BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE,

PILANI
HYDERABAD CAMPUS
ME F320, SECOND SEMESTER 2022-2023
PROJECT TITLE & PLAN OF WORK

Particle Swarm Optimisation

Group 8: Chris Alex Mattam

Introduction
Throughout history, nature has provided humanity with a plethora of inspiration
for various inventions and scientific discoveries. One of these natural
phenomena, Swarm Intelligence, in which schools of fish or a flock of birds
travel in certain formations to identify potential food sources mutually
benefiting each other, was proposed to be used to create a mathematical
model to find solutions to complex problems by applying the same theory. A
function collects a particular set of points called ‘particles’, and local maxima/
minima are identified in the vicinity of these particles. This process is reiterated
until a satisfactory or suitable solution for the optimisation problem is found.
Particle Swarm Optimization(PSO)[1] algorithm can be implemented to search
for optimum solutions in a solution space as proposed by Kennedy and
Eberhart in 1995. Since then, much work has been done to improve the
algorithm's efficiency, as detailed in the following sections.

1
 Need for the Study

In a recent paper published[2], a new proposal has been put forward to implement
the power law function to improve the weight value(w) of the particle swarm
algorithm, recognising the fact that the function plays an instrumental role in the
optimisation result of the algorithm.

The particle swarm optimisation algorithm updates the velocity and position by the
following formulas:

Formula: w is the inertia weight, C1 and C2 are the learning factors, r1 and r2 are
random numbers between [0, 1], vid (k ) is the search speed of the i particle at time k
in d dimensions, and xid (k ) is the search speed at time k and i The position of each
particle in the d dimension, pid (k) is the optimal position of the i particle in the d
dimension at time k, and pgd (k) is the optimal position of the entire particle swarm in
the d dimension at time k.

The weights are adjusted such that a larger optimisation speed is achieved in the
initial iterations of the algorithm in order to increase the optimization speed. And in
the subsequent iterations of the optimization, a smaller optimisation speed is
applied in order to improve the search accuracy. This is achieved by utilising the
power law in the weight value w.

2
The above graph[1] compares a linear-based weight to a power law function-based
weight in terms of the number of iterations required to nd the desired solution in
the solution space.

In another paper, researchers have tried to inculcate a dimensional learning

strategy[3], in order to overcome many of the shortcomings of PSO such as the
phenomena of “oscillation” and “two steps forward, one step back”. The paper
adopts a dimensional learning strategy (DLS) for discovering and integrating the
promising information of the population best solution according to the personal best
experience of each particle, has been proposed.

Hence the paper adopts a two-swarm learning PSO based on di erent learning
strategies. In the words of the paper, “One of the subpopulations constructs the
learning exemplars by DLS to guide the local search of the particles, and the other
subpopulation constructs the learning exemplars by the comprehensive learning
strategy to guide the global search”.

Prior to the paper, Shi and Eberhart [4] rst proposed the inertia weight and Clerc
and Kennedy [5] proposed the contraction coe cient. However, these constraints
only balanced the exploration and exploitation abilities by controlling the
convergence tendency of the algorithm. However, they do not address the issue of
oscillation and two steps forward, one step back.

The phenomenon of oscillation can best be explained in the words of the paper
itself, “When the personal best position and population best position are located in
two opposite directions of the current position, after particle i approaches, it will
move closer to in the next generation because the di erence is large. A particle can
always wander between the personal best position and population best position,
which will cause “oscillation” and reduce the search e ciency of the algorithm”

To take into account the useful information from the particles, the paper proposes a
‘dimensional learning strategy’.

The following example from the paper should exemplify the case, the immediate
image shows the denotations.

3
fi
ffi
ff
ffi
fi
ff
 The nal updated velocity equation is denoted in the following:

4
fi
 Objectives

Considering the theory and its applications in my degree in Mechanical

Engineering. I can already identify various fields of interest where this theory
could greatly improve- material sciences, nuclear energy production, etc.
However, from my personal interests, I would like to apply the theory in the
field of Cooling Systems and Refrigeration for large, high-density data centres.

Data centres are the backbone of the internet, their uptime and smooth
functioning are crucial for the various businesses whose data, websites and
internal systems are hosted in these large computer centres. Within Data
centres as well, many sections of the systems require much more varied
cooling capabilities on the basis of their loads and network traffic they are
dealing with. Hence, These computer systems have to be maintained at
specific temperature and humidity conditions as subjected to their
manufacturer’s specifications, making cooling systems of incredible
importance to their safe operations.

With the current age of IoT and AI model training, etc. Servers with much
greater capabilities and load are required in the Industry. Conventional CRAC
units which provide cooling to an entire data centre or server room can no
longer meet the cooling requirements of these high-density data centres.
Hence cooling needs to be provided on a rack to rack basis, which needs to
be optimised by predicting on the load that each rack of servers is dealing with
in the whole data centre, and then accordingly engaging the respective
CRAC(Computer Room Air Conditioner) units to specifically provide cooling to
their respective server racks.

The PSO algorithm can be implemented in an ML model to predict the cooling

requirements for individual server rack units and hence provide cooling to
where its required while cutting costs to provide the same cooling.

5
 References
[1]Kennedy J,Eberhert R.(1995)Particle swarm optimization.IEEE International
Conference on Neural Networks,pp.1942-1948.

[2]Shouya Wang et al 2021 J. Phys.: Conf. Ser. 1827 012151

[3]Xu, G., Cui, Q., Shi, X., Ge, H., Zhan, Z., Lee, H. P., Liang, Y., Tai, R., &
Wu, C. (2019). Particle swarm optimization based on dimensional learning
strategy. Swarm and Evolutionary Computation, 45, 33-51. https://doi.org/
10.1016/j.swevo.2018.12.009

[4]Y. Shi, R. Eberhart

A modified particle swarm optimizer

Evolutionary Computation Proceedings, 1998. IEEE World Congress on

Computational Intelligence., the 1998 IEEE International Conference on
(1998), pp. 69-73

[5]M. Clerc, J. Kennedy

The particle swarm-explosion, stability, and convergence in a

multidimensional complex space
IEEE Trans. Evol. Comput., 6 (1) (2002), pp. 58-73