Figure 1. General architecture for object in the model.
Abstract
Controlling crowds in airports, train terminals, sporting events, etc., is a complex problem. This particular problem has a great deal interaction between the entities themselves (i.e. among the individual members of the crowd) and the crowd (or individuals) with the environment in which the crowd is placed. These interactions are best observed when the simulation output has a graphical component. Therefore, the combination of a simulation model with a good graphical representation will facilitate the understanding of the behavior of this complex system. This paper focuses on the construction of such a model able to reproduce some of the characteristics of complex crowds. Our goal is to provide a useful tool for the operations managers of areas (such as the ones mentioned above) where the control of the ebb and flow of a crowd is an important part of daily procedure. However, we also view this need not to be limited to single site, but to several sites physically separated in the domain. Hence, a system designed for this problem should be easily available and easy to use for several people over a networked environment (this is not a description of a parallel simulation, but of a simulation model that has networked access, an easy human interface and a graphical display. To this end, we have developed a prototype crowd control simulation system using Java as test case for verifying the computational capacity and the efficiency on web environment for this kind of problem. The authors present an object oriented architecture able to reproduce the behavior of a crowd, including subgroup behaviors, interactions, stochastic decisions of single units etc. The model has been applied to a typical test case for validation and verification. This example involves the simulation of the people visiting a museum with different attractions and possible paths. The model involves interactive features and 2D graphics. It also allows the possibility for emergency test procedures (i.e. fires and alarms) in order to identify the efficiency of evacuation procedures. In effect, this model allows us to study the layout of any public or industrial facility. We also imagine that this prototype could be applied to training or education. In effect , the web based simulation would provide a large audience the possibility to interact and experiment with large complex models independently in a well known, user-friendly platform (i.e. the Web). [8]. The next step is to integrate this existing model with a VRML 2.0 environment in order to provide a three dimensional interface. The purpose of this future development is to understand the technological restrictions, cost, performance and current efficiencies of such an endeavor. This work would help simulationists understand the complexities involved in implementing a new kind of simulation environment. As a further development, the authors look to the constructions of a web server that allows to the users to interact with the model in a cooperative environment.
The Problem of Modeling Crowd Control
Crowd control can be applied to many different areas, such as
police operations [10], shopping center design, public
park re-organization [5], rail station re-engineering
[3] and epidemic diffusion. Generally, we can model
the applications with the following techniques (which we briefly summarize):
For many years, the computational work load was to high to apply particle dynamics to real problems involving crowd control. However, with the advances in computer speed and efficiency, it has become feasible to model many real world problems without requiring a supercomputer. In fact, it is possible to simulate an environment ranging from a few hundred people (conference room, small station) to some thousands (large rail-way station, theater). Unfortunately, analyzing crowd behavior using particle dynamics in a complex environment such as a large stadium (where the combination of the size of the crowd and the interactions of individuals entities in the crowd are computationally expensive) is still difficult. Therefore, using a combination of approaches, both particle dynamics and particle queuing, we can model many real world problems of various sizes.
In our research ,which we describe in this paper, we are mostly interested in combining particle modeling (both dynamic and queuing) to support the layout design of public facilities. In particular, we wish to develop a dynamic model that designers could use to create environments that address safety issues, efficiency issues, and quality of service issues. We look at safety issues since many regulations for complex environments (such as large stadiums) are complex and ultimately expensive. Efficiency plays a major role in facilities design since moving people into and out of the facility has a direct effect on the usability of the facility (crowds leaving and entering a stadium are examples of this). Finally, quality of service plays an important role in facility layout. The more enjoyable to the user (i.e. the crowd) a facility is, the more used the facility will be.
To accomplish the above, the approach we will describe requires
the modeling detail provided by particle dynamics and the computational
efficiency that is characteristic of particle queuing. That's the
motivation for combing the two into one model. In this paper we will
describe how we apply this combined approach to crowd analysis and control
in a museum. However, the approach is applicable to any crowd control
problem, and is scalable in size to fit most real world problems. We are
also interested in creating a platform that is easy to use, distributed,
and still computationally sound. The possibility to integrate these model
in a Web-Based environment by using Java applets was very appealing.
This platform is potentially an efficient way to distribute Java
in a complex simulation.
Our Web-Based Model for Crowd Control
As we mentioned, we combined particle dynamics and particle queuing
by applying particle dynamics to monitoring specific entities that
required more detail investigation and applying particle queuing to entities
that needed less detailed analysis. For example, the point services (ticketing,
single man gate. etc.) can be modeled efficiently by using particle queuing
since queuing is efficient for handling traditional multi-lane, multi-server
facilities where the details of the individuals in the queue is not important.
However, it's more appropriate to use particle dynamics for reproducing
the detail of the crowd behavior after the individuals leave the queue.
For example, particle queuing can be used for modeling the queue(s) entering
a museum, but then particle dynamics is used to model the individuals as
they move from exhibit to exhibit. The change from particle dynamics to
queuing is implemented by changing the focus of the object when it reaches
an area of interest. For example, when an object reaches a point
services it joins to the queue related to point service.
Using the definition in [11] , we view approach as
multimodeling. We based our global model of the crowd and the museum
on many submodels. each using an appropriate simulation methodology.
For example, particle diffusion was used for the detail behavior of the
crowd while discrete event logic was used to model certain events that
can be stochastically determined (such as waiting for elevator doors to
open or the time it takes to contemplate a sculpture). In this way,
the multimodel approach provides a platform to simulate both interactions
in the crowd and sequences needed to move the crowd in and around the museum.
The entire model is finally implemented using standard object oriented
design and analysis. Finally, used combined stochastic simulation
to representing the particle dynamics. [2]
Each entity of the crowd is modeled as an object with the following attributes:
Figure 1. General architecture for object in the model.
The general architecture of the entities is proposed in Figure 1 above. The figure demonstrates the hierarchical nature of our design. For example, we have an object that is the group of men subject to some specific methods (init, drawing, alarm for a group of men, changing the behavior of the men instead of the walls or of the attractions). Thus each instance of men, called man, have different properties. The interactions are based on the following rules:
A) Internal Forces: This is related to goal of each entity. The final
destination (or target) attracts each
entity based on a predefined path (sequence of points that lead to
the target) with the following force:
Where:
B) External Forces: This force is related to other entities, layout
objects or external force fields and is composed of three parts: (I) collision
avoidance, (II) group attraction and (III) external forces.
I) Avoiding collision with other entities
Where:
Where:
Where:
The interaction with external objects in the model includes collision
with layout obstacles (for example walls). Since an entity is not
allowed to cross obstacles, we have included in our model the ability of
entities to recognize a collision and to react appropriately. Figure
2 shows the interaction areas of an entity with other entities and obstacles.
Figure 2 - Interaction among entities
Experimental Results
We implemented our model using Java applets in a standard Web environment.
The application of our multimodel is a simple two floor museum. On
the ground floor are several emergency exits to model an evacuation of
the museum in case of fire or other alarm situation. The crowd consists
of thousands of individuals operating as either independent individuals
or individuals in a specific group. Thus the members of the crowd move
individually or collectively or both. The individuals' goal is to
view several attractions that hang on the walls or are in cases on each
floor of the museum. The floors are connected by a single stairway.
The layout of the museum can be seen in Figure 3 below.
Figure 3. The Scenario Layout interface of the simulator
as viewed in Netscape.
We ran experiments with our Web simulation using the following target functions:
Where:
In order to verify and validate the model, we applied standard Analysis
of Variance (ANOVA ) and Mean Square pure Error (MSpE temporal evolution
analysis) techniques. We summarize our results in Figure 4. . The
experimental analysis is based on design of experiments central composite
design 3k with 2 independent factors as presented in Table I.
The response surfaces obtained by the experimental analysis are presented
in Figure 5 (Evacuation Time) and Figure 6 (Quality of the Service). A
simplified release of the model (excluding the modules for loading and
saving data from files for security reasons) is included Figure 7. We have
compared our model to that of a 3rd polynomial regression for verification
and validation. Our model was statistically significantly as compared
to the 3rd polynomial regression and using a standard fitness test showed
that fitness of the model to the data is good. The results, for the statistical
validations are summarized below:
| Sum Square | D.o.F. | Mean Square | F<> | Reference Value | |
| Regression | 3112.09 | 9 | 3445.79 | 52.74 | < 4.1 Significance test passed |
| Error | 39.34 | 6 | 6.56 | ||
| Lack of Fit: | 37.31 | 2 | 18.66 | 36.81 | 6.94Lack of Fit test Ok |
| MSpE: | 2.03 | 4 | 0.51 |
| Table I Independent Variables | |||
| People |
|
|
|
| Emergency Exit Dimension |
|
|
|
Figure 4 Verification and Validation of the Model
Figure 5 - Evacuation Time with different crowd distributions in function of People and Exit Dimension
Figure 6 - Quality of the Service and Confidence Band as function of
People in the Museum
Figure 7 Simulation Applet
Conclusions and Future Work
The study demonstrates the effectiveness of modeling public facilities for improving service and supporting federally regulated safety levels. In effect, the model demonstrates the ability to process over several thousand entities using personal computers. Our multimodel approach, combing particle dynamics and particle queuing in a combined continuos and discrete event simulation allowed us to effectively apply the model to a real case study: crowd control in a museum setting. However, the model used can easily be applied to other real world scenarios such as rail stations, stadiums, opera houses, passenger ships, etc..
The potential of this work is very high since the approach allows simulators to obtain accurate quantitative information easily and quickly. This information would then be used to support improved design alternatives. Additionally, since the model is implemented as a Web based applet, the simulation is available to all members of the design team. Thus this model could be part of a distributed collaborative design effort win areas where facilities need to control crowd movement (especially in areas where, due to cost and time constraints, traditionally a qualitative approach was used). We see this model as the foundation of a set of tools that the design team utilizes in constructing such facilities. This Web model is a accessible tool to aid the design team in focusing on specific issues of crowd control and design cost effective solutions to deal with these issues.
The following step in our research is to integrate this existing model with a VRML 2.0 environment in order to provide a three-dimensional shape for the simulation. The purpose of this development is to offer users a sort of visual verification and validation of this simulation. This type of visual enhancement is becoming more popular and is often an important part of the verification and validation phase of model building. [1][7] In our initial efforts to add 3D visualization, we realized that VRML 2.0 is still too slow on too low level platform for integration of scenarios with only a few hundred entities. Our initial effort can be seen in Figure 8. Thus the integration of the model with three-dimensional representations based on VRML 2.0 demonstrates computational problems for large systems. However, as workstations become more powerful and as advancements to products such as VRML continue, we would expect to reach satisfactory levels in this area in few years.
As further developments the authors look ahead to the construction of
a web server that allows to the users to interact with the model in a co-operative
environment.
Figure 8 The Cosmo Player visualizing entities of the model in 3D by VRML 2.0
References
2 Booch G (1994) "Object-Oriented Analysis and Design", Addison Wesley, Menlo Park, CA
3 Bouvier E., Cohen E. (1995) "Simulation of Human Flow with Particle Systems", Proceeding of Simulators International XII, Phoenix, AZ, April 9-13
4 Bruzzone A.G. & Rivarolo D. (1996) "An Innovative Vehicular Traffic Combined Simulation Tool: from a Detailed Analysis to a Synthesis of Results", Proc. of 29th Annual Simulation Symposium IEEE, New Orleans LA, April 8-11
5 Bruzzone A.G., Rivarolo D. (1997) "Virtual Reality as the Starting Point for the Functional Analysis of Public Plant Design Services", Proceedings of High Performance Computing, HPC'97, Atlanta, Georgia, April 6-10
6 Langton C. (1997) "Artificial Life an Overview", MIT Press, Cambridge, MA
7 Lea R., Matsuda K, Miyashita K. (1996) "Java for 3D and VRML Worlds", New Riders Publishing, Indianapolis
8 Mosca R., Giribone P. & A.G.Bruzzone (1995) "Simulator Object-Oriented Modeling for Educational purposes in the Industrial Plant Sector", Proceedings of WMC95, Las Vegas, Nevada (USA), January 15-18
9 Mosca R., Giribone P. & A.G.Bruzzone (1995) "AI Techniques to Optimize the distribution of Fire-Fighting Squads during an Accident", Proc. of Simulators International XII, Phoenix, April 9-13
10 Varner D., Scott D.R. Micheletti J., Aicella G. (1998) "USMC Small Unit Leader Non-Lethal Trainer (SULNT)", Proc. of ITEC98, Lausanne, 28-30 April
11 Fishwick, P. (1995) "Simulation Model Design and Execution", Prentice Hall, New Jersey
Biographies
Agostino Bruzzone is a member of the Department of
Production Engineering of the University of Genoa and he is Genoa Center
Director for the McLeod Institute of Simulation Science and president of
the Liophant Simulation Club. He has published extensively in all areas
of simulation theory and applied simulation.
Robert Signorile is currently an Associate Professor in the Computer Science Department of Boston College. His research interests include multimodel simulation, simulation in business, networks and distributed computing. He has published regularly in applied simulation and simulation methodology.