You are on page 1of 4

Simulation Modelling of Group Dynamics

Chandraneil Basa
31746985
University of Southampton, UK
cb1e20@soton.ac.uk

Abstract By comparing a metric provided in Equation 3 with another


random student, students adjust their approach for the next
The research paper delves at two different approaches semester. If the selected student’s performance measure is
to tackling a simulation modelling problem about how stu- higher than the student’s current measure, the student imi-
dents choose and modify techniques over certain time pe- tates that student’s method.
riod, as well as an approach to potential extension to the
discussion. The very first approach examines into using π =m−a∗s (3)
differential equations to predict changes in student dynam-
where, m = the student’s grade
ics. The second strategy looks into using agent-based mod-
a = Initially selected parameter
elling. The outcomes of both methods are reviewed and
S = S=H or S=L, Depending on the strategy
cross-referenced.
The following are the primary questions to be answered
for this problem:
1. Introduction
Evolutionary game theoretic models were used to exam-
• What would the demographic distribution be in the
ine how populations of biological sentient creatures evolved
long-term run?
by adapting and pursuing diverse strategies, making evo-
lutionary game theory an important component of biology. • How long does it take to reach stable equilibrium?
What distinguishes evolutionary game theory from classical
game theory is that strategies adapt, and it also attempts to • How much would the following variables affect the
determine which strategies are the most efficient. The sub- outcome:
ject addressed in this study is an evolutionary game theory
– Demographics of the population
problem in which an initial student population attempts to
complete assignments in clusters of n students. Individuals – Group size (n)
must elect between two tactics at the start of each course: – Cost of effort (a)
”H” hard worker or ”L” lazy worker.The following is how – Types of workers’ contributions to joint effort (H,
students’ grades are calculated based on their total collabo- and L)
rative effort:
The remainder of the report is organized in the following
e=h∗H +l∗L (1) manner: Section 2 provides demonstration about how to de-
where, e: Total Collective Effort scribe this problem using differential equations and how to
h: Group’s number of hard workers solve it by both analytically and mathematically. The im-
l: Total number of freeloaders in the group plementation of an agent-based simulation that replicates
H: Effort expended as a hard worker the constraints and interactions depicted in the definition of
L: Effort expended as a freeloader. the problem is specified in Section 3. Section 4 presents
By dividing the entire group collaborative effort by the n a modification with an possible extension to the problem
number of participants in the classroom, all students in the and investigates how this modification will impact the solu-
group receive the same grade. tions. The discoveries of the generated simulated scenarios
and the obstacles of applying them are discussed in Section
m = e/n (2) 5.

1
2. Methods be used to explain this. As a result, for n = 2 and γ= 3, the
following matrix describes the distributions of techniques
2.1. DIFFERENTIAL EQUATIONS MOD- used by students in each group type:
ELLING
   
Let M = H, L, to be the set of strategies that students PH(H,H) PH(L,L) PH(H,L) 2/2 0/2 1/2
=
can select from each course. We define xi where i ∈ M , to PL(H,H) PL(L,L) PL(H,L) 0/2 2/2 1/2
be the proportion of students following strategy i. We can (6)
say that the rate of change of population xi with respect to The population state is expressed as the ratio of strategies
time t equals the outflow of students converting from strat- at a given moment in time, where x ∈ ∆m , ∆m = xH , xL ,
egy Xi plus the inflow of students converting from the other where XH and XL are the distributions of students who
strategy to strategy xi . So to model this problem we need follow strategy H and L, respectively. A crucial phrase to
two differential equations, one for each strategy. The first understand is group state, which is defined as z ∈ ∆γ where
simple equation to describe this is: zg represents the fraction of groups of type g. A matching
rule function, f : ∆m → ∆γ , is defined to link the popula-
∂xi tion state with the group state at a given moment.
= −Inf low + Outf low (4)
∂t X
Equation 4 behaves similarly to the generic replication xi = Pig ∗ fg (x) (7)
equation described in 5 for an evolutionary demographic g∈G

scenario, in which a participant’s payoff has an effect on


Finally, wig (x) represents the likelihood that a student of
how the populace will resemble in the next semester, and
type i belongs to the type g group.
this is fully implemented by replicator dynamics based only P ∗f (x)
on moderated payoff. wig (x) = ig xig

∂xi 2.1.2 Group Payoffs


= xi (πi − π̄) (5)
∂t
The matrix A ∈ Rn∗γ describes the payment that students
where, πi = i Strategy’s payoff receive for being in a certain group, where Aig is the pay-
π̄= The population’s average payout out for a student adopting strategy i in group type g. For
The key distinction observed is the payoff’s fluctuating instance:
over time depending on the membership configuration of AH(H,H) = m − 12 = ne − 12 = nh − 12 = 22 − 21 = 12
the groups. This section will examine and see how we might We may improve the chances of that student being in that
include it into the basic replicator equation. group by the total payout for a student adopting method i in
a group type g and get the following:
2.1.1 Groups Formation X
Fi (x) = ig (x) ∗ Aig (8)
Individuals are randomly randomised into groups of size w
n at the start of each course, as previously outlined. The
payout that each student can obtain is determined by the Assuming students are placed in clusters at random, the
composition of students in the groups, and the aggregate uniformly random pairing criteria for n = 2 is as follows:
performance of the population determines how the percent-
age of students changes in the following course, according f(2,0) (XH , XL ) = Xh2 (9)
to the replicator equation. The method used to calculate
f(1,1) (XH , XL ) = 2 ∗ Xh ∗ XL (10)
the changing group fitness over time was inspired by the
work of Jensen et al, in his research Evolutionary games f(0,2) (XH , XL ) = XL2 (11)
and matching rules.
The set of n students in a group is specified as a group
2.1.3 Model Dynamics
type, where gi is a member of group g who is following
strategy i which is either H or L. Depending on the number Since the structure of groups and student payoffs change
of iterations, a student types can evolve in a group of size n and evolve with time, several of the equations developed in
with two methods is defined as γ= (n+2−1)!
n! . For example, earlier phases must be updated. The population configura-
we have γ value as for n = 2, This implies we have three tion at time t is specified as x t, which is then translated
different sorts of groupings: ”H,H,” ”L,L,” and ”H,L,” re- to the distribution of group types using the matching rules
spectively. Pig = gi /n denotes the percentage of students of function f (x )t. Equation 8 becomes as follows under these
class i in group type g. The matrix P ∈Rmγ might likewise conditions:

2
X grows when a strategy payout exceeds the average payoff
FiA,f (xt ) = wig (x) ∗ Aig (12) of the two populations, and decreases when the converse is
g∈G true. In accordance w ith the earliest circumstances stated
The average return on invested payback period for both in the original problem, lethargic workers students dom-
techniques is as follows: inated the population after a reasonably small number of
runs which was approximately around 40 iterations for all
A,f
2
X values of group size n ¿ 2. Figure 2 depicts this behaviour.
F (xt ) = xti ∗ FiA,f (xt ) (13) In the fourty fourth run, 74 percent of the population were
i=1 sluggish employees, while just 26% were dedicated work-
When the payoffs of the student groupings are entered ers. The proportions of the population fluctuated for a long
into the replicator Equation 5, the following two expres- time in group size n = 2, but after running it more than
sions explain the rate of fluctuations of the hard and lazy 25,000 times, it ultimately obtained the same outcome as
worker populations over time: previously, where lethargic workers dominated the popula-
tion. I suppose the explanation for this is because groups
∂xH A,f A,f of two with two lazy employees have a measure of zero,
= xtH (FH (xt ) − F (xt )) (14)
∂t and if the population is made up of such group formations,
lazy workers will always have a lower measure than other
∂xL A,f hard workers. This will only change if a specific amount of
= xtL (FLA,f (xt ) − F (xt )) (15)
∂t group types comprising hard and lazy employees are em-
These equations work for any group size value n number, ployed to map the population, causing the lazy workers to
but the matching function f must be adjusted to account for have a value greater than zero.
the change in students’ constitution with in the group. When the value was growing towards 1, the cost of ef-
fort (a) with group size n ¿ 2 had no effect. When the value
3. AGENT-BASED MODELLING falls below a particular threshold, (a 0.24), the proportions
reverse, and hard workers take over the population. Modi-
Modelling the problem was considerably easier to ex- fying the effort of being a hard or lazy worker (H,L), with
ecute using an agent-based approach. It was built from group size n > 2, lazy workers always ended up taking the
the initial scratch concept using the Python programming whole population as long as L is smaller than H. The pro-
software. To construct the simulation of this situation, two portions are inverted when L ¿ H, and hard workers take
classes were required. The first class (Agent) is associated over the population. When H and L have the same cost, the
with the student and contains data which include the agent population will remain the same and will not change.
strategies (hard or lazy worker), the student’s grade, and fi-
nal payoff performance. Switching techniques, as well as 4. Problem Expansion
establishing and measuring the student’s performance, re-
quired many approaches. The second class (Group) kept Whenever a simulation is performed for a long stretch of
track of which agents were assigned to which groups. The time period, lethargic employees will dominate the popula-
approaches involved estimating the collective effort and tion, as can be seen in Agent-based modelling solution. This
then determining the appropriate performance metric for means that students will never be capable of passing a cur-
each agent.The following processes were implemented in riculum again in the future. The alteration recommended in
each simulated virtual environment runs: first, agents were this study is that students will remember their previous run’s
selected at random to be assigned groups based on the group mark which they scored, and if it was not a passing mark
size n. Then every group’s payoff exertion was evaluated, (below 0.5), they will switch strategies 50% of the time.
and each student’s grade and performance were recorded. Similarly to the prior challenge, group sizes of 2 behaved
Afterwards when agents conducted a random piece-wise differently than those with the values greater than 2. For n
pairing interaction with several other agents in which they ¿ 2, this solution resulted in a more evenly distributed per-
compared their performance evaluation measurements and centage of hard and lazy employees throughout time, with
only tried switching strategies with the probability of 50% a trend to have more hard workers in the end.
if the resulting payoff of theirs was inferior than the other Figure 4 demonstrates how the population of sluggish
comparing agent. To begin with the experimentation, The employees becomes extinct when the group size is n = 2.
population was initially divided in 50%, among which Half Both simulation computations were repeated 1000 folds
were dedicated workers, while the other half were freeload- with a demographic community of 10,000 students and in-
ers. deed with the identical baseline parameters as the original
The findings acquired using this method reflected the challenge. Only the agent-based model continued to ob-
replicator equation’s behaviour, in which a population tain updates with an the extension. It was unclear how

3
this new behaviour might be reflected on its performance
in the differential Equations 14,and 15. In the agent-based
method, however, this was quite simple to implement be-
cause the state of the previous run can be simply preserved
and verified when constructing new groups in the following
run.This new modified extension is more in accordance with
how students actually behave when faced with the choice
of working hard or slacking behind, as most students will
base their decision on their previous grades, hoping to avoid
making the same mistakes by switching to a hard worker
strategy and securing a higher grade.

5. Conclusion
In conclusion, it was clear that the agent-based method
was far easier to model and execute than working out the
model’s differential equations. Both procedures are ob-
served to provide considerably similar results, especially
when the simulation is conducted for a long enough period.
Modifying and expanding the agent-based technique is eas-
ier, and it allows you to consider additional circumstances,
which allows you to better represent real-world behaviours.

6. Future Scope
References
[1] R. F. Antonio Torralba and W. T. Freeman, “Computer
science and artificial intelligence laboratory technical
report,” CSAIL, 2007.

[2] G. Amato and F. Falchi, “Knn based image classifica-


tion relying on local feature similarity,” 01 2010, pp.
101–108.
[3] H. Le, T. Le, S. Tran, H. Tran, and N. Thuy, “Image
classification using support vector machine and artifi-
cial neural network,” 05 2012.
[4] A. Karim and R. Sameer, “Image classification using
bag of visual words (bovw),” pp. 76–82, 12 2018.
[5] T. D. Science.

You might also like