How Policy Decisions A ect Refugee Journeys in South Sudan: A Study using Automated Ensemble Simulations

Forced displacement has a huge impact on society today, asmore than 68million people are forcibly displaced worldwide. Existingmethods for forecasting the arrival of migrants, especially refugees, may help us to better allocate humanitarian support andprotection. However, few researchers have investigated the e ects of policy decisions, such as border closures, on the movement of these refugees. Recently established simulation development approaches have made it possible to conduct such a study. In this paper, we use such an approach to investigate the e ect of policy decisions on refugee arrivals for the South Sudan refugee crisis. To make such a study feasible in terms of human e ort, we rely on agent-based modelling, and have automated several phases of simulation development using the FabFlee automation toolkit. We observe a decrease in the average relative di erence from 0.615 to 0.499 as we improved the simulation model with additional information. Moreover, we conclude that the border closure and a reduction in camp capacity induce fewer refugee arrivals and more time spend travelling to other camps. While a border opening and an increase in camp capacity result in a limited increase in refugee arrivals at the destination camps. To the best of our knowledge, we are the first to conduct such an investigation for this conflict.


Introduction
. A civil war makes people vulnerable and leads them to migrate, in search for a secure and stable location. The choice of destination determines whether fleeing individuals are internally displaced people (IDPs) or refugees. IDPs seek safety within their own country and do not cross borders to neighbouring nations while refugees have been forced to flee their home countries due to war or violence (UNHCR ). There are more than million people forcibly displaced worldwide, of which million are refugees (UNHCR a). These fleeing individuals are the unfortunate victims of civil wars and internal conflicts, who make decisions to migrate at the times of distress. To understand the causes of forced displacement, researchers establish three concerns faced by migrants, namely, the choice to stay or flee, the choice to flee internally or across borders, and the choice of destination (Salehyan ). Their decisions are o en based on economic and political push-pull factors in sending and receiving countries. Especially, Schmeidl ( ) states that economic and political instabilities, poverty, violence and insecurity in the origin countries push people to flee. In contrary, economically and politically stable and safe countries pull forcibly displaced people to receiving areas. Thus, we can consider the economic and political conditions, security, the challenges and expenses of moving internally or across borders as causes of forced displacement. .
Unfortunately, forced displacement has enduring consequences on the refugees, as well as on both sending and receiving countries. For instance, civil war and violence within the origin countries may spread across borders. Similarly, receiving nations may interfere in internal conflicts and wars occurring in sending countries to prevent .
Forecasting forced displacement is challenging as many forced population data sets are small and incomplete, data sources have too little information; statistical methods are outdated and do not consider refugee arrival estimations (Edwards ; Disney et al. ). Yet, forced population predictions are essential to save refugee lives, to investigate the consequences of a nation closing its border for forced population, and to help complete incomplete data collections on refugee movements (Groen ). Improvements in data collection may be a possible solution to overcome data issues, but we require an enhanced logical framework to capture forced displacement thoroughly. .
In this paper, we investigate the e ect of policy decisions in the South Sudan conflict on the camp arrival rates of displaced persons. Policy changes may occur at short notice in crisis situations, and correctly predicting arrivals of displaced people in these cases is essential to prevent shortages in food, water and shelter in refugee camps.
. The recent study by Gilbert et al. ( ) examine the role and applicability of agent-based modelling (ABM) when experimenting with policy decisions, where ABM provides an understanding and knowledge to governments, stakeholders and policymakers by modelling complex systems, such as human movement in this paper. Our ABM approach examines the e ects of policy implications under various scenarios. It also provides a new perspective and helps researchers and other organisations in forecasting refugee movements, and inform policy decisions related to forced displacement. This was previously much more di icult and ine ective due to incomplete data and outdated statistical analysis.

.
To enable this investigation, we extend the simulation development approach (SDA) to support counterfactual scenarios. The original SDA, initially proposed by Suleimenova et al. ( a), adopts an ABM approach to estimate how refugees reach destination camps. As part of the Verified Exascale Computing for Multiscale Applications (VECMA) project, we seek systematic approaches to validate and analyse the sensitivity of our simulations, to investigate output variability and to generate more actionable results (Groen et al. b). To do this, we incorporate integrated sensitivity analysis and exploration of policy decisions in our SDA. This helps us to better understand refugee behaviour and better assist policymakers with their decision-making process. .
Moreover, we add value by automating parts of the simulation development process, from construction to execution. Here, we propose an automated policy exploration toolkit, together with the sensitivity analysis, which is an essential step towards enabling users to create refugee arrival forecasts within days of a new conflict erupting. To showcase our approach, we present simulation results for seven runs for South Sudan and discuss how changes in the policy of forced displacement a ect the refugee arrival rates in camps. Forecasting refugee arrival rates in camps is crucial since governments and NGOs can use this information to better allocate humanitarian resources and provide humanitarian protection to forcibly displaced people (Groen ). .
In the remainder of this paper, Section discusses computational modelling techniques of forced displacement, describes a generalized simulation development approach for refugee modelling and presents an automated toolkit for policy explorations. We apply our proposed approach to the South Sudan conflict. We then discuss simulation results in application to policy decisions of camp or border closure, camp capacity changes and forced redirection (Section ). Finally, Section concludes and o ers final remarks on refugee modelling and policy decisions.

Computational Modelling of Refugee Movements
. In recent years, there has been a gradual increase in the use of computational techniques, both machinelearning based (Sfyridis et al. ; Quinn et al. ) and simulation-based (Hassani-Mahmooei & Parris ; Sokolowski et al. ; Hebert et al. ), to provide migration forecasts. In the case of simulations, one of the more widely adopted approaches is ABM, a computational approach that provides an opportunity to model complex systems with individual heterogeneity. It consists of agents that represent animals, humans, organisations or any other types of entities interacting with each other and within their environment. Particularly, the use of ABM allows to model how agents and their environment vary across time and space. Agents are autonomous and o en unique, meaning that each agent is distinct in terms of size, location and other attributes (Macal & North ).

Recent
In the context of refugee arrival predictions, Suleimenova et al. ( a) propose a generalized simulation development approach forecasting the distribution of refugee arrivals across destination camps. To understand the significance and generalisation of the proposed approach, the authors for the first time successfully modelled three African countries experiencing refugee emergencies, namely Burundi, Central African Republic and Mali, using a single approach. Their generalized approach relies on an ABM, where refugees are agents, and each time step represents one day since the validation data has a granularity of a single day and cannot be used to validate patterns on an intra-day timescale. The simulation starts by inserting a number of refugees (obtained from the United Nations High Commissioner for Refugees (UNHCR) database) in their conflict locations (extracted from the Armed Conflict Location and Event Data Project (ACLED)) that can be presented using a network-based ABM model. Each refugee can traverse from zero to more links during each simulation step. The probability of an agent's movement depends on the move chance, where the move chance of . represents agents in conflict and between locations, .
for refugee camps and . for all other locations. Suleimenova et al. ( a) provide a detailed flowchart of algorithm assumptions and agent parameters used by a simulation code -Flee, which can be found at https://github.com/djgroen/flee-release. It is optimised for its simplicity and flexibility, and it can be adapted to most scenarios involving escaping refugees.

Description of a generalized simulation development approach with the FLEE code .
To facilitate rapid and consistent simulation development, Suleimenova et al. ( a) suggest a generalized SDA, which enables rapid construction, execution and validation of refugee counts in conflict scenarios. We present a revised generalized SDA in Figure , which contains the same six phases of the original SDA (refer to Suleimenova et al. ( a)), including situation selection, data extraction, model construction, model refinement, simulation execution and analysis, but also enhanced to fit the focus of this paper. Specifically, this revised version incorporates changes in policy decisions (e.g. camp and border closures, camp capacity changes and forced redirection) in the refinement phase and introduces both an ensemble of simulation executions and sensitivity analysis of simulation runs in the simulation execution phase. .
Currently, the construction and execution of simulations are mostly done manually, which is both ine icient and time-consuming. For instance, an extraction of input data, construction of network maps and initial models for Burundi and CAR required -weeks of manual work. While refugee predictions need quick construction and execution as there is a prediction urgency of refugee crises or multiple conflict scenarios to simulate on JASSS, ( ) , http://jasss.soc.surrey.ac.uk/ / / .html Doi: . /jasss.
short time period (Suleimenova et al. b). Hence, we automate several phases of the SDA, namely, the construction, an instantiation and execution of ensemble runs using a unified approach. In the next section, we describe the automation of each phase of a generalized SDA. Figure : A generalized simulation development approach forecasting the distribution of refugee arrivals across destination camps. We use the same assumptions as given in Suleimenova et al. ( a) for our simulations (except where this is mentioned otherwise for individual runs).

Automation of simulation development using FabFlee
.
Manual routine tasks in model construction and simulation execution can be simplified using automation tools. Automation is essential in simulation development since it provides time e iciency to modellers, reduces the probability of human error, simplifies and accelerates process activities and delivers a highly transparent and customised programme to users. Suleimenova et al. ( b) comprehensively discuss existing automation tools, as there is an extensive number of languages, open-source so ware and automation tools that facilitate the development of computational research. Groen et al. ( a) perform an analysis of added value for a range of coupling tools, including several automation tools. In both works, FabSim is recognised as a toolkit that helps to curate and simplify simulation research at the simulation deployment, execution and optimisation stage. Based on these findings, we chose to use the FabSim toolkit, which is an improved version of FabSim. Among other things, FabSim simplifies organising input and output files, user and machine configurations, and application executions (Groen et al. ). Currently, FabSim contains an integrated test infrastructure, more flexible customisation options using a plug-in system, and in-code documentation and examples to improve usability. It is distributed under a BSD -clause license.
. A FabSim -based FabFlee toolkit is one of the plug-in applications, which predicts the distribution of incoming refugees across destination camps under a range of di erent policy situations (Groen et al. b). FabFlee is a partially automated implementation of our SDA, and provides an environment for researchers and organisations to construct and modify refugee simulations, instantiate and execute multiple runs for di erent policy decisions, as well as to validate and visualise the obtained results against the existing data. In Figure , we present the SDA phases with automated functionalities from model construction to analysis. Specifically, we aim to construct the initial model using a comma-separated values (csv) formats, refine the model with a new set of parameters or policy range decisions, execute an automated ensemble of runs and analyse the obtained results with the use of automated plotting tools. In the next section, we provide a detailed description of automation applied to each phase of the SDA. Figure : Phases of our simulation development approach,given in arrow boxes, and automation implemented in FabFlee for each phase, described in the ovals.

Model construction .
To start with, we simplify the model construction phase by creating reader modules for csv formats for input data. Three formats of csv files, namely locations.csv, routes.csv and closures.csv, are integrated with FLEE's input interface. The initial idea of introducing these files is described in Suleimenova et al. ( b). For this paper, we revised their outline to reduce data collection time and implemented these csv file formats in the model construction phase. We create these csv files manually according to the formats demonstrated in Tables , and for the conflict scenario and store them under the base conflict data in a specified conflict directory (e.g. https://github.com/djgroen/FabFlee/tree/master/conflict_data/SSudan). name county country latitude longitude location_type conflict_date* population/capacity conflict population of location town -camp -camp capacity forwarding_hub -- Table : locations.csv contains all locations with the properties required for simulation construction, such as name and geographical information of locations, and populations (for non-camp locations) or capacities (for camp locations). Note: conflict_data is given as an integer, counting the number of days a er the simulation start. The value of indicates the start, while -indicates the end date of the simulation. Table : routes.csv specifies distances between two locations. Note: forced_redirection refers to redirection from source location (can be town, camp or forwarding_hub) to destination location (mainly camp). The value of indicates no redirection, indicates redirection from location to location and corresponds to redirection from location to location . closure_type* name name closure_start* closure_end* location country Table : closures.csv provides camp closure event specifying locations names or border closure event requiring country names to name and name respectively. Note: closure_type can be two types: location corresponding camp closure and country referring to border closure. closure_start and closure_end are given as integers, counting the number of days a er the simulation start. The value of indicates the start, while -indicates the end of the simulation.
A er the generation of locations.csv, routes.csv and closures.csv files, we follow the FabFlee workflow diagram (see Figure ). As a start, we load a base conflict data which includes csv files and the source data of a conflict scenario using load_conflict command. This, in turn, duplicates all existing files from a base conflict directory to a working directory, namely active conflict data. The load command also generates a text file (i.e. commands.log.txt) that records command logs of commencing activities. Moreover, to refine the model, we examine policy implications through parameter explorations for policy decisions related to a refugee emergency. We have developed several parameter exploration commands to modify a range of parameters illustrated in Table . JASSS  .
Following the refinement phase, we duplicate parameter changes of the model by running the instantiate command. The instance is then saved in a new directory, which can include run name, version and date of instantiation on users insert choice. Now that we have our simulation input, we can proceed with the fi h phase of our SDA and run execution command triggering the FLEE code and producing results. Next, we visualise and validate the obtained results with graphs for each camp in a neighbouring country by running plot_output command. .
To create a clean slate for future work, we can clear the active conflict directory using fab localhost clear_active_conflict, upon which we can reload the conflict and change other parameters (and instantiate and run a new simulation). Indeed, phases four to six in Figure can be iterative and produce additional results as we extend our policy and parameter exploration. Similarly, we can conduct sensitivity analysis for each instantiated model by running test_sensitivity function (see Table for   a). For many years, Sudan experienced a civil war from which South Sudan declared independence on the th July . However, the authorities of South Sudan failed to deliver the basic needs (Reid ), and in December , a conflict between the government and rivals broke out. .

Specifically, the civil war in South Sudan started on December , following fierce fighting between rival units of the Sudan Peoples' Liberation Movement (SPLM) and the Sudan People's Liberation Army (SPLA) in the capital, Juba (UNHCR
). Subsequently, South Sudan's president Salva Kiir announced that former vice president Riek Machar had attempted a coup. Machar escaped from Juba and became the leader of an armed opposition movement, namely the 'SPLM/A in Opposition'. Violence and fighting spread to other parts of the Jonglei, Upper Nile and Unity states, as well as other regions of South Sudan (ICG ). This forced people to flee internally and across neighbouring countries.

.
Our South Sudan model has a simulation period of days starting from the th December to the th August , during which . million refugees were known to escape the country. We run the simulation for camps (listed in Table )

Setup of simulation execution for South Sudan .
A er selecting our conflict country and the simulation period, we then extract data from the sources according to the SDA. Next, we construct our initial model for South Sudan with default settings using the discussed three csv file formats, namely locations.csv, routes.csv and closures.csv. The initial constructed model, which is the third phase of SDA, is then refined with additional information obtained from reports (fourth phase of SDA). In Figure , we demonstrate the layout of our simulation tests for the South Sudan conflict. This includes refinements to determine how policy decisions, such as camp and border closures, changes in camp capacities and redirection between camps, can a ect the distribution of refugee counts and simulation results. Using our approach, we also automatically create and perform sensitivity analysis study for each of our scenarios. Bearing in mind, we set our default setting to the refugee move speed is equal to km per day and the awareness of surrounding is link.

Description of the base scenarios .
A er constructing the initial South Sudan model (ssudan_default), we executed and obtained the initial results. Next, we determined level refugee registrations from the source data and included them to improve the initial model. We named the second model as ssudan_reg and executed to observe changes in the results. We further refined the ssudan_reg model using additional information obtained from publicly available online reports. Specifically, the UNHCR ( ) report declares that refugees arrived at Ethiopian camps on foot, due to the lack of roads. To accommodate this fact, we modified our simulation assumptions, and we incorporated specific "o -road links" from conflict zones to Ethiopian camps in a modified simulation setup named ssudan_links. To reflect the fact that o -road routes are likely to result in slower travel speeds, we multiplied the coordinate point-by-point distances by for all walking routes. We also incorporated additional information in regards to later camp openings and closures, which was derived from the UNHCR reports (run ssudan_ccamp).

Figure :
Overview of the geographic network model for South Sudan. This includes conflict zones (red circles), refugee camps (dark green circles) and other major settlements (yellow circles). Interconnecting roads walking routes are given with lines, with adjacent numbers used to indicate their length in kilometres (blue for roads and brown for walking routes). Background maps are courtesy of carto.com created using OpenStreetMap data. Figure : Setup of simulation execution for South Sudan. For each execution, we perform ensemble runs for sensitivity analysis. The structure of these ensembles is given in the bottom grey panel.

Results
. In Figure , we demonstrate the averaged relative di erence for four simulations (ssudan_default, ssudan_reg, ssudan_links and ssudan_ccamp). Despite the same levels before day , the average relative di erence for these runs persistently lessens respectively from .
to . over the simulation period and the refinement of the South Sudan model as we incorporated additional details. Overall, ssudan_ccamp is the most refined with the lowest average relative di erence in the aggregate level. We calculate the average relative di erence using the equation below: where, the number of refugees found in each camp x of the set of all camps S at time t is given by n sim,x,t based on the simulation predictions, and by n data,x,t based on the UNHCR data. The total number of refugees reported in the UNHCR data is given by N data,all (Suleimenova et al. a).
. Moreover, we perform a range of sensitivity analysis tests to identify the important input variables in an awareness level and agents' movespeed of the simulation outputs. To begin with, we executed replicas of ssu-dan_ccamp with default settings to determine the range of the output due to the probabilistic nature of the simulations. Over these executions, the average relative di erence ranged between . and . . In addition, we perform a sensitivity analysis for each run by varying the level of agent awareness range and a speed limit of refugees. Here, the awareness range represents the level of knowledge of refugees about nearby locations. They may know only the distance to the adjacent locations in the graph (path distance only), or also the type of location for adjacent locations ( link away), or also the location type of locations adjacent to those ( links away). We present the results of this analysis in Table . For the most refined scenarios, the averaged relative di erence is lowest when agents are aware of locations link away, though the di erence is marginal compared to simulations with an awareness range of links away. Our simulations are clearly sensitive to the maximum refugee move speed parameter, and in particular move speeds below km/day result in significantly higher validation errors. This parameter sensitivity is in line with our simulations of previous conflicts (Suleimenova et al. a).   Table : Averaged relative di erence values, averaged over time and all four base type of simulations using di erent agent awareness ranges, and di erent speed limits for agents. Note that we present results from separate executions of the default type run: in the first data row, the third data row (labelled " link away") and the ninth data row (labelled " ").
. We present ssudan_ccamp simulation results for all camps validated against the UNHCR refugee registration data in Figure . The most populous camp in our simulation is Adjumani with more than , refugees over the simulation period and slightly overpredicted a er days into simulation compare to the data. The reason being that it is the closest camp for refugees fleeing from the South Sudan conflict. The forecast refugee counts in Kiryandongo and Kakume (at the start prior to days) are in close agreement with the UNHCR data, while our simulations underpredict for Kule, Jewi and Khartoum camps. South Sudan has a record of being in conflict prior to our simulation start date. Kakuma ( ), Pugnido ( ), Rhino ( ) and Kiryanongo ( ) camps had registered number of refugees fled prior to the simulation start; these, therefore, do not count towards the refugee arrival numbers.  There are no arriving refugees at the start of simulation period for several camps, namely Tierkidi, Kule and Jewi, illustrated in Table , as they opened a er the conflict has commenced according to the UNHCR data. For instance, the Tierkidi camp has no arrivals prior to days of simulation, but refugee counts increase over the simulation period and overpredict UNHCR data by the end of simulation period. In addition, the Jewi, Kule and Khartoum camps show slowly increasing and underpredicted refugee counts. Whereas, the Pugnido, West Kordofan and Rhino camps are considerably overpredicted according to simulation results by almost refugees decreasing to refugees for each camp by the end of simulation period.

Examining policy decisions .
There are various real-world policy implication instances, which have changed the course of refugee movements. To demonstrate, the Dadaab camp in Kenya, which was opened in , currently hosts more than Somali refugees (Cannon & Fujibayashi ). Despite its populated occupancy, in , the Kenyan authorities attempt to shut this camp, but a high court judge ruled out the authority's decision and allowed refugees to remain in the Dadaab camp. However, if the authority closed the camp, it would have forced refugees to return to their violent home country, hide illegally in Kenya or flee to other neighbouring countries. Another instance is refugee camps in the Gambia, which were planned to reopen in since there was an increase in refugee numbers. Nevertheless, the authority was hesitant to site camps near borders due to armed opposition groups who opposed a danger on refugee lives. Hence, the authority has decided to place Casamance refugees in the old camp at Bambali in the central Gambia. Since this camp was farther from borders, refugees refused to travel and settled in nearby villages of Gambia (Grant ). In this case, it is uncertain how refugees have impacted the Gambian villages and their residents, as well as how refugees managed themselves in the neighbouring country. Moreover, in May , Jordan closed its borders with Syria to stop the influx of refugees, which instead increased illegal crossings into the country (Hargrave et al. ). It is an instance of a single country, while there are many other countries that have closed their borders to refugees, such as European countries, and increased illegal crossings and human tra icking. .
These instances illustrate that policy decisions have influenced refugee movements towards destination camps. However, we do not know how these decisions a ect refugees, their movement and overall refugee counts. Hence, we aim to investigate and understand the implications of these policy decisions on refugee arrivals, as well as inform other authorities and policymakers on the consequences of their decisions. We model policy decisions using four di erent scenarios of the South Sudan conflict.
. First, we compare the refugee arrivals in camps between three scenarios, which are a model without camp and border closures (ssudan_links), a model with camp closures (ssudan_ccamp), and a model containing an additional border closure between South Sudan and Uganda, enforced until day that is halfway into the simulation (ssudan_cborder). We present our comparison results in Figure , and find no significant di erences between ssudan_links and ssudan_ccamp. However, we do find di erences between these two scenarios and ssudan_cborder, which results in % fewer refugee arrivals on day . This implies an increasingly long travel time for refugees up to day , the day that the border is again reopened. In addition, the delaying e ect of border closures lingers in our simulation results a er borders have been reopened, with approximately % fewer arrivals on day , for instance. This emergent behaviour can by definition not be validated against reality (we're examining a counterfactual). However, explanations for such delays are possible. For instance, refugees may fear that recently opened borders are more likely to be closed again, or may not be immediately aware that a previously closed border has again reopened. .
Second, to explore how changes in camp capacities a ect simulation results, we changed the capacity of the most populous camp, namely Adjumani. For the first instance, we decreased the original capacities of refugees by half. The second instance involved an increase in the original capacity by %. In Figure , we present the number of refugees for Adjumani camp ssudan_adjumani (capacity: refugees) and ssu-dan_adjumani (capacity: refugees). We find that a reduction of capacity in Adjumani results in up to % fewer refugee arrivals in camps, which implies considerably longer refugee travel times. However, increasing the capacity at Adjumani by allocating more resources appears to only result in a very limited increase in refugee arrivals (< %). Based purely on these results, we find that, in a setting where aid resources are heavily constrained, the default capacity of this camp is close to optimal. .
Finally, we explore how the enforced redirection of arriving refugees from one camp to another can a ect the distribution of refugees across all the camps. As an exemplar, we created a scenario (ssudan_redirect) where all refugees arriving in Kule, Jewi and Pugnido are redirected to the Tierkidi camp, which has its capacity increased accordingly, creating a counterfactual situation where Tierkidi is the single central camp in Ethiopia receiving refugees from South Sudan. This kind of centralised management of incoming refugees has been known to occur in some other conflict situations, such as Mauritania (Mbera camp) in the North Mali conflict in Suleimenova et al. ( a). .
We present a comparison of arrivals across seven camps in both scenarios in Figure . Here, Kule, Jewi and Pugnido are excluded from the comparison, as they do not host any refugees in the modified simulation. In comparison to the ssudan_ccamp simulation results for individual camps, we attain di erent distribution of refugees across camps in ssudan_redirect. By Day , Tierkidi has received twice as many arrivals in ssudan_redirect Figure : Comparison of the number of refugees in seven camps as forecast by our ssudan_ccamp and ssu-dan_redirect simulations for the South Sudan conflict. (a-g) Graphs are ordered by camp population size, with the most populous camp on the top to the smallest one on the bottom. than in ssudan_ccamp, while the other six camps retain similar arrival rates. However, a er Day the number of refugees in the other six camps becomes lower in ssudan_redirect than in ssudan_ccamp, while the number of refugees in Tierkidi remains considerably higher. This behaviour can primarily be attributed to the Pugnido camp, which reaches full capacity around Day in ssudan_ccamp (see Figure ), but which is redirected to Tierkidi in ssudan_redirect, a camp with a higher (combined) capacity.

Conclusion
. Forecasting forced displacement, especially refugee movements, is both very important and very challenging. Forecasting the distribution of refugee arrivals to potential destinations, as governments and NGOs can e iciently allocate humanitarian resources and provide protection to vulnerable refugees. Through the use of computational modelling and the automation approach presented here, we are able to systematically explore the possible impact of specific policy decisions while accounting for the sensitivity to at least some of individual JASSS, ( ) , http://jasss.soc.surrey.ac.uk/ / / .html Doi: . /jasss. parameters and assumptions in the model. To achieve this, we have extended the simulation development approach by Suleimenova et al. ( a) and used it to forecast refugee arrivals in camps in the South Sudan crisis. Our approach, which relies on the FabSim -based FabFlee toolkit, publicly available as part of the EU-funded VECMA project (https://github.com/djgroen/FabFlee). Though the runs in this paper were all performed on local resources, the FabSim toolkit has been applied extensively to execute simulations on supercomputers. We aim to enable this functionality for FabFlee and perform much larger parameter and policy explorations in the near future using it. .
We demonstrated our automated ensemble simulation approach by analysing the e ect of policy decisions on refugee journeys in the South Sudan conflict. This conflict is relatively di icult to simulate, primarily due to the lack of roads and di icult food circumstances. While investigating the latter aspect requires new model development and is beyond the scope of this paper, we did update the model to include several walking routes, and were able to achieve a much lower validation error (averaged relative di erence) as a result. All policy decisions presented here are purely hypothetical, and largely derived from having observed similar decisions being made in the three conflicts we analysed previously in Suleimenova et al. ( a). .
In terms of policy decision examples, we incorporated camp and border closures, two camp capacity changes and a forced redirection. As expected, border closure and a reduction in camp capacity result in fewer refugee arrivals as more refugees end up travelling to other camps. Likewise, an increase in camp capacity results in a limited increase in refugee arrivals at the destination camps. However, we also found several unexpected behaviours, such as a lingering e ect in prolonged refugee journey times once a border is again reopened, and a clear boost in refugee arrivals when refugees are redirected to a reduced number of camps with larger capacities. These findings help to understand the e ects of policy decisions on refugee arrivals and inform other similar conflict situations. We believe these policy decisions in particular warrant more in-depth investigation, using simulation and data analysis approaches that take into more relevant factors and circumstances, and can also leverage the benefits from the automation approach we presented here.