Understanding the collective dynamics of crowd movements during stressful emergency situations is central to reducing the risk of deadly crowd disasters. Yet, their systematic experimental study remains a challenging open problem due to ethical and methodological constraints. In this paper, we demonstrate the viability of shared three-dimensional virtual environments as an experimental platform for conducting crowd experiments with real people. In particular, we show that crowds of real human subjects moving and interacting in an immersive three-dimensional virtual environment exhibit typical patterns of real crowds as observed in real-life crowded situations. These include the manifestation of social conventions and the emergence of self-organized patterns during egress scenarios. High-stress evacuation experiments conducted in this virtual environment reveal movements characterized by mass herding and dangerous overcrowding as they occur in crowd disasters. We describe the behavioural mechanisms at play under such extreme conditions and identify critical zones where overcrowding may occur. Furthermore, we show that herding spontaneously emerges from a density effect without the need to assume an increase of the individual tendency to imitate peers. Our experiments reveal the promise of immersive virtual environments as an ethical, cost-efficient, yet accurate platform for exploring crowd behaviour in high-risk situations with real human subjects.
In a new study published in PNAS, we studied how people form and revise their perception of a risk, and how their evaluation of a danger is influenced by the judgment of their peers.
For risk experts, the assessment of danger results from a very pragmatic calculation: Simply multiply the probability that an accident would happen with the damage that this accident would cause, and you will obtain a number that describes how serious is a risk. In such a way, plane crashes or nuclear explosions are low risk because their probability is so low that it compensates the potentially high number of loss when they happen, but car accidents, smoking, or over-exposure to the sun are high risk because they are very frequent although they affect just a few persons at a time.
Yet, public perception of risk is often at odd with experts’ assessments, because people do not calculate such probabilities. Instead, they form risk judgments in a much more intuitive and emotional way, on the basis of various features of the risk.
In this study, we wanted to know to what extend people’s perception of danger is influenced by the judgment of their friends and relatives, and how this process of social influence plays out in large populations of interacting people.
For this, we used a nice experimental paradigm that was pioneered by Frederic Bartlett more than 80 years ago: The telephone game. The experiment is very simple: The first participant receives a message and is instructed to communicate it to a second individual, who in turn communicates this information to a third person, and so on until the end of the chain (10 subjects in our experiment), as illustrated by this famous painting of Norman Rockwell.
The interesting variation in our experiment, is that the communicated message was not just a neural story, but a comprehensive set of information about the benefits and harms of a controversial and widely used antibacterial agent called Triclosan. Therefore, participants intuitively combined this message with their own subjective judgment of the risk before passing it to the next subject. In such a way, we managed to mimic how a rumor would spreads in a social network, directly in the laboratory.
With no surprise, we found that the initial message was seriously distorted as it spread down the chain. We ran 15 chains in parallel. The message that arrived at the last participant was almost completely different between the chains, sometimes focusing environmental damages, on breast-feeding, or on Greenpeace protests. At the end end the chain, the message was furthermore much shorter than the initial imput, and contained a lot of wrong information.
The more important finding is that the signal of the message was also distorted. The degree of risk that the message carries became gradually more extreme over repeated social transmissions. As it spreads down the chain, the message became either more alarmist or more reassuring, the direction of the change being determined by the initial judgmen of the subjects. In other words, people who believe that the risk is low tend to tone down the dangers during their communication (e.g. they may forget to mention some critical side effects), while those who are initially worried tend to amplify the risk (e.g. they may forget to mention some existing regulations).
In other words, people act as a filter during risk communication, by ‘coloring’ the message according to what they initially thought about the risk. Therefore, instead of changing people’s mind, the message mutates to fit people’s judgment.
With this social process, the worst-case scenario happens when a risk message spreads within a community of like-minded people – which is arguably also the most common situation. In that case, any risk message would mutate rapidly to fit the opinion of the community. Pieces of information that contradict the view of the community will disappear, while those that support their judgment will be highlighted.
Ultimately, the message can have a counter-intuitive polarizing effect on the population: After a few rounds of distortion, the message can strengthen the existing bias of the group, even though it initially supported the opposite view.
This study demonstrates that people’s risk perception is – at least in part – the outcome of a social process.
Following an experimental study that we have recently conducted about the mechanisms of opinion change, two important trends have been identified. First, our computer simulations show that the presence of a very confident person (let’s call him an expert) in a group of lay people has a strong attractive power on the others. Most of the time, simulated group discussions end up with the entire group following the expert’s opinion. The second mechanism is more subtle. In the absence of experts, the group tend to converge toward the opinion that is shared by the majority of people, even though these individuals have very low confidence levels. In fact,if many people have a similar judgment, it comforts the others following them.
Therefore, two factors can have a significant influence during a group discussion: A few experts, or a large majority. Now the interesting question is: What happens when both attractors are simultaneously present in a group?
To answer this question, we have conducted specific computer simulation with a large majority of lay people facing a small minority of highly confident experts. The results show that the group outcome actually depends on the relative proportion of experts, as indicated in the figure below:
If the group is constituted by less than 10% of experts, the majority wins. When the proportion of experts is higher than 20% they manage to attract all the others. In between, a conflicting situation occurs, in which the group split in two clusters. Therefore, if you’d like to convince a group of 20 people, you would need at least 4 of them expressing a their judgment with high certainty.
M. Moussaïd, et al.
Social influence and the collective dynamics of opinion formation
PLoS ONE, 2013
Following the experiment I described in a previous post, one side question that was raised concerns the relevance of the confidence level as an indicator of accuracy. That is, we wanted to know if it’s a good strategy to imitate people expressing a high level of confidence.
To answer this question, we have asked participants in the lab a series of 32 factual questions, such as “To you opinion, what is the distance between Berlin and London ?”. Additionally, participants were instructed to evaluate their level of confidence on a scale ranging from 1 (very unsure), to 6 (very sure).
Now the question is: How strong is the relation between the confidence level of the participants and their accuracy? The following figure gives some first answers.
So in the above figure, the colors indicate the confidence level, from 1 to 6, and the height of the bars indicate the probability to get a very good, a good, a bad, or a very bad answer. The maximum confidence level 6 is indeed a good indicator of the quality of the answer. It leads to a good or very good estimate about 80% of the time. However, lower confidence levels are less reliable. For example, the second highest confidence value C=5 has already 40% chance to be a bad or very bad estimate – not far from a random chance! Moreover, the confidence level C=4 – which is still above the average – actually has a greater chance to be a bad or very bad answer (53%) than good or very good (47%). Finally, the lowest confidence values C=1 and C=2 do not even differ from each other.
Therefore, the confidence level is a bad indicator of quality. Only people who are absolutely certain of what they say are reliable, not the others.
(and by the way…. Berlin is 910 km away from London).
M. Moussaïd, et al.
Social influence and the collective dynamics of opinion formation
PLoS ONE, 2013
Like many people, I believe that my personal opinions are the results of a careful, rational and independant thinking process. Well, like many people, I’m partly wrong…
Past research in social psychology has many times demonstrated that our judgments about political ideas, new technologies, or commercial products are strongly influenced by the opinions of others. Solmon Asch’s conformity experiment is probably one of the most studied example of social influence. Likewise, a nice experiment by Myers & Bishop shows how people discussing issues about racial prejudice influenced each other in the course of the discussion. Examples are numerous.
In a recent project, we conducted this very simple experiment. We asked experimental subjects in the lab to answer a series of factual questions, such as “To your opinion, what is the aluminium’s melting temperature, in celsius (°C)?“.
Do you have any idea? Try to guess before looking at the figure below… In addition, participants were instructed to evaluate their confidence level on a scale ranging from 1 (very unsure) to 6 (very sure) : “How sure are you about that?“. As you can see from the figure below, most people underestimate the correct answer for this particular question:
Well, but the interesting part of the experiment comes now. Each participant was then exposed to the opinion and confidence of another randomly selected person. For example we told them: “Ok, now another participant has answered 150°C with a confidence level 4. Would you like to revise your opinion now?“. Depending on the information they received, some people prefered to ignore the feedback and maintained their initial opinion, while some others choose to revise their initial view accordingly.
How strong was the social influence? The behaviour of 59 participants over 15 different questions is summarized in the influence map below (this is a simplified representation, see the original article for more precise results).
This result tells us a lot about the mechanisms of social influence. For example, we can see that the upper part of the map is all blue, which means that most people choose to ignore the feedback if they are initially more confident than the other person. In the bottom of the map, however, a green zone appears, and becomes red when the other person has a much larger confidence level.
On the other hand, the level of agreement is another important variable. When participants have similar opinions in the first place, they keep it (left-side of the map). When they strongly disagree, they tend to ignore the opinion of other person as well, unless a very strong confidence level is provided (right-side of the map). In between, we see a zone where the strength of social influence is maximized.
The next question that we will adress is whether the confidence level of the participants is a good indicator of quality. To be continued…
M. Moussaïd, et al.
Social influence and the collective dynamics of opinion formation
PLoS ONE, 2013
- This project is an experimental investigation of the lane formation phenomenon. Lane formation is a self-organized pattern that appears spontaneously in crowds: When you have two flows of people moving in opposite directions (for example in a crowded commercial walkway), a sort of “pedestrian highway” sets up naturally, where people moving in the same direction occupy one half of the street and those moving in the other direction occupy the other half.
When it has been first observed in the 1970th and first reproduced by computer simulations in 1995, lane formation has been quickly labeled as a smart collective behavior, a nice illustration of the wisdom of crowds. In fact, this collective pattern is an efficient way of organizing counter flows to minimize avoidance maneuvers and frictions among people moving in opposite direction, just like cars in highways. The major difference with cars, however, is that highways of pedestrians are self-organized: They don’t require any specific centralized traffic rule. They appear naturally, when it’s necessary… That’s the magic of self-organization!
So, to answer that question, we have designed a big lab experiment to characterize the features of lane formation. For this, we hired up to 60 participants and put them in a large ring-shaped corridor. Half of them were instructed to walk clockwise, and the other half to move anti-clockwise. Participants were randomly located in the corridor at the beginning. At the starting signal, they started to move in their walking direction. As expected, it took about 20 to 30 seconds before the two opposite flows segregate almost perfectly, without giving the participants any specific instructions about that. A very nice emergence of order out of disorder, as it has been theoretically described in the study of self-organized systems. The video below shows an example of what we observed during the experiment, after the tracking of pedestrians.
These results can suggest some real-life applications to enhance the traffic efficiency and the walking comfort in crowded cities. For instance, dividing the pavement into a ‘‘fast lane’’ and a ‘‘slow lane’’ will reduce the speed differences among pedestrians, and therefore maintain stable and well-organized traffic flows. I heard about a urban project like that in Oxford street, London, but in practice, it is not sure that people will obey such a constraining walking rule…
M. Moussaïd, et al.
Traffic instabilities in self-organized pedestrian crowds
Plos Computational Biology, 2012 (Online Article)
Here we go! It’s Christmas again! And as usual, commercial streets, shops, and malls are getting totally crowded over the week-end.
Bad news? Not for everybody…For me, it’s the best moment to go fishing some data for my research.
Here is a short sample of what I got today: a very dense bidirectional flow of shoppers coming in and out of the Alexa center in Berlin. Admire the very nice formation of lanes! Unfortunately, I just recorded a few minutes of this before I was kicked out by the security guards of the mall. I tried to explain them that it was my job, but apparently I was too suspicious for them and had to stop the camera… Too bad!
Today, I would like to talk about my very first research project. I undertook it during the first year of my PhD thesis, at the University of Toulouse and at the ETH Zürich. The main question we addressed in this work is very simple: How do people avoid each other?
In fact, if you think about it, avoiding a collision with another pedestrian when walking in the street is a pretty complex cognitive task: It requires 1/collecting various information about the walking behavior of that person, 2/ anticipating what he or she is going to do, 3/ processing all these information in real time, and 4/ adapt you walking behavior accordingly by choosing an avoidance side and changing your walking speed and direction. And most of the time, you just have a few seconds to take your decision.
Interestingly, this task comes down to a coordination problem. The pedestrian has to choose a passing side: “Should I avoid that person on the right- or the left-hand side? ” If both individuals choose the same passing side, the avoidance is a success. If they choose different ones, the risk of collision is increased. OK, but you don’t know what the other one will decide… and he is actually facing the same problem! So how do we do? And why do we choose correctly most of the time, without even being focused on the problem?
To investigate on this issue, we have conducted a very simple pedestrian experiment. In the lab, we built a simple experimental corridor and hired about 40 participants. Then, we selected randomly two of them, put them at each end of the corridor, and instructed them to walk to the other end at the starting signal… and repeated this procedure many times, for each pair of participant. Of course, they had no idea that we were actually observing how they avoided each other in the narrow corridor. The experiment looks like this:
The analyses of the video recordings highlighted a clear bias in the avoidance side. Without talking to each other and even looking at each other, 81% of the participants (they didn’t know each other) spontaneously choose to evade on the right-hand side. This behavioral bias is called the side preference. It’s a simple cognitive shortcut that helps to reach a rapid decision without thinking too much: “When you are not sure, go to the right!” And because the majority of people apply the same strategy, the coordination is achieved successfully almost all of the time.
The side preference is a behavioral convention, or social norm. In other words, it’s an implicit behavioral rule that is shared by the majority of people without being explicitly imposed by any central authority.
Does the side preference of pedestrians derive from car traffic rules? In fact, the side preference is a cultural feature of crowds: In some countries, pedestrians avoid on the right-hand side (like in France, where the experiment has been conducted), and in some others they prefer the left-hand side (like in Japan). However, it’s not always consistent with car traffic. In Great Britain for example, and London in particular, people walk on the right-hand side but drive left (as shown in the photo of Oxford street, London).
In the end, it is likely that car traffic rules have some kind of influence on pedestrians’ habits, but it is probably not the main mechanism. The most plausible theory, to my opinion, is a simple self-reinforcing learning process: Initially, people choose at random (as young children do). If, by chance, they choose the same side, the probability for choosing that side again in future interactions is increased, otherwise it is decreased. With this process, over repeated interactions, most people will develop the same preference in the end. As both sides are equivalent in the beginning, the theory predicts that different preferences emerge in different regions of the world, as it is actually observed.
Interestingly, the side preference is not just “a detail” of crowd dynamics. It has some importance consequences. In dense bidirectional flows of people this bias is amplified, leading to the emergence of a sort of “pedestrian highway”, where all people moving the same direction occupy on half of the street (on the preferred side), and those moving in the opposite direction occupy the other half. This self-organized phenomenon actually enhances the traffic efficiency and can be easily observed in crowded streets.
M. Moussaïd, D. Helbing, S. Garnier, A. Johansson, M. Combes & G. Theraulaz
Experimental study of the behavioural mechanisms underlying self-organization in human crowds
Proceedings of the Royal Society B: Biological science, 2009 (Online Article)
Modeling crowd behavior ? What does that mean?
Well it’s pretty simple, modeling consists in finding a proper description of how a pedestrian behaves in a crowd: How he navigates, avoids obstacles, or when he starts to panic. In particular, a “good” model should be able to predict the emergence of known collective crowd behavior, such as lane formation or crowd turbulence. Elaborating a good and reliable model is essential for two very important reasons:
- First, the model is an essential research tool. It helps to understand why a particular pattern emerges, under which conditions, and which behavioral variables affect its features.
- Second, the model is also an important planning tool from an applied perspective. It can be used by urban planners to predict in advance the behavior of crowds around a stadium or during a music festival. It can also help for the planning of evacuation strategies, or the assessment of urban layouts.
Therefore, a good model is the key to many theoretical and applied insights.
All right, so how do we build a model?
Today, much of the existing models are based on a small trick: an analogy with physical systems. The main idea is to consider that the movements of a pedestrian in the crowd are similar to movements of a particle in a gaz. Based on this assumptions, it is possible to make use of tools from Newtonian mechanics to describe the behavior of a pedestrian by means of attractive and repulsive forces. The pedestrian is attracted toward a destination point, but at the same time repulsed by other pedestrians. And it works pretty well. Today, force-based models are probably the most dominant in the scientific literature.
Great! So what’s the problem?
The problem is that pedestrians are actually not particles… To describe precisely the movements of a pedestrian with Newtonian forces, one usually needs pretty sophisticated equations of motion, which are hard to calibrate. Moreover, the movements of simulated pedestrians during computer simulations look pretty artificial and sometimes obviously not realistic.
Recently, we though it could be worth spending some time to elaborate another approach for modeling pedestrians’ behavior. And the idea was pretty simple: Instead of treating human beings as if they were molecules, let’s treat them as if they were human beings…
The model we suggested is pretty simple. The first step is to describe mathematically the visual information a pedestrian has in a crowd. For this, we defined the limits of the vision field and compute the distance before collision with all other individuals and obstacles. This constitutes the main visual input used by the pedestrian to navigate.
Next, we assumed that the pedestrian rely on simple movement rules to adapt his walking speed and direction, based on the visual information previously calculated. The first one: choose a walking direction that minimizes the coverage of the vision field without deviating too much from the destination point. The second one: adjust the walking speed to keep some safety distance with the closest obstacle or individual.
These rules are called heuristics – a term of cognitive science describing rapid decisions that people make without much thinking about their behavior. Finally, in order to reproduce realistically crowd motion in situations of overcrowding, these navigation rules are combined with the physical forces that occur during unintentional body contacts among people in dense crowds.
Numerical simulations of this model show that this simple framework is able to generate a large variety of collective behaviors, such as the spontaneous separation of opposite flows of pedestrians in bidirectional traffic (as shown in the movie below).
Interestingly, the model also predicts that above a threshold crowd density (around 3 people per square meter), a transition from smooth flows to stop-and-go waves happen, as it has been observed in real crowds. Moreover, as the density increases, congestions emerge, followed by chaotic crowd instabilities, similar to the phenomenon of crowd turbulence previously observed during crowd disasters. The model could help urban planners to design better exit routes for evacuation of large crowds from buildings and to adapt the environment for a safe planning of mass events. For example, it allows for the identification of zones where the occurrence of crowd congestions is likely, as shown in this last illustration.
All right. So your model is the best one, right?
Not at all! This model has several drawbacks as well. First of all, the model is hard to program. Even through the human brain can easily compute expected collision times with moving obstacles when we are walking in a crowd, it’s pretty hard to implement in a computer program. I did manage it finally, but it requires heavy computational power and is prone to programming errors. This technical limitation hinders the future development of the model. Moreover, we are not sure that the behavioral rules assumed in the model are the ones actually used by pedestrians. And finally, just like the social force model, there are some simulation inconsistencies from time to time..
Well, I would say there are some advantages and disadvantages in using it, but the model is pretty young. It require to be polished, confronted to empirical data, optimized, etc.. Hopefully, it will be improved in the near future!
M. Moussaïd, D. Helbing, and G. Theraulaz
How simple rules determine pedestrian behavior and crowd disasters
Proceeding of the National Academy of Science, 2011 (Online Article)
It is pretty surprising to notice that current research about pedestrian flows always assumes that people are lonely walkers.
Simulated pedestrians move from one place to another, avoid each other, avoid obstacles, form all kinds of collective patterns, but they are always isolated: they have no friends to walk with, and they never move together with other people… in short, there is no social interactions among them… Sad, isn’t it? But don’t worry, that’s just the theory. In a recent article, we have shown that in practice things are different…
More than 50 years ago, the sociologist James Coleman already noticed that pedestrians moving within small groups composed of two of three individuals are more frequent than people walking alone (although it depends on the social context). Can we make pedestrian simulations more reliable and more realistic by including social interactions among individuals?
The starting point of this work is a set of video recordings that we made together with my frolleague (friend+colleague) Simon Garnier. In order to have a nice point of view, we went on the roof of this nice building in Toulouse (which is nothing else that the city hall, a very nice place to visit by the way :)).
From there, we had a good view on the passers-by for our analyses. After hours of recordings, and additional painful days of manual tracking on the computer, we finally got this dataset of about 1500 moving pedestrian groups.
In good agreement with Coleman’s findings, we measured that up to 70% of the pedestrians actually walk in small social groups, such as friends, couples, or familly members moving together. Group size vary from 2 to 4 persons, while groups of made of 5 members and more are rare.
Moreover, further analyses have shown that these groups – technically called “medium-scale aggregated structure” 🙂 – have very specific walking configurations: At low density, group members tend to walk side by side, forming a line perpendicular to the walking direction. As the density increases, however, the linear walking formation is bent forward, turning it into a V-shapped walking pattern, where people in the middle of the group tend to move back and those on the side get closer to each other.
This walking configuration may look completely natural to you, but for our scientific theories of pedestrian movements, that was a real bad news… In fact, current theory of crowds is based on analogies with Newtonian mechanics: it assumes that people in a crowd behave like particles in a gaz or a fluid. But according to the law of aerodynamics, a small aggregate of particles would never form a V when moving ahead! It forms a reverse-V , like a flock of migrating birds or a group of cycling competitors during the Tour de France. Damned! The theory has reached its limits…
So why do people deliberatly violate the laws of physics and break down 30 years of scientific work? Well, the answer is pretty simple. This is not about physics, this is about social interactions. People adopt a concave walking configuration because this is the only way for every group member to see all his friends with a simple head movement. Imagine that you are trying to talk to some friends located behind you.. That would be pretty unconfortable, obvisouly.
Modelling this behaviour was rather easy. Assuming people have a 180° vision field, and can turn their head left and right (seems like a reasonable assumption…), then one very simple rule is enough to reproduce the observed V walking patterns: “If you can’t see your friends, slow down“.
However, let’s come back to the physics of pedestrian flows. If the most efficient way to move ahead is the reverse-V walking configuration, then how bad is the observed V-shaped configuration? In fact, computer simulations taking into account pedestrian groups demonstrate that the traffic efficiency is actually decreased by an average of 17% as compared to a situation with isolated people only!
Groups reduce the flow capacity… but at least now we can chat with our friends!
The only solution to enhance the traffic efficiency without forbidding people to talk in the street is to provide pedestrians with larger sidewalks and a more confortable urban environment. One day maybe!
M. Moussaïd, N.Perozo, S. Garnier, D. Helbing, and G. Theraulaz
The walking behaviour of pedestrian social groups and its impact on crowd dynamics
PLoS ONE, 2010, 5(4):e10047 (Online Article)