Elementary modelling and behavioural analysis for emergency evacuations using social media

Social media usage in evacuations and emergency management represents a rapidly expanding ﬁeld of study. Our paper thus provides quantitative insight into a serious practical problem. Within this context a behavioural approach is key. We discuss when facilitators should consider model-based interventions amid further implications for disaster communication and emergency management. We model the behaviour of individual people by deriving optimal contrarian strategies. We formulate a Bayesian algorithm which enables the optimal evacuation to be conducted sequentially under worsening conditions.


Introduction
There has been increasing attention paid to evacuation problems in recent years (see e.g.Bish and Sherali, 2013;Bretschneider & Kimms, 2012;Lim, Zangeneh, Baharnemati, & Assavapokee, 2012).In addition to several high profile events (Schadschneider et al., 2008) this has been accompanied by increases in the frequency and severity of natural disasters coupled with population growth in various high-risk areas of the world (Bozorgi-Amri, Jabalameli, Mirzapour, & e Hashem, 2013).One of the key themes that has emerged in recent years is that emergency evacuations may be greatly aided by social media.
As evidenced by recent testimony to the UK Parliament (Preston, Branicki, & Binner, 2014a) social media usage in emergency evacuations represents an important practical problem to which our paper adds quantitative insight.The use of social media to co-ordinate emergency evacuations is already in its operational infancy (Chen & Xiao, 2008;Nakajima, Yamane, Hattori, & Ishida, 2008;Preston, Binner, Branicki, Ferrario, & Kolokitha, 2011;Preston, Binner, Branicki, Galla, & Jones, 2014b).Further real world examples include survey evidence of social media usage to gain information about emergencies (American Red Cross, 2010) and the Personal Localised Alerting Network (PLAN) in New York that has emerged out of a collaboration between the federal communications commission, the federal management agency and industry.
Within emergency evacuations a behavioural approach to OR is key (Hämäläinen, Luoma, & Saarinen, 2013;Tamura, 2005).Human aggression and terrorism (Annlin Jin & Paramasivan, 2012;Oh, Agrawal, & Rao, 2011), panic in response to perceived dangers ("flight panics") and greedy behaviour ("acquisitive panics" or "crazes") may all have important implications for real evacuations (Schadschneider et al., 2008).In our context behavioural and social processes are extremely important (Drake, Gerde, & Wasieleski, 2009).Social media has been variously used to document the London 7/7 bombings and the Fukushima radiological disaster amid several more localised tragedies (Preston et al., 2014b).Further, anecdotal evidence from practitioners at the UK cabinet office suggests that individuals continue to use social media, and may do so excessively, even in the throes of an emergency evacuation.
Both behavioural OR (Hämäläinen et al., 2013) and evacuations (Zheng, Zhong, & Liu, 2009) present interesting multidisciplinary challenges for which multiple methodologies are needed.On a similar theme our paper presents a network flow model in Sections 2-3, a game theoretic model in Section 4 and computational modelling in Sections 5-6.More intricate modelling of specialised evacuations problems is possible (Abdelghany, Abdelghany, Mahmassani, & Alhalabi, 2014) but our model is specifically based on discussions we had with practitioners about how real crowds were managed at major events in London.Our model thus seems to apply very generally.For additional discussion and model justification see Sections 2-3.

N evacuees
Fig. 1.Schematic of the model described in the main text.The illustration is for the special case of B = 4 exit routes, each characterised by a processing capacity, K i , and an offset waiting time, J i .
in such evacuations.The main contributions of this paper are thus as follows.First, we provide several ways to measure the usefulness of information from social media in the context of emergency evacuations -to date an under-explored issue.Second, our hand-calculation model provides a blueprint for those, such as civil engineers, social planners, etc, tasked with dealing with the mechanics of real-life evacuations.Third, we incorporate behavioural aspects in our discussion.We discuss optimal model-based interventions.We address the issue of over-crowding.We describe optimal contrarian strategies in which an individual may be best advised to eschew busy popular routes in favour of less popular but theoretically slower routes.This basic optimality result may also have further implications for psychological aspects of evacuations (Muir, Bottomley, & Marrison, 1996).We provide a Bayesian algorithm allowing a facilitator to continuously update the optimal evacuation strategy as new information arrives.The layout of this paper is as follows.In Section 2 we give a mathematical description of the evacuation problem.In Section 3 we discuss model-based interventions.Section 4 models individual behaviour.Section 5 derives an optimal online allocation algorithm under worsening conditions.Section 6 discusses an empirical application pertaining to the potential emergency evacuation of the City of Sheffield in the United Kingdom.Section 7 concludes.A mathematical appendix can be found at the end of the paper.

A mathematical description of the evacuation problem
In this section we consider a macroscopic version of the evacuation problem.A related statistical mechanics model is considered in Galla (2011).However, rather than explicitly modelling a flow process we consider a model with a group of N agents who arrive simultaneously in a central location and are awaiting evacuation through one of several exit routes.Fig. 1 shows a model with B = 4 such routes or branches, but any number is possible.Fig. 2 indicates how this model may be tailored to a more realistic practical setting.
In the sequel we label branches by i = 1, . . ., B. A branch, i, is defined by two parameters: a capacity K i and a 'baseline' journey time to a place of safety, J i .The parameter J i subsumes information pertaining to distance/average speed -both of which should have a clear physical interpretation in the context of a practical problem.The quantity J i is an offset, and reflects the time needed to traverse the exit route.K i is the number of evacuees which can be 'processed' by exit route i per unit time, reflecting for example the width of the corresponding exit or other operational constraints.This simple formulation is intended to capture the observation that whilst bottlenecks (represented by decreasing K i ) and delays (represented by increasing J i ) are extremely important in real evacuations often the exact geometry of the bottleneck is of only minor importance (Schadschneider et al., 2008).If n i people are allocated to Route i then the evacuation time of the last individual exiting through route i becomes i.e. the cohort of evacuees waits a total of n i /K i units of time before all are processed.Safety is then reached J i periods of time later.Evacuees are processed sequentially, i.e. the first agent evacuating through exit i reaches safety at time 1/K i + J i , the second agent at time 2/K i + J i and so on.An allocation of evacuees to exit routes is The n i denote the number of agents evacuating through exit i.Here, we make a continuum approximation under the assumption that the total number of evacuees is large, N 1.
In addition to minimising the final evacuation time there is also interest in two related solution concepts (Fry, Galla, & Binner, 2014;Jarvis & Ratliff, 1982): (i) The number of people evacuated via Route i by time t is given by (2) (ii) The total exposure prior to evacuation is given by (3) Eqs. ( 1 congestion effects etc.However, presenting the model in this way tallies with discussions we had with practitioners about how crowds were managed at major events in London.Further, we envisage that using information from social media parameter values may be updated sequentially thus capturing any nonlinear effects (Zobel, 2014).Moreover, even if our model appears unreallistic research in humanitarian logistics suggests that despite their inherent unpredictability preliminary preparations for such events may still be crucial in securing a faster response (Balcik & Beamon, 2008;Campbell & Jones, 2011;Yi & Kumar, 2007).
Proposition 1 shows how the optimal evacuation process may be achieved.Further, this solution appears to hold a degree of robustness (Fry et al., 2014).
Proposition 1 (Fry et al., 2014).The allocation n 1 , n 2 , . . ., n B defined by Suppose instead that the J i and K i are assumed to be unknown parameters.In this case an alternative solution can be found by appealing to Eq. (3) and minimising the expected total exposure

Expected Total Exposure
This leads to the following solution.
Proposition 2 (Fry et al., 2014).Suppose that J i and K i are randomly distributed.The expected total exposure prior to evacuation is minimised by A related result is available that serves as a basic proof of concept and illustrates how social media may be used to improve emergency evacuations (see Proposition 3).Suppose that as in (Chen & Xiao, 2008;Nakajima et al., 2008;Preston et al., 2011) the J i and K i have to be estimated by a facilitator using information from social media.In the sequel we label these estimates Ĵ i and Ki .Throughout this process we assume only that evacuees have access to social media via a suitable handheld device e.g.mobile phone/tablet and make no assumptions regarding the platform (e.g.Facebook, Twitter, news websites etc.) that individuals may use.This fits well with the rise of so-called citizen journalism (Allan & Thorsen, 2009) practised by non-professional journalists via social media platforms.We envisage that facilitators may be able to use social media to gain information regarding delays and congestion via evacuees (in similar spirit to citizen journalism) or via strategically placed agents.This may be made possible in applications using sensory or cellular phone data (Chiu, Zheng, Villalobos, & Gautam, 2007), gate counts (ODA, 2011) or via automated probability elicitation techniques (see e.g.Gosling, Oakley, & O'Hagan, 2007).In addition to a plethora of recent applications, the dynamic and imperfect nature of real evacuations (Schadschneider et al., 2008) reinforces the potential significance of social media in these situations.The following proposition shows that, in principal, if social media can be used to accurately track system parameters then the final evacuation time can be reduced Please cite this article as: J. Fry, J.M. Binner, Elementary modelling and behavioural analysis for emergency evacuations using social media, European Journal of Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.05.049

ARTICLE IN PRESS
JID: EOR [m5G;June 11, 2015;11:29] Proposition 3. Better information reduces the final evacuation time in the following sense.Using information from social media the difference between the actual and optimal final evacuation time satisfies the inequality where λ and λ denote Lagrange multipliers defined by Eqs. ( 5) and ( 9) Proof.The allocation made using information from social media sat- The optimal final evacuation time is given by λ in Eq. ( 5).As a consequence of the triangle inequality the difference between the two allocations satisfies The first term on the RHS of Eq. ( 8) shows that as knowledge of the optimal system processing capacity increases the actual final evacuation time becomes closer to the optimum value.The second term on the RHS of Eq. ( 8) shows that as knowledge of the true journey times and route capacities improves the actual final evacuation time also gets closer to the optimum value.
In view of the above, some further discussion is needed regarding the ease with which data can be gathered from social media.Inter alia location data from mobile phones allows for a comprehensive study of mobility patterns (Bekhor, Cohen, & Solomon, 2013).In road networks smart devices can be used to obtain high-quality flow and travel-time data.Roggen, Wirz, Troster, and Helbing (2011) use mobile-phone location data to model the behaviour of pedestrian crowds.Sakaki, Matsuo, Yanagihara, Chandrasiri, and Nawa (2012) and Kosala and Adi (2012) discuss the harvesting of real-time traffic information from Twitter to inform users about traffic congestion.Thompson, White, Dougherty, Albright, and Schmidt (2010) consider the use of smartphones to detect car accidents.Mohan, Padmanabhan, and Ramjee (2008) detail the use of mobile smartphones to monitor more general road conditions.Recent work in Chen, Fan, and Sun (2015) and Ni, He, and Gao (2014) extracts statistically significant signals from big social media datasets.Indeed, research presented to the US Congress suggests that it is the volume, rather than the paucity, of information on social media systems that presents the real challenge (Lindsay, 2011).Finally, empirical evidence of social media usage to track road closures in real-life emergency evacuations is documented in Bird, Ling, and Haynes (2012) and Palen (2008) together with other uses (Oxendine & Waters, 2014).
Within our specific setting there are also further complications as real-time analysis of social media is difficult to manage and subject to abuse (Preston et al., 2014a).One possibility is that networks may be maliciously seeded with poor quality information (Annlin Jin & Paramasivan, 2012).A further complication is that terrorists may themselves have access to social media (Oh et al., 2011).Amidst such uncertainty how might facilitators proceed and decide if information from social media is sufficient to indicate that a change in evacuation strategy is required (Chen & Xiao, 2008;Nakajima et al., 2008).From a network perspective (Vojnovi ć, Gupta, Karagiannis, & Gkantsidis, 2008) a natural benchmark is an oblivious strategy such as a uniform evacuation strategy.If using information from social media outperforms such a strategy then the information it provides clearly has a degree of value.Proposition 4 thus provides a simple benchmark to help decide when information from social media suggests that a change in strategy is merited.

Proposition 4 (Comparison with uniform allocation strategies). Suppose that E[
denote the total exposure corresponding to a uniform evacuation strategy with n i = N B , and T opt denote the unknown optimal strategy obtained using n i = K i (λ − J i ).We have the following: There exists a threshold such that if then evacuating using information from social media performs better on average than a uniform allocation strategy.
then evacuating using information from social media performs worse on average than a uniform allocation strategy where and T opt denotes the optimal minimum exposure , where n i is the optimal allocation satisfying Eq. ( 12).Using social media the expected total exposure is Formulas ( 10)-( 11) follow by comparison noting that under a uniform evacuation strategy with n i = N/B the total exposure is given by Proposition 4 contributes, at least in part, to wider debates on resilience and information quality (Kolfal, Patterson, & Yeo, 2013;Zobel & Khansa, 2012).If Eq. ( 10) applies then evacuating using information from social media performs better on average than a uniform allocation strategy and provides tangible benefits in terms of added control of the evacuation process.If Eq. ( 11) applies then a uniform allocation strategy is better on average and using information from social Please cite this article as: J. Fry, J.M. Binner, Elementary modelling and behavioural analysis for emergency evacuations using social media, European Journal of Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.05.049

ARTICLE IN PRESS
JID: EOR [m5G;June 11, 2015;11:29] media may be detrimental in extreme cases.If var( ni ) = σ 2 then the threshold in Eq. ( 11) simplifies to (13) Eqs. ( 10)-( 13) remain a little complicated to apply directly but do at least suggest that they may be applied in practice by using automated probability elicitation techniques (see e.g.Gosling et al., 2007).
The condition E[ Ki λ] = K i λ relates to the ability to accurately assess knowledge of the true optimal route allocation and true route capacities since Ki λ = ni + Ĵ i Ki .It is easy to show that this condition is satisfied if Ki = K i a.s. and only the Ĵ i have to be estimated.However, the condition and only the Ki have to be estimated.This suggest that Eqs. ( 10)-( 11) are less likely to apply if evacuations are subject to considerable uncertainties regarding congestion.Over-confidence in assessing congestion may thus be a serious problem in real-world evacuations and there may be interesting parallels here with over-confidence in financial markets during bubbles (Fry, 2012).A more general version of Eq. ( 11) assuming only Finally, Eqs. ( 10)-( 14) also offer a slight refinement upon recently published results in Fry et al. (2014).
When is it worthwhile to use incoming information from social media to update previous strategies?If Proposition 4 answers this question relative to a naive uniform evacuation strategy Proposition 5 answers this question relative to existing pre-planned strategies that do not make use of the most recent information available.This is thus intended to be closer in spirit to real-world applications where some emergency pre-planning has already been undertaken (Balcik & Beamon, 2008;Campbell & Jones, 2011;Yi & Kumar, 2007).
Proposition 5 (When to update evacuation strategies).Define n i = λ(K i − J i ) and suppose that the true underlying parameter values J i and K i have now changed to J i, new and K i, new .Suppose further that E[ then using updated information is optimal.
then it is optimal to retain use of the original pre-planned strategy.
Proof.Using the original optimal strategy without the benefit of incoming information from social media gives n i = λ(K i − J i ).The total exposure is given by i J i,new n i + . Using an updated strategy the expected total exposure becomes and Eqs. ( 15)-( 16) follow by comparison.

Modelling individual behaviour
It is clear that during evacuations it is not just the authorities who are directing the evacuation who can access information from social media.Individual evacuees can also access this information freely and make their own decisions accordingly.This may be particularly important in relation to over-crowding and jamming which appear to be key features of both real evacuations (Schadschneider et al., 2008) and computer simulation results (Smyrnakis & Galla, 2012).Emergency evacuations take place in socio-technical systems (Bonen, 1979;Cliff & Northrop, 2012;Hollnagel, Pariès, Woods, & Wreathall, 2011) whereby interactions between human agents and new (social media) technologies may have wide-ranging consequences beyond the technical specifications of infrastructure.For instance, suppose during an emergency evacuees observe information from social media and all rush to the same exit.How might individual evacuees and emergency planners best proceed?
Within the context of this study we analyse individual behaviour as follows.We consider optimal contrarian strategies by which an individual may be able to reduce their expected evacuation time based on how the rest of the crowd will act.Such contrarian strategies are well-documented in finance and investing (Chan, 1988) but may also lead to optimal evacuation strategies here. 1 We view the evacuation problem from the perspective of an individual known as A who has to compete with a crowd of size m.This results in a one-shot two-player game in which we can analyse A's best strategy response.Suppose the crowd is of size m.Of this crowd m i elect to evacuate through Route i.If A chooses to evacuate through Route i the m i + 1 evacuees are stacked and are randomly ordered.A's waiting time prior to being processed and traversing the route is distributed according to U (0, m i + 1) and the expected waiting time is (m i + 1)/2.Thus the expected final evacuation time via route i is A seeks to evacuate as quickly as possible and so must choose Rational expectations (Muth, 1961), Eqs.( 17)-( 18) and physical considerations mean that we may assume that m i and 1/K i are negatively correlated since if 1/K i decreases then Route i becomes wider and a more attractive choice of exit.Similarly, if 1/K i increases then Route i becomes narrower and a more difficult route from which to escape.We have the following Proposition which illustrates the tradeoff between choosing more populous faster routes and slower less populous routes faced by individual agents.
Proposition 6. Optimal contrarian strategies (i) A's optimal route of exit satisfies where s.d denotes standard deviation.
1 As a simple illustration the authors are regular train commuters and we often find ourselves adopting ad hoc contrarian strategies to avoid busy carriages and stairwells etc. Doubtless the reader can think of similar times when they adopt such contrarian strategies in their own daily lives.
Please cite this article as: J. Fry, J.M.
Proof.From Eq. ( 17) it follows that Optimal Exit Route = min Eqs. ( 20)-( 21) follow by using Cov(m i , 1 Proposition 6 shows that A's optimal choice of exit should address the physical parameters of the problem -the E(J i ) and E(1/K i ) terms -but should also include a penalty for popular choices of route -the 1/2E(m i )E(1/K i ) term.The Cov(m i , 1/K i ) term is a measure of how crowds respond to congestion.Eq. ( 19) gives the most general result.Eq. ( 20) describes the idealised scenario whereby crowds rationally incorporate information about congestion into their collective actions.Eq. ( 21) describes a zero-intelligence or stampede scenario when the behaviour of crowds is not moderated by such congestion and may be closer to the dynamics of non-adaptive behaviour in real-life evacuations (Schadschneider et al., 2008).
Beyond narrow mathematical definitions of rationality Propostion 6 also highlights several further potential issues.A combination of heuristics and other psychological factors (Muir et al., 1996) may mean that the behaviour of evacuees in real evacuations systematically departs from Eq. ( 19).Inter alia there may be links to insight problems which, though analytically tractable, may still mislead large numbers of people (Mayer & Davidson, 1995).If real evacuees do tend to behave according to Eq. ( 19) there may still be important implications for the communication of risk during evacuations so that bottlenecks may be avoided (Smyrnakis & Galla, 2012).Individual outcomes and the system-level behaviour as a whole ultimately depend on the actions of everybody else.If agents located close to each other on the network form similar opinions and replicate each other's actions this may lead to bottlenecks and congestion which is something that emergency planners should pay close attention to.

On-line allocation algorithm
In this section we construct an algorithm to discuss the on-line monitoring of an evacuation problem -so that evacuees may be better directed using updated information on capacities (over-crowding) and delayed journey times.Thus, this represents on alternative treatment of a mass-evacuation problem considered in Chiu et al. (2007) pand complements existing studies on the simulation and analysis of evacuations (Schreckenberg & Sharma, 2002), pedestrian motion (Schadschneider, Pöschel, Kühne, Schreckenberg, & Wold, 2006) and road traffic (Kerner, 2004).Further, if social media usage leads to congestion and jamming then the algorithm presented here shows how evacuees may be better directed.This is significant as real-time analysis of social media is extremely challenging and subject to abuse (Preston et al., 2014a).
The parameters K i and J i are assumed to be random variables as under the Bayesian paradigm (Bernardo & Smith, 2000).Uncertainty about J i and K i is expressed by the posterior distributions which can be updated sequentially given information on the number of evacuees n t j at time t j that have evacuated along route i.
It is assumed that the most likely values for J i and K i are Ĵ i and Ki respectively.In the context of a real evacuation it is assumed that the Ĵ i and Ki represent optimal values so that J i ≥ Ĵ i and K i ≤ Ki .We choose the prior distributions Under the Specification ( 22) we have that E[J i ] = 2 Ĵ i and E[1/K i ] = 2/ Ki and minimising the expected total evacuation time yields the same optimum strategy as assuming that J i = Ĵ i and K i = Ki a.s.If we choose n i to minimise the expected total exposure the functional to be minimised becomes i.e. the original functional up to a constant of proportionality.
Unforeseen delays are liable to be an intrinsic feature of real evacuations (Schadschneider et al., 2008) and it is unlikely that evacuees can move faster during evacuations compared to conventional performance levels (Pauls, 1995).Hence, we assume the conditional density (likelihood): This means that the number of evacuees n t j observed at time t j cannot exceed K i (t j − J) and that no more than K(t j+1 − t j ) individuals can evacuate via Route i during the interval (t j , t j+1 ).Our basic online algorithm is as follows: 1. Observe n t j the number of evacuees through Route i by time t j 2. Update the distributions of J i and K i given the data n t j

Calculate the expectations E[J i |n t j
] and E[1/K i |n t j ] and allocate evacuees according to Eq. ( 7) -the optimal evacuation strategy under parametric uncertainty In the sequel we suppress the i subscript and replace J i , K i by J, K. Eq. ( 24) shows that the parameters for each route can be updated in parallel -the calculations for each route are essentially independent.This feature of our algorithm is important and may lead to significant time savings in applications.From the likelihood function in Eq. ( 23) we can use Bayes' Theorem (Bernardo & Smith, 2000) to form a posterior probability distribution for J and K given the observed evacuees n t j at time j: where n t j = n t j − n t j−1 .Further, we have that required as part of the optimal allocation according to Proposition 2, may be calculated as follows: Proposition 7 (Posterior expectations).Let t j = t 1 , . . ., t T .We have that where Please cite this article as: J.
Proof.See the Appendix.

Empirical application
Sheffield is England's third largest metropolitan authority and fourth largest city with an estimated population of 555,500.The geography of the city renders it vulnerable to flooding and the city has experienced very bad flooding in recent years, especially in 2007 (Sheffield City Council, 2013).
A heuristic map of the Sheffield Inner Ring Road that encompasses Sheffield city centre is shown in Fig. 2. For the purposes of this study we assume that A-roads are dual carriage ways with an average speed of 64 kilometers per hour (40 miles per hour) and Broads are single carriage way with an average speed of 48 kilometers per hour (30 miles per hour).In line with some of the themes expressed in Sheffield City Council (2013) we assume that the purpose is to evacuate the city centre to a distance of 5 kilometers.This gives the parameters J A = (60×5)/64 = 4.6875 minutes and J B = (60×5)/48 = 6.25 minutes.Similarly, we can use data on stopping distances published in Highway Code (2014) to estimate the K i .For A-routes one car requires 40 meters of road (36 meters stopping distance at 40 miles per hour plus 4 meters for the average car).The distance travelled in 1 minute is 64, 000/60 = 1066.667meters giving K A = 1066.667/40×2 Carriage Ways = 52.For B-routes one car requires 27 meters of road (23 meters stopping distance at 30 miles per hour plus 4 meters for the length of the average car).The distance travelled in 1 minute is 48, 000/60 = 800 meters giving K B = 800/27 = 29.
Given Sheffield's size and the combined capacity of the city's two football stadia these figures appear reasonable.We have that K i = 9(52) + 7(29) = 671 and K i J i = 9(52×4.6875)+ 7(29×6.25)= 3462.5.Under Scenario 1. we have from Eq. ( 5 In this case 1187 cars should be allocated to each A road and 617 to each B road.In the sequel we will refer to both solutions as the "Original Strategy". In the sequel we introduce random error into the system.Suppose K i,new = K i , so that the conditions satisfying Propositions 4-5 are satisfied, and suppose further that J i,new = J i v i where v i ∼ U(1, T end ) and the estimator Ĵ i = J i,new e Z i where Z i ∼N(−σ 2 /2, σ 2 ).Set up in this way the error terms satisfy E(e Z i ) = 1.In computer simulations we compare the effect of using information from social media against the original strategy under Scenarios 1-2.above and a naive uniform evacuation strategy.
Simulation results are shown in Table 1.First, for both perfect information (σ 2 = 0) and low noise (σ 2 = 0.03) updating strategies using information from social media perform best and may lead to significant time savings.For noisier observations (σ 2 = 0.06, σ 2 = 0.15) updating strategies perform worse for less severe disruptions (T end low) but may still be very worthwhile when disruptions are more severe (T end high).For higher values of σ 2 using information from social media is not as effective as the original strategy but still out performs the naive uniform strategy.Throughout the simulations the original strategy out-performs the uniform strategy.This simple result nonetheless reaffirms the importance of emergency planningeven in the presence of manifold uncertainty (Balcik & Beamon, 2008;Yi & Kumar, 2007).

Conclusions, policy implications and discussion
This paper explores evacuation problems in a way that to date has been fundamentally under-explored -namely modelling the effects of social media.As evidenced by recent testimonies to the UK Parliament this represents an important practical problem to which we lend quantitative insight.Our models identify several key features of real evacuations, may purposefully inform policy debates and may ultimately guide the construction of more elaborate and applicable models.However, the contributions of this paper extend beyond the purely operational and we provide an algorithm through which the optimal network allocation may be updated online.We will consider the real-world application of our models in later work.In future work it might also be interesting to compare results with a different form of information spreading e.g. in the form of evacuation assistants (Schadschneider, Eilhardt, Nowak, Wagoum, & Seyfried, 2011).
There is an important behavioural dimension to evacuation problems (Hämäläinen et al., 2013).Issues highlighted by our study include over-confidence, psychological factors, heuristic decision making and insight problems.Information quality is an important issue associated with real world networks and evacuations (Preston et al., 2014b;Preston et al., 2014a).We can show that as information quality increases the potential benefits brought by social media increase.However, when information quality from social media is less good its use comes with risks attached and other strategies may prove more effective.This leads to a complex model-based intervention problem (Luoma et al., 2010).This is particularly pertinent as both simulation results (Smyrnakis & Galla, 2012) and empirical evidence (Schadschneider et al., 2008) emphasise the role of over-crowding and jamming in evacuations.We offer two different solutions to this problem.From an individual perspective we derive an optimal contrarian strategy which enables an individual to offset technically faster routes against their propensity for over-use by other agents in the system.This approach may also hold some wider significance for a behavioural treatment of evacuation problems.From a facilitator's perspective we derive an online allocation algorithm for updating the optimal evacuation strategy.Results from a computer simulation Please cite this article as: J. Fry, J.M. suggest that when delays are sufficiently large even noisy information from social media may still be very valuable.Finally, irrespective of the potential impact of social media, results from the computer simulations reaffirm the importance of emergency planning (Balcik & Beamon, 2008;Campbell & Jones, 2011;Yi & Kumar, 2007).
In order to maximise the benefits of social media various capacity issues need to be addressed (Preston et al., 2014a).This requires people in the right places having the appropriate mix of technical and soft skills.However, in seeking to build new capacities it is important to recognise the conflicting aims facing businesses in terms of profit maximisation and resilience planning (Hamel & Välikangas, 2003).It is important to understand how people behave in emergency situations (CSaP, 2014;Schadschneider et al., 2008) and in large crowds (ODA, 2011).For instance the issue of poor quality signage appears to be commonly over-looked (ODA, 2011).Further, communication with a diverse public about emergency situations remains a challenging issue (CSaP, 2014).From a behavioural perspective it is interesting to note that social media can also be used to support purely emotional crises -see e.g.Simm, Ferrario, Gradinar, and Whittle (2014) who design a digital anxiety device to support autistic adults.
Empirical evidence of real-world evacuation problems leads to several real-world policy implications (CSaP, 2014;Schadschneider et al., 2008).Congestion and blockages tend to occur at "hotspots" in the network (ODA, 2011).This may have important implications for both passenger flow control (discussed here) and the structural design of networks and buildings.There may also be important administrative challenges that need to be overcome in terms of coordinating action between different agencies and between different branches of the transport network (ODA, 2011).The amount of data involved in evacuation problems can be very large (ODA, 2011).Thus, future behavioural analyses of emergency evacuations using social media will require tools and techniques from Big Data analytics.
Our paper offers an intriguing blue print which we hope will allow others to better conduct real-world emergency evacuation planning once the relevant physical and geographical parameters are included.This potential utility underpins the value of our contribution.Evacuation problems are well-studied although many outstanding problems remain (Schadschneider et al., 2008).Here, our ultimate objective remains to create a reliable tool that allows emergency planners to leverage social media to protect the public at large -enabling smarter evacuations.The current authors are already actively engaged with policymakers, scientists and businesses to investigate the calibration of our model and its real world implementation (Binner et al., 2013;Busayawan, Whittle, Binner, & Lawlor-Wright, 2015;Preston et al., 2014b;Schmidt and Binner, 2014).Ultimately, the value of work such as ours occurs when these models prove directly responsible for lives being saved.
(29) and the RHS of ( 29) is an increasing function of J we must also have Hence, it follows that π (K|n t j ) ∝ K 1−2T over the range min n t 1 t 1 − Ĵ , min j>1 n t j t j , K . (30) We have that 2T − 1 .
For T = 1 we have that Please cite this article as: J.  [m5G;June 11, 2015;11:29] Similarly, it follows that the posterior density for J can be written as π (J|n t j )∝J −3 (t 1 − J) −2 on t 1 − n t 1 ξ , t 1 − n t 1 K .
Next, we need to reapply the Recursion (33) in order to find the normalisation constant.From (33) we have that and I −1 3,2 is the appropriate normalisation constant.Using partial fractions Combining Eqs. ( 35)-(36) Eq. ( 28) follows.

J
J|n t j ) is a normalized probability density we need to find the normalisation constant in order to calculate E[J|n t j ].It follows from (31)-(33) with m = 3, n = 2, b = −1 and c = t 1 that −2t 1 I m,n = 1 )-(3) thus present a linear model which might be deemed physically unrealistic in relation to nonlinearity brought about by Please cite this article as: J. Fry, J.M. Binner, Elementary modelling and behavioural analysis for emergency evacuations using social media, European Journal of Operational Research (2015), http://dx.doi.org/10.1016/j.ejor.2015.05.049Heuristic model of the Sheffield Inner Ring Road that encompasses Sheffield City Centre.