossible to reconstruct the decision process of human beings. Forthis reason we concentrate on the path finding abilities of the agentsand rely on predefined goals. Another limitation of our method isthe calculation time. The enhanced complexity of the urban layoutincreases the calculation time.

11. Conclusion

This project presents a novel dynamic urban environment, whichconnects the physical environment to the movement and utilizationpatterns of its occupants. This is a crucial part to better understand themovement of large crowds within a city. This dynamic environmentguides agents to POIs they would have chosen from the start — butresembles a more natural path finding process, where the path andthe goal get adapted to the environment constantly. The behaviorpatterns of pedestrians and their decision processes on the small scalehave an effect on all scales. This increased knowledge about theinteraction between actors in a city and the emerging movementpatterns on a large scale enables planners and decision makers to findmore subtle ways to adjust the urban fabric. It offers the opportunityto investigate parts of a city and to optimize corresponding aspectswith minimal interventions on the urban level.

12. Future work

This attempt of bridging the gap between function, behavior, andstructure is still only automated from structure to function and

hands over the behavior model to a different system. Therefore, anautomatic iteration of the loop is still far from being an integrated,interactive system. A closer integration of function and behaviorreduces calculation time and will lead to a practical integration.

Acknowledgments

We would like to thank the participants of the elective course‘Reclaim the Public Space’ at the Chair for Information Architecture,autumn semester 2009, ETH Zürich. Gratitude goes to Martina Jacobi,who designed the model of the case studies and did the preprocessingwork for the simulation.

References

[1] R. Arnott, K.A. Small, Urban spatial structure, Journal of Economics Literature Vol.XXXVI (1998) 1426–1464.

[2] G. Aschwanden, J. Halatsch, G. Schmitt, Crowd simulation for urban planning,eCAADe proceedings, Antwerpen, 2008, pp. 45–82.

[3] M. Batty, Y. Xie, Modeling inside GIS: part 2. Selecting and calibrating urbanmodels using ARC-INFO, International Journal of Geographical InformationScience 8 (1994) 451–470.

[4] G. LeBon, Psychologie des Foules, Alcan, Paris, in: R.A. Brooks (Ed.), 1991,Behaviour-based Humanoid Robotics, 1895.

[5] J.J. Gibson, The ecological approach to visual perception, HoughtonMifflin, Boston,MA, 1979.

[6] S. Guy, J. Chhugani, C. Kim, N. Satish, M.C. Lin, D. Manocha, P. Dubey, Clearpath:Highly parallel collision avoidance for multi-agent simulation, Proc. Of ACMSIGGRAPH/Eurographics Symposium on Computer Animation, 2009, pp. 177–187.

[7] J. Halatsch, A. Kunze, R. Burkhard, G. Schmitt, ETH Value Lab — A Framework forManaging Large-scale Urban Projects, 7th China Urban Housing Conference,Faculty of Architecture and Urban Planning, Chongqing University, Chongqing,2008.

[8] J. Halatsch, A. Kunze, G. Schmitt, Using shape grammars for master planning, ThirdConference on Design Computing and Cognition (DCC08), Atlanta, 2008.

[9] D. Helbing, P. Molnar, Social force model for pedestrian dynamics, Physical ReviewE 51 (May 1995) 42828 © 2008 The Americn Physical Society.

[10] B. Hillier, A theory of the city as object: or, how spatial laws mediate the socialconstruction of urban space, Urban Design International 7 (3) (2002) 153–179.

[11] W. Jager, R. Popping, H. van de Sande, Clustering and fighting in two-partycrowds: simulating the approach-avoidance conflict, Journal of Artificial Societiesand Social Simulation 4 (3) (2001).

[12] D. Thalmann, N. Magnenat-Thalmann, An ontology of virtual humans, VisualComputing 23 (2007) 207–218, doi:10.1007/s00371-006-0093-48 Springer-Verlag.

[13] P. Müller, P. Wonka, S. Haegler, A. Ulmer, L. Van Gool, Procedural modeling ofbuildings, Proceedings of ACM SIGGRAPH 2006/ACM Transactions on Graphics(TOG), Vol.25, ACM Press, 2006, pp. 614–6238, No. 3.

[14] S.R. Musse, D. Thalmann, Hierarchical model for real time simulation of virtualhuman crowds, IEEE Transactions on Visualization and Computer Graphics 7 (2)(2001) 152–164.

[15] R. Narain, A. Golas, S. Curtis, M.C. Lin, Aggregated dynamics for dense crowdsimulation, ACM Transactions on Graphics (2009)8 (Proc. Of ACM Siggraph Asia).

[16] Y. Parish, P. Müller, Procedural modeling of cities, ACM (2001) 301–308.[17] Y.I.H. Parish, P. Müller, Procedural modeling of cities, in: E. Fiume (Ed.),

Proceedings of ACM SIGGRAPH 2001, ACM Press/ACM SIGGRAPH, New York,COMPUTER GRAPHICS, Annual Conference Series, ACM, 2001, pp. 301–308.

[18] A. Turner, M. Doxa, D. O'Sullivan, A. Penn, From isovists to visibility graphs: amethodology for the analysis of architectural space, Environment and Planning B:Planning and Design 28 (1) (2001) 103–121.

310 G.D.P.A. Aschwanden et al. / Automation in Construction 20 (2011) 299–310

[19] A. Penn, A. Turner, Space syntax based agent simulation, in: M. Schreckenberg,S.D. Sharma (Eds.), Pedestrian and Evacuation Dynamics, Springer-Verlag,Berlin, 2001.

[20] C.W. Reynolds, Flocks, herds and schools: a distributed behavior model, ACMSIGGRAPH Computer Graphics 21 (1987) 25–34.

[21] Reynolds C. W., Steering behavior for autonomous characters. Game DevelopersConference, 1999.

[22] N. Pelechano, J.M. Allbeck, N.I. Badler, Virtual Crowds: Methods, Simulation andControl, Morgan and Claypool Publishers, 2008.

[23] Procedural Inc., Zurich, Switzerland; http://www.procedural.com.[24] J. Surowiecki, M.P. Silverman, The wisdom of crowds, American Journal Physics

Volume 75 (2007) 1904.[25] A. Ulmer, J. Halatsch, A. Kunze, P. Muller, L. Van Gool, Procedural Design of Urban

Open Spaces, vol 25, 2007, pp. 351–358.[26] Still, G. K.: 2000, Crowd Dynamics, PhD thesis, Warwick University.[27] A. Treuille, S. Cooper, Z. Popovic, Continuum crowds, Proc. Of ACM Siggraph, 2006,

pp. 1160–1168.[28] Y. Umeda, H. Takeda, T. Tomiyama, H. Yoshikawa, Function, behavior and

structure, Applications of Artificial Intelligence in Engineering V, ComputationalMechanics Publications/Springer Verlag, 1990, pp. 177–193.

[29] F. Tecchia, Y. Chrysanthou, Real-time rendering of densely populated urbanenvironments, Proc. Eurographics Rendering Workshop, 2000.

[30] UN-Habitat, Planning Sustainable Cities: Global Report onHuman Settlements 2009,United Nations, Human Settlements Programme, 2009, http://www.unhabitat.org/downloads/docs/GRHS2009/GRHS.2009.pdf.

[31] J. Van den Berg, Seawall, M.C. Lin, D. Manocha, Virtualized traffic: reconstructingtraffic flows from discrete spatio-temporal data–temporal data, Proc. of IEEEVirtual Reality Animation, 2004.

[32] Van de Berg, J. Guy Stephan, M. Lin, D. Manocha, Reciprocal n-body collisionavoidance, International Symposium of Robotics Research, 2009.

[33] C. Vanegas, D. Aliaga, B. Beneö, P. Waddell, Visualization of simulated urbanspaces: inferring parameterized generation of streets, parcels, and aerialimagery, IEEE Transactions on Visualization and Computer Graphics (2009)424–435.

[34] C.A. Vanegas, D.G. Aliaga, B. Benes, P. Waddell, Interactive design of urban spacesusing geometrical and behavioral modeling, ACM Transactios on Graphics 28 (5)(2009)8 (also in Proceedings SIGGRAPH Asia).

[35] P. Waddell, Modeling urban development for land use, transportation, andenvironmental planning, Journal of the American Planning Association 68 (2002)297–314.

[36] P. Waddell, G. Ulfarsson, Introduction to urban simulation: design anddevelopment of operational models, Handbook of Transport Geography andSpatial Systems 5 (2004) 25–78.

[37] M. Wegener, Operational urban models state of the art, Journal of the AmericanPlanning Association 60 (1994) 17–29.

[38] F. Wu, C.J. Webster, Simulation of natural land use zoning under freemarket andincremental development control regimes, Computers, Environment and UrbanSystems V. 22 (Issue 3) (1998) 241–256.

[39] G.K. Zipf,HumanBehavior and thePrincipleof Least Effort, Addison-WesleyPress, 1949.[40] Cai Jing, Zhigang Jiang, Changing of energy consumption patterns from rural

housholds to urban households in China: an example from Shaanxi Provice, China,Renewable and Sustainable Energy Reviews 12 (6) (2008) 1667–1680.

[41] S. Patil, J. van den Berg, S. Curtis, Ming C. Lin, D. Manocha, Directing crowdsimulations using navigation fields, RapidPost (2009)8 ISSN: 1077–2626.

[42] R.A. Brooks, Intelligence Without Reason, Proceedings of the 1991 internationaljoint conference on artificial intelligence, Sydney (1991) 569–595.

[43] P. Wadell, G.F. Ulfarsson, Introduction to urban simulation: design anddevelopment of operational models (2009).