Agent-Based Modeling and Simulation of

Mosquito-Borne Disease Transmission

Akshay Jindal Shrisha Rao

International Institute of Information Technology, International Institute of Information Technology,
Bangalore Bangalore
akshay.jindal@iiitb.org shrao@ieee.org

ABSTRACT tions for their containment—particularly, solutions capable

Mosquito-borne diseases, such as chikungunya, dengue, and of tracking the trajectory of the infection evolution at a very
malaria, are re-emerging and expanding to new and formerly fine granularity. However, modeling the complicated rela-
unaffected places, leading to a need for models which can tionships among human-mosquito-pathogen continues to be
track their evolution and thus help with public policy and a challenging problem faced by all epidemiologists and pub-
epidemiological studies. Such diseases’ evolution is driven lic health experts.
by the interactions between hosts and vectors, and is thus Traditionally adopted mathematical approaches for this
heavily dependent on factors like host and vector population purpose are the differential equation models [21] and mean-
distributions and mobility, and geographical and weather field type models [30], which fail to track spatial and tempo-
conditions. Traditionally used mathematical models fail to ral factors like population density and dynamics, and ignore
capture such issues, thereby creating a gap between what the spatial implications within the system, thereby creat-
epidemiologists and disease modelers can provide, and what ing a huge gap between what epidemiologists can deliver
public health policy requires. We give a generalized agent- and and what public health policy makers would like [15,
based model (ABM) which overcomes these limitations by 41]. This neglect of spatial and temporal components in
careful integration of geographic information (GIS) and cen- epidemic model formulation can be resolved by the use of
sus data to account for the spatial movement of infections, the agent-based modeling paradigm [44]
and climate data to capture the temporal nature of an epi- We give an agent-based model which overcomes the above
demc. It captures the disorganized interactions of hosts and limitations of prior mathematical models, to adequately cap-
vectors at a micro-scale by explicitly modeling each human ture the spread of vector-borne diseases at a fine granu-
and mosquito to simulate the complex trajectories of dis- larity, and show how it can contribute to low-level policy
ease outbreaks (even those have yet to occur), and makes it design such as testing intervention strategies in epidemic
possible to test the efficacy of various public health policies. life-cycle. It captures the complex relationship between hu-
This model also suggests that it is possible to estimate hard- mans, mosquitoes, and pathogens by modeling them indi-
to-determine parameters about vectors (e.g., a mosquito’s vidually at a micro scale using simple adaptive rules, and
sensing distance), through simple model calibration. Unlike letting them interact. The environment is structured using
previous solutions, our model is trained and validated us- multi-layered GIS data to address the space implications in
ing real data from a 2013-14 chikungunya epidemic in the disease proliferation, is realistically initialized using census
Caribbean and is seen to give accurate results. data to account for variable population density, and driven
by climate (precipitation + temperature) data to correctly
capture disease vector dynamics. Each of these three agent
Keywords populations have been designed to imitate the characteristic
Chikungunya; Computational Epidemiology; Mosquitoes; Pub- behavior of the vector-borne diseases:
lic Health; Simulation
1. Our human agent population approximates the mo-
bility patterns across various population sub-groups
1. INTRODUCTION while being sensitive to the spreading infection to ade-
Vector-borne diseases, of which mosquito-borne diseases quately capture the spatial movement of the infection.
are the largest subset, account for more than 17% of all This overcomes the usual drawback of the aggregated
infectious diseases, and cause more than 1 million deaths nature of mathematical models and helps us analyze
every year [49]. Adding to their high fatality rates, they various intervention strategies.
also have high attack rates, and are very prone to muta- 2. To account for the seasonal nature of an epidemic,
tion, leading to recurrent infections and a lack of single vac- the mosquito agent population is made to be heav-
cine [50]. This creates the need for more proactive solu- ily dependent on precipitation data and availability
of resources as these are known to be major factors
Appears in: Proc. of the 16th International Conference on in disease propagation [17, 4]. This captures the in-
Autonomous Agents and Multiagent Systems (AAMAS 2017), herently disordered nature of such epidemics. One
S. Das, E. Durfee, K. Larson, M. Winikoff (eds.), of our novel contributions here is the incorporation
May 8–12, 2017, São Paulo, Brazil. of the climate sensitive complex reproduction cycle of
Copyright © 2017, International Foundation for Autonomous Agents
and Multiagent Systems (www.ifaamas.org). All rights reserved. mosquitoes within the mosquito agent. Every stage of

426
 a mosquito is separately modeled with its age being to study disease propagation, disregarding other dominant
a linear function of temperature and time to closely factors which severely undermines the prediction accuracy
resemble the mosquito population in nature. of such models. Also, none of these models are validated us-
ing real data making their results wildly unpredictable and
3. Our infection model is designed as a state machine
impossible to compare with other models.
with variable transition probabilities, making it eas-
In summary, our novel contributions are:
ily generalizable to a large subset of mosquito-borne
diseases. (i) A framework for incorporating climate sensitive com-
plex reproduction cycle of mosquitoes in the mosquito
The epidemic emerging from our model is calibrated and agent.
validated using data from the 2013-14 chikungunya epidemic (ii) A generalized mechanism for an epidemic ABM’s cal-
on some Caribbean islands, as a proof of concept. The model ibration and validation.
predictions are 93.3% accurate on the test data of the re-
ported cases, with the small gap suggesting that there were (iii) A method for estimating hard-to-determine vector pa-
some cases which went unreported, which matches with re- rameters like mosquito sensory range, using ABMs.
ality, where a lot of chikungunya cases go unreported due to (iv) Methods for testing different policies like intervention
the non-fatal nature of the disease. We also show how our strategies using our model. (We also have a compar-
model can be utilized to estimate some vector parameters ative analysis of existing ABMs for modeling vector-
like mosquito sensory range, mortality rate, etc. which are borne diseases and our model.)
hard to determine directly. Furthermore, we test two pre-
vention strategies, LSM and ITN+IRS, and the results show In Section 2, we give a detailed explanation of our model
that for a small urban tropical area, LSM is more effective in design and its intrinsic workings. Here, we also describe our
terms of cost, mosquito control, and infection elimination. simulation approach and the data-set used for validation.
Also, considering that in most cases a mosquito’s life cycle In Section 3, we explain our model parameterization and
is not greatly dependent on the presence or absence of the calibration, and discuss our model’s performance on the past
infection due to their short lifespan [19], our model can be epidemic data. In Section 4, we show how our model can
easily adapted for other mosquito-borne diseases with minor be used for policy formulation by comparing two popular
modifications. To adopt the model for some other mosquito- mosquito prevention techniques. Finally, we conclude with
borne disease, we only need to modify the values of relevant the current state of our model and some thoughts for future
parameters based on the new vector’s and pathogen’s be- directions in Section 5.
havior, and calibrate others using data from past epidemic
of the same disease. 2. MODELING DISEASE TRANSMISSION
There are some ABMs proposed to model the spread of The modeling of the effects of climate and topology on
vector-borne diseases. Dommar et al. [17] developed an vector behavior, and the patterns of human mobility, is a
ABM to model chikungunya outbreaks, and indicate that desideratum in the epidemiology of vector-borne diseases.
topology and precipitation are dominant factors in vector- Here, we present our agent-based model which successfully
borne disease propagation. Teng et al. [16] also provide a captures these effects in a disease’s evolution at an adequate
similar model for simulating dengue spread. However, nei- scale. The simulation strategy we adopted to qualitatively
ther of these accounts for human mobility patterns, thereby analyze our model is also explained here.
neglecting the major source of spatial movement of infec-
tions. Arifin et al. [33] provide a framework for effectively us- 2.1 Agent-Based Model
ing multi-layered GIS data in an ABM-based malaria study. The agent-based model presented here represents the cu-
Illangakoon et al. [27] explore the efficacy of ABMs for study- mulative effects of the behaviors of individual humans and
ing malaria prevalence and transmission, and show how hu- mosquitoes in an urban environment while noting the bio-
man mobility patterns have a strong affect on disease trans- logical trajectory of the spreading infection. It is composed
mission. Ying et al. [4] study the effect of spatial heterogene- of three parts: The first is a small modification of the popu-
ity on mosquito populations. All of these only focus on the lar SEIR model [43] to describe the infection life cycle. The
spatial nature of the epidemic and are inconsiderate of the second describes the vector distribution, interaction capabil-
climatic conditions, a leading contributor to vector dynam- ities with the immediate environment and their dependency
ics. Mniszewski et al. [34] leverage the efficacy of differential on the climatic conditions. The third presents the laws gov-
models by proposing a “hybrid network patch model” to give erning the daily behavior/movements of human agents, and
insights into the effect of variable probabilities in infection is used to stratify the human population. The scale consid-
model on the ABM; and Isidoro et al. [28] show the utility ered is at city level, and agent granularity is at the level of an
of ABMs to try out different mosquito control strategies. individual human and an individual mosquito. The follow-
However, both of these operate on random topologies and ing sections describe the three agents and the environment
thus do not correctly capture the host-vector interactions. they interact in.
Miksch et al. [32] show the potential of ABMs to recreate
real life epidemics through model parameterization and cal- 2.1.1 Infection Model
ibration but in a manner very limited to a particular disease Mosquito-borne diseases are transmitted human-to-mosquito
and location. Our work is a significant improvement over and mosquito-to-human. Symptoms vary but are generally
them as most of these models operate at a large scale with classifiable into two stages: intrinsic incubation period, and
granularity ranging from human crowd to cities network [16, infectious state. Incubation is when a mosquito has suc-
17, 33, 34], making them infeasible for low level policy de- cessfully infected a human and the pathogen has started
sign. They also focus on a very narrow set of parameters multiplying inside the host body. Depending on the disease,

427
a person may or may not be contagious during this stage. and sensory range (Sr ). Flying speed and maximum dis-
The infectious state is when the pathogen has multiplied tance bound the spatial movements of a mosquito agent.
and reached a certain threshold, and the human has possi- Target locations are assigned to a mosquito within a cir-
bly started to show symptoms. The probability of pathogen cle of radius Fr at the start of the active period, and the
transmission is very high during this period. Considering mosquito moves from one location to another with speed Fs .
that a lot of vector-borne diseases are known to have sig- The active period As to Ae are the hours between which the
nificant incubation periods (malaria, dengue, chikungunya, mosquito is actively searching for a blood meal or oviposi-
etc.), the SEIR model (Susceptible → Exposed → Infectious tioning sites. Maximum number of meals Mm bounds the
→ Recovered) was chosen as its exposed state addresses the number of bites a mosquito is going to take in a day. Mor-
incubation period. The susceptible state is when an individ- tality rate Mr is the probability of a mosquito dying on a
ual is vulnerable to infection by the pathogen. An individual given day due to some natural circumstance. Ovipositioning
is said to be exposed when he is carrying infection but is not characteristics include Pm : probability of an adult female to
yet contagious. The infected state is when the individual has mate successfully; and Oc → [0,1]: whether the mosquito
started showing symptoms and is ready to infect a mosquito. lays egg on a single ovipositioning site or spreads them over
Once the infection has passed, the individual enters the re- multiple such sites. Sex ratio (Pf ) is the probability that the
covered state. Whether a recovered person gains a lifetime hatching egg will be female. The sex ratio is relevant as only
immunity is determined by the disease. Figure 1 shows the female mosquitoes bite humans and lay eggs; males have no
transition of an individual between the four states and their direct role in the spread of infectious diseases. Sensory range
respective transition probabilities and average time spent determines how far a mosquito is aware of its surroundings.
in each state. The arrows in Figure 1 represents transition If a human agent or a water source comes within a radius of
probabilities between connected states. Chronic recovered is Sr of mosquito agent, only then may a mosquito target it.
the state when the person has recovered from the infection All these parameters were determined to be essential to
and is no longer contagious, but still has symptoms of the disease propagation (as they directly affect the mosquito
disease. It is given as an extension to the recovered state be- population and their behavior relative to the environment)
cause a lot of mosquito-borne diseases are known to exhibit and may vary across mosquito species. For example, female
such behavior [34]. Anopheles mosquitoes are generally active during night, while
Infection in mosquitoes follows a similar cycle, except they Aedes mosquitoes are day-biting. Similarly, Aedes albopic-
do not recover once they have been infected. When a sus- tus have more meals as compared to Aedes Aegypti because
ceptible female mosquito bites an infected human, it gets in- of their rapid bites, which generally keep their blood meals
fected with a non-zero probability. If the mosquito acquires short [7, 29].
the pathogen, it multiplies in the mosquito till it reaches Figure 2 shows a flow chart of how all the above parame-
enough strength to infect some other susceptible human, ters are connected together. A mosquito agent is created in
thereby completing a cycle. This period is called the ex- one of the water sources in the environment. There appears
trinsic incubation period. to have been no research with definitive results to mathemat-
ically model mosquito movements. Some evidence suggests
that simple random walk simulations give remarkably good
approximation of real data acquired through mark-release-
recapture field trials [45]. Thus, this study models mosquito
movements as random walks in a radius of Fr /day with a
survival rate of (1-Mr )/day. If it is the active period of
a mosquito, it starts moving randomly till a human agent
comes within its sensory range. Then it follows a targeted
approach and feeds on the human. This is the only stage
when the infection propagates between existing agents. If a
mosquito bites an infected human then it gets itself infected
with a probability β, and when such an infected mosquito
bites a susceptible human, it infects the human with prob-
ability α. If the mosquito has reached its Mm limit, it
rests. Every day, with probability Pm , a mosquito’s state is
changed to carrying eggs. If the mosquito is carrying eggs,
it continues to feed on human agents till its eggs are mature
enough to be laid. A mosquito needs at least one meal after
mating for the eggs to mature [7]. The maturation period τ
is determined as a function of the temperature θ in °Celsius
Figure 1: Infection Model: Infection progression states in (equation (1)).
humans τ = 3 + |θ − 21|/5 (1)
2.1.2 Mosquito Agents Three days is the time taken to mature eggs at the ideal
The mosquito agents move through the geographical space temperature of 21°C [13, 10]. The maturation period was
and are characterized by the following parameters: flying approximated to be decreasing linearly with respect to the
speed (Fs ), maximum distance (Fr ), active period (As ,Ae ), difference of temperature from ideal temperature, and the
maximum number of meals in a day (Mm ), mortality rate slope was derived from the graph of Ae. aegypti (the pri-
(Mr ), ovipositioning characteristics (Pm ,Oc ), sex ratio (Pf ), mary vector in validation case study) at different tempera-

428
tures [10, 46, 20]. Once a mosquito mates successfully, the
time required for the batch of eggs to mature is set to τ as
calculated by that day’s temperature.
Once the eggs have matured, the mosquito starts looking
for ovipositioning sites (water sources) and lays 100 eggs at
once or over time at different sites determined by Oc . For
simplicity, all the three aquatic stages of mosquito life cycle
(egg, larva and pupa) have been modeled as eggs, and will
be referred as such throughout this article. The time taken
by eggs to turn into an adult mosquito µ, is also a function
of temperature θ, but eggs may die before hatching if their
water source dries up and is not refilled [13]. The function
(equation (2)) was approximated using the average life cycle
of Ae. aegypti (aquatic phases) [7] and effects of temperature
on them [35].
µ = µegg + µlarve + µpupae
= (2 + |(θ − 25)|/2) + (2) + (4 + |(θ − 25)|/2|) (2)
= 8 + |(θ − 25)|
Once a batch of egg is laid, the time each egg will spend
in aquatic state before evolving into adult mosquito is set to
µ, calculated according to that day’s temperature.
Due to lack of enough published empirical evidence, both
the above equations ( (1) and (2)) were derived with the
assumption that there is no drastic change of temperature
(< ±5o C) between consecutive days. Eggs turn into adult
female mosquitoes with a probability Pf ; male mosquitoes
are discarded by the model as they do not directly contribute
to disease propagation.

2.1.3 Human Agents

The ABM tries to realistically represent the average daily
behavior of an individual in an urban landscape. Consid-
ering that mosquitoes move in a relatively short range, the
long-distance spatial movement of infection through the city
happens mostly as a result of displacement of human agents.
Thus, human agents were divided into four general cat-
egories based on their mobility patterns, based on where
they gather in groups and how often they go near potential
mosquito breeding grounds like lakes (as shown in Figure 3):
• Type 1: These human agents leave their assigned build-
ings at 7:00 a.m. for school and may take up to 30 min- Figure 2: Mosquito Agents: Movement, feeding and re-
utes to reach the school. They stay there till 3:00 p.m. production behaviors of a mosquito
and come back to their respective dwellings. From
there, with probability Pp1 an agent goes to a park at category is determined from the age distribution in the cen-
4:00 p.m. and comes back at 7:00 p.m., after which it sus data. The agents commute using the road network using
stays in place till next morning. Overall commuting the shortest path to their target location. Each agent has
takes at most 2 hours. an attribute called state which depicts which stage (SEIR)
it is in. Once a susceptible agent is bitten by an infected
• Type 2: Similar to Type 1, but the initial destination mosquito, it enters the exposed state with probability β.
is office and the probability of going to a park is Pp2 , An exposed agent carries on its daily activity and is mainly
which is less than Pp1 . responsible for the spatial movement of infection through
• Type 3: These are agents that are continuously on the the city. Once the agent has entered the infectious state, it
move. They leave their assigned buildings at 7:00 a.m. stays in its home till it has recovered. It may stay in recov-
and move in between buildings every two hours. They ered (immune) state or move to susceptible state depending
come back to their dwellings at 7:00 p.m. and stay on the disease.
there till next morning.
2.1.4 Environment
• Type 4: These are stationary agents who do not in-
volve themselves in any dynamic activity; they stay in In this model, an infected mosquito is considered to lay
their specific buildings throughout the day. infected eggs, as has been observed in the case of dengue
and chikungunya [8]. Such infected eggs develop into in-
During initialization of the model, each agent is assigned fected adult mosquitoes. Figure 4 gives a flow diagram to
a category and a building. The number of agents in each describe the infection propagation among vector and human

429
 The model uses two-layered geo-referenced GIS data of a
city to realistically represent spatial movements of an indi-
vidual in an urban setting. The first layer consists of build-
ings which include homes, offices and schools, parks and sta-
tionary water sources like lakes which are an integral part of
a vector’s life cycle. The second layer depicts the road net-
work connecting these buildings, and all human agents move
on this network. To simplify the model, the distribution of
human agents in first layer is initialized randomly. A small
subset of buildings are assigned as schools and offices, and
one park with lake/pond is introduced. The model com-
bines these GIS data with the weather information of the
location which is integral in the study of vector agents. It
consists of daily temperature and precipitation data. Us-
ing these parameters, the development cycles of mosquitoes
and pathogens are determined. Small water patches, which
may serve as mosquito breeding grounds, are introduced af-
ter every rainfall and their number depends on the extent
of rainfall. (Small puddles cease to exist in less than a week
if it does not rain in between, while large water bodies can
persist for months or more.)
For the purpose of this study, GIS data of Luneray, a
commune in Haute-Normandie region in northern France,
was used to model a general urban setting, and was initial-
ized with human population of 1000, mosquito population
Figure 3: Human Agents: Daily activities of human of 2000, and 500 mosquito eggs. This relatively low ratio
agents of mosquito/human was necessary to keep the model com-
putationally tractable, but was accounted for by tuning the
agents. An infection is transferred to a human from an in- mosquito mortality rate. It was then adapted and scaled to
fected mosquito with probability α, and infected human to the case studies using the corresponding location’s census
mosquito with probability β. data, population density, and climate information. A park
with a lake was separately added as it was not a part of the
available GIS data. For the purpose of subjective analysis of
the model, an animated visualization of the simulation was
generated using GAMA platform.

2.2 Simulation
As a proof of concept, an application of the model was
used for simulation of a case study of a past epidemic of
a mosquito-borne disease. The model was used to predict
the trajectory of the epidemic and the results were checked
against real data. Two intervention strategies were also
tested and analyzed. For the purpose of the demonstra-
tion, some model parameters needed to be tuned to fit that
particular scenario. We used a training data set containing
the data from a previous epidemic of the same disease in a
similar location to the testing site. Once all parameters are
determined which provided reasonably accurate results on
training data, the simulation is ran for the testing site.
(a) Mosquito Feeding
2.2.1 Case Study
For this study, simulation of outbreak of chikungunya in
the Caribbean region is used to illustrate the given frame-
work. For training the model, data from the 2013-14 chikun-
gunya epidemic in Saint Martin was used. Saint Martin,
with a population of fewer than 75,000, was the epicenter of
the Caribbean epidemic. In December 2013, two laboratory-
confirmed non-imported cases were reported for the first
time in the Caribbean in the district of Ocean Pond, close
(b) Mosquito (c) Egg Hatch- to the border of the Dutch side Sint Maarten. The model
Ovipositioning ing trained on this data was tested on epidemic data from an-
other Caribbean island, St. Barthélemy. As with the rest
Figure 4: Disease Transmission Cycle: Possible scenar- of the Caribbean, residents of St Barthélemy had never en-
ios of disease transmission

430
countered the virus before, and had no existing immunity. the area is likewise smaller, the model was initialized with
The female Ae. aegypti mosquito was taken as the vector 500 human individuals and 500 mosquitoes during testing.
in the simulation, as it was the primary vector of the 2014 This relatively low ratio of mosquito/human was accounted
CHIKV outbreak in the Caribbean [36]. for by significantly reducing the mosquito mortality rate to
0.05/day (in comparison to the typically accepted rate of
2.2.2 Data 0.2/day).
Chikungunya epidemic data were taken via the sentinel Uni-variate sensitivity analysis was conducted, i.e., model
network in Guadeloupe, Martinique, Saint Barthélemy and outcomes were analyzed with respect to one parameter at a
Saint Martin [47]. It was compiled together and made pub- time. The first parameter was a mosquito’s sensory range
licly available on a weekly basis by PAHO WHO [39]. Most Sr , which was tested for six different values and error in
epidemics only last for a few months. Also, only 3 months fitting was recorded for each value as shown in Figure 5.
of data on confirmed cases was made publicly available by
PAHO WHO. Therefore, we simulated a duration of 3 months
starting from January, 2013 for validation purpose. Weather
information was taken from [38] and interpolated to get daily
temperature and rainfall data. Lunerays’s GIS data were
made openly available by GAMA [2]. Census information
was taken from [6].

2.2.3 Simulation Toolkit

For the simulation of the ABM, GAMA was chosen as the
platform due to its data-driven ability and intuitive agent-
based language [42, 23].

3. MODEL VALIDATION AND RESULTS

Figure 5: MSE: Mean-Squared Error recorded for different
3.1 Parameterization and Initialization Sr values (Mm = 2)
Summarizing the above model, the following are parame- The error metric used was mean-squared error Equation
ters that drive the model: Mosquito Parameters (Fs , Fr , 3 (where, n is the total number of epidemiological weeks, i
As , Ae , Mm , Mr , Pm , Oc , Pf , Sr ), Human Parameters is the current week in simulation, Ȳi is the number of newly
(Pp1 ,Pp2 ), Infection Cycle Parameters (Ne , Ni , Prc , Pirc ) reported cases in real life and Yi is the number of new cases
and Infection Transmission Cycle Parameters (α, β). Of in simulation).
1X
n
these parameters, Sr , Mm , Pf , Pm , α and β were tuned M SE = (Ȳi − Yi )2 (3)
during the training of the model due to the lack of any con- n i=1
crete empirical evidence of their values, while the value of
Sr parameter was found to be significantly sensitive, and
others were compiled together from various sources. Some
gave best results when tuned to 3 meters. The second pa-
detailed description of the parameters and their sources is
rameter for sensitivity analysis was Mm , a mosquito’s max-
provided as Table 1. The number of schools, offices and
imum number of meals in a day. It was varied for three
households were initialized using census data [6]. Buildings
different values and errors were recorded (Figure 6). Mm
with the largest areas in the GIS data were assigned to be
was then fixed to 1 to minimize the error. Pm , Pf and Mr
schools and offices randomly.
were likewise fixed to 0.2, 0.5 and 0.05 respectively as they
3.2 Training on average gave approximately constant mosquito popula-
tion.
Model parameters were tuned using trial and error to give
a good fit on training data. The results were animated as
a simulation and were graphically represented. Error was
derived with respect to the historical/training data. Sen-
sitivity analysis was conducted to determine the effect of
each parameter on the model’s output and the more sen-
sitive parameters were made sufficiently accurate to reduce
the error [25].
A slight error margin was left while training because train-
ing data only consists of reported cases, while due to its non
fatal nature, some number of chikungunya infections go un-
reported [1, 3].
The entire population of even a small city cannot be taken
into consideration due to computational in-feasibility. The
model simulation was accomplished considering 1000 human Figure 6: Mm : Mean-Squared Error recorded for different
individuals and 2000 mosquitoes involved in a chikungunya Mm values
epidemic and interacting at a city scale, in case of training
location. Figure 7 shows how the ABM was fit into the historical
Considering that the population density of the testing lo- data. An error of ±9 cases was left to account for unreported
cation is only about half that of the training location, and cases.

431
 Table 1: ABM Parameters
Parameter Description Value Reference
Mosquito Parameters:
Fs Flying speed 0.0...1.0 km/hr [5]
Fr Maximum distance 350m [45]
As ,Ae Active period 7:00 a.m., 6:00 p.m. [8]
Mm Maximum number of meals in a day 1 Trained
Mr Mortality rate 0.05/day* [45]
Pm Probability of an adult female to mate successfully 0.2 Trained
Oc Ovipositioning behavior (single (0) or spread over multiple sites (1)) 1 [8]
Pf Probability that the hatching egg will be female 0.5 Trained
Sr Sensory range 3m Trained
Infection Transmission Cycle Parameters:
α Transmission probability of infection from mosquito to human 0.6 [17]
β Transmission probability of infection from human to mosquito 0.275 [17]
Infection Cycle Parameters:
nb infected init Initial number of infected people 2 [47]
Ne Number of days a human spends in exposed state 2...6 days [9]
Ni Number of days a human spend in infected state 4...7 days [9]
Prc Probability of transiting from recovered state to susceptible state 0 [48]
Pirc Probability that passing infection leaves human in chronic state 0.95 [9]
Human Parameters:
Pp1 Probability of a Type 1 agent of going to park at 4:00 p.m. 0.5 Trained
Pp2 Probability of a Type 2 agent of going to park at 4:00 p.m. 0.1 Trained
* [45] proposes 0.2/day mortality rate which was reduced considering the relatively low mosquito/human ratio in the model.

Figure 7: St. Martin Epidemic (Training): Results Figure 8: St. Barthélemy epidemic (Testing): Results
from training the model on the St. Martin epidemic from testing the model on the St. Barthélemy epidemic
3.3 Testing
real epidemic and used it for analyses of different prevention
Figure 8 shows the results when the trained model was strategies on the test location.
ran to check the effects of a chikungunya epidemic in Saint
Barthélemy and results were compared with the actual 2014 4.1 Strategies
chikungunya epidemic of the same. The error encountered No vaccine exists to prevent chikungunya virus infection.
was ± 0.54 cases, nearly insignificant, indicating that our It can only be prevented by avoiding mosquito bites either
model scales well to changing population size. The initial at the individual level or by reducing the mosquito popula-
results are a little less than the reported results which may tion [12]. Keeping this in mind, two popular paradigms of
very well be from the fact that by the time the first case mosquito prevention were implemented, analyzed and com-
was recorded, the infection had already spread through the pared using the above model:
environment in the real world, while the model is initialized
as infection-free. It also suggest estimates on certain Ae. 1. Centralized Prevention Strategies: In this case, there
aegypti parameters like sensory range (3 meters) and average is a centralized organization dedicated to reducing the
feeding frequency (1 meal/day). mosquito population through methods like detection
The results obtained show that ABMs can provide very and elimination of breeding places, proper covering of
accurate results when used to model the complex epidemiol- persistent water sources, and reliable water supply [37].
ogy of mosquito-borne diseases. They also provide estimates We choose LSM (Larval Source Management), which
on certain mosquito parameters like sensory range, feeding is economical and very popular, as a strategy. Its per
frequency, etc. which are difficult to measure otherwise. sq km associated annual cost when adjusted for 2016
prices [11], comes to US$6000 [26].
4. PREVENTION STRATEGIES 2. Decentralized Prevention Strategies: Here people adopt
As the model performed well in the validation phase, we prophylactic measures like use of mosquito repellent
assumed that it is a reasonably effective representation of the creams, liquids, coils, mats, etc., or full body coverings

432
 to prevent mosquito bites. One person can be pro-
tected for an year at a cost of US$10 with insecticide-
treated nets (ITNs); indoor residual spraying (IRS)
costs US$180/building [22, 24].

All costs mentioned above are as quoted by private agen-

cies. They may be brought down significantly by conducting
large-scale mosquito- control programs, and with relevant
government subsidies.
For comparison of the two strategies, the two iterations
were initialized with same resources, i.e., the same initial
budget of US$ 20,000. Considering it would cost approxi-
mately US$ 30,000/year to cover a town of 5 sq km com-
pletely for a larvicide program, we assume that only 66% of
the larval habitats are found and treated. This is consistent Figure 10: Mosquito Population (Strategy 1): Number
with the fact that larval habitats may be small and widely of Mosquitoes and Eggs per Day
dispersed, making it difficult to locate all of them [14]. For
the second strategy (ITN+IRS), resources worth 20,000 US$
were distributed randomly among the population (220 US$
per household) helping individuals in the population avoid
mosquito bites in their households.

4.2 Outcomes
Strategy 1 (LSM) was found to be more effective at com-
pletely eliminating the infection from both the human and
mosquito populations within 5 weeks. Strategy 2 took 7
weeks to do the same. Not trying to contain the infection
kept it going even after 13 weeks (Figure 9). This differ-
ence in effectiveness may be from the fact that Strategy 2 is
mostly useful during the night time when most individuals
are back to their households, while A. Aegypti is a day-biting Figure 11: Mosquito Population (Strategy 2): Number
mosquito. of Mosquitoes and Eggs per Day
quantitative analyses of mosquito-borne epidemics. This
and other such models carry high practical value as they can
be used to quickly try out different combinations of strate-
gies, study the effects of infection on different population
groups, find high risk groups, plan vaccination programs,
and categorize locations with respect to their seasonal risk
factor. They can also be used to arrive at estimates of vec-
tor parameters (such as the mosquito’s sensing range) that
are not easy to measure directly. The case study demon-
strated how the model was accurately able to trace the evo-
lution of the infection, given data from a past epidemic in
a similar location is provided. It also predicted that on a
low budget, centralized strategies like larval source manage-
ment are more effective compared to individual measures
Figure 9: Prevention Strategies: Results from testing the
like insecticide-treated bed nets and indoor residual spray-
two strategies
ing. Though the above model has been tested on tropical
Strategy 1 also had a much greater effect on mosquito pop- climate of Caribbean Islands, it can be easily extended to
ulation, reducing it by 70% while Strategy 2 only reduced it other locations provided relevant training data are available.
by 25% (Figure 10 and Figure 11). The results of both the The model can also be extended to consider mutations
strategies were found to be consistent with past studies [31, of infection [18], actual human mobility patterns, multiple
40]. mosquito species, and other factors like wind velocity.
From the fact that in most cases a mosquito’s life cycle
is not greatly dependent on the presence or absence of the Acknowledgements
infection due to their short lifespan [19], it can be reasoned
that the above results will hold true for other diseases spread The authors would like to thank Barbara Han for giving
by A. Aegypti, like the presently prominent Zika virus. insights on mosquito behavior.

5. CONCLUSION AND DISCUSSION REFERENCES

Diseases like chikungunya and dengue which currently have [1] Chikungunya fever guide.
no vaccine, need proactive measures to prevent or contain http://www.chikungunya.in/, 2016. Accessed:
their breakout. Our ABM provides a good mechanism for 2016-05-06.

433
 [2] Importation of gis data. https://github.com/ [18] J. Z. Farkas, S. A. Gourley, R. Liu, and A.-A. Yakubu.
gama-platform/gama/wiki/LuneraysFlu_step3, 2016. Using Mathematics at AIM to Outwit Mosquitoes.
Accessed: 2016-5-15. Notices Amer. Math. Soc., 63(03):292–293, 2016.
[3] D. Almond. What is the state of the Caribbean [19] H. M. Ferguson and A. F. Read. Why is the effect of
chikungunya epidemic now?, 2015. Accessed: malaria parasites on mosquito survival still
2016-05-06. unresolved? Trends in parasitology, 18(6):256–261,
[4] S. M. N. Arifin, G. J. Davis, and Y. Zhou. A spatial June 2002.
agent-based model of malaria: Model verification and [20] S. Fischer, I. S. Alem, M. S. De Majo, R. E. Campos,
effects of spatial heterogeneity. Int. J. Agent Technol. and N. Schweigmann. Cold season mortality and
Syst., 3(3):17–34, July 2011. hatching behavior of Aedes aegypti L. (Diptera:
[5] I. Bargielowski, C. Kaufmann, L. Alphey, P. Reiter, Culicidae) eggs in Buenos Aires City, Argentina.
and J. Koella. Flight performance and teneral energy Journal of Vector Ecology, 36(1):94–99, 2011.
reserves of two genetically-modified and one wild-type [21] S. Fu and G. Milne. Epidemic modelling using cellular
strain of the yellow fever mosquito Aedes aegypti. automata. In Proc. of the Australian Conference on
Vector-Borne and Zoonotic Diseases, Artificial Life. Citeseer, 2003.
12(12):1053–1058, July 2012. [22] GiveWell. Mass distribution of long-lasting
[6] M. Barrientos and C. Soria. Saint Barthelemy age insecticide-treated nets (LLINs).
structure. http://www.indexmundi.com/saint_ http://www.givewell.org/international/
barthelemy/age_structure.html, 2015. Accessed: technical/programs/insecticide-treated-nets,
2016-04-12. 2016.
[7] D. Bleijs. Aedes albopictus. http://www. [23] A. Grignard, P. Taillandier, B. Gaudou, D. A. Vo,
chikungunyavirusnet.com/aedes-albopictus.html, N. Q. Huynh, and A. Drogoul. GAMA 1.6: Advancing
2014. Accessed: 2016-04-12. the art of complex agent-based modeling and
[8] D. Bleijs. Aedes aegypti. http: simulation. Lecture Notes in Computer Science, pages
//www.denguevirusnet.com/aedes-aegypti.html, 117–131, 2013.
2016. Accessed: 2016-04-12. [24] Hicare.in. Make your home safe today. expert
[9] D. A. Bleijs. Chikungunya signs clinical symptoms. mosquito control service.
http://www.chikungunyavirusnet.com/ http://hicare.in/mosquito_control/, 2016.
signs-a-symptoms.html, 2016. Accessed: 2016-04-12. Accessed: 2016-5-15.
[10] O. J. Brady, M. A. Johansson, C. A. Guerra, S. Bhatt, [25] Y. Huang, X. Xiang, G. Madey, and S. E. Cabaniss.
N. Golding, D. M. Pigott, H. Delatte, M. G. Grech, Agent-based scientific simulation. Comput. Sci. Eng.,
P. T. Leisnham, and R. e. a. Maciel-de Freitas. 7(1):22–29, 2005.
Modelling adult Aedes aegypti and Aedes albopictus [26] ICMR. Urban mosquito control - a case study. ICMR
survival at different temperatures in laboratory and Bulletin, 30(3), 2000.
field settings. Parasites |& Vectors, 6(1), Dec. 2013. [27] C. Illangakoon, R. D. McLeod, and M. R. Friesen.
[11] U. I. Calculator. Us inflation calculator. Agent based modeling of malaria. In Humanitarian
http://www.usinflationcalculator.com/, 2016. Technology Conference - (IHTC), 2014 IEEE Canada
Accessed: 2016-5-15. International, June 2014.
[12] CDC. Chikungunya virus. [28] C. Isidoro, N. Fachada, F. Barata, and A. C. Rosa.
http://www.cdc.gov/chikungunya/prevention/, Artificial life model of dengue host-vector disease
2016. Accessed: 2016-05-06. propagation. In International Joint Conference on
[13] CDC. Dengue and the Aedes aegypti mosquito. Computational Intelligence, pages 243–247, Oct. 2009.
http://www.cdc.gov/dengue/resources/30Jan2012/ [29] N. Kamaladhasan, B. K. Tyagi, P. S. Swamy, and
aegyptifactsheet.pdf, 2016. Accessed: 2016-05-06. S. Chandrasekaran. Studies on the maintenance of
[14] CDC. Larval control and other vector control ’self-sustained’ mosquito vector population in Vaigai
interventions. river, South India. Current Science, 110(1):57–68,
http://www.cdc.gov/malaria/malaria_worldwide/ 2016.
reduction/vector_control.html, 2016. Accessed: [30] A. Kleczkowski and B. T. Grenfell. Mean-field-type
2016-05-06. equations for spread of epidemics: The small world
[15] R. Connell, P. Dawson, and A. Skvortsov. Comparison model. Physica A: Statistical Mechanics and its
of an agent-based model of disease propagation with Applications, 274(1):355–360, 1999.
the generalised sir epidemic model. Technical report, [31] E. Majambere, SilasWorrall. Potential and
DTIC Document, 2009. costeffectiveness of lsm, 2007.
[16] C. Deng, H. Tao, and Z. Ye. Agent-based modeling to [32] F. Miksch, P. Pichler, K. J. P. Espinosa, K. S. T.
simulate the dengue spread. In Proceedings of SPIE - Casera, A. N. Navarro, and M. Bicher. An agent-based
The International Society for Optical Engineering epidemic model for dengue simulation in the
7143, Nov. 2008. philippines. In 2015 Winter Simulation Conference
[17] C. J. Dommar, R. Lowe, M. Robinson, and X. Rodó. (WSC), pages 3202–3203, Dec 2015.
An agent-based model driven by tropical rainfall to [33] S. M.Niaz Arifin, R. Reaz Arifin, D. De, A. Pitts, and
understand the spatio-temporal heterogeneity of a G. R. Madey. Integrating an agent-based model of
chikungunya outbreak. Acta Tropica, 129:61–73, Jan. malaria mosquitoes with a geographic information
2014. system. In The 25th European Modeling and

434
 Simulation Symposium (EMSS 2013), Sept. 2013. [43] N. H. Shah and J. Gupta. SEIR model and simulation
[34] S. M. Mniszewski, C. A. Manore, C. Bryan, S. Y. for vector borne diseases. Applied Mathematics,
Del Valle, and D. Roberts. Towards a hybrid 4(08):13, 2013.
agent-based model for mosquito borne disease. In [44] S. Swarup, S. G. Eubank, and M. V. Marathe.
Proceedings of the 2014 Summer Simulation Computational epidemiology as a challenge domain for
Multiconference (SummerSim ’14), pages 10:1–10:8, multiagent systems. In Proceedings of the 2014
2014. international conference on Autonomous agents and
[35] A. Mohammed and D. D. Chadee. Effects of different multi-agent systems, pages 1173–1176. International
temperature regimens on the development of Aedes Foundation for Autonomous Agents and Multiagent
aegypti (L.) (Diptera: Culicidae) mosquitoes. Acta Systems, 2014.
Tropica, 119(1):38–43, Apr. 2011. [45] C. J. Thomas, D. E. Cross, and C. Bøgh. Landscape
[36] L. Mowatt and S. T. Jackson. Chikungunya in the movements of Anopheles gambiae malaria vector
Caribbean: An epidemic in the making. Infect Dis mosquitoes in Rural Gambia. PloS one, 8(7):e68679,
Ther, 3(2):63–68, 2014. July 2013.
[37] M. of Health and G. o. I. Welfare. Vector control [46] S. Thomas, U. Obermayr, D. Fischer, J. Kreyling, and
measures. C. Beierkuhnlein. Low-temperature threshold for egg
http://www.nvbdcp.gov.in/dengue12.html, 2016. survival of a post-diapause and non-diapause
Accessed: 2016-05-06. European aedine strain, Aedes albopictus (Diptera:
[38] L. Osborn. Caribbean weather in December. Culicidae). Parasites |& Vectors, 5(1), May 2012.
http://bit.ly/1Tn8zAj, 2016. Accessed: 2016-04-12. [47] W. Van Bortel, F. Dorleans, J. Rosine, A. Blateau,
[39] PAHO. Chikungunya: Statistic data. D. Rousset, S. Matheus, I. Leparc-Goffart, O. Flusin,
http://bit.ly/1KA6cUE, 2016. Accessed: 2016-05-06. C. Prat, R. Cesaire, et al. Chikungunya outbreak in
[40] S. Pang, L. Chiang, C. Tan, I. Vythilingam, the Caribbean region, December 2013 to March 2014,
S. Lam-Phua, and L. Ng. Low efficacy of and the significance for Europe. Eurosurveillance,
delthamethrin-treated net against singapore aedes 19(13), Apr. 2014.
aegypti is associated with kdr-type resistance. Tropical [48] WHO. Chikungunya.
Biomedicine, 32(1):140–50, Mar. 2015. http://www.who.int/denguecontrol/arbo-viral/
[41] L. Perez and S. Dragicevic. An agent-based approach other_arboviral_chikungunya/en/, 2016. Accessed:
for modeling dynamics of contagious disease spread. 2016-04-12.
International journal of health geographics, 8(1):1, [49] WHO. Vector borne diseases. http:
2009. //www.who.int/mediacentre/factsheets/fs387/en/,
[42] S. F. Railsback, S. L. Lytinen, and S. K. Jackson. 2016. Accessed: 2016-04-12.
Agent-based simulation platforms: Review and [50] V. Wiwanitkit. Vaccination against mosquito borne
development recommendations. Simulation, viral infections: current status. Iran J Immunol,
82(9):609–623, Mar. 2006. 4(4):186–196, 2007.

435