A Computational Study of the Station Nightclub Fire Accounting for Social Relationships

Using agent-based modeling, this study presents the results of a computational study of social relationships amongmore than four hundreds evacuees in The Station Nightclub building in Rhode Island. The fire occurred on the night of February 20, 2003 and resulted in 100 fatalities. A er summarizing and calibrating the computational method used, parametric studies are conducted to quantitatively investigate the influences of the presence of social relationships and familiarity of the building floor plan on the death and injury tolls. It is demonstrated that the proposed model has the ability to reasonably handle the complex social relationships and group behaviors present during egress. The simulations quantify how intimate social a iliations delay the overall egressprocess and show theextentbywhich lackof knowledgeof abuilding floorplan limits exit choices and adversely a ects the number of safe evacuations.


Introduction
. There is widespread consensus that people participate in social gathering during emergency egress (Aguirre et al. a,b; Chu & Law ; Moussaïd et al. ; Pluchino et al. ). Group members are o en connected through pre-existing social relationships, e.g. familial or friendship, and their behavior is significantly a ected by such social a iliations (Santos & Aguirre ; Moussaïd et al. ; Aguirre et al. b; Chu & Law ). Participants tend to interact with each other and stay together, potentially increasing the dangers they collectively face (Johnson et al. ; Cornwell ). Yet, of the many egress models that have been published to date, only a few are able to adequately handle social interaction and social emergence involving groups of evacuees (Santos & Aguirre ; Aguirre et al. a). Moreover, aside from a few cases, most existing models lack validation of their simulated results by real-world processes (Aguirre et al. a). .
To address such gaps, this paper employs the agent-based egress simulation tool, EgressSFM, developed by Fang et al. ( . This study o ers the results of a numerical study using EgressSFM of social relationships among the more than four hundreds evacuees of The Station Nightclub fire. The theory behind the computational platform is first summarized and background about the social organizational features of the gathering at the venue is presented. Key modeling parameters in EgressSFM are calibrated to detailed data from The Station Nightclub event. A er calibration, parametric studies are conducted to quantitatively investigate the influences of the presence of social relationships and familiarity of the building floor plan on the death and injury tolls. Field information first collected by Aguirre et al. ( b) about the persons and groups that were present at the Station during the fire is used to evaluate the validity of the results of this study.

Background
The Scalar Field Method (SFM): Theory and validation . The SFM developed by Fang et al. ( , ) assumes that the behavior of each evacuee is controlled by a rational thinking process. Agents representing evacuees can perceive surrounding entities, evaluate di erent potential avenues of action, desired goals and social relationships, and represent those factors through locomotion. These goals may comprise the evacuee's need to escape through an exit, avoid collision with walls and obstacles, move towards related agents, keep given spacing to other agents, and respond to social relationships. By assuming that agents are analogous to charged particles in an electrical field, the SFM quantitatively evaluates these e ects as a series of scalar quantities, termed virtual potential energies (VPEs). The VPEs from various sources can be directly added together to form a comprehensive field around an agent that signifies the additive or subtractive e ects of issues of importance to the agent. Based on charged-particle-in-field analogy, an agent will seek to minimize its VPE, i.e. the lower the value of VPE, the greater will be the desire of the agent to take action, and vice versa. .
The VPEs are computed through a series of functions of distances to other agents or objects in the environment. While detailed equations can be found in Fang et al. ( ), some governing equations are shown next for the sake of completeness: where E 1 , E 2 and E 3 are the VPEs of the goals to exit a building, preserve private space, not collide with other agents and with walls; d 1 , d 2 and d 3 are the distances between agent and exit, other agent and wall, respectively. c 1 , c 2 and c 3 are strength constants that are assigned to be , , . ∆θ 1 in Equation is the absolute value of the angle di erence between the forward facing orientation of an agent and the direction pointing to a target object. D 20 and D 30 are influence distances in Equations and , respectively. Agents and other entities within the influence zone can interact together in a VPE sense; otherwise they are unable to influence one another. D 1a is m, and D 1b is . m and are associated with the orientation of an agent; R a is the radius of an agent in the direction of interest. To simplify calculation of R a , an agent is assumed to be enclosed by an ellipse with principal radii R T and R T + R S , where R T and R S are the sizes of the torso and shoulder respectively with R T = . m and R S = .
m. R T , other is the size of the torso of the other agent in Equation . E 2 ,counter is a term that accounts for an agents' dodging behavior in a counter-flow situation, where agents attempt to prevent face-to-face situations as they are approaching other oncoming agents. .
Along with the desired goals, two categories of social relationships are outlined: kin-relationship (such as spouses and dating partners) and friend-relationship (such as friends and co-workers). The former is assumed to be effective over distances and stronger than the latter. The friend relationship is assumed valid for a limited distance beyond which it is considered ine ective. The governing equations for these two cases are as follow: where E 4 and E 5 are the VPEs of kin-relationship and friend-relationship; d 4 and d 5 are distances between kinrelated agents and between friend-agents, respectively. c 4 and c 5 are strength constants that are assigned to be and -(negative sign for attractive e ect). D 4b is the distance within which agents can communicate and decide on their collective action: they can stop moving towards one another and seek to exit as a group. d 4 is a term employed to ensure that an agent achieves the correct orientation, towards its target. D 50 is the influence distance of E 5 in Equation . .
The computational process is described in Fang et al. ( ). Each agent in the simulation processes a sequence of algorithmic steps of "decision-making" in every time-increment: observe and update perception; refresh sampling points for VPE computation; compute an evacuation route; estimate others' movements; calculate VPEs to reach a locomotion decision; and execute the decision. In the second-to-last step, an agent's locomotion is decomposed into translation and rotation. The agent needs to first consider whether to rotate or not and a erwards translates when an orientation decision is made. Both rotation and translation decisions are dependent on VPEs computations.

.
The value of the parameters used in EgressSFM and detailed validation studies can be found in Fang et al. ( , ). The validation exercises undertaken in Fang et al. ( , ) include comparison of simulation results to those of field experiments and other refined models.

The Station Nightclub Building Fire
. Once pyrotechnics ignited polyurethane foam lining the walls and ceiling of the band platform and dance floor, the fire spread aggressively. Film shows that at seconds a er initiation, the dense black smoke layer was near the floor, while ". . . the entire club was engulfed in flames within minutes of initiation (Gill & Laposata )". The fire occurred on February , during a heavily attended night. The building, which had accumulated a number of risks, some quite severe, over the years mainly through the practice of grand fathering the structure from recent and safer municipal building code requirements (Barylick ), was a single-story wood frame building shown in Figure . It was comprised of multiple spaces or ecologies, including a dance floor and a raised platform in front of it for the performers, a sunroom, a dining room, main bar, kitchen, dart room, bathrooms and o ice. It had four exit accesses: front door entrance, main bar side, kitchen side, and platform side. The crowd began to evacuate as soon as it became clear that the fire was occurring, or about twenty five seconds a er ignition. The last person to escape came out minutes and seconds a er ignition, although most survivors got out during the first seconds (other details of the fire and a timeline are in Tubbs & Meacham ( ). For the purposes of this study, the simulation's timeline count starts the moment the crowd began to evacuate (considered to be seconds a er ignition). . The platform's side exit (see Figure , west side of building) was blocked by the spreading fire and became impassable about seconds a er ignition. Only occupants escaped through it, while on the other side of the building only escaped through the kitchen door, probably because of its very poor signage and lack of visibility. The remaining survivors evacuated through the front entrance and the main bar side exit, and respectively. When other would-be evacuees clogged the spaces near the main hall, main bar and the corridor of the front entrance, some of the other occupants at the back of these queues searched for alternative egresses, eventually breaking the windows of the main bar room and sunroom about seconds a er the fire started. In this manner another occupants escaped the fire. Clearly, attendees were not blindly following others to the main entrance in a herd-like manner but showed initiative and creativity as they tried to exit the building. Nevertheless, occupants died from severe burns and smoke injuries, making it the fourth deadliest fire in the nation's history.
The social organization of the Station Nightclub the night of the fire .
The Station Nightclub in West Warwick, Rhode Island was a popular dance hall for people in the city and region. percent of the patrons the night of the fire had visited the nightclub previously. An even greater percent ( percent) saw the sparks that started the fire. The gathering was composed of older than average concert goers (median age = years) and had an unusually high degree of sociality, amity and goodwill among its members (for an in-depth, a ecting although tragic reconstruction of the o en intimate relations among the people in attendance see Barylick ( ). Only percent of the people in the Station that evening were by themselves. The rest were members of groups. percent of them were in groups of persons, percent were in groups of and , and percent were in groups of or more members. Almost half ( percent) of these groups were made up by coworkers and friends, dating partners, and kin and spouses (Aguirre et al. a,b, Unpublished; Torres ; Best ). percent of the members of groups were in close proximity of each other when the fire started (with the average distance of group members to each other less than linear feet.) Size of group and distance among group members are highly statistically correlated (Pearson R . ). .
The social cohesion produced by the norms shared by members of these groups can be measured "in extremis" even if ghoulishly, by examining the extent to which group members stayed with other members of their groups in the midst of this fire even if by doing so they augmented their chances of death and injury, showing the strengths of systems of social control that operated in this instance. For, as with injury (Pearson R . between size of group and the chance of injury), the mean number of dead increased almost monotonically with the sizes of the groups (Pearson R . ). Thus, the groups of persons had a . mean number of dead persons; groups of had a . mean number of dead; groups of had a mean of . ; groups of a mean of . ; groups of a mean of . ; groups of a mean of . ; groups of had a . mean number of dead; and groups of or more persons had a . mean number of dead members (similar finding are reported, among others, by Cornwell ( )). A somewhat unusual characteristic of this gathering is that when the fire struck, there tended to be a division by space and gender inside the building, for males separated from the other members of their groups tended to congregate at or near the bar while their female counterparts congregated in the dance floor. The result is that once the fire commenced, there was a good deal of movement in opposite directions of men and women searching for each other and unintentionally creating "knots" of people who blocked the paths of other would be evacuees, in an environment that was deteriorating very rapidly as flames engulfed the building. .
Human density (number of persons per square foot) of the ecologies inside the building was also an important predictor of deaths and injuries (Pearson R . ). The highest death rate occurred in the ecology to the north of the main bar in front of the bar windows (. death per square feet). Perhaps many of those who perished in this area migrated to the space trying to reach the main entrance nearby and were overtaken by smoke and fire due to delays in evacuating caused by the large number of people in front of them who were also hoping to exit through the front door. The high percent of dead and injured in this fire is partly the result of these social organizational features. Many victims lost precious seconds in the search for their group members, while others were inconvenienced by the knots of people that formed in the middle of the building. For these and other reasons, the resulting delays in evacuating placed many of these victims in the back of the throngs of people who were also trying to escape the fire. .
In the next sections, the Scalar Field Method is presented and the relationships of these groups are comprehensively modeled and quantitatively analyzed.

Assumptions and Model Implementation
Environmental hazards . Environment hazards are harmful to an evacuee's health. In particular, fire can lead to burn injuries and fatality, and the toxic e ects of smoke will reduce an evacuee's stamina (Bryan ; Best ). An agent's mobility is related to whether or not its "health" is impaired (Pauls ; Klote et al. ). To describe an agent's health, stamina is quantified as a scalar number termed energy level, or EL (terminology adapted from Aguirre et al. a; Best ; Aguirre et al. Unpublished), not to be confused with the VPE used by agents to model their rational thinking process. EL is a non-negative quantity, and the agent's mobility is assumed to be dependent on its EL. The lower the energy level is, the more injured the agent is and the less likely it can move and exit the building. Once the energy level is zero, the agent is assumed to have died. .
The building and environment model of the EgressSFM takes into account fire and smoke hazards. Fire is presented herein as a series of rectangular areas with stochastic sizes and start times.

An Agent's Energy Level
.
Before the fire occurs in the simulation, each agent is assumed to have an initial EL based on occupant demographics with a stochastic element added to account for variability. The initial EL values are taken from Aguirre et al. ( b), Torres ( ), and Best ( ). A er the simulation begins, each agent in the building su ers smoke damage over time, manifested by a reduction in energy level, until it either evacuates or is killed. An agent's energy level is computed as follows (based on Best ( )): . When an agent does not move from an active fire region, its energy level drops instantaneously to zero. This signifies that it is deceased.
. Smoke leads to a gradual reduction in an agents' energy levels in all building spaces except as noted next.
The EL changes at the rate of -. , -. , and -. EL/second during the time periods of -second,second, and a er second, respectively.
. Based on an analysis of oxygen volume fractions conducted by Grosshandler et al. ( ), as shown in Figure , agents in the main bar room are assumed to su er damage at a lower rate ( % of values specified above) because: ) this room is far away from the fire, and the fire and smoke are impeded by the walls of the front entrance corridor and kitchen; and ) this room has access to one side exit and multiple windows that can provide more fresh air than other rooms.
. When an agent is present in an oxygen zone, the damage rates of EL are divided by a factor of -. (Gill et al. ) to recognize the beneficial e ects of oxygen. As a result, the EL gradually increases in oxygen zones. .
An injured agent is assumed to su er mobility loss that is linearly dependent on the ratio of its current energy level to its initial energy level. If the energy level is equal to or higher than % of the initial amount of energy, the agent's maximum velocities are not influenced. Otherwise, the agent's maximum velocities in various directions are lowered in a linear manner with the remaining energy level, as shown in Equation .
max.v original max.v = 0.2 + energy level initial energy level

Egress model implementation .
As modeled in EgressSFM (Fang et al. , ), the Station Nightclub building model is comprised of a collection of exits, doors, windows, and interior spaces. Agents that reach exits are considered to have safely exited. Each exit has an open and close time that determines whether this exit is available (passable) or not, respectively. Application of such open/close times is necessary to account for dynamic conditions during the fire, e.g. the side exits became impassable as the fire progressed. Windows are a special set of exits that are normally impassable. They can switch functions to enable egress a  Table . Such times are based on the simulation timeline starting when the crowd begins to evacuate, as initially estimated by Grosshandler et al. ( ).

Front Entrance None Platform Side Exit Main Bar Side Exit
None Kitchen Side Exit None All Windows None Table : Time of opening and closing of exits and windows.
. The agent's normative behavior is controlled by the Scalar Field Method as discussed earlier. The agent model is implemented to address the demographic and interview data of the Station Nightclub fire as follows: . Personal demographic information of age, gender, initial energy level, and prior visit experience are considered. The term 'prior visit experience' pertains to whether the agent has visited the building before the night of the fire, i.e. it accounts for familiarity with the floor plan, which presumably facilitates successful evacuation from the building.
. Initial location and orientation. Initial location of each occupant is determined based on coding of survivor interviews (Aguirre et al. b; Torres ). Each agent's initial orientation is randomly selected for each simulation.
. Social a iliation. The majority but not all agents are members of one of the social groups, which are characterized by a specified type of relationship: they were either alone, with co-workers, or with friends, dating partners, family members, and multiple group types. The first term refers to an individual without pre-existing relationship to others. The last term means an agent is in more than one group type.

. Group leader. A social group can have a leader that influences other members' decisions in this group.
In the case of the Station Nightclub scenario, group leaders were identified and coded based on survivor interview data (Aguirre et al. Unpublished; Torres ; Best ). .
Age determines an agent's mobility before being injured. The maximum speeds of each agent are dependent on its age category. Adult agents (ages to ) are assumed to have a maximum forward speed that is randomly selected from a range of . m/s to . m/s to reflect the stochastic nature of moving individuals. Agents in the "children + seniors" category have a maximum forward speed in the range between . m/s to . m/s. The lateral speed limit is selected as . m/s and the backward limit . m/s for the "adults" and as . m/s and . m/s for "children + seniors". The maximum rotational capability is randomly determined between rad/s to rad/s for the "adults" and half of that value for "children + seniors". These speeds are based upon on information in previous studies (Tang & Ren ; Thompson ). The initial orientation of each agent is allocated randomly.

.
Prior visit experience influences an agent's awareness of side exits, so that if they had never visited the Station Nightclub they lacked awareness and would have a higher probability of missing an exit near to them. The data generated by Torres ( ) and Aguirre et al. ( a,b) show that close to half of the evacuees had no prior visit experience. Grosshandler et al. ( ) mentions that approximately percent of the occupants believed the main entrance to be the only exit. In this study, prior visit experience is assumed to determine an agent's knowledge of the floor plan when the evacuation starts: an agent without prior visit experience is aware of the front entrance exit and main bar side exit only, and is assumed to be unaware of other side exits. An agent who visited the building previously is assumed to know all the exits. However, an agent can learn from the surrounding environment and updates its knowledge and considers other exits as alternative potential destinations. .
For simplicity, this study assumes a dichotomous coding of social relationships, either friend-or kin-related. Thus, each agent has the same type of social relationships to other group members in the same group. Spouses and dating partners are interpreted as kin-related in the SFM, and co-workers and friends are categorized as friend relationships. Members of more than one group are also assumed to be friend-related. If a group leader is specified in a strongly bonded relationship like spouses, the group leader is responsible for leading the group, and the other group members are assumed to follow the leader. To do so, the leader establishes kin-related interactions with each of the other group members, but non-leader members only set up a social relation with the leader. In addition, the non-leader members duplicate the leader's decision to follow a specific escape route.
. An agent has multiple potential choices of destination for egress, since there are four exits and two walls with windows. Selecting the exit, particularly the platform exit, was discussed by previous researchers such as Grosshandler et al. ( ) and Best ( ). Both studies assumed the occupants to always select the closest exit and applied algorithms to control their decisions. The former used two so ware packages, buildingEXO-DUS and Simulex. In the simulation with buildingEXODUS, the platform exit was assumed to be impassable a er s, and the front entrance was blocked at s. In the Simulex simulation, Grosshandler et al. ( ) first calculated number of occupants who would use the platform exit, which resulted in people, and then made the platform exit only visible to these occupants. The latter study conducted by Best ( ) assumed that only occupants were aware of the existence of the platform exit and that / of the occupants believed that the main entrance was the only exit. .
In this study, the agent generally selects one exit to which the travel distance from the agent's current location is the shortest, although the agent is not forced to use it. The final choice of exits is dependent on availability of exits, prior visit experience, and group leadership, as discussed previously. To address the fact that only a limited number of people escaped through the platform exit, a penalty is added to each agent's perception of this particular exit's distance, so as to make it less desirable as an exit. This empirical approach is motivated by two facts: ) this exit door swung inwards rather than outwards, and hardware on the door was broken (Grosshandler et al. ), and ) the exit was close to the fire and covered by heavy smoke shortly a er the fire ignited. A -meter penalty is selected to use as shown by the parametric study shown in Figure , in which the number of agents using the platform exit is simulated with various penalty distances. As can be seen, the correct number of agents using the exit corresponds to the use of a m penalty. Figure : Parametric study of the 'penalty' distance to the platform exit.

Simulation and Hypothetical Investigations of Social Traits
. The egress scenario in the Station Nightclub Building Fire is modeled in EgressSFM. The simulation results are shown in Table . Because of the stochastic nature of the simulations, twenty simulations are conducted and average values and standard deviations are reported. The computed number of occupants using each exit and people dying are compared to the actual values as reported by Aguirre et al. ( a,b) (see also Best ( )). Note that the sum of 'actual' occupants in Table adds up to and not the agents simulated herein. The di erence is due discrepancies in the published literature about the actual number of patrons in the nightclub, e.g. as reported in Grosshandler et al. ( ). Nevertheless, the di erence between the sum of actual and simulated people is less than %. As shown in Table , the simulation results match the actual statistical data reasonably well. The implications of this favorable match are discussed later on.

Bar exit Kitchen exit Platform exit Windows Deceased
Actual Simulated Standard Deviation Table : Simulated and actual data of escaped and deceased occupants. .
To give an impression of how the simulation progresses, snapshots at the initial starting point and a series of intermediate times during one run of simulation are taken and presented in Figure . An important observation is that the egress process lasted for most victims less than seconds due to the extreme severity and rapidity of the fire and smoke that enveloped the building. During the egress process, pre-existing social relationships took place and influenced agents' decisions and behaviors. Group behavior driven by strong interactions, influenced neighboring agents and led to clogging and delays in egress (c.f. Aguirre et al. Unpublished). For example, agents in social groups are involved in kin-related interactions at a particular instant of time rather than in prompt evacuation thereby delaying themselves and others. In the remainder of Figure , the color variation of the agents signifies di erent degrees of impairment, specifically an agent changes color gradually from green (and its variations) to yellow to red based on the remaining energy level compared to the initial level. Dead agents are colored light gray. Generally, the main exit, bar exit, and windows played primary roles for egress, and other side exits were ignored by the majority of the agents, perhaps due to loss of visibility. Moreover, the toxic e ect of smoke impaired agents' health and lowered injured agents' mobility and their ability to egress from the building. .
Two areas, as highlighted in Figure , are found to be critical for overall egress e iciency of the occupants in the building. These areas are the connection between the main bar and main hall and the connection between the front entrance and main hall. Along with the corridor of the front entrance exit, these areas are filled with agents and become problematic because of the presence of strong social bonding, e.g. spouses and dating partners. Agents driven by such interactions tend to congregate with their groups, and such gatherings lead to tra ic congestion in the connection areas. As a result, as initially reported by Best ( ) and Aguirre et al. (Unpublished) and confirmed herein, the overall egress is delayed by these bottlenecks. .
To investigate the influence of social traits in a quantitative manner, two series of parametric simulations are conducted. The first is based on "break down", which is defined as the degree by which an agent ignores its social a iliations. The second focuses on the e ect of prior-visit experience. Each simulation shown hereon is conducted twenty times to account for the stochastic nature of the problem.

Break down of social relationships .
As shown in Table , the numbers of agents using various exits and those that are deceased are compared in a sensitivity study of "break down" probabilities of %, %, %, %, %, and %. In particular, the case of % assumes that every agent responds to its pre-existing relationships, and the case of % assumes that all agents ignore their social a iliations and egress alone as individuals. Four plots are drawn in Figure to showcase the tendencies of using front entrance exit, main bar exit, window exit, and deceased agents versus the "break down" probability, respectively. As shown, they are generally linearly dependent on the break down probability. More agents successfully evacuate through the front entrance exit and main bar exit as the break down probability increases, i.e., as more of them become free agents. Table : Sensitivity study of e ect of the "break down" probability.

Probability Main exit Bar exit Kitchen exit Platform exit Windows Deceased
.
An example of the case of % (pre-existing relationships fully active) is given in Figure a,    in Figure b. In Figure a, like-colored agents are members of the same group. As they attempt to congregate, they slow down all other agents by blocking their path. In Figure b, the variation in green color implies agents with di ering EL, the darker the shade, the lower the EL. These variations in mobility contribute to an overall slowdown in the evacuation process. Figure : Sensitivity study of the e ect of the "break down" probability. . The number of agents using windows to evacuate decreases because the number of remaining agents in the building decreases when the windows become passable at seconds. As a result, the number of deceased agents decreases and is almost half of the % condition when every agent drops its social relationships. Clearly, the presence of social relationships increases potential risk and delays the overall egress. This result is consistent with many previous studies, e.g.

Prior visit experience .
An agent who has no prior visit experience is considered to be only aware of the main and bar exits and unaware of others such as the kitchen and platform exits. To explore the influence of such limitations, a set of control tests, which are comprised of % and % "break down" cases are conducted under a hypothetical situation in which all agents are assumed to have prior visit experience and awareness of the full floor plan. The simulation results are drawn in four pie charts as shown in Figure , in which the numbers of agents using various exits and the number of deceased agents are divided by the total number of agents and presented as di erent components. Figure a and  % "break down" conditions. As expected, significantly more agents evacuate through the platform exit and kitchen exit, so the deceased agents are fewer. On the other hand, the number of agents who use the front entrance exit and the main bar exit are not a ected.

Meaningfulness and Implications of the Computational Study
. Even though EgressSFM was extensively validated in Fang et al. ( , ), it was not validated under the same conditions for which it was exercised in this work. The Station Nightclub scenario modeled is complex and incorporates many levels of multi-dimensional interactions that occur between numerous actors in the simulation, i.e. agents, physical building components such as walls and exits, fire regions and oxygen zones. Each of these interactions are modeled based on key assumptions as outlined in the manuscript. Thus, it is naturally di icult to draw strong conclusions about the fidelity of the simulation results. Yet, the observed reasonable comparison to field data lends credence to the simulation model and suggests that it is capable of capturing some key aspects of the event. Obviously, this is not rigorous validation of the model given the extent of uncertainties and assumptions. However, it is acceptable since the methodology employed is the only way available at present to carry out ethical studies of these crisis evacuations and conduct quantitative parametric research on a uniquely complicated and multi-disciplinary problem that has implications for life safety within facilities.

Summary and Conclusions
. This paper reports on the use of the EgressSFM platform that used the Scalar Field Method (SFM) to model a historical egress scenario, the Station Building Fire. The platform is modified to incorporate environmental hazards of fire and smoke, and computes each agent's stamina as an energy level, which impacts the agent's mobility. The study considers demographics and social relationships of the occupants in the building when the fire happened. When calibrated, the simulation captures the realism of the actual data, and shows EgressSFM's ability to reasonably handle the complex social relationships and group behaviors present during egress. The parametric simulation exercises show in a quantitative manner that the presence of intimate social a iliations delay the overall egress, and that lack of knowledge of the building floor plan limits exit choices and adversely a ect the number of safe evacuations.