Agent-Based Simulation of Mass Events Egress using Group Dynamics

Agent-Based Simulation of Mass Events Egress using Group Dynamics Duarte Almeidaa, Fernando Brito e Abreua, Inês Boavida-Portugalb aISTAR-Iscte, Instituto Universitário de Lisboa, Av. das Forças Armadas, 1649-026, Lisbon, Portugal bIGOT-UL, R. Branca Edmée Marques, 1600-276, Lisbon, Portugal Abstract Context: Public mass events require thorough planning in order to avoid or minimize possible safety issues and allocate the right amount of resources, such as emergency, cleaning, and security forces. Agent-based simulations can assist in the decision-making of event organization. The literature has shown group behavior plays a role in pedestrian flow models. Objective: Since mass events are often occasions where participants gather in social groups, we wanted to assess how their dynamics affected the egress of 90% of the event visitors and total inter-group collisions. Method: We have implemented an agent-based simulation model, based on an existing extension of the Social Force Model (SFM) considering group behavior, to describe non-urgent egress from a public mass event, using the GAMA platform. Simulations of scenarios of different group sizes in a digital twin of the historic square in Lisbon, where mass events often take place, were run. An ODD protocol was produced to describe that model to improve its clarity and facilitate its deployment and replication. Results/Conclusion: The model was able to describe group cohesion and repulsion interactions. Simulations have not shown a significant impact of average group sizes in egress times, for the given pedestrian density. Stochasticity may affect model results, so repeated simulations of larger pedestrian densities should be performed. Keywords: agent-based modeling, pedestrians simulation, egress, evacuation, social force model, groups Email addresses: [email protected] (Duarte Almeida), [email protected] (Fernando Brito e Abreu), [email protected] (Inês Boavida-Portugal) Preprint submitted to arXiv July 27, 2022 1. Introduction Cities have integrated in their tourism development strategies the realization of crowded events in public spaces, such as music festivals, fireworks and video mapping. Other public events such as political rallies, protests, celebrations or commemorations can attract an immense number of visitors. Choosing the adequate public space for each of these events requires knowing the expected number of visitors, acoustic conditions, spatial context and accessibility. For the success of these events, adequate planning is required for the allocation of security, emergency and cleaning resources. Thus, it is necessary to undertake risk analysis based on the carrying capacity and egress time from the public space hosting the event. The safety of urban public space events depends on a wide set of factors, such as the scope of the event, the characterization of the expected visitors and the geometry of the space. For instance, a protest of senior citizens over the degradation of their pensions would have different risk factors when compared to a concert for younger crowds. A rally for a tolerance-preaching political party would not be as risky as an extremist radical party rally. A quick event would probably not potentiate as much damage as a long-duration event. In case of the appearance of order disrupting situations, such as natural disasters, explosions, or terrorist attacks, the geometry of the public space must allow for quick and efficient evacuation, in order to avoid or keep casualties to a minimum. In this context, a simulation model can be an effective planning tool by allowing to testing of multiple scenarios without requiring empirical observation. To obtain realistic and meaningful results, the model should describe a variety of possible microscopic phenomena of the different partaking agents must be described in the model. One such relevant phenomenon is that these events are usually attended by social groups of pedestrians who walk together (e.g., friends, family). The interactions within each group and among members of different groups have an impact on the crowd behavior dynamics [14], which could also affect the efficiency of event egress. Although this is an important aspect of pedestrian dynamics, it has been often neglected in pedestrian flow models [3]. 2 Through the use of OpenStreetMap (OSM) 1data and of the GAMA modeling and simulation development environment 2, we developed an agentbased model (ABM) to simulate non-urgent egress dynamics of pedestrians from mass events in public open spaces, implementing an existing group behavior model based on the extended Social Force Model (SFM), proposed in [14]. Using the presented model, we perform simulated experiments to assess the effects of group interactions, group size and inter-person collisions in egress times. This paper is organized as follows: in section 2 we recap the SFM and its extension to support group dynamics behavior; then, in section 3, we review some of the most relevant related works; in section 4 we present our agent-based model, covering the simulation platform, the description of how the geographical location was mapped in the simulation environment, how agents behavior was modeled and some additional model assumptions; on section 5 we describe the used simulation scenarios and present the obtained simulation results and discuss them on section 6; finally, on section 7, we provide some conclusions and outline the future work. 2. Background The classic SFM, presented by Helbing et al. [6], is a microscopic mathematical model that represents pedestrian dynamics and interactions, based on the concept of social forces, presented by Lewin [11] as psychological forces from the environment and other pedestrians, which collectively act upon the decision-making of the movement of pedestrians, resulting in the manifestation of a decided velocity. In the classic SFM from Helbing et al., there are 3 main terms regarding the assessment of the driving force of the actual velocity of a pedestrian: (i) The force towards the desired velocity, dependent on the desired walking speed and the directions towards the destination, (ii) the repulsive forces from other pedestrians and obstacles (e.g., borders, walls, statues), to avoid and keep distance from them and (iii) the attractive forces from other possible sources (e.g., a shopping window on a way to a destination). The SFM has been widely used in pedestrian behavior models and simulations, with 1https://www.openstreetmap.org/about 2https://gama-platform.org/ 3 some modifications (e.g. the attractive forces are often neglected) and improvements to consider additional phenomena such as panic behavior [5]. While the classic SFM does not implement group behavior, there are some relevant proposals for modeling social group behavior in pedestrian dynamics, based on SFM. Moussaïd et. al [14] presented extensions to the classic SFM, to represent inter-group and intra-group interactions of repulsion and attraction. Those authors extended the classic SFM with the following components: (i) intra-group interactions based on gaze, (ii) the attraction of group members to the group’s center of mass, responsible for group cohesion, and (iii) the repulsion interactions between group members to avoid collisions. In [7], the authors present a more advanced extension to the previously referred group SFM. This model adds the following components: (i) the repulsion forces between groups and (ii) subgroup coordination forces. 3. Related Work 3.1. Publications Many pedestrian flow models have been developed, using multiple modeling types (e.g., ABM, Cellular Automata (CA)), some applied to evacuation scenarios, some used for the study of isolated pedestrian dynamics. Since we used an ABM with groups behavior in the context of evacuation simulation, that will be the scope of this brief literature review, that will be presented in chronological order. The inclusion criteria for selected papers, besides topic coverage, were: •Published in a journal •Indexed by Scopus 3, one of the most important scientific indexing databases [18] •Written in English •Community recognition (at least an average of 4 citations per year since publishing, according to Google Scholar 4 3https://www.scopus.com/ 4https://scholar.google.com/ 4 Qiu et al. [15] presented an ABM based on the Reynold’s Boids algorithm to simulate group behavior in crowd evacuation, assessing the influence of group size, intra-group structure, and inter-group relationships. The authors concluded that the factors in study are important in crowd evacuation. Lemercier et al [10] proposed an RVO2-based ABM to assess the influence of the perceptions of pedestrians on their attitudes in a crowd evacuation. RVO2 is a collision avoidance model proposed in [21]. By simulating leaderfollower behavior, grouping, or not grouping, the authors concluded that a more realistic crowd behavior can be modeled by combining these different behaviors. Ma et al [13] proposed an extension of the classic SFM to simulate leaderfollower behavior in crowd evacuation, incorporating the visibility range of the area. The authors concluded that the impact of leadership on evacuation time depends on the range of visibility of the area and crowd size. Li et al [12] also presented an extension to the classic SFM by Helbing and Mòlnar for crowd evacuation, incorporating social groups and leader-follower behavior. Based on the results, the authors concluded group behavior causes more efficiency in egress than the classic SFM model. Turgut et al. [20] presented an extended SFM model with 2 types of group behavior: leader-centered and group-centered behavior. Through the analysis of simulations in evacuation scenarios, the authors concluded that in (i) small groups, evacuation time is lower in leader-centered than in groupcentered behavior (ii) the increase of leaders reduced evacuation time and (iii) with multiple exits, group-centered behavior resulted in less evacuation time and better exit choice balancing than leader-centered behavior. Most of these studies focus on leader-following or herding effect of emergency evacuation, and not groups derived from social ties in non-urgent egress. These studies also do not apply simulations in large scale, real environments. Table 1shows a summary of the selected related studies and attributes, in comparison with the one presented here. 4. Agent-based model This section presents an abbreviated version of our proposed ABM. We used the Overview, Design concepts, and Details (ODD) protocol, which was proposed as a standardized format for providing a consistent, logical, and readable account of the structure and dynamics of ABMs [4]. Many examples 5 Table 1: Related work summary Reference Non-urgent egress Real life scenario simulation ABM Platform/ Library [15] No No OpenSteer [10] No No RVO2 [13] No No - [12] No No Multiple platforms [20] No No AnyLogic Our model Yes Yes GAMA based on ODD can be found online at the CoMSES Model Library 5, a digital repository that enforces reproducibility, and reuse of ABMs [9]. Our aim was to simulate the non-urgent egress of visitors from mass gathering events in public open spaces, namely, squares or plazas, in order to understand how the different built environments of the open spaces, crowd density, exit choosing scenarios and group behavior affect the egress times of the event visitors. In this paper, we solely focus on the effects of group behavior. We used an OSM fragment representing the chosen open urban space, upon which we added rectangular geometry data for representing the exit gates. The leaving pedestrians were modeled as autonomous agents who interact with each other and the environment in order to progress towards the exit gates, simulating crowd maneuvering. The simulations of the model in various scenarios were run on a remote Virtual Machine (VM), generating data files and simulation snapshots for the analysis of the results. The main outcome variables obtained from simulations of the model were (i) the remaining pedestrians at each simulation cycle, from which is derived the total egress time of 90% of the attendees, and (ii) the total collisions detected in each cycle. A collision was considered to occur when 2 pedestrians from different groups were less than 0.5 meters apart. 5https://www.comses.net/codebases/ 6 4.1. Simulation platform The presented model was implemented using version 1.8.1 of the GAMA platform 6. It is an Eclipse-based 7Integrated Development Environment (IDE) for spatially explicit, agent-based and agent-oriented modeling and simulation [19]. The platform is designed to develop and run complex, largescale models, with a wide support for integrating and visualizing geographic information systems (GIS) data. It also provides a headless mode, allowing to run simulations in dedicated servers without the overhead of a user interface. 4.2. Environment The model environment is composed by the buildings that delimit the open space, and the exit geometries, which are modeled as polygons that cover the entrances to the exiting streets, bordering the open space polygon. We used QGIS 3.16 LTS 8to manually model the exit geometries and open space limits polygon, and to download and process OSM building geometry data to model the built environment of the public open spaces. 4.3. Agent modeling In this model, the behavior of the pedestrian agents is orchestrated by the Model agent (an observer agent), in order to harness the experimental usage of concurrent processing of agents, present in this version of the GAMA platform, to mimic their autonomy. This agent is also responsible for calculating the result statistics (i.e., average pedestrian speed, pedestrian egress, collisions). The active agents in the model are the pedestrians who egress from the open space, the Model agent, and the Group agents. Pedestrian agents walk toward the chosen exit using a simplified version of the extended SFM for group behavior presented in [14], based on the model of Helbing et. al. [6]. For every simulation cycle, each pedestrian agent evaluates the repulsive social forces other pedestrians and obstacles act upon them, and attractive forces representing group cohesion. Group agents store references of their members to improve efficiency in the calculation of the intra-group forces and calculate the group’s centroid (center of mass), 6https://gama-platform.org/ 7https://www.eclipse.org/ 8https://qgis.org/en/site/ 7 which is required both for path calculation and in visualization. Building geometries are also considered agents but do not have behavior, serving as obstacles for the SFM submodel. Group agents: The number of groups is calculated from the total number of pedestrians and group size. At the beginning of a simulation, a group center location is randomly created to guarantee a uniform distribution across the environment area. The pedestrians are then spawned at a set distance to the center of their respective groups. At each cycle, a group agent behaves as follows: 1. The group agent checks if a group member has left the simulation. If that happens, it is removed from the list of members. 2. The group agent calculates the group’s centroid based on the current position of all its current members. 3. If there are no members left, the group agent is removed from the simulation. Pedestrian agents: When the pedestrians spawn, they are assigned to their groups, and all members of a group are placed at a given distance (2 meters) from the center of their group. Then, all members of a group choose an exit, randomly distributed by the weight of the exit width (i.e., a wider exit is more likely to be chosen than a narrow one). At each cycle, the pedestrian agents follow the following steps: 1. The agent calculates the repulsive forces exerted by its group members and the attraction forces to the group’s centroid, using the respective equations from [14]. 2. The agent calculates the repulsive forces exerted by pedestrians within a defined distance using the classic SFM motion equation [6]. 3. The agent displaces according to the speed obtained for the final sum of forces. 4. If the agent reaches its chosen exit geometry, it is removed from the simulation. 5. Simulation Scenarios The following simulation scenarios are set on a digital twin of the Praça do Município square, Lisbon, which has been a host of numerous mass events 8 (a) Aerial view of the real twin (Google Maps) (b) Digital twin on GAMA with 2600 pedestrian agents Figure 1: Simulation environment (Lisbon town-hall square) 9 Figure 5: Simulation egress time versus collisions (data points and nonlinear regression curve) that may be considered as potential internal validity threats. We address each one separately. All area without buildings inside the bounding polygon is considered usable area. In urban open space mass events, roads are often blocked to vehicles, only allowing pedestrians or special vehicles traffic (e.g., ambulances, police cars, event staff vehicles). In our example we omitted the latter (that can be seen to the right in Figure ??), as well as the central statue and the round kiosk bar in one of the corners of the town-hall square. In urgent egress situations, individuals in panic often adopt an herd mentality and choose the closest exit to evacuate as quickly as possible [22]. In the, fortunately much more recurring, non-urgent egress situations, it is not trivial to predict where the crowds go, as it will depend on their final destination, means of transportation, and even on the time of the day when the event occurred. In the absence of empirical data, the best alternative we have devised was assigning the choice of exit gate for each group based on a weighted randomization (random quotas) decision, where the weights were the exit widths. In other words, wider exits were more likely to be assigned than narrow ones. On average, the same number of pedestrians is expected to flow out per unit of exit width. To study the influence of group size, we blocked group size variability 16 from scenario to scenario. In other words, for each simulation scenario group sizes were homogeneous. Non-homogeneous groupings should be analysed as well, but their actual distribution requires further research on related work. Our own empirical evidence is that the size of groups can vary widely and also be dependent on the type of event. For instance an event targeted at families is likely to have an average group size smaller than an event for youngsters, like an hard-rock concert. Finally, pedestrian flow beyond the exits is not considered, being assumed as non restricting to egress flow. Any constraints to the flow of pedestrians beyond the exits would influence the egress rates, but that is unlikely in most real case scenarios. 6.3. Simulation performance All simulations were deployed and run on a VM created on top of OpenStack 11, a free, open standard cloud computing platform used by INCD to provide computing and storage services to the Portuguese scientific community 12. The machine runs as operating system the Linux distribution Ubuntu 20-04, having 8 virtual CPU and 128 GB of RAM, allowing the running of a suitable number of parallel simulations with decent memory allocated to each. The simulations showed poor performance with a great number of agents, as each agent needs to perform multiple calculations considering all the neighbour pedestrians within a defined range. The simulation speed decreases with the increase of group size. An attempt to harness parallelization feature of the GAMA platform was done for parallel SFM computing for each agent, but it was not enough to increase performance. 7. Conclusions and Future Work The research presented in this paper is a continuation of our previous work [2], analyzing the egress time for multiple scenarios of crowd density in the open spaces using a derived model for egress flow rates in exits, from the Society of Fire Protection Engineers (SFPE) Handbook [8], and [1], where the initial version of the simulation model is introduced, without the implementation of group behavior. 11https://www.openstack.org/ 12https://www.incd.pt/?p=servicos/cloud&lang=en 17 In this article, we presented an ABM to simulate the non-urgent egress of pedestrians from mass events, considering group behavior. Using the GAMA platform, we simulated the egress of 2600 pedestrians from the Praça do Município square in Lisbon, defining different group size scenarios. The full description of our ABM model, following the ODD protocol, can be found at the aforementioned CoMSES repository. The source code for our GAMA implementation can be found in GitHub 13. Since we aim at developing digital twinning models for extreme crowding situations, we need to generate simulations with a higher pedestrian density, allowing a more precise assessment of the social group behavior effect on non-urgent mass event egress. However, with the current setup, simulation time increases almost quadratically with the number of agents, making it unfeasible beyond a certain limit. To mitigate this problem, we plan to explore the usage of GPU-enhanced ABM tools such as Flame GPU [16] to increase concurrency in agent calculations, allowing us to simulate scenarios with a massive number of agents and lower execution time, possibly allowing real-time or faster-than-real-time visualization. To improve model realism, therefore mitigating internal validity threats, we also plan to incorporate unto the ABM model the more advanced group SFM presented in [7]. Still, observation and studies of mass crowd nonurgent egress conditions should be performed to allow calibration of model parameters and to detect additional macroscopic and microscopic behavior to be modeled in the agents. Acknowledgments This work was partially funded by the European Commission RESETTING project (COS-TOURINN 101038190) and the Portuguese Foundation for Science and Technology (FCT), under ISTAR-Iscte projects UIDB/04466/2020 and UIDP/04466/2020. Agent-based simulations in this work were executed in a cloud-based infrastructure (computing and storage) provided by the INCD, which is funded by FCT and FEDER under the project 01/SAICT/2016 nº022153. We are particularly grateful to INCD’s technical team for their timely support of our research needs. 13https://github.com/duartealmeida/pedestrian_egress_abm 18 Special thanks go also to Doctor Patrick Taillandier, one of the main GAMA platform developers, for his invaluable help in understanding platform intricacies and solving related issues. References [1] Duarte Almeida. 2021. Urban tourism crowding dynamics: Carrying capacity and digital twinning. Master’s thesis. http://hdl.handle. net/10071/24242 [2] Fernando Brito e Abreu and Duarte Almeida. 2021. Capacidade de Carga e Tempo de Evacuação para Eventos nos Espaços Públicos das Cidades. IntelCities (apr 2021). https://doi.org/10.5281/ZENODO. 4600967 [3] Lin Cheng, Ragamayi Yarlagadda, Clinton Fookes, and Prasad Yarlagadda. 2014. A review of pedestrian group dynamics and methodologies in modelling pedestrian group behaviours. World Journal of Mechanical Engineering 1, 1 (2014), 1–13. [4] Volker Grimm, Uta Berger, Donald L DeAngelis, J Gary Polhill, Jarl Giske, and Steven F Railsback. 2010. The ODD protocol: A review and first update. Ecological Modelling 221, 23 (2010), 2760–2768. https: //doi.org/10.1016/j.ecolmodel.2010.08.019 [5] Dirk Helbing, Illés Farkas, and Tamás Vicsek. 2000. Simulating dynamical features of escape panic. Nature 407, 6803 (2000), 487–490. https://doi.org/10.1038/35035023 arXiv:0009448 [cond-mat] [6] Dirk Helbing and Péter Molnár. 1995. Social force model for pedestrian dynamics. Physical Review E 51, 5 (1995), 4282–4286. https://doi. org/10.1103/PhysRevE.51.4282 arXiv:9805244 [cond-mat] [7] Lin Huang, Jianhua Gong, Wenhang Li, Tao Xu, Shen Shen, Jianming Liang, Quanlong Feng, Dong Zhang, and Jun Sun. 2018. Social force model-based group behavior simulation in virtual geographic environments. ISPRS International Journal of Geo-Information 7, 2 (2018). https://doi.org/10.3390/ijgi7020079 19 [8] Morgan J Hurley, Daniel T Gottuk, John R Hall Jr, Kazunori Harada, Erica D Kuligowski, Milosh Puchovsky, John M Watts Jr, CHRISTOPHER J WIECZOREK, and Others. 2015. SFPE handbook of fire protection engineering. Springer. [9] Marco A Janssen, Lilian Na’ia Alessa, Michael Barton, Sean Bergin, and Allen Lee. 2008. Towards a community framework for agent-based modelling. Journal of Artificial Societies and Social Simulation 11, 2 (2008), 6. [10] Samuel Lemercier and Jean-Michel Auberlet. 2016. Towards more behaviours in crowd simulation. Computer Animation and Virtual Worlds 27, 1 (2016), 24–34. [11] Kurt Lewin. 1951. Field theory in social science: selected theoretical papers. The Annals of the American Academy of Political and Social Science (1951). https://doi.org/10.1177/000271625127600135 [12] Yan Li, Hong Liu, Guang peng Liu, Liang Li, Philip Moore, and Bin Hu. 2017. A grouping method based on grid density and relationship for crowd evacuation simulation. Physica A: Statistical Mechanics and its Applications 473, 88 (2017), 319–336. https://doi.org/10.1016/ j.physa.2017.01.008 [13] Yi Ma, Richard Kwok Kit Yuen, and Eric Wai Ming Lee. 2016. Effective leadership for crowd evacuation. Physica A: Statistical Mechanics and its Applications 450 (2016), 333–341. https://doi.org/10.1016/j. physa.2015.12.103 [14] Mehdi Moussaïd, Niriaska Perozo, Simon Garnier, Dirk Helbing, and Guy Theraulaz. 2010. The walking behaviour of pedestrian social groups and its impact on crowd dynamics. PLoS ONE 5, 4 (2010), 1–7. https: //doi.org/10.1371/journal.pone.0010047 [15] Fasheng Qiu and Xiaolin Hu. 2010. Modeling group structures in pedestrian crowd simulation. Simulation Modelling Practice and Theory 18, 2 (2010), 190–205. https://doi.org/10.1016/j.simpat.2009.10.005 [16] Paul Richmond and Mozhgan K. Chimeh. 2017. Flame GPU: Complex system simulation framework. In Proceedings of the 2017 Inter20 national Conference on High Performance Computing and Simulation (HPCS’2017). 11–17. https://doi.org/10.1109/HPCS.2017.12 [17] Nagui M. Rouphail and D. Patrick Allen. 1998. Recommended Procedures Chapter 13, "Pedestrians," of the Highway Capacity Manual. Highway capacity Manual February (1998), 1–56. [18] Stephan Stahlschmidt and Dimity Stephen. 2020. Comparison of Web of Science, Scopus and Dimensions databases. Technical Report. German Centre for Higher Education Research and Science Studies (DZHW), Hannover, Germany. [19] Patrick Taillandier, Benoit Gaudou, Arnaud Grignard, Quang Nghi Huynh, Nicolas Marilleau, Philippe Caillou, Damien Philippon, and Alexis Drogoul. 2019. Building, composing and experimenting complex spatial models with the GAMA platform. GeoInformatica 23, 2 (2019), 299–322. https://doi.org/10.1007/s10707-018-00339-6 [20] Yakup Turgut and Cafer Erhan Bozdag. 2022. Modeling pedestrian group behavior in crowd evacuations. Fire and Materials 46, 2 (2022), 420–442. https://doi.org/10.1002/fam.2978 [21] Jur Van Den Berg, Stephen J. Guy, Ming Lin, and Dinesh Manocha. 2011. Reciprocal n-Body Collision Avoidance. Springer Tracts in Advanced Robotics 70, STAR (2011), 3–19. https://doi.org/10.1007/ 978-3-642-19457-3_1 [22] Jinhuan Wang, Lei Zhang, Qiongyu Shi, Peng Yang, and Xiaoming Hu. 2015. Modeling and simulating for congestion pedestrian evacuation with panic. Physica A: Statistical Mechanics and its Applications 428 (2015), 396–409. https://doi.org/10.1016/j.physa.2015.01.057 21