Professional Documents
Culture Documents
Abstract
The choices of a population to apply social distancing are modeled as a Nash game, where the agents determine their
social interactions. The interconnections among the agents are modeled by a network. The main contribution of this
work is the study of an agent-based epidemic model coupled with a social distancing game, which are both determined
by the networked structure of human interconnections. The information available to the agents plays a crucial role.
We examine the case that the agents know exactly the health states of their neighbors and the case they have only
statistical information for the global prevalence of the epidemic. The agents are considered to be myopic, and thus,
the Nash equilibria of static games for each day are studied. Through theoretical analysis, we characterise these Nash
equilibria and we propose algorithms to compute them. Interestingly, in the case of statistical information the equilibrium
strategies for an agent, at each day, are either full isolation or no social distancing at all. Through experimental studies,
we observe that in the case of local information, the agents can significantly affect the prevalence of the epidemic with
low social distancing, while in the other case,they can also affect the prevalence of the epidemic, but they have to pay the
burden of not being well informed by applying strict social distancing. Moreover, the effects of the network structure,
the virus transmissibility, the number of vulnerable agents, the health care system capacity and the information quality
(fake news) are discussed and relevant simulations are provided.
Keywords: Social distancing, Games on networks, Epidemics on networks, Nash game, Information patterns
∂ 2 Jˆi X
piu (k) ∈ [0, 1]. (25) = G i (1 − p x )(1 − p r ) pju pc px (1 − pr )
(∂piu )2
j∈Ni
We then drop k in order to proceed with the analysis of X
k c
Y
one step of the game. In order to study the equilibria · pu p px (1 − pr ) (1 − piu plu pc px (1 − pr )),
of this game we need the expectation of the state of the k∈Ni \{j} l∈Ni \{j,k}
j∈Ni
one of them. Indeed, all the other quantities in (31) are
(26) strictly positive before the end of the epidemic (when the
next we compute the expectation of the previous condi- epidemic ends px = 0). Particularly, Gi (1−px )(1−pr ) > 0,
tional expectation over all the states: px pc > 0, and 1 − piu plu pc px > 0. Let pj∗ u 6= 0. Then, a
value p̃iu , with 0 < p̃iu < 1, can be a best response for
Y player i, only if:
Ex,r E{x+ (1−uij uji pc px (1−pr )),
i | x, r} = 1−(1−px )
j∈Ni Jˆi (piu = 1) = Jˆi (piu = 0) = 0. (32)
(27)
and thus the criteria have the following form: Thus:
pj∗ c
u [si − Gi (1 − pr )(1 − px )p px ] = Gi (1 − pr )px , (33)
X
Ji = si uij uji and:
j∈Ni ∂ Jˆi i
(p = 0) = 0, (34)
Y ∂piu u
+ Gi (1 − px ) (1 − uij uji pc px (1 − pr )) − Gi (1 − pr ),
which implies:
j∈Ni
(28) pj∗ c
u [si − Gi (1 − pr )(1 − px )p px ] = 0. (35)
where the strategies are random and uniform for all the It is then obvious that, while the epidemic continues (px >
neighbors of an agent according to eq.(24), so we have to 0), the equations (33) and (35) are contradictory. So, piu
compute the expected criteria, given the probabilities piu has to be in {0, 1} for all i.
of the uniform strategies:
In order to characterize the Nash equilibria of this game
X we observe that it is strategically equivalent to the follow-
Jˆi = E{Ji |piu , pju , j ∈ Ni } = si piu pju + Gi (1 − p0x )
ing one:
j∈Ni
Y
piu pju pc p0x )
· (1 − − Gi (1 − pr ) (29) X Y
J˜i (piu , p−i i
u ) = ai pu pju + (1 − bpiu pju ), (36)
j∈Ni
j∈Ni j∈Ni
si 4
ai = , b = pc px (1 − pr ), (37)
Gi (1 − px )(1 − pr )
3
f(mi)
and piu
∈ {0, 1}, for all i. 2
We proceed with the calculation of the best response
1
for each agent. From Proposition 3 we know that each
agent plays piu = 0 or piu = 1, so we rewrite the payoffs of 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
the agents as functions of the number of their neighbors mi
playing pju = 1. We denote this number mi .
Figure 1: The function f (mi ), for several values of ai and ln(1−b) =
J˜i (piu , mi ) = ai mi piu + (1 − bpiu )mi (38) −1.
and
Proof. It is easily observed that fi (m) is convex and fi (0) =
J˜i (0, mi ) = 1, (39) 1 for each i. So, if fi0 (0) ≥ 0 ⇒ fi (m) > 1, ∀m, in this
case m0 = 0. Else if fi0 (0) ≤ 0 ⇒ ∃!m0 ∈ R∗+ : f (m0 ) = 1
J˜i (1, mi ) = ai mi + (1 − b)mi . (40)
and ∀m > m0 , m ∈ N : f (m) > 1 due to the convexity of
Thus, her best response depends solely on mi . To study fi (m).
this dependence, we define the following functions:
Proof of the Proposition 4. : Due to this lemma, begin-
fi (m) = J˜i (1, m) = ai m + (1 − b)m = ai m + em ln(1−b) ning with the maximum feasible value for mi (which is di )
(41) the changes in the agents actions from 1 to 0 can result
only in the decrease of their neighbors mj ’s and thus it
The best response of each agent is: is possible to happen only one change of action (1 → 0)
for each agent until the algorithm converges. To see this
1 , if fi (mi ) > 1
BRi (mi ) = (42) observe that if f (mi ) ≤ 1 as mi becomes smaller f (mi )
0 , otherwise
cannot become larger than 1. Moreover, due to this obser-
So, we propose Algorithm 2 for the computation of the vation, in the worst case the ‘while-loop’ will run N times
actions corresponding to a Nash equilibrium. and so the algorithm will converge in O(N 2 ) steps.
The point that the algorithm converges is a Nash equilib-
Algorithm 2 Computation of the NE strategies for the rium of the game, since the agents actions are their best
game with information for the distribution of the states responses to their active contacts numbers mi ’s and for
this profile of mi ’s no agent will be benefited from a uni-
1: The optimal strategies pi∗
u
lateral deviation from her action.
2: Set piu = 1, ∀i
Furthermore, we should point that, since the algorithm is
3: Compute fi (mi ), ∀i (mi = di )
in fact a descent on the possible mi -profiles, i.e. it initial-
4: while ∃fi (mi ) ≤ 1 do
izes with all the contacts being active (mi = di , ∀i) and
5: if fi (mi ) ≤ 1 then
each mi decreases or stays the same, the Nash equilibrium
6: Set piu = 0
that the algorithm converges is the one corresponding to
7: end if
the maximum possible sociability for the agents.
8: Compute new mi , ∀i
9: Compute new fi (mi ), ∀i Remark 9. Each player i, to implement Algorithm 2,
10: end while needs to know the number of neighbors intended to play
pju = 1 i.e., mi . So after each iteration of the algorithm
we assume that each player broadcasts to her neighbors her
Proposition 4. There exists a Nash equilibrium of the intended action, and finally chooses the actual piu to play
game with statistical information for the distribution of the after Algorithm 2 converges.
states. Furthermore, Algorithm 2 converges to the Nash
equilibrium in O(N 2 ) steps. Remark 10. If for each agent i it holds that si di + Gi (1 −
px )(1 − pr )[(1 − pc px (1 − pr ))di − 1] > 0 then the strategy
To prove this proposition we firstly prove the following profile piu = 1, ∀i is a Nash equilibrium of that game.
lemma:
Proposition 5. The strategy profile piu = 0, ∀i is again
Lemma 1. For the functions fi (m), defined in (41), there a Nash equilibrium, since it results to indifference between
exists a unique m0 ∈ R+ such that f (m0 ) = 1 and for all the unilateral changes of each agent.
m > m0 , m ∈ N : f (m) > 1.
7
The analysis of this section, and especially Proposition
3, indicates a rather interesting fact: in the statistical in- 0.5
Infected
formation game the agents choice is either full isolation or
agents
no social distancing at all. This phenomenon can be at-
tributed to the fear of the agents due to the lack of knowl-
edge about their neighbors’ health states. If the preva- 0
0 10 20 30 40 50 60 70 80 90 100
lence of the disease is high and the agent considers herself
1
Recovered
to be vulnerable it is probable to be afraid to have any
social interaction and choose full isolation. On the other
agents
0.5
hand, if the agent considers herself non-vulnerable or the
prevalence of the disease to be low it is rather probable to
0
continue her daily activities without applying social dis- 0 10 20 30 40 50 60 70 80 90 100
tancing.
Active social
1
contacts
5. Numerical studies 0.8
8
Average Non-vulnerable Vulnerable
payoff agents agents
4 4 3.5
0.5 3
Infected
3 3
2.5
agents
2
2 2
1.5
1
1 1
0 0.5
0 10 20 30 40 50 60 70 80 90 100 0 0 0
1 2 1 2 1 2
1
Recovered
Figure 5: Comparison of the average payoffs of the agents for the two
agents
0.5 games. The case 1 stands for the game with perfect local information
and the case 2 for the game with statistical information
0
0 10 20 30 40 50 60 70 80 90 100
of the varying parameters on the behavior of the agents
Active social
1
and on the outspread of the epidemic. All the results are
contacts
0.2 1
Infection
0.5
peak
0
0 10 20 30 40 50 60 70 80 90 100
1 0
Recovered
1
agents
0.5
outspread
0 0.5
0 10 20 30 40 50 60 70 80 90 100
Active social
1 0
1.5 2 2.5 3 3.5 4 4.5 5 5.5
active contacts
contacts
0.8 1
Minimum
0.6
0.5
0 10 20 30 40 50 60 70 80 90 100
Days 0
1.5 2 2.5 3 3.5 4 4.5 5 5.5
Figure 4: Infection, recovery and sociability curves for the game with Average degree
1 1
statistical information: r = 100 (dotted line), r = 300 (dashed line),
1 1 Figure 6: The effects of the average degree of the graph topology for
r = 1000 (dashed-dotted line) and r = 2000 (solid line)
the case of no social distancing (solid line), the case of the game with
perfect local information (dashed line) and the case of the game with
statistical information (dashed-dotted line)
they are unable to choose rationally their social interac-
tions and it is more probable for them to get infected. So,
the vulnerable agents pay a greater burden for not being gree i.e., the average number of neighbors of each agent, on
well informed. the epidemic peak, on the total epidemic size and on the
maximum social distancing i.e., the minimum active so-
cial contacts. We considered random graphs of 100 agents
6. Case Studies and Discussion with varying adjacency probabilities padj = 0.03, ..., 0.15
resulting in average degrees d¯ = 3, .., 15. The increase of
In this section, we consider several variations of the ini- the average degree, which results in a better mixing of
tial problem and examine, through simulations, the effects the population results in the increase of the total infection
9
outspread for both games. Moreover, in the case of the
game with statistical information a small average degree 0.2
Infected
agents
results in almost isolation of the agents during the social
distancing. 0.1
Subsequently, we present the effects of the graph topol-
ogy. We consider four different graph topologies: a ran- 0
0 10 20 30 40 50 60 70 80 90 100
dom graph [22], a stochastic block graph i.e., a coalition
1
Recovered
of 5 blocks (random graphs) with higher adjacency prob-
agents
ability for the agents belonging to the same block, a scale
0.5
free graph [7] and a small world graph [44]. In every case
we have chosen the network parameters in a way that the
0
graphs have almost the same average degree (d¯ ≈ 10), in 0 10 20 30 40 50 60 70 80 90 100
order to avoid the consequences of different degrees shown
Active social
1
contacts
in Figure 6. In Figure 7, we present the effects of the
topology in the case of no social distancing game. In Fig- 0.9
ure 8, we examine the case of the game with perfect local
information and in Figure 9, the case of the game with 0.8
statistical information. 0 10 20 30 40 50 60 70 80 90 100
Days
0.6 Figure 8: Epidemic dynamics for a random graph (solid line), a
stochastic block graph (dashed line), a scale free graph (dashed-
Infected
agents
0.4 dotted line) and a small-world graph (dotted line) for the game with
perfect local information
0.2
0
0 10 20 30 40 50 60 70 80 90 100 0.2
Infected
agents
1
0.1
Recovered
agents
0.5
0
0 10 20 30 40 50 60 70 80 90 100
1
Recovered
0
0 10 20 30 40 50 60 70 80 90 100
agents
Days
0.5
Figure 7: Epidemic dynamics for a random graph (solid line), a
stochastic block graph (dashed line), a scale free graph (dashed- 0
dotted line) and a small-world graph (dotted line) for the case of 0 10 20 30 40 50 60 70 80 90 100
Active social
no social distancing 1
contacts
0.8
In every case we observe that the topology affects both
the epidemic outbreak and the agents behavior. The seg- 0.6
mentation of the population into ill interconnected blocks
0 10 20 30 40 50 60 70 80 90 100
(stochastic block graph) results in the diminishment of the
Days
outbreak in every case. The scale free property i.e., the
existence of central nodes with significantly higher degree, Figure 9: Epidemic dynamics for a random graph (solid line), a
results in an early high peak of the epidemic and the con- stochastic block graph (dashed line), a scale free graph (dashed-
sequent need for strict social distancing during this period. dotted line) and a small-world graph (dotted line) for the game with
statistical information
Finally, the small world property i.e., the existence of edges
that reduce the graph diameter, results in lower peaks but
in extended duration of the epidemic and thus it results to of the parameter R0 on the epidemic outspread and on
the need for an extended ”soft” social distancing. the agents behavior for both games, with perfect local and
with statistical information. We consider R0 = 1.5, ..., 5.5.
6.2. Virus transmissibility From Figure 10 we observe that in the case of no social
A very important characteristic of every epidemic dis- distancing and in the case of the game with perfect local
ease is its transmissibility. In the compartmental models information the disease prevalence (peak and total size) is
the transmissibility is incorporated in the basic reproduc- increasing with respect to R0 . However, in the case of the
tion number R0 . So, in this subsection we study the effects game with statistical information the disease prevalence
10
0.25 0.7
0.2 0.6
0.5
peak
0.15 0.5
0
1.5 2 2.5 3 3.5 4 4.5 5 5.5
Total infection
1 0.1 0.4
0 0.1 0.2 0.3 0 0.1 0.2 0.3
outspread
1
Next, we study a variation of the game with statisti-
Minimum
0.5
cal information, where the vulnerability parameters of the
agents (Gi ) depend on the level of infection in the commu-
0
nity. This is a realistic scenario, since the health systems
1.5 2 2.5 3 3.5 4 4.5 5 5.5 worldwide have finite and usually small capacity, so if the
Virus Transmissibility (R0) number of infected agents who need health care pass a
certain level it is not probable for the next agents who
Figure 10: The effects of the virus transmissibility for the case of no
will get infected to have access in the necessary facilities.
social distancing (solid line), the case of the game with perfect local
information (dashed line) and the case of the game with statistical We model this phenomenon considering the vulnerability
information (dashed-dotted line). parameters to be proportional to the infection ratio:
0.7
agents
0.1 0.2
0.6
0.15
0
0 50 100 150 0.5
0.1
1
Recovered
0.05 0.4
agents
0.8 7. Conclusion
epidemic’s outspread.
0.2 Future work in this direction may include the follow-
ing. At first, in the game analysed in section 4 may arise
0 more equilibria, so it is interesting, if possible, to charac-
0 10 20 30 40 50 60 70 80 90 100
1
terize all the equilibria. Secondly, the study of the case of
Recovered
1
contacts
12
[9] Pierre-Alexandre Bliman and Michel Duprez. How best can networks. arXiv preprint arXiv:2011.00445, 2020.
finite-time social distancing reduce epidemic final size? Journal [29] Ashish R Hota, Tanya Sneh, and Kavish Gupta. Impacts
of Theoretical Biology, 511:110557, 2020. of game-theoretic activation on epidemic spread over dynam-
[10] Joseph Frédéric Bonnans and Justina Gianatti. Optimal control ical networks. SIAM Journal on Control and Optimization,
techniques based on infection age for the study of the covid-19 60(2):S92–S118, 2021.
epidemic. 2020. [30] Yunhan Huang and Quanyan Zhu. Game-theoretic frameworks
[11] Colin J Carlson, Ana CR Gomez, Shweta Bansal, and Sadie J for epidemic spreading and human decision-making: A review.
Ryan. Misconceptions about weather and seasonality must not Dynamic Games and Applications, pages 1–42, 2022.
misguide covid-19 response. Nature Communications, 11(1):1– [31] William Ogilvy Kermack and Anderson G McKendrick. A con-
4, 2020. tribution to the mathematical theory of epidemics. Proceedings
[12] Arun G Chandrasekhar, Paul S Goldsmith-Pinkham, of the royal society of london. Series A, Containing papers of a
Matthew O Jackson, and Samuel Thau. Interacting re- mathematical and physical character, 115(772):700–721, 1927.
gional policies in containing a disease. Available at SSRN, [32] Johannes Köhler, Lukas Schwenkel, Anne Koch, Julian
2020. Berberich, Patricia Pauli, and Frank Allgöwer. Robust and
[13] Sheryl L Chang, Mahendra Piraveenan, Philippa Pattison, and optimal predictive control of the covid-19 outbreak. Annual
Mikhail Prokopenko. Game theoretic modelling of infectious Reviews in Control, 51:525–539, 2021.
disease dynamics and intervention methods: a review. Journal [33] Wonjun Lee, Siting Liu, Hamidou Tembine, Wuchen Li, and
of biological dynamics, 14(1):57–89, 2020. Stanley Osher. Controlling propagation of epidemics via mean-
[14] Samuel Cho. Mean-field game analysis of sir model with social field control. arXiv preprint arXiv:2006.01249, 2020.
distancing. arXiv preprint arXiv:2005.06758, 2020. [34] Lauren Ancel Meyers, MEJ Newman, and Babak Pourbohloul.
[15] Wongyeong Choi and Eunha Shim. Optimal strategies for social Predicting epidemics on directed contact networks. Journal of
distancing and testing to control covid-19. Journal of theoretical theoretical biology, 240(3):400–418, 2006.
biology, 512:110568, 2020. [35] Cristopher Moore and Mark EJ Newman. Epidemics and perco-
[16] Wongyeong Choi and Eunha Shim. Optimal strategies for vac- lation in small-world networks. Physical Review E, 61(5):5678,
cination and social distancing in a game-theoretic epidemiologic 2000.
model. Journal of theoretical biology, 505:110422, 2020. [36] Mark EJ Newman. Spread of epidemic disease on networks.
[17] Oliver M Cliff, Nathan Harding, Mahendra Piraveenan, E Yag- Physical review E, 66(1):016128, 2002.
mur Erten, Manoj Gambhir, and Mikhail Prokopenko. Investi- [37] Balázs Pejó and Gergely Biczók. Corona games: Masks, social
gating spatiotemporal dynamics and synchrony of influenza epi- distancing and mechanism design. In Proceedings of the 1st
demics in australia: An agent-based modelling approach. Sim- ACM SIGSPATIAL International Workshop on Modeling and
ulation Modelling Practice and Theory, 87:412–431, 2018. Understanding the Spread of COVID-19, pages 24–31, 2020.
[18] Mohammadali Dashtbali, Alaeddin Malek, and Mehdi Mirzaie. [38] Michael Renardy and Robert C Rogers. An introduction to
Optimal control and differential game solutions for social dis- partial differential equations, volume 13. Springer Science &
tancing in response to epidemics of infectious diseases on net- Business Media, 2006.
works. Optimal Control Applications and Methods, 41(6):2149– [39] Ronald Ross. An application of the theory of probabilities to the
2165, 2020. study of a priori pathometry. Proceedings of the Royal Society
[19] Ceyhun Eksin, Jeff S Shamma, and Joshua S Weitz. Disease of London. Series A, Containing papers of a mathematical and
dynamics in a stochastic network game: a little empathy goes physical character, 92(638):204–230, 1916.
a long way in averting outbreaks. Scientific reports, 7(1):1–13, [40] LM Sander, CP Warren, IM Sokolov, C Simon, and J Koopman.
2017. Percolation on heterogeneous networks as a model for epidemics.
[20] Romuald Elie, Emma Hubert, and Gabriel Turinici. Contact Mathematical biosciences, 180(1-2):293–305, 2002.
rate epidemic control of covid-19: an equilibrium view. arXiv [41] Hamidou Tembine. Covid-19: Data-driven mean-field-type
preprint arXiv:2004.08221, 2020. game perspective. Games, 11(4):51, 2020.
[21] Joshua M Epstein, Jon Parker, Derek Cummings, and Ross A [42] FMO Toxvaerd. Equilibrium social distancing. 2020.
Hammond. Coupled contagion dynamics of fear and disease: [43] Zhishuang Wang, Chengyi Xia, Zengqiang Chen, and Guanrong
mathematical and computational explorations. PLoS One, Chen. Epidemic propagation with positive and negative preven-
3(12), 2008. tive information in multiplex networks. IEEE transactions on
[22] Paul Erdős and Alfréd Rényi. On the evolution of random cybernetics, 51(3):1454–1462, 2020.
graphs. Publ. Math. Inst. Hung. Acad. Sci, 5(1):17–60, 1960. [44] Duncan J Watts and Steven H Strogatz. Collective dynamics
[23] Geoffrey P Garnett and Roy M Anderson. Sexually transmitted of ‘small-world’networks. nature, 393(6684):440, 1998.
diseases and sexual behavior: insights from mathematical mod-
els. Journal of Infectious Diseases, 174(Supplement 2):S150–
S161, 1996.
[24] Jana Gevertz, James Greene, Cynthia Hixahuary Sanchez
Tapia, and Eduardo D Sontag. A novel covid-19 epidemiologi-
cal model with explicit susceptible and asymptomatic isolation
compartments reveals unexpected consequences of timing social
distancing. medRxiv, 2020.
[25] Marko Gosak, Moritz UG Kraemer, Heinrich H Nax, Matjaž
Perc, and Bary SR Pradelski. Endogenous social distancing and
its underappreciated impact on the epidemic curve. Scientific
reports, 11(1):1–10, 2021.
[26] Clara Granell, Sergio Gómez, and Alex Arenas. Competing
spreading processes on multiplex networks: awareness and epi-
demics. Physical review E, 90(1):012808, 2014.
[27] Gregory Gutin, Tomohiro Hirano, Sung-Ha Hwang, Philip R
Neary, and Alexis Akira Toda. The effect of social distancing
on the reach of an epidemic in social networks. arXiv preprint
arXiv:2005.03067, 2020.
[28] Ashish R Hota, Tanya Sneh, and Kavish Gupta. Impacts of
game-theoretic activation on epidemic spread over dynamical
13