Deep Learning in Gaming and Animations

Deep Learning in Gaming
and Animations
Explainable AI (XAI) for Engineering Applications
Series Editors: Aditya Khamparia and Deepak Gupta
Explainable AI (XAI) has developed as a subfield of artificial intelligence, focusing

on exposing complex AI models to humans in a systematic and interpretable manner.
This area explores, discusses the steps and models involved in making intelligent
decisions. This series will cover the working behavior and explains the ability of
powerful algorithms such as neural networks, ensemble methods including random
forests, and other similar algorithms to sacrifice transparency and explainability for
power, performance, and accuracy in different engineering applications relates to
the real world. Aimed at graduate students, academic researchers, and professionals,
the proposed series will focus key topics including XAI techniques for engineering
applications, Explainable AI for deep neural network predictions, XAI for machine
learning predictions, XAI-driven recommendation systems for automobile and man-
ufacturing industries, and XAI for autonomous vehicles.
Deep Learning in Gaming and Animations: Principles and Applications

Vikas Chaudhary, Moolchand Sharma, Prerna Sharma, and Deevyankar Agarwal
For more information about this series, please visit: https://www.routledge.com/

Explainable-AI-XAI-for-Engineering-Applications/book-series/CRCEAIFEA
Deep Learning in Gaming
and Animations
Principles and Applications
Edited by
Vikas Chaudhary, Moolchand Sharma, Prerna Sharma,
and Deevyankar Agarwal
First edition published 2022
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
and by CRC Press

2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
© 2022 selection and editorial matter, Vikas Chaudhary, Moolchand Sharma, Prerna Sharma, and
Deevyankar Agarwal; individual chapters, the contributors
First edition published by CRC Press 2022
CRC Press is an imprint of Taylor & Francis Group, LLC
Reasonable efforts have been made to publish reliable data and information, but the author and pub-
lisher cannot assume responsibility for the validity of all materials or the consequences of their use.
The authors and publishers have attempted to trace the copyright holders of all material reproduced in
this publication and apologize to copyright holders if permission to publish in this form has not been
obtained. If any copyright material has not been acknowledged, please write and let us know so we may
rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or here-
after invented, including photocopying, microfilming, and recording, or in any information storage or
retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access www.copyright.com
or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-
750-8400. For works that are not available on CCC please contact [email protected]
Trademark Notice: Product or corporate names may be trademarks or registered trademarks and are
used only for identification and explanation without intent to infringe.
Library of Congress Cataloging‑in‑Publication Data
Names: Chaudhary, Vikas, editor. | Sharma, Moolchand, editor. | Sharma, Prerna,

editor. | Agarwal, Deevyankar, editor.
Title: Deep learning in gaming and animations : principles and applications / edited by
Vikas Chaudhary, Moolchand Sharma, Prerna Sharma and Deevyankar Agarwal.
Description: Boca Raton : CRC Press, 2022. | Series: Explainable ai (xai) for
engineering applications | Includes bibliographical references and index.
Identifiers: LCCN 2021030248 (print) | LCCN 2021030249 (ebook) |
ISBN 9781032126098 (hardback) | ISBN 9781003231530 (ebook)
Subjects: LCSH: Object-oriented programming (Computer science) | Artificial
intelligence.
Classification: LCC QA76.64 .D437 2022 (print) | LCC QA76.64 (ebook) |
DDC 005.1/17–dc23
LC record available at https://lccn.loc.gov/2021030248
LC ebook record available at https://lccn.loc.gov/2021030249
ISBN: 9781032126098 (hbk)

ISBN: 9781032139302 (pbk)
ISBN: 9781003231530 (ebk)
DOI: 10.1201/9781003231530
Typeset in Times LT Std

by KnowledgeWorks Global Ltd.
Dedication
Dr. Vikas Chaudhary would like to dedicate this book to his father
Sh. Rajendra Singh and his mother Smt. Santosh for their constant
support and motivation; and his family members, including his
wife Ms. Amita Panwar, his daughter Astha Chaudhary, and
his sons Shivansh Chaudhary and Anmol Chaudhary. I would
also like to give my special thanks to the publisher and my
other co-editors for having faith in my abilities. Before all and
after all, the main thanks should be to the Almighty God.
Mr. Moolchand Sharma would like to dedicate this book to his
father Sh. Naresh Kumar Sharma and his mother Smt. Rambati
Sharma, for their constant support and motivation, and his family
members, including his wife Ms. Pratibha Sharma and his son
Dhairya Sharma. I would also like to give my special thanks to the
publisher and my other co-editors for having faith in my abilities.
Ms. Prerna Sharma would like to dedicate this book to
her father Mr. Vipin Sharma, mother Ms. Suman Sharma,
husband Mr. Parminder Mann and her family for their
support and constant motivation. Specially dedicated to
her beloved sons Pratyaksh, Kairav & Kevin Mann
Mr. Deevyankar Agarwal would like to dedicate this book
to his father Sh. Anil Kumar Agarwal and his mother Smt.
Sunita Agarwal, and his wife Ms. Aparna Agarwal, and his
son Jai Agarwal for their constant support and motivation.
I would also like to give my special thanks to the publisher
and my other co-editors for having faith in my abilities.
Contents
List of Figures and Tables..........................................................................................ix
Preface.................................................................................................................... xiii
Editors....................................................................................................................... xv
Contributors��xvii
Chapter 1 Checkers-AI..........................................................................................1
Priyanshi Gupta, Vividha and Preeti Nagrath
Chapter 2 The Future of Automatically Generated Animation with AI.............. 19

Preety Khatri
Chapter 3 Artificial Intelligence as Futuristic Approach for Narrative Gaming..... 37

Toka Haroun, Vikas Rao Naidu, and Aparna Agarwal
Chapter 4 Review on Using Artificial Intelligence Related Deep Learning

Techniques in Gaming and Recent Networks..................................... 65
Mujahid Tabassum, Sundresan Perumal, Hadi Nabipour
Afrouzi, Saad Bin Abdul Kashem, and Waqar Hassan
Chapter 5 A Review on Deep Learning Algorithms for Image Processing

in Gaming and Animations................................................................. 91
Sugandha Chakraverti, Ashish Kumar Chakraverti, Piyush
Bhushan Singh, and Rakesh Ranjan
Chapter 6 Artificial Intelligence in Games........................................................ 103

Abhisht Joshi, Moolchand Sharma, and Jafar Al Zubi
Chapter 7 A Framework for Estimation of Generative Models Through

an Adversarial Process for Production of Animated Gaming
Characters.......................................................................................... 123
Saad Bin Khalid and Bramah Hazela
Chapter 8 Generative Adversarial Networks Based PCG for Games................ 137

Nimisha Mittal, Priyanjali Pratap Singh, and Prerna Sharma
Index....................................................................................................................... 157
vii
List of Figures and Tables
FIGURES
Figure 1.1 An example of payoff matrix for a zero-sum two-player game........... 4
Figure 1.2 Code implementation of the heuristic function...................................5
Figure 1.3 Binary Search tree example.................................................................6
Figure 1.4 A Minimax tree...................................................................................7
Figure 1.5 Algorithmic representation of a Minimax tree example.....................7
Figure 1.6 Alpha-Beta pruning illustration...........................................................8
Figure 1.7 Alpha-Beta pruning illustration...........................................................9
Figure 1.8 Alpha-Beta pruning illustration......................................................... 10
Figure 1.9 Flowchart depicting the implementation algorithm.......................... 11
Figure 1.10 Before single capture of a piece......................................................... 12
Figure 1.11 After single capture of a piece........................................................... 13
Figure 1.12 Before multiple capture of pieces...................................................... 13
Figure 1.13 After multiple capture of pieces........................................................ 13
Figure 1.14 U
pgradation of a piece into king on reaching the end side of
the opponent�� 14
Figure 1.15 Displaying the result of the game...................................................... 14
Figure 1.16 Use case flowchart............................................................................. 15
Figure 1.17 System environment diagram............................................................ 16
Figure 2.1 Setting up keyframes for timing the bouncing balls.........................20
Figure 2.2 Classification of autonomous agents.................................................. 22
Figure 2.3 A topology of agent........................................................................... 23
Figure 2.4 ive emerging technologies that will change the world in next
F
five years.��26
Figure 2.5 Attractive opportunities in the 3D animation market........................34
Figure 3.1 The MDA framework........................................................................ 49
Figure 3.2 Proposed functionality diagram........................................................ 52
ix
x List of Figures and Tables
Figure 3.3 Player character class in UE4............................................................ 53

Figure 3.4 First person game view...................................................................... 53
Figure 3.5 Inspect item prompt........................................................................... 54
Figure 3.6 Player picks item................................................................................ 54
Figure 3.7 Item inspection.................................................................................. 55
Figure 3.8 User interface..................................................................................... 55
Figure 3.9 Main menu......................................................................................... 56
Figure 3.10 Journal blueprint................................................................................ 56
Figure 3.11 Player hunger and stamina................................................................. 57
Figure 3.12 Blueprint of AI based functionality................................................... 57
Figure 3.13 Overview of the scene....................................................................... 58
Figure 3.14 User interface of game design platform – Unreal Engine................. 59
Figure 3.15 AI behavior tree for the two stages of Q-learning............................. 61
Figure 4.1 Difference of M2M and IoT.............................................................. 70
Figure 4.2 Bluetooth Communication................................................................. 71
Figure 4.3 IoT operating layer model.................................................................. 73
Figure 4.4 Top 10 IoT involved industries........................................................... 73
Figure 4.5 IoT application and industries............................................................ 76
Figure 4.6 First order model framework............................................................. 77
Figure 4.7 Wireless network growth................................................................... 79
Figure 4.8 Overview of AI..................................................................................80
Figure 4.9 Machine learning algorithm.............................................................. 81
Figure 4.10 Decision tree example........................................................................ 82
Figure 4.11 Data Analytics................................................................................... 85
Figure 5.1 Computer “vision” via data...............................................................96
Figure 5.2 Classification of object detection, here, the self-driving car.............96
Figure 5.3 xamples of different transfer functions applied to an image:
E
(a) original, (b) logarithmic, (c) histogram equalization, and
(d) negative�� 98
Figure 6.1 Comparison of gaming industry size with other industries............. 105
Figure 6.2 Making AI model to adapt the physics laws.................................... 109
List of Figures and Tables xi
Figure 6.3 Unique experience of VR................................................................ 109

Figure 6.4 AR used in Pokémon Go................................................................. 110
Figure 6.5 AR vs. VR........................................................................................ 110
Figure 6.6 Extend reality with AR, MR, and VR............................................. 111
Figure 6.7 Evolution of game characters........................................................... 111
Figure 6.8 Evolution of graphics in Need for Speed......................................... 112
Figure 6.9 Reinforcement learning................................................................... 113
Figure 6.10 AlphaGo Zero beating Ke Jei.......................................................... 113
Figure 6.11 Minimax tree for Tic-Tac-Toe.......................................................... 116
Figure 6.12 The finite state machine for a real time strategy worker NPC........ 117
Figure 6.13 Implementation of an A* graph to find the best way....................... 117
Figure 6.14 Dota 2 hero AI for long-horizon learning........................................ 119
Figure 7.1 Graph of loss of Generator and Discriminator in CGAN................ 135
Figure 8.1 Evolution of GAN in video games................................................... 143
TABLES
Table 2.1
Agent Properties Based on the Coordination, Planning, and
Cooperative Ability��24
Table 2.2
Summary of AR Frameworks............................................................28
Table 2.3
A Summary of Game Engines........................................................... 29
Table 3.1
Levels of Fear (See Ntokos, 2018)..................................................... 47
Table 4.1
Overview of ML Algorithm used in IoT Networks........................... 83
Table 4.2
Overview of IoT Applications, Protocols and Algorithm.................. 86
Table 8.1
Main Contributions of the Above Explored Algorithms................. 147
Table 8.2 Popular Datasets in Video Game Content Generation.................... 150
Preface
We are delighted to launch our book, “Deep Learning in Gaming and Animations:
Principles and Applications.” Artificial intelligence has been a growing resource
for video games for years now. Most video games—whether they are racing games,
shooting games, or strategy games—have various elements controlled by AI, such
as the enemy bots or neutral characters. Even the ambiguous characters that do not
seem to be doing much are programmed to add more depth to the game and give you
clues about your next steps. Today’s modern world is currently under a significant
influence on innovative technologies such as artificial intelligence, deep learning,
machine learning, and IoT. This book aims to present the various approaches, tech-
niques, and applications that are available in the field of gaming and animations.
It is a valuable source of knowledge for researchers, engineers, practitioners, and
graduate and doctoral students working in the same field. It will also be helpful for
faculty members of graduate schools and universities. Around 25 full-length chap-
ters were received. Amongst these manuscripts, eight chapters have been included
in this volume. All the chapters submitted were peer-reviewed by at least two inde-
pendent reviewers, provided with a detailed review proforma. The comments from
the reviewers were communicated to the authors, who incorporated the suggestions
in their revised manuscripts. The recommendations from two reviewers were taken
into consideration while selecting chapters for inclusion in this volume. The exhaus-
tiveness of the review process is evident, given the large number of articles received
addressing a wide range of research areas. The stringent review process ensured that
each published chapter met the rigorous academic and scientific standards.
We would also like to thank the authors of the published chapters for adhering to
the schedule and incorporating the review comments. We wish to extend my heart-
felt acknowledgment to the authors, peer-reviewers, committee members, and pro-
duction staff whose diligent work shaped this volume. We especially want to thank
our dedicated team of peer-reviewers who volunteered for the arduous and tedious
step of quality checking and critique on the submitted chapters.
Vikas Chaudhary,
Moolchand Sharma,
Prerna Sharma,
Deevyankar Agarwal,
Editors,
October 5, 2021
xiii
Editors
Vikas Chaudhary is a professor in the Computer Science & Engineering depart-
ment at JIMS Engineering Management Technical Campus, Greater Noida. He
has 18 years of teaching and research experience. He obtained a Doctorate from
the National Institute of Technology, Kurukshetra, India, in Machine Learning/
Unsupervised Learning. He has published various research papers in the International
Journals of Springer, Elsevier, Taylor & Francis. Also, he has published various
papers in IEEE International Conferences and national conferences. He is a reviewer
of Springer Journal as well as of many IEEE conferences. He has written a book on
Cryptography & Network Security. His research area is machine learning, artificial
neural networks.
Moolchand Sharma is an Assistant Professor in the Department of Computer Science

and Engineering at Maharaja Agrasen Institute of Technology, GGSIPU Delhi. He
has published scientific research publications in reputed International Journals and
Conferences, including SCI indexed and Scopus indexed Journals such as Cognitive
Systems Research (Elsevier), Physical Communication(Elsevier), Intelligent Decision
Technologies: An International Journal, Cyber-Physical Systems (Taylor & Francis
Group), International Journal of Image & Graphics (World Scientific), International
Journal of Innovative Computing and Applications (Inderscience) & Innovative
Computing and Communication Journal (Scientific Peer-reviewed Journal). He has
authored/co-authored chapters with such international publishers as Elsevier, Wiley,
and De Gruyter. He has authored/edited three books with CRC Press and Bhavya
Books. His research area includes artificial intelligence, nature-inspired comput-
ing, security in cloud computing, machine learning, and search engine optimiza-
tion. He is associated with various professional bodies like ISTE, IAENG, ICSES,
UACEE, Internet Society, etc. He possesses teaching experience of more than nine
years. He is the co-convener of the ICICC-2018, ICICC-2019& ICICC-2020 springer
conference series and also the co-convener of ICCRDA-2020 Scopus Indexed IOP
Material Science & Engineering conference series. He is also the reviewer of many
reputed journals like Springer, Elsevier, IEEE, Wiley, Taylor & Francis Group and
World Scientific Journal, and many springer conferences. He is currently a doctoral
researcher at DCR University of Science & Technology, Haryana. He completed his
Post Graduate in 2012 from SRM University, NCR Campus, Ghaziabad and gradu-
ated in 2010 from KNGD Modi Engg. College, GBTU.
Prerna Sharma is an Assistant Professor in the Department of Computer Science

and Engineering at Maharaja Agrasen Institute of Technology, GGSIPU Delhi. She
has authored/co-authored SCI-indexed journal and Scopus indexed journal articles in
high ranked and prestigious journals such as Journal of Supercomputing (Springer),
Cognitive Systems Research(Elsevier), Expert Systems (Wiley) International Journal
of Innovative Computing and Applications (Inderscience). She has also authored book
xv
xvi Editors
chapters with International level publishers (Wiley and Elsevier). She has extensively
worked on Computational Intelligence. Her area of interest includes artificial intel-
ligence, machine learning, nature-inspired computing, soft computing, and cloud
computing. She is associated with various professional bodies like IAENG, ICSES,
UACEE, Internet Society, etc. She has a rich academic background and teaching
experience of 8 years. She is a doctoral researcher at Delhi Technological University
(DTU), Delhi. She completed her Post Graduate in 2011 from USIT, GGSIPU, and
Graduate in 2009 from GPMCE, GGSIPU.
Deevyankar Agarwal works as a lecturer in the Engineering Department–EEE

Section (Computer Engineering) at the University of Technology & Applied Sciences
(Public University), Muscat, Oman. He has 20 years of teaching and research
experience. He is currently a doctoral researcher at the University of Valladolid,
Spain. He has published various research papers in the International Journals of
Springer, Elsevier, Taylor & Francis. Also, he has published various papers in IEEE
International Conferences and national conferences. He is a reviewer for Springer
Journals as well as of many IEEE conferences. He is a Doctoral researcher from
University of Valladolid, Spain. He completed his M.TECH (Computer Science
& Eng.) in 2008 from GGSIP University, M.Sc (Computer Science) in 2001 from
Agra University (State University), Agra & B.Sc (Computer Science) in 1998 M.J.P
Rohilkhand University (State University), Bareilly.
Contributors
Hadi Nabipour Afrouzi Waqar Hassan
Faculty of Engineering, Computing and Research and Finance,
Science, Barani Institute of Sciences, Pakistan &
Swinburne University of Technology, JV Arid Agriculture University,
Sarawak, Malaysia Pakistan
Aparna Agarwal Bramah Hazela

Department of Engineering, Amity School of Engineering &
Middle East College (OMAN), Technology Lucknow,
Muscat Amity University,
Uttar Pradesh, India
Jafar Al Zubi
Computer Engineering department, Abhisht Joshi
Al-Balqa Applied University, School of Department of Information Technology,
Engineering, MAIT,
Jordan Delhi, India
Ashish Kumar Chakraverti Saad Bin Abdul Kashem

Department of Computer Science and Qatar Armed Forces – Academic
Engineering, Bridge Program,
PSIT College of Engineering, Qatar Foundation,
Kanpur UP, India Qatar
Sugandha Chakraverti Saad Bin Khalid

Department of Computer Science and Amity School of Engineering &
Engineering, Technology Lucknow,
PSIT College of Engineering, Amity University,
Kanpur UP, India Uttar Pradesh, India
Priyanshi Gupta Preety Khatri

Department of Computer Science and Information Technology,
Engineering Department, Institute of Management Studies,
Bharati Vidyapeeth’s College of Noida, India
Engineering,
New Delhi Preeti Nagrath
Department of Computer Science and
Toka Haroun Engineering Department,
Department of Computing, Department Bharati Vidyapeeth’s College of
of Engineering, Engineering,
Middle East College (OMAN) New Delhi
xvii
xviii Contributors
Vikas Rao Naidu Prerna Sharma

Department of Engineering, Department of Computer Science &
Middle East College (OMAN), Engineering,
Muscat Maharaja Agrasen Institute of
Technology,
Nimisha Mittalella New Delhi, India
Department of Computer Science &
Engineering, Piyush Bhushan Singh
Maharaja Agrasen Institute of Department of Computer Science and
Technology, Engineering,
New Delhi, India PSIT College of Engineering,
Kanpur UP, India
Sundresan Perumal
Faculty of Science and Technology, Priyanjali Pratap Singh
Universiti Sains Islam, Department of Computer Science &
Malaysia Engineering,
Maharaja Agrasen Institute of
Rakesh Ranjan Technology,
Department of Computer Science and New Delhi, India
Engineering,
PSIT College of Engineering, Mujahid Tabassum
Kanpur UP, India Faculty of Science and Technology,
Universiti Sains Islam,
Moolchand Sharma Malaysia
Department of Computer Science &
Engineering, Vividha
Maharaja Agrasen Institute of Department of Computer Science and
Technology, Engineering Department,
New Delhi, India Bharati Vidyapeeth’s College of
Engineering,
New Delhi
1 American Checkers
Checkers-AI
Game Using Game

Theory and Artificial
Intelligence Algorithms
Priyanshi Gupta, Vividha and Preeti Nagrath
CONTENTS
1.1 Introduction....................................................................................................... 1
1.2 Related Work..................................................................................................... 2
1.3 Methodology......................................................................................................3
1.3.1 Game Theory.........................................................................................3
1.3.2 Zero-Sum Game.................................................................................... 3
1.3.3 Heuristic Function.................................................................................4
1.3.4 Search Tree............................................................................................6
1.3.5 Minimax Approach...............................................................................6
1.3.6 Alpha-Beta Pruning...............................................................................8
1.3.7 Minimax vs Alpha-Beta Pruning.......................................................... 8
1.4 Implementation..................................................................................................9
1.4.1 Game Algorithm....................................................................................9
1.4.2 Graphical User Interface..................................................................... 12
1.5 Utility and Application.................................................................................... 14
1.5.1 System Environment............................................................................ 15
1.6 Conclusion and Future Scope.......................................................................... 16
1.6.1 Conclusion........................................................................................... 16
1.6.2 Future Scope........................................................................................ 16
References................................................................................................................. 17
1.1 INTRODUCTION
For years, the application of artificial intelligence (AI) in video games has been
growing steadily. The bulk of video games have multiple controllable traits, such as
racing games, fighting games, etc. [1]. AI is one of the computer sciences fields that
can be used to construct a variety of intelligent games, whether board games, video
DOI: 10.1201/9781003231530-1 1
2 Deep Learning in Gaming and Animations
games, or educational games, that would respond to a human being and that a com-
puter machine would not be able to react to them [2].
The suggested game checkers are one of the most played board games that
involve thought and tactics to win the game, or if it was not played properly and used
the methods that would allow the player to win, it would be easy to lose. Checkers
is a series of two-player strategy board games that include diagonal movements
of game pieces [3]. American checkers are the most common form of checker. In
the popular “checkerboard” pattern, with 12 parts per hand, it is played on an 8 ×
8 board of light and dark squares. It is played by two players, as with all types of
checkers, taking turns on opposite sides of the board. Black, red, or white are the
typical pieces. By hopping over them, enemy bits are captured. The interactions
between the pieces are very less in checkers as compared to chess due to different
types of pieces. In AI, the heuristic solution definition is based on awareness or
study of the reasoning of people. In order to solve a problem, a heuristic approach
is faster than traditional methods. The goal of the heuristic is to find a solution
that is adequate within a suitable amount of time to solve the problem. A way of
evaluating the quest for the target is the heuristic approach. This chapter is divided
into several sections. The first section gives an introduction of checkers games and
its probable computations along with rules. The second section is the literature
survey which tells about the existing computer programs to solve checkers and its
history associated with it. The third section is the methodology section which gives
details about the methods and fields used in the project. The fourth section is the
implementation section in which the applicative view of the algorithm is briefly
explained along with implementation of graphical user interface (GUI) in the game
and its snapshots. The fifth section is the utility and application section which tells
the basic uses, features, and the system environment of the game developed. The
last section is the conclusion section in which results and findings along with future
scope of the project is mentioned.
1.2 RELATED WORK

Christopher Strachey developed the first computer software for the American check-
ers in February 1951, at the National Physical Laboratory (NPL) in London [4].
It ran on NPLs Pilot Automatic Computing Engine (Pilot ACE) for the first
time on 30 July 1951. IBM researcher Arthur Samuel wrote the second machine
program in 1956. In the research process, we discovered that a search tree of board
positions which can be viewed from the current state was the key feature of the
software. Samuel applied Alpha-Beta pruning [5], although he had only a very
small amount of machine memory available. A score mechanism was established at
any given time, depending on the location of the board, instead of looking for each
direction before it came to the end of the game. On each hand, this feature tried
to determine the probability of winning on every turn. Many factors were taken
into consideration such as the number of pieces each player has or the number of
kings and the position of the pieces to the kings were taken into consideration. The
program chose its move based on the Minimax method [6]. In 1989, a team at the
University of Alberta headed by Jonathan Schaeffer published the most powerful
Checkers-AI 3
software named Chinook. It was the first computer machine to win the title of
world champion in the game for checkers against a human. The Chinook software
program contains an opening book, an archive for starting moves from games
played by grandmasters with a deep search algorithm which is an efficient motion
evaluation feature; and an eight or less piece end-game database for all positions.
Because many previous inventions in the gaming industry have been done using
the search tree and heuristic function technique, we chose to use Alpha-Beta prun-
ing algorithm as a major emphasis for this article. Different search algorithms
were developed before the Alpha-Beta pruning aimed to minimize the tree search.
In other games, such as chess, checkers, tic-tac-toe, Isola, and many more, this
algorithm has been used.
1.3 METHODOLOGY
The different techniques studied and discussed to understand and implement the
checkers game are as follows.
1.3.1 Game Theory

While checkers game is a “solved game,” which means a computer has to be pro-
grammed in a way that will never lose a game. The method of decision-making has
to involve decision-makers, actors, environmental variables, priorities, policies, and
parameters. The efficiency of the decision-making process in competitive markets
relies on assessing all environmental considerations and analyzing them according to
objectives. The aim of policy makers is to identify optimum solutions for competing
priorities. Game theory is a mathematics-based methodology in which player tactics
are reciprocally tested by considering environmental effects [7]. Games theory pro-
vides methods for analyzing situations in which interdependent decisions are taken
by individuals, called players. This interdependence enables both players to predict
the other player’s possible choice and strategies to formulate one’s strategic move.
As a solution to a game, player’s preferences and the outcomes that may result from
these choices are represented. The science of games has been extended to a wide
range of circumstances in which players’ decisions interfere to influence the result.
The theory both supplements and goes beyond the classical theory of probability in
emphasizing the strategic dimensions of decision-making, or aspects regulated by
the players rather than by pure chance.
1.3.2 Zero-Sum Game

A zero-sum game is a mathematical definition of a circumstance in which the benefit
or loss in profit of each player is exactly accounted for by the losses or gains in utility
of the other players in game theory. The key difference between a general game and
a zero-sum game is that the agents have opposite utilities in a zero-sum game as
opposed to the independent utilities in a general game or a non-zero-sum game. If
the participants’ net gains are added and the combined losses are subtracted, they
will add up to zero. Non-zero-sum, on the other hand, describes a situation where the
FIGURE 1.1 An example of payoff matrix for a zero-sum two-player game.
net gains and losses of the involved parties can be less than or greater than zero. A zero-
sum game is sometimes called a purely competitive game, whereas non-zero-sum
games can be competitive or non-competitive. Zero-sum games are most commonly
solved using the Minimax theorem, which is directly connected to the duality of linear
programming [7], or Nash equilibrium. Many persons have a cognitive predisposition,
known as zero-sum tendency, to see circumstances as zero-sum. The easy representa-
tion for a match is a payout matrix. Then the choices are disclosed and the total points
of each player are influenced according to the payout for those choices. Consider the
two-player zero-sum game illustrated in Figure 1.1. The order of play is as follows: one
of two secret acts, 1 or 2, is chosen by the first player (red); one of three secret deeds,
A, B, or C, is selected by the second player (blue), unaware of the first player’s deci-
sion. Red will choose Action 2 and Blue will choose Action 1. Red wins 20 points by
allocating prizes, while Blue loses 20 points. Rather than agreeing on a single action
to be taken, the two players allocate probabilities to their respective actions, and use a
random process that selects an action for them according to these probabilities.
Each player tests the odds in order to decrease the estimated overall point loss,
regardless of the approach of the opponent. This leads to a linear problem of pro-
gramming with each player’s best solutions. For all zero-sum games with two teams,
this Minimax approach will possibly determine optimal strategies. The final chance
of Red preferring Action 1 is 4/7, and it turned out to be 3/7 for Action 2. For Blue,
the odds are 0, 4/7, and 3/7 respectively for acts A, B, and C. Red is the winner with
an average of 20/7 points per session.
1.3.3 Heuristic Function
Heuristic, AI, and mathematical optimization in computer science is a concept intended
to solve a problem faster when the classical search is too long. For optimality, complete-
ness, accuracy, or speed accuracy, this is done by trading. This could be named, in a
way, a shortcut. A heuristic function is a function that, at each branching step, ranks
alternatives in search algorithms based on the available knowledge to decide which
branch to pursue, through which, the exact solution [8] may be approximated. There
is a different sense of heuristic functions used in single-agent search; they return an
approximation of the distance of a given state from a target. The heuristic function in
this project is used to make the search in the algorithm easier. The heuristic, like the
full-space search algorithm, initially tries all possibilities at each point. However, if the
present probability is potentially worse than the better alternative discovered, the search
Checkers-AI 5
can be stopped at any moment. For the heuristic function proposed in this chapter, we
theorize that the perfect weight for the king piece can be found, but assigning the king
a weight much greater than a standard chip will make the player give the king too much
priority, leading to situations where the player potentially damages their odds of win-
ning by losing a king’s regular chips. Thus, we have created a balance between the value
of a pawn and the value of the king by creating a function called PieceSumDifference.
Pawn’s Value = 1
King’s Value = 3
As we can see in Figure 1.2, the function is described and the number of kings and
pawns each player has in the current scenario is calculated in order to assign a value
and evaluate the position at which the current player is standing.
FIGURE 1.2 Code implementation of the heuristic function.

The player that maximizes is the red player and the player that minimizes is the
yellow player. Therefore, the heuristic value returned is the difference between the
values of the player maximizing and the player minimizing.
1.3.4 Search Tree

A search tree is a type of tree data structure which is used within a set to locate
specific keys. In order for a search tree to work, the key must be greater for each
node than any subtree key on the left and smaller than any subtree key on the right
[9]. The key idea is to simulate, using self-play, several thousands of random games
from the present place. In a search tree, new positions are added and each tree node
includes a value that predicts who is going to win from that position. There are many
search-tree data structures, some of which often allow elements to be easily added
and removed, and operations must then retain the balance of the tree. The Binary
Search tree depicted in Figure 1.3 is implemented in this project.
The Binary Search tree is a node-based binary tree data structure in which every
left child is smaller than node and right child is greater than the node.
1.3.5 Minimax Approach
In AI and game theory, Minimax is a decision rule used to minimize the future
loss while optimizing the expected advantage. Before applying this strategy, the two
assumptions made for the game are the fact that the human player plays optimally
and tries to win. The other is that the game should be a game that is purely strategic.
The Minimax algorithm is the modified version of the backward-induction algo-
rithm. It can also be considered to maximize the minimum gain called maximin.
In two cases, players make simultaneous moves and opposite moves, each player
minimizing the other player’s overall potential gain. A player therefore maximizes
his own minimum advantage, becoming a zero-sum game. Minimax and Maximin,
however, are not identical. In non-zero-sum game conditions, Maximin can be used.
FIGURE 1.3 Binary Search tree example.

Checkers-AI 7
FIGURE 1.4 A Minimax tree.
In this project, the Minimax [10] algorithm is used as a recursive algorithm to

determine the moves made by the two players, usually black and white. The two
players are competing strategically and usually they have the complete information
about the game. In checkers game, there is an integer line and a number line, and
any integer which is positive favors white, and any integer which is negative, favors
black. So positive infinity would be a win for white and negative infinity would be
a loss for black. There is a value assigned to every position of the game which is
measured by the heuristic feature defined. The players make moves to maximize the
minimum position value based on the assumption that the opponent will follow the
predicted moves. If it’s white’s turn to pass, white provides each of their legal moves
with a value. The algorithm conducts what we in computer science call a “depth-first
search,” as shown in Figure 1.4.
This means that it explores the whole length of the tree vertically, until it reaches
the terminal node and then works its way back up. Secondarily, or among other
sibling nodes, the algorithm moves horizontally. This is in contrast to a “breadth-
first search,” which does exactly the reverse. It mostly passes between fellow sibling
nodes horizontally, scans at a time for a whole “level,” and makes its way down the
tree secondarily. The algorithmic representation is shown in Figure 1.5 [11].
FIGURE 1.5 Algorithmic representation of a Minimax tree example.

1.3.6 Alpha-Beta Pruning
Minimax approach amounts to exhaustive search of solution space. In a realistic game
such as checkers, the search space is large. It is therefore more efficient to heuristically
reduce the search space in case of Minimax solutions. Alpha-Beta pruning algorithm
provides a mechanism to decrease the number of counts of the nodes thus reducing
the search space in the Minimax tree [12]. Apart from checkers, it is also used in other
two-player games. Further, the search is stopped once it encounters a branch where
at least a single value has been found that confirms that the branch would be worse
in comparison to earlier payoffs. Therefore, these moves can be avoided in the game
to reduce unnecessary computation. Alpha-Beta pruning does the same. It discards
the branches of the search tree without affecting any of the decisions to return the
same result in spite of pruning [13]. Apart from reducing the search time, it is pos-
sible to have a deeper search allowing a greater depth of subtree to be scanned. Two
values are used in the algorithm, Alpha and Beta. The highest (lowest) score that the
optimizing player has is stored in the Alpha variable. Beta value keeps the full score
that the player who minimizes is guaranteed to win. The original Alpha value shall
be equal to −infinity and Beta value is equal to +infinity. This gap gets smaller as the
search continues. When the Beta value gets smaller than the Alpha value, it means
that it is possible to limit the search beyond the current node.
In Figure 1.6, Alpha-Beta pruning cuts the gray subtrees while moving from left
to right and it is not appropriate to investigate them at every step since the category
of subtrees yields the value as a whole of an equal subtree or worse leaving the end
result unaffected. The implementation of this algorithm is demonstrated as shown in
Figures 1.7 and 1.8.
1.3.7 Minimax vs Alpha-Beta Pruning

As we know, the Minimax algorithm works with a game tree where the two agents
are Max and Min. Max tries to maximize the gain of the game and Min tries to
minimize the loss of the game. They are mutually recursive, i.e., Max calls Min, and
Min calls Max, and this process takes place all the way to the bottom of the search
tree. The Alpha-Beta pruning is introduced as an opportunity to remove branches of
the search tree [14], thus, shortening the time to find the “most promising” subtree,
FIGURE 1.6 Alpha-Beta pruning illustration.

Checkers-AI 9
and allowing a greater search to be done at the same time [15]. It is a member of the
algorithm branch and binding class, unlike its predecessor.
Usually, the subtrees are temporarily dominated by either a first player advantage
during Alpha-Beta pruning in the case when many first player movements are good,
but all second player moves are calculated in order to find a comeback, or vice versa.
If the ordering of the move is incorrect, this gain will repeatedly switch sides during
the search, it will lead to inefficiency each time. Any step closer to the current posi-
tion decreases exponentially with the number of searched positions, significant effort
is worth focusing on sorting early moves.
1.4 IMPLEMENTATION
The algorithmic implementation and the GUI are described in the following sections.
1.4.1 Game Algorithm

The game algorithm is implemented using JavaScript. Figure 1.9 describes the
implementation in the form of a flowchart. The procedure is as the player completes
its turn; a search algorithm is called which allows the program to look ahead at the
possible future positions before deciding what move it wants to make in that current
position. It’s white turn to move now. In every move, there are only two possible
moves to choose. We can visualize these moves as two separate branches at the end
of which are two new positions, of course, it is black’s turn to move now. We continue
expanding these moves till either we reach the end of the game or we decide to stop
because going deeper would take too much time. Either way, at the end of the tree we
now have to perform the static evaluation on these final positions. The static evalua-
tion means try to estimate how good the position is on one side without making any
more moves. Large values would favor white and small values would favor black. For
this reason, white constantly tries to maximize the evaluation hence known as the
Maximizing player. In addition, black is always trying to minimize the evaluation,
hence known as the Minimizing player. We start by evaluating the positions on the
bottom left. In the previous position, it was white’s turn to move, and since white will
always choose the value that leads to the maximum evaluation, we assign the value to
Checkers-AI 11
FIGURE 1.9 Flowchart depicting the implementation algorithm.

the node accordingly and complete the evaluation of that node using the right branch
as well. Now, black will try to minimize the evaluation function, so we assign the
position the lower value by comparing. In addition, we go up the tree hence return-
ing the maximum-minimum gain and getting the higher probable position for white
to move. This is where pruning comes into the picture. It would take a lot of time
to go down all the branches to get the best value. Without exploring all the nodes
through the branch, it tries to get the least or the max value from the already evalu-
ated position. This would result in the player knowing that he already has a better
option available and that he won’t have to go down the other branch. These checks
are made through Alpha-Beta parameters. This observation concludes that we don’t
have to waste any computation in evaluating the final position. Hence, we’ve pruned
that position from the tree.
1.4.2 Graphical User Interface

The GUI is built using HyperText Markup Language (HTML) and Cascading Style
Sheets (CSS). The primary reason for building the algorithm in JavaScript was
due to the fact that it is one of the primary building blocks of web development.
Therefore, the whole game is implemented as a web page. The board game is given
a real checkerboard like feature to simulate real-time game playing making it visu-
ally more attractive as shown in Figures 1.10–1.13. Besides the board, the front end
contains rules and tutorials so that the user with no previous knowledge can also
play the game without any difficulty. A little brief about the history of checkers
board is also mentioned so that the user can know about the origin of the check-
ers’ game. Lastly information about chinook is also given. The rule-based features
such as the upgradation of a piece into king is also depicted in a visually appealing
form by depicting a crown on the piece (see Figures 1.14 and 1.15).
FIGURE 1.10 Before single capture of a piece.

Checkers-AI 13
FIGURE 1.11 After single capture of a piece.
FIGURE 1.12 Before multiple capture of pieces.
FIGURE 1.13 After multiple capture of pieces.

FIGURE 1.14 Upgradation of a piece into king on reaching the end side of the opponent.
FIGURE 1.15 Displaying the result of the game.
1.5 UTILITY AND APPLICATION

The project is an implementation of a strategic board game checkers. It can be used
to play checkers in an interactive environment. Figure 1.16 displays the use case
flowchart of the project. The game’s algorithm is developed using JavaScript which
is one of the core technologies of web development. To provide a real-time playing
simulation, the GUI of the game is built using HTML and CSS. The game is more
accessible and user-friendly as it is built as a website. The frontend of the web page
contains the instructions and rules of the game so that the user can play the game
even without having any previous knowledge.
Checkers-AI 15
FIGURE 1.16 Use case flowchart.
The game is a one-player checkers game where the user can play with the AI and
test their skills in strategic and logical gaming. Some benefits of playing checkers
game are stated below:
• It develops concentration skills and promotes confident decision-making.

• It reduces stress and serves as a fun way to overcome boredom.
• It boosts problem-solving and pre-mathematics skills.
1.5.1 System Environment

By giving input, gamers will interact with the device (selecting the pieces and drag-
ging them to the available tile). System produces the script with such inputs, if some
change takes place (if the value is changed), the object is sent to render to display
FIGURE 1.17 System environment diagram.
the content (a character can change its place). Figure 1.17 depicts the system environ-
ment diagram.
1.6 CONCLUSION AND FUTURE SCOPE

1.6.1 Conclusion
• The use of Minimax without optimization is not effective, although it can
be a good solution with them.
• Alpha-Beta pruning improves exponentially with respect to Minimax as the
range increases.
• Simple algorithms as a random player have no chance at depth greater than
1 against Alpha-Beta.
• There are several other solutions to zero-sum play, but Minimax with opti-
mization using heuristic function and Alpha-Beta pruning tends to be a
successful one.
• The introduction of different features will significantly boost heuristics.
• The depth of the game tree affects significantly the com-player quality.
1.6.2 Future Scope

• Different levels based on the depth of the tree can be introduced in the game
such as easy, medium, and hard.
• Design a much more efficient heuristic function.
• Improve the GUI.
Checkers-AI 17
• Built a data management system to store the details of the user such as
name, number of wins and losses.
• Add new features such as two-player games, different types of checkers.
• Using the same algorithms to make more two-player games such as tic-tac-
toe and chess.
REFERENCES
1. Idzham, K., Khalishah, M., Steven, Y., Aminuddin, M.S., Syawani, H., Zain, A.M.,
and Yusoff, Y., Study of Artificial Intelligence into Checkers Game using HTML and
JavaScript. IOP Conf. Ser.: Mater. Sci. Eng., 864, 012091, 2020. doi:10.1088/1757-899X/
864/1/012091
2. Alkharusi, S., Checkers Research Paper Based on AI (2), 1, 7, 2020. https://www.
researchgate.net/publication/339337169_checkers_research_paper_based_on_AI_2
3. Masters, J., Draughts, Checkers – Online Guide. www.tradgames.org.uk
4. Sutton R., Samuel’s Checkers Player. In: Sammut C. and Webb G.I. (eds.), Encyclopedia
of Machine Learning. Boston, MA: Springer, 2011. https://doi.org/10.1007/978-0-387-
30164-8_740
5. Arthur, S., Some Studies in Machine Learning Using the Game of Checkers. IBM J.
Res. Dev., 3 (3), 210–229, 1959. CiteSeerX 10.1.1.368.2254. doi:10.1147/rd.33.0210
6. Myerson, R., Game Theory: Analysis of Conflict. Cambridge, MA; London, England:
Harvard University Press, p. 1, 1991. doi: 10.2307/j.ctvjsf522.15
7. Pearl, J., Heuristics: Intelligent Search Strategies for Computer Problem Solving.
Reading, MA: Addison-Wesley Pub. Co., Inc., p. 3, 1984.OSTI 5127296.
8. Black, P. and Vreda, P., “Search Tree”. Dictionary of Algorithms and Data Structures
Figure 2. 2005. https://levelup.gitconnected.com/an-into-to-binary-search-trees-
432f94d180da
9. Binmore, K., Playing for Real: A Text on Game Theory. New York, 2012. Oxford
Scholarship Online: http://dx.doi.org/10.1093/acprof:osobl/9780199924530.001.0001
10. Kuo, Jonathan C.T., Artificial Intelligence at Play — Connect Four (Mini-max
Algorithm Explained), 2020. https://medium.com/analytics-vidhya/artificial-intelli-
gence-at-play-connect-four-minimax-algorithm-explained-3b5fc32e4a4f
11. Russell, S.J. and Peter, N., Artificial Intelligence: A Modern Approach (2nd ed.). Upper
Saddle River, NJ: Prentice Hall, 2003, pp. 163–171. ISBN 0-13-790395-2 Figure 3: By
Nuno Nogueira (Nmnoguera) http://en.wikipedia.org/wiki/Image:Minimax.svg, cre-
ated in Inkscape by author, CC BY-SA 2.5, https://commons.wikimedia.org/w/index.
php?curid=2276653d=2276653
12. Russell, S.J. and Norvig, P., Artificial Intelligence: A Modern Approach (3rd ed.).
Upper Saddle River, NJ: Prentice Hall, 2010.
13. McCarthy, J., Human Level AI Is Harder Than It Seemed in 1955. 2005. Retrieved
2006-12-20.
14. Edwards, D.J. and Hart, T.P., The Alpha–Beta Heuristic (AIM-030). RLE and MIT
Computation Center: Massachusetts Institute of Technology, 1961. hdl:1721.1/6098.
15. Knuth, D.E. and Moore, R.W., An Analysis of Alpha-Beta Pruning. Art. Intel., 6 (4),
293–326, 1975. doi:10.1016/0004-3702(75)90019-3 S2CID 7894372
2 The Future of
Automatically Generated
Animation with AI
Preety Khatri
CONTENTS
2.1 Introduction..................................................................................................... 19
2.2 Ai’s Role in Animation.................................................................................... 21
2.2.1 How AI Replaces Animation............................................................... 21
2.2.2 Various Agents in AI........................................................................... 22
2.3 AI Latest Techniques in Animation................................................................ 23
2.3.1 Latest AI Technology..........................................................................25
2.3.2 AR Technology....................................................................................26
2.3.3 VR Technology....................................................................................28
2.4 The Traditional and Modern Animation......................................................... 30
2.4.1 Traditional Animation......................................................................... 31
2.4.2 Stop-Motion......................................................................................... 32
2.4.3 Modern Animation.............................................................................. 33
2.5 Future Aspects of Animation with AI............................................................. 33
2.6 Conclusion.......................................................................................................34
References................................................................................................................. 35
2.1 INTRODUCTION
Computer animation refers to the addition of something new and improved to the
conventional method of animation. A technique that uses a sequence of images in
frames to create the illusion of movement. It is a technique that creates the illusion
of movement by viewing images on a screen and capturing a sequence of individual
states of an active scene using a recording device. An animation can be described as
a movie made up of a series of rendered images. Various features such as file size,
file format, frames per second (fps), compression, output size, and so on have been
used to monitor the quality of the images in the picture or frames [1, 2]. The most
popular form of animation is keyframing, in which the animation is generated at
several points throughout the animation. At the same time, the computer creates all
of the transition frames between the two keys. Changing the position, rotation, and
scale of objects are all examples of animation techniques.
Setting the length of the animation in frames and fps is very important when
animating. We can place the keys inside the frames with the aid of keyframes [3, 4].
DOI: 10.1201/9781003231530-2 19
To alter, transfer, rotate, or resize an object, the key is placed at the beginning and
end of the desired path. Consider this scenario: if an object moves from point X to
point Y in 4 seconds and you have 50 fps, put four keys 100 frames apart. Real-time
animation allows you to give different objects physical properties. To control them,
it employs a variety of controls and features. In the x, y, and z planes, you can cre-
ate different objects, alter masses, create actors, control friction, and control forces.
Architectural walk-throughs can be constructed using a variety of real-time anima-
tion and three-dimensional (3D) games.
There are two approaches to taking computer animation and evolution into
account. The first approach involves combining traditional animation techniques
with the use of a computer. The second approach focuses on simulation models
based on physics and dynamics laws [5]. Consider the following scenario: Traditional
methods allow us to build 3D characters with enhanced gestures, while simulation
methods are used to model human actions accurately. Take bouncing balls, for exam-
ple, where the motion of the balls can be enhanced by adding squash and stretch [3].
When an object is squashed, it expands and flattens out, indicating that it is made
of a pliable and soft material. Various traditional animators use this tool. This does
not provide a practical simulation, but it gives the audience an idea. A bouncing ball
motion can also be fully simulated by computer using mechanics such as quantum
conservation and Newton’s laws, as shown in Figure 2.1 [2].
In the multimedia industry, autonomous virtual actors and real-time animation
are critical because immersive use of functionality is a direct advantage. Each tele-
vision and film producer will be eager to develop new programs and features to
participate interactively [6]. Real-time animation capabilities are needed by editors,
publishers, and writers of interactive TV programs, CD-ROMs, and CD-Is that are
becoming increasingly interactive [3].
Computer animation has been regarded as a cutting-edge medium for visual
effects (VFX) and advertising in films for many years. The rapid development of
powerful super workstations in recent years has given rise to new areas such as video
gaming, virtual reality (VR), and multimedia. The use of real-time animation and
FIGURE 2.1 Setting up keyframes for timing the bouncing balls.

The Future of Automatically Generated Animation with AI 21
digital technology has become a significant concern. The audience can only choose
which programs to watch on traditional television [7]. With the latest advances in
multimedia products and digital and interactive television, viewers will communi-
cate with programming. This creation would pave the way for personalized pro-
gramming for each viewer [6].
In animation, we can look at the following developments: computer animation
began with very basic methods derived from conventional animation and keyframes
[3]. There have also been developed several time-consuming rendering methods.
Importing inverse dynamics and kinematics from robotics paves the way for more
advanced simulation methods. Computer animation has traditionally focused on
dynamic simulation and physics methods, particularly in collision detection and
deformations [5].
With the advent of super workstations and VR applications, brute force approaches
such as radioscopy are likely to resurface. In the future, real-time complex animation
systems will be built using simulation and VR devices [8]. Autonomous actors and
actual actors with motion captured by sensors are calculated by the machine using
real-time behavioral simulation with complex physics-based interactions with the
environment. Long-distance partners could be assigned to these dynamic scenes.
2.2 AI’S ROLE IN ANIMATION

Computers have left an indelible impression on art in a variety of ways. They may
recognize their structures by recognizing pictures or photographs. On the other
hand, computers affect animated videos or motion graphics because they have simi-
larities with other pieces of similar work. The automatic engine gives its integrated
software and base instructions [9].
A network that has been used for artificial intelligence (AI) is the phase feature
neural. As a result, AI is being used to make humans seem more realistic in animated
films and video games. However, it is improving at a faster rate than most animators.
The impressiveness of the AI takeover of this new invention can be attributed to the
animation’s accuracy. However, there has been much backlash against individual cre-
ative expression. When speaking with real animators about the case, for example [10].
Art is a type of human speech that is subjective. As a result, when a machine takes
away the fundamental concept of creative expression, society feels challenged. This
non-human entity is challenging to have in a space that was not designed for it, and
if AI continues to advance, humanity could be in grave danger. Even if none of it
reverberates with the public, they would prefer to animate and build [11]. This is why
the inventors have proposed that AI be limited to the function of association, where
it will alter and improve processes while allowing humans to influence creative pro-
duction and design.
2.2.1 How AI Replaces Animation

Nowadays, technology has replaced many occupations, including creatives, result-
ing in a dynamic job market. The animators are the creatives who are obsessed with
their work. Their talent appears to be on the verge of a significant change, as with
most workers, which are both boring and menial. To engage in such activities, most
people need some motivation. Inventors use technology to make their lives simpler
in any way. There are already computers and appliances that perform our every-
day tasks, such as washing machines and microwaves. There are several automation
tools available that can perform most of the tasks that animators can perform with
AI. As a result, we can conclude that AI is now the most dominant industry and that
technology is performing better than humans [12].
2.2.2 Various Agents in AI
One of the most critical areas in AI is agent, which is primarily concerned with
intelligent systems [13]. In terms of AI, an agent is a machine that exists in a given
environment and makes its own decisions. It uses sensors to perceive the environ-
ment and actuators to function in the environment. Intelligent agents, robotics, and
other technologies are examples.
As seen in Figure 2.2, which shows how autonomous agents are classified. The
term “real-life agents” refers to living animals such as mammals, reptiles, fish, and
birds. Robotic agents of the mechanical kind, on the other hand, are also agents from
the AI perspective—for example, the robot rovers used for NASA’s Mars Rover mis-
sions. Computer agents are agents that only exist in a virtual or web-based world [10].
Agents that process human language, agents that gather various information,
agents that are informative, agents that are intelligent and learn, and agents that are
programmed for a specific purpose such as entertainment, e.g., in games, special
VFX, 3D animation, and so on are all examples of software agents. Software agents
FIGURE 2.2 Classification of autonomous agents.

FIGURE 2.3 A topology of agent.
and artificial life agents are examples of computational agents. Computer agents
are further classified as quest agents, learning agents, preparation agents, conversa-
tional, intelligent agents, and so on [13].
Figure 2.3 shows a topology of intelligent agents, collaborative agents, collab-
orative learning agents, and interface agents based on their learning, coordination,
autonomy, and other characteristics. Agents function at a higher level than symbols
and require high-level communications.
Agents use a variety of bots, including chatterbots, which are used for web chat-
ting. Annoybots, which are used to disrupt chat rooms and newsrooms. Spambots,
which are used to generate junk mails after collecting web email addresses. Mail
bots, which are used to manage and filter email, and spider bots, which are used to
scrape content from the internet.
As a result of the above, we may conclude that agents are those who act or exert
force. Anything that produces or is capable of producing an effect is referred to as an
agent. An agent is allowed to act for or in the role of another person, such as a gov-
ernment delegate, emissary, or official who engages in undercover activities; it may
also be a business representative, as shown in Table 2.1 [14]. Based on the coordina-
tion, planning, and cooperative ability, the agent’s properties can be classified as:
2.3 AI LATEST TECHNIQUES IN ANIMATION

Adobe has introduced the most recent techniques, which include 3D object movement,
morphing, kinetics, and the addition of voice recording. It took some time for this
skill to build. It restricts what a human animator can accomplish [15]. Most of these
mediums are being reformed and expanded, which has enormous potential. In the
case of network traffic, about 60% of it originates from internet-connected comput-
ers. Animation videos demonstrate the effectiveness and productivity of the animation
industry, which is growing in terms of conversion rate and variety [1]. There are many
ways to understand how AI helps animators. The following are some of the features:
• Better Workflow: AI can help animation studios enhance communication

by improving essential functions: the strategy development and decision-
making process for the AI-assisted project. The essential advantage is that
TABLE 2.1
Agent Properties Based on the Coordination, Planning, and Cooperative
Ability
Agent Property Description

Ability to set goals The agent has a purpose
Autonomy The agent exercises control over its actions and run asynchronously
Temporal continuity The agent is a continuity running process
Social ability The agent can communicate in a complex manner with other agents, including
people, in order to obtain information
Proactivity The agent responds in the best possible way to possible future actions that are
anticipated to happen
Reactivity The agent responds in a timely fashion to changes in the environment and
decides for itself when to act
Versatility The agent can have multiple goals at the same time
Persistency The agent will continue steadfastly in pursuit of any plan
Rationality The agent makes rational, informed decisions
Mobility The agent can transport itself around its environment
managers can make better choices in their workflows and activities as a

result [16].
• The Ideas about Streamlining: Customer support and chatbots are pro-
grammed to work together in the same way that humans do. These digital
portals are appealing in terms of improved communication methods and
changing the way artists think about their projects. The dedication entails
putting concepts and thoughts associated with projects on hold before an
incredible new concept emerges. Not only can it be saved without human
intervention with the aid of the virtual assistant, but it will also be achieved
more effectively due to the artist’s simple mind frame.
• Information Control: Regularly, animators work with a large number
of different types of data. Animators must consider coloring, set design,
motion graphics, and scale charts within a single picture. This info, which
serves as a digital assistant for artists, is now handled by AI technology.
Prioritization of urgent texts, instructions, and the importance of style
accessible to designers so that they can focus on their specific tasks. In
creative settings, this hands-on assistance approach has proven to be highly
beneficial [17].
• Diminishing Workload: Experts agree that AI has significantly reduced
workload, allowing developers to have lighter workloads. Voice-activated
personal assistants like Alexa and Siri are assisting animators in doing their
jobs more effectively. AI is now taking over business activities that are not
part of the creative process [14]. Artists can be more creative and perform
smarter as a result of this. Since AI follows scripted paths, the animator’s
artistic integrity is not jeopardized.
• Patterns Prediction: The chatbots will be designed to have honest conver-

sations using machine learning (ML). With the help of Alexa or Google,
this nonlinear advantage will allow predictions about employee prefer-
ences and workplace trends. The virtual assistant is evolving into an actual
assistant capable of answering complicated questions about production and
design. The animators avoid tedious study activities while keeping their
previous work organized and saved for future reference [8].
AI algorithms will easily do data-driven work without taking up the time that
people might. This new technology can now accomplish what teams of anima-
tors will take weeks to accomplish. AI’s true success is making seamless edits,
nuanced characterization, and significant VFX, in addition to providing a solid
track record. Even though these algorithms are very costly and are used by large
companies such as Disney. The animators are uncertain about their future careers
as a result of this.
To make human lives more straightforward, the tricky part is that not every artist
contributes to this mentality. If not alarming, the intrusion of technology into their
artistic work is frustrating. Filmmakers have not yet captured the entire AI move-
ment where these training devices are present to generate worlds and animations.
This is primarily due to a lack of communication between the filmmaker and the
invention.
2.3.1 Latest AI Technology
As we all know, digital technology is rapidly evolving. We use various new technol-
ogy regularly, such as voice assistant Alexa, Siri, Google Maps, and so on. AI in
animation is currently being developed and used to some degree by animators to
speed up time-consuming tasks. One of the key benefits of using AI to automate
those stages in animation [18] is that you have more time to focus on other activities
when you speed things up.
Companies like Adobe have already developed features for their standard ani-
mate suite that connect up characters’ mouth movements to sound. Moreover, others
are going even further by leveraging AI to create solutions that completely automate
characters’ movements and facial expressions as they speak, based on data collected
through ML, effectively automating this delicate, time-consuming task.
When it comes to cutting-edge technology, ML is one of the AI systems in which
computers are not directly programmed to perform those tasks. Instead, they auto-
matically learn and develop as a result of their experiences. Deep learning is a
form of ML that uses artificial neural networks to make predictions. Unsupervised
learning, supervised learning, and reinforcement learning are examples of ML
algorithms.
The algorithm in unsupervised learning does not use personal data to work on it
without any guidance. It deduces a function from the training data, which consists
of collecting an input object and the desired output in supervised learning. Machines
use reinforcement learning to take appropriate actions to maximize the incentive in
order to find the best option that should be considered.
FIGURE 2.4 Five emerging technologies that will change the world in next five years.
(Source: Forrester Research, Inc., unauthorized production, citation, or distribution prohibited.)
As seen in Figure 2.4, five new technologies will change the world in the next five
years. Based on parameters including systems of engagement technologies, systems
of insight technologies, and supporting technologies [19], various emerging technol-
ogies such as Internet of Things (IoT), virtual and augmented reality (VR and AR),
hybrid wireless, AI/cognitive, intelligent agents, real-time interaction management,
spatial analytics, cloud-native application frameworks, insight platforms, and so on
have been elaborated in this figure.
2.3.2 AR Technology
In AR systems, the consumer wears a head-mounted display (HMD) and uses a
device, which involves many boundaries. The H/w was bulky, and the HMD
interfered with the user’s standard view [20]. It is also susceptible to causing discom-
forts such as dizziness or nausea. As a result, the performance of the AR systems on
that particular hardware necessitated a significant amount of effort and time from
the developers. These drawbacks have been alleviated by the appearance and rapid
evolution of smartphones. There are variously powerful smartphones on the market
based on HMD’s technology. As a result, the consumer does not have to keep the
hardware and their versatility is enhanced. As a result, there is now various smart-
phone AR applications available on the market that digital layer knowledge over the
physical world [21].
We can see the digital data visible with the aid of a smartphone camera and position
it in a specific location. Knowledge may be positioned in a variety of ways, including
by the use of a mobile device. The new smartphones have Global Navigation Satellite
System (GNSS) orientation sensors, localization systems, and other features. The
orientation sensor processes the unprocessed sensor data from the accelerometer to
obtain information. The orientation sensor provides pitch (degrees of rotation about
the x-axis) and roll (degrees of rotation about the y-axis). As a result, the device’s
direction is determined by the pitch and roll values. When a smartphone is posi-
tioned at a specific location with known coordinates and orientation, the digital
information is visualized using the camera. The accuracy of the GNSS positioning
is a disadvantage of this system.
The accuracy of the signal decreases in areas with tall buildings or a large number
of trees, and the signal bounces off. Furthermore, this device is ineffective inside
houses. The use of markers is a different way of displaying digital data. A marker
is a square frame with a light-colored, usually white, center and a dark-colored,
usually black, outer frame [22]. Each marker is made using a unique pattern, mak-
ing them one-of-a-kind. In the marker recognition process, the smartphone camera
gathers and processes images in real-time intending to identify a pattern. The color
detection, known shapes, repeated and geometries patterns image recognition tech-
niques are focused on finding patterns of color detection, known shapes, repeated
and geometries patterns.
When a creator is discovered, the virtual knowledge on the camera overlaps.
Natural feature tracking (NFT) is a technique that uses artifacts or pictures instead
of markers to monitor natural features. As a result, NFT enables users to recognize
and monitor natural features on objects and images. Digital data such as two-dimen-
sional (2D) images, 3D models, video, audio, text, animation, etc. This opens up the
possibility of developing a wide range of applications, such as:
1. In the case of ads, it is often used to promote new goods online.

2. In entertainment and education, AR helps develop cultural learning games
(educational aim games).
3. In medicine, AR applications are being created to help surgeons achieve
a minimally invasive operation with increased dexterity, precision, and
visualization.
It is now much easier to create AR applications. The primary advantage is that

anyone can build applications for similar hardware. This allows for the best app
TABLE 2.2
Summary of AR Frameworks
Location- Marker-
Based Operating System NFT Based
Droid AR x Android ---------- X
AR Toolkit ----------- iOS, Android, Linux, Windows, Mac OS X, Unity 3D x X
Beyond AR x Android --------- ------------
Vuforia ----------- Android, iOS, Unity 3D x X
Mixer ----------- Android & iPhone x X
progress [20] due to many growing software structures. A framework usually pro-
vides some basic features that can be used to build more complex applications. The
most significant advantage of using outlines is that they include general application
structure, they aid developer relationships, and many available resources and librar-
ies can be used for frameworks. There are many platforms available these days for
quickly creating AR applications [21].
Table 2.2 summarizes the existing state of free AR framework functionality.
Many structures are no longer maintained, and as a result, the list is constantly
changing. AR Toolkit and Vuforia are the most user-friendly and comprehensive
tools, but if you want to build a position-based app, you will need to use one of the
other frameworks [2].
2.3.3 VR Technology
Even though AR and VR are closely linked technologies, they all depict peculiar real-
ities. When we talk about AR, we are talking about adding elements to reality, while
VR creates a new reality that is not actual. VR is an artificial world developed with
specialized software. VR displays a 3D image that can be explored interactively using
controls such as a game console, a computer mouse, or sensor-equipped gloves [23].
VR systems include using a HMD, such as glasses, to learn about the virtual
world. The first VR device had poor graphics quality and needed a complicated
HMD. The HMD has advanced dramatically in recent years, and there are now regu-
lating solutions on the market, such as Oculus Rift. Oculus VR [7] is working on a
VR headset similar to the Oculus Rift.
Stereoscopic vision is used in the treatment. Stereoscopic vision is a method of
gathering 3D visual information that gives a picture the illusion of depth. Due to
their separation, the eyes in natural stereo vision generated two images with minor
variations between them. To construct depth perception, the brain processes these
variations. The HMD projects a stereoscopic view. Two images are shown on the
projector, one for each eye. A small controller is used to monitor the interpupillary
distance based on the display.
Individuals and elements vary in their eye separation, resulting in a realistic ste-
reoscopic vision-logic. The HMD function and sensors that monitor the user’s head
movements and change the picture are known as a virtual surround sound system.
For the time being, this gadget will connect your machine to your smartphone. Some
users report experiencing headaches or motion sickness as a result of the immersive
3D vision. To avoid this, it is necessary to adjust the lens to each individual’s vision.
However, prolonged use can cause anxiety.
Imagining VR with an HMD is a little annoying at times. However, creating
a semi-immersive app in which a 3D model is shown on the smartphone monitor
seems promising. Throughout the touch screen, the user will collaborate with it.
Users can load 3D models and collaborate with them using those frameworks. These
frameworks can be used to build simple semi-immersive applications. However, if
you want to work with 3D models for animations, rendering, and other purposes, a
game engine is the best choice.
A software framework is a game engine developed for the production and creation
of video games that can be played on a variety of platforms. Loading, rendering,
object collision detection, animation, physics, inputs, graphical user interface (GUI),
and AI are the engine’s key components. The game engine also has other tools for
creating the actual game, such as terrains, characters, real-world object behaviors,
and so on. The game engine includes all of the resources needed to create a VR app.
Many game engines also support stereoscopic vision [24].
Mind3d, a lightweight 3D framework for Android based on OpenGL ES v1.0/1.1,
is one of the several 3D model frameworks available. This system includes tools for
loading and modifying.m2d, 3ds, and. Obj files. It cannot be sustained at this time.
The disadvantage is that these previously mentioned frameworks are only compat-
ible with Android, making it difficult for beginner programmers to create apps. On
the other hand, the game engine is the best way to build a virtual world and expand
basic functionality.
Unreal Engine 4, CryEngine, and Unity are the most common game engines at
the moment. These game engines are prevalent, and although each has its own set of
advantages and disadvantages, they are all compatible with VR glasses [25]. Unity is
the simplest to use and is compatible with all mobile devices, but the graphics quality
is lower, and real-time simulation is impossible.
Unreal Engine 4 has incredible graphics capabilities, allowing for the development
of hyper-realistic scenes, but it is only compatible with iOS and Android. However,
it is easy to use. Finally, because of the engine’s steep learning curve, CryEngine is
better suited to experienced developers, though the graphics quality is excellent. The
main characteristics of these game engines are summarized in Table 2.3 [25].
TABLE 2.3
A Summary of Game Engines
Mobile Platforms Languages Support Other Platforms

Unity Android, iOS, JavaScript, C#, Boo PC, PlayStation, Xbox, Wii, 3DS,
Windows VR, TV, Web
Unreal Engine 4 Android, iOS C++ PC, PlayStation, Xbox, VR
CryEngine [25] C++, Lua, C# PC, PlayStation, Xbox, VR
2.4 THE TRADITIONAL AND MODERN ANIMATION

The illusion of movement and extrude in a speedy display of a series of still, static
images with slight vast modifications is the technical concept of what generalized
animation is. Since its first appearance in France, animation has come a long way
in a short period. The primary known second of excellent animated film in records
on the Musee Grevin is projecting the first lively excellent animated film in Paris’s
Theatre Optique. Since then, animation has grown into a massive industry with a
diverse range of art styles, techniques, a significant starting point for storytelling,
technological advancements, job opportunities in design and animation, and much
more. The list could go on and on.
We had to go back to the beginnings of how live-action movies first became popu-
lar, and we looked at it in the same way we looked at the beginnings of traditional
art, i.e., cave paintings. Hand-drawn drawings and movie rolls have become the
opposites of traditional art’s cave paintings in animation.
Sketches, drawings, paintings, and practically any handcrafted recognizable piece
created with worldly bodily materials made up of traditional animation. When the
world’s first full-length animated feature film, Snow White and the Seven Dwarfs,
was released in 1937, a few years after Walt Disney’s 1928 Steamboat Willie cartoon,
no one could have predicted how famous, and influential this new type of entertain-
ment would become in the years to come.
When Toy Story was released in 1995, it was the first computer-generated mental
imagery film. Not only did ancient invigorating patterns blend with digital medium
models, but Hollywood produced such a significant flip. Computer-generated imag-
ery (CGI) films, on the other hand, were a massive success with audiences after that.
A significant new investment for big-name studios with big budgets, as well as an
opportunity for animators to try out new mediums. Thousands of people will have
new work and career prospects.
CGI creates animated images using computer graphics that use both dynamic
and static images to generate scenes. Though computer animation consists solely
of moving images, the technical concept of View CGI is another amazing way to
explore new styles of art and storytelling through miles of data on an electric operat-
ing system. The current controversy about whether conventional or digital animation
is superior is only partially true.
It is because many people overlook the significance of the two. It is similar to under-
standing the distinction between video and animation. The physical and virtual dif-
ferences in the room, equipment, and materials are a simple statement [11]. The video
consists of the anatomy of movement into individual images, while animation starts with
individual images that are then shaped together to produce the illusion of continuous.
Traditional animation requires a story that should and will always be valued,
loved, and cherished, as it is the foundation for all forms of animation. It does not
matter whether we are talking about digital, stop-motion, chuckimation, puppetry,
claymation, sand animation, typography, painting on glass, drawing on film, or
something else. Because of all of these various animation styles, we would have
come from making this point legitimate to be divisive, but it is absurd to pick one
side or claim one is better than the other [4].
It is the equivalent of ignoring the roots of animation and film, which are tradi-
tional art and photography. Both are intertwined and would not exist without the
advancement of technology and artistic movements. Computer animation is why so
many people are upset to be protective because it is a different way for thousands of
artists, designers, and engineers to have more freedom to work that is cared for your
likeness, not just that it is a more open and structural workflow.
When it comes down to it, it is all about personal preference. It might also come down
to what is best for the project. People should consider that creating digital animations
opens up a world of possibilities and access that conventional animation cannot match.
For instance, having various controls in any environment where you want to work
physically and artistically is a good example. Only animation studios had the physi-
cal resources to create and screen animated films. Anyone now has unrestricted
access to make an animated film and distribute it as they see fit. Because of all the
possibilities for animating [9], the field is far more available in terms of creativity.
There is no way to animate in a single way or a single form. It is shocking when
someone criticizes it because it is so extensive and accessible.
Whether 2D or 3D is better nowadays has almost nothing to do with technological
properties, but instead with which one can work better. Recently, there have been some
fantastic 2D films, such as Iron Giant and Lilo & Stitch, as well as some terrifying 3D
films, such as Final Fantasy. Spirit, Treasure Planet, and Sinbad failed simply because
they were terrible movies in plot and pacing, not because of the medium in which they
were produced. And some of the most popular 3D films include Shark Tale (2003),
Planes (2013), Norm of the North (2016), and Alpha and Omega (2010). Essentially,
3D is yet another method to aid in the filmmaking process. We use it and want to use
it to be most appropriate for the project, but the medium does not determine the result.
Ideas, pencils, and paper are the best starting points for a film. The way you play your
story has a significant impact on its success. What best suits the project is decided by
finding out the suitable media and resources that will work with the story.
At this point in the history of animation and its technology, having the best of all
worlds, physical and digital, is critical. Technology has simply moved into a more
modern reality. However, it has rarely replaced key elements that capture the heart of
animation, which many people fear when they believe technology is about to make
a significant shift in any field.
Many people are afraid of being replaced one day by a computer or someone who
has the tools to do three jobs for one low price.
To counter this misconception, there will always be new employment that will
most likely replace old ones, yes. However, it will also lead to more creative content
in other fields that defiantly require the physical labor of a human mind. Suppose the
future doesn’t manage to invent a computer with such an advanced artificial mind
to produce fantastic animation without being villains someday. In that case, we will
have something to be concerned about [17].
2.4.1 Traditional Animation

Traditional animation is also known as hand-drawn animation or classic animation.
Each frame is hand-drawn on a physical medium in this animation technique. This
technology was prevalent in the cinema before the advent of computers. A sequence of
sketches on transparent pages is used to create this animation technique. Traditional
animation starts with the creation of a plot. Following the selection of the plot, the
artist creates a storyboard that resembles a comic book [4]. The storyboard depicts
the filming sequences of camera angles and frames. Before the director approves a
scene in the storyboard stage, the animator may need to repeat it several times.
Traditional animators work in batches, drawing one picture or frame at a time. In
conventional animation, the pencil is the primary weapon. Artist creates a drawing
on a sheet of transparent paper that can be inserted into your desk’s connector strip.
The peg bar is a traditional animation method used to hold drawings in place. The
artist creates the character solely with a pencil, which is then shot or scanned and
synced with the necessary soundtracks. Before the animation is sent to the supervi-
sors, the artist will check and develop his work [4].
Each pencil-drawn frame of the animation is transferred to the animation.
Background images are painted and superimposed on cels, cels, or celluloid pages.
Background artists typically paint the sets on which the animated sequence’s action
takes place. Acrylic paint was commonly used to create the backdrop: animators
study models, dolls, puppets, real-life figures for inspiration. The main animator in
an animation studio creates a character’s key or mainframes [4]. Twining is a tech-
nique in which the primary animator draws the main points of the action while the
junior animator completes the intermediate or incomplete frames.
The animator uses keyframes to monitor the movement of the character’s arms,
eyes, and mouth. A supervisor animator, a small group of leading animators, and
many assistant animators work together in a large budget animation production group.
Key animators decide the main action of the character. After that, the accepted pencil
animation graphics are photographed with a black and white animation camera.
Traditional animation includes 2D cell animation and stop-motion animation,
even though both can use digital recording techniques in the end. The process used
to create the animation is what matters most. Cell animation typically includes hand
drawing, hand inking, and hand painting each structure on actual paper and cells. At
the same time, stop-motion involves dealing with physical designs and objects shot
with the camera one frame at a time. You get your simple cartoon movie, like Mulan,
with traditional animation. It is a method of drawing pictures, photographing them,
and then animating them so that the original drawings are slightly different.
2.4.2 Stop-Motion
Stop-motion animation has been a staple of film special effects for almost as long as
films have existed. Stop-motion animation is an enjoyable and straightforward ani-
mation technique. Even if you do not know it, you have probably seen stop-motion
animation in advertisements, music videos, TV shows, and movies. Although it is
popular to think of stop-motion as a single form [12], it is not. Stop-motion tech-
niques can create a wide variety of film forms, not just sound animation. You are on
your way if you combine parallel parts of digital camera, machine, and imagination.
Though computer-generated animation is more flashy, stop-motion animation has its
own rich culture.
2.4.3 Modern Animation
The effect is created electronically using either 2D or 3D models in computer ani-
mation. Virtualization of the traditional 2D animation workspace is ordinary in 2D
computer animation, taking pen and paper into the digital realm to redesign cartoon
animation types and workflows. 3D computer animation workflows usually mix con-
ventional timelines and workflows adapted to work in 3D virtual space [2].
In any case, you are dealing with computer animation if you are animating on
video. Avatar and Antz, for example, contain 3D animation, while Cartoon Network
and Nickelodeon will use 2D for the majority of their cartoon features. Pixar’s ani-
mation style is 3D computer animation. It entails animating with computers in a 3D
world.
Computer graphics and animation are now used in almost every audiovisual
media, from movies to commercials. This is so that objects, environments, and char-
acters can be created that would be impossible or difficult to construct in actual life
animation. It has now become an integral part of our everyday lives. It is used for
movies, education, and a variety of other commercial purposes.
2.5 FUTURE ASPECTS OF ANIMATION WITH AI

Since many experts are scrutinizing ML and AI, it is only logical that their imple-
mentation is explained at a basic level. Only by incorporating AI into unique areas of
everyday life will animators and video makers embrace this new AI era. The usage
of smartphones and the advent of the new millennium after the tumultuous 1990s are
two examples. Engineers, designers, and actors will understand the neural network
and use it in their creative processes as soon as these types of applications appear [19].
The day does not come as soon as we spend too much time in a movie theater
watching a film that AI, performed by robots, conceptualized and animated using
deep learning algorithms. Because of the AI-powered animation automation, it is
reasonable to assume that algorithms, AI software, and robots will begin to domi-
nate the industry, based on their approach to factories and, as a result, customer
service tasks [26].
Some may even consider the Associate in animation advice to be an art form.
Some tasks should no longer be performed manually, and that AI can easily auto-
mate, but there is a greater need for qualified people to train deep learning algo-
rithms. Routine tasks, such as creating a digital character, would encourage creative
artists to spend less time on the time-consuming frame-by-frame method of writing
and instead focus on more exciting things. So that animators do not have to draw
frame by frame, AI automates animation-only tasks. Advanced AI-based algorithms
can automate the rendering of advanced VFX.
As shown in Figure 2.5, which shows opportunities in the 3D animation market.
From the analysis, we can summarize that:
• The global 3D animation market is projected to expand at a compound

annual growth rate (CAGR) of 11.9 percent during the precast era, from
USD 12,010 million in 2017 to USD 21,050 million in 2022 [17].
FIGURE 2.5 Attractive opportunities in the 3D animation market. (Source: Markets and
markets analysis.)
• Increased use of VFX technology and 3D applications in media and games

and 3D imagery mapping and geospatial coordinates are driving market
development.
• Market growth is attributed to the increasing use of VFX technology and
#D application in media and games and 3D imaging in mapping and geo-
spatial coordinates.
• The emergence of VR and AI technologies provides growth opportunities
for the 3D animation market [23].
• Emerging economics in various regions, such as APAC, offer new
opportunities.
2.6 CONCLUSION
The future of automatically generated animation with AI was addressed in this chap-
ter. The effect of AI on animation and how modern animation differs from tradi-
tional animation are discussed in this chapter [8]. This chapter examines whether AI
can replace animators in animation, which is a significant concern these days.
The latest technologies for 3D models were discussed in this chapter, and these
technologies provide users with authentic and enriching experiences when visual-
izing 3D objects. Some users may be bothered and inconvenienced by these apps,
especially those that require the use of VR glasses. As a result, some users can prefer
AR apps or using the mobile screen to view 3D models [27].
As AI technology improves and spreads across the industry, animators, filmmak-
ers, and designers are jumping in and making high-quality animations with fewer
people in less time. AI is also having a positive effect on animation and motion
graphics, pushing the boundaries of what can be done with animation. Disney is
known for using AI to create storyboard animations solely from scripts that include
terms like “turn right” and make the character turn in that direction in the animation.
The first move is to build a photorealistic 3D model. Image-based photogram-

metric techniques may be used to create complex and extensive models. Nonetheless,
the resulting point clouds are so large that the hardware of a cell phone does not
support them. As a result, optimizing 3D models is critical if photorealistic models
for smartphone apps are developed. VR and AI technologies are paving the way for
potential growth in the 3D animation industry [23].
REFERENCES
1. Agrawal, A., Joshua G., and Avi G. eds., Introduction to: “Economics of Artificial
Intelligence.” In Economics of Artificial Intelligence. Toronto: nber.org, 2018. http://
www.nber.org/chapters/c14005.pdf.
2. Funge J., Making Them Behave: Cognitive Models for Computer Animation, Ph.D.
thesis, Department of Computer Science, University of Toronto, 1998.
3. Burtnyk N, and Wein M., Computer-generated Keyframe Animation, J. SMPTE, 80,
149–153, 1971.
4. Lasseter J., Principles of Traditional Animation Applied to 3D Computer Animation,
Proc. SIGGRAPH ‘87, Computer Graphics, 21 (4), 35–44, 1987.
5. Amiguet-Vercher J., Szarowicz A., and Forte P., Synchronized Multi-agent Simulations
for Automated Crowd Scene Simulation, AGENT-1 Workshop Proc., IJCAI, 1, 4–10,
2001.
6. Mayer, R.E., and Moreno, R., Animation as an Aid to Multimedia Learning. Educational
Psych Rev., 14 (1), 87–99, 2002.
7. CSIRO Mathematical and Information Sciences (n.d.), Virtual Reality for Teaching Anatomy
and Surgery, 2004. Retrieved from http://www.siaa.asn.au/docs/CSIROscopestudy.pdf.
8. Boden, M.A., AI: Its Nature and Future. Oxford, New York: Oxford University Press,
2016.
9. Park O.C., and Gittelman S.S., Selective use of Animation and Feedback in Computer-
based Instruction, Educ. Technol. Res. Dev., 40, 27–38, 1992.
10. Fikes R., and Nilsson, N., STRIPS: A New Approach to the Application of Theorem
Proving to Problem-Solving, Artif. Intell., 2, 189–208, 1971.
11. Milheim, W.D., How to Use Animation in Computer-assisted Learning, Br. J. Educ.
Technol., 24 (3), 171–178, 1993.
12. Palmer, S., and Elkerton, J., Animated Demonstrations for Learning Procedural
Computer-based Tasks, Hum. Comput. Interact., 8, 193–216, 1993.
13. Funge J., Tu X., and Terzopoulos D., Cognitive Modeling: Knowledge, Reasoning, and
Planning for Intelligent Characters. Computer Graphics Proceedings: SIGGRAPH 99,
1999.
14. Bass, A.S., Non-Tech Businesses Are Beginning to Use Artificial Intelligence. Financial
Times, 2018.
15. Rieber, L.P., Animation as Feedback in Computer-Based Simulation: Representation
Matters, Educ. Technol. Res. Dev., 44, 5–22, 1996.
16. Long D., The AIPS-98 Planning Competition, AI Mag., 21 (2), pp 13–33, 2000.
17. Reynolds C.W., Computer Animation with Scripts and Actors, Proc. SIGGRAPH’82,
289–296, 1982.
18. Russell S., and Norvig P., Artificial Intelligence, A Modern Approach. London:
Prentice-Hall, 1999.
19. Tversky, B., Morrison, J. B., and Bétrancourt M., Animation: Can it facilitate?, Int. J.
Hum. Comput. Stud., 57, 247–262, 2002.
20. ARTOOLKIT, Open source augmented reality SDK. Artoolkit.org, 2016. [online]
Available from: http://artoolkit.org.
21. Carmigniani, J., Furht, B., An Isetti, M., Ceravolo, P., Damiani, E., and Ivkovic, M.,
Augmented Reality Technologies, Systems, and Applications, Multimedia Tools Appl.,
51 (1), 341–377, 2011.
22. Schnotz, W., Böckheler, J., and Grzondziel, H., Individual and Cooperative Learning
with Interactive Animated Pictures, Eur. J. Psychol. Ed., 14, 245–265, 1999.
23. Beier K.P., Virtual Reality: A Short Introduction, 2004. Retrieved from http://www-vrl.
umich.edu/intro/.
24. Zeltzer D., Towards an Integrated View of 3D Computer Animation, The Visual
Computer, 1 (4), 249–259, 1985.
25. CRYENGINE. Cryengine. 2016. [online] Available from: https://www.cryengine.com/
features.
26. Fikes R., Hart, P., and Nilsson, N., Learning and Executing Generalized Robot Plans,
Art. Intell., 3, 251–288, 1972.
27. Draganov, I.R. and Boumbarov, O.L., Investigating Oculus Rift Virtual Reality Display
Applicability to Medical Assistive System for Motor Disabled Patients. The 8th IEEE
International Conference, 2015.
3 Artificial Intelligence
as Futuristic Approach
for Narrative Gaming
Toka Haroun, Vikas Rao Naidu,
and Aparna Agarwal
CONTENTS
3.1 Introduction..................................................................................................... 37
3.2 Related Works.................................................................................................. 38
3.2.1 AI for Computer Games...................................................................... 38
3.2.2 AI for Adaptive Computer Games....................................................... 41
3.2.3 AI in Video Games: Toward a Unified Framework............................. 41
3.2.4 Narrative in Video Games................................................................... 42
3.2.5 Narrative Game Mechanics................................................................. 43
3.2.6 Interactive Narrative............................................................................44
3.2.7 Adventure Games and Puzzle Design.................................................44
3.2.8 The Horror Genre and Video Games.................................................. 45
3.3 Player Experience............................................................................................ 48
3.4 Methodologies used in Game Development.................................................... 51
3.5 AI Elements in the Proposed Narrative Gaming Model................................. 58
3.6 Q-Algorithm for AI in Gaming.......................................................................60
3.7 Conclusion....................................................................................................... 62
References................................................................................................................. 62
3.1 INTRODUCTION
Artificial intelligence (AI) in video games comprises the systems designed and
developed to create the choices and actions of non-player characters (NPC) within
a game. Video games that are developed currently, deliver an awfully fascinating
ground for testing and researching new ideas for AI. These games integrate diverse
and detailed environments with systems that are developed to allow dynamic, com-
plex, and smart real-time choices. In various video game genres such as strategy,
action, and role-playing games, the NPC are developed to have a rule-based system
with variations according to different scenarios. However, machine learning (ML)
methods are sometimes applied to allow NPC to adapt and learn from their interac-
tion with the player character according to their success or failure. Although ML
can be used to improve the NPC’s overall performance, it is not usually applied in
video games.
DOI: 10.1201/9781003231530-3 37
Social isolation measures due to COVID-19 has affected many people espe-
cially young adults causing loss of a sense of community, negative impacts on
learning and development, and increase in impersonality (Sikali, 2020). Computer
games can offer entertainment and stress relief during the pandemic (Ferguson,
2020).
Narrative and storytelling games can help players in social isolation by offering
an immersive storytelling experience where the player can explore and interact with
the game environment to reveal the plot of the game.
This research is aimed at helping people during social isolation especially young
adults with emotional wellbeing (Anderton Kevin, 2018). Providing the players with
an immersive experience that is unlikely to be experienced in the real world, where
the players need to solve puzzles and explore a thrilling and mysterious environment
from the safety of their homes (Butler, 2016).
This chapter will also contribute to the independent game development commu-
nity by using the recent studies and tools in game development to create an immer-
sive storytelling experience for player that is accessible online.
Acquiescent to gaming objectives of AI, generally seven goals or on target by the
game developers for delivering an enjoyable and thrilling gaming experience to the
players:
1. Unpredictable behavior: The game moves should not be predictable by

the players.
2. Cheating should not be allowed: AI in games must be dynamic devoid of
trickery.
3. Utilizing the game surroundings: AI in the game should make use of the
game surroundings in a smart manner, such that no characteristic of the
environment goes untraceable or unutilized.
4. Imagination: novel solutions should be generated by the game AI for unan-
ticipated game situations.
5. Apparent inferior behavior: AI in games should not reveal evident sub-
standard behavior against the player, reason being not to be easily defeated
by the player.
6. Identical human behavior: The behavior displayed by game AI should be
comparable in intricacy to the human behavior.
7. Self-correction: To prevent the mistakes being repeated by the game AI,
self-correction in AI will be needed in the game development.
3.2 RELATED WORKS

3.2.1 AI for Computer Games
AI has grown into a vital component of several games. It oversees how the game
works including the behavior of non-playable characters and enemies attacking
skills. It is often thought as a computer-generated mind that makes decisions, learns,
and adapts to the player changing the gameplay constantly. This idea of AI is incor-
rect in video games (Gameranx, 2016).
Artificial Intelligence as Futuristic Approach for Narrative Gaming 39
Although AI is defined as the capability to obtain and understand knowledge

and skills, it is not the default nature of the technology especially in the context of
video games. What gets defined as AI in video games is a set of conditions created
by a series of “if” statements that gets run by the computer. The actions performed
by the AI are the result of the conditions being met which gives the illusion of
intelligence; that the game or the entities inside the game are reacting, thinking,
and learning when in reality the AI only understand the given conditions without
consideration of anything else. For example, if the player is doing x, react by per-
forming x action.
For instance, if there is an AI for an entity that has the role of patrolling a hall-
way, certain conditions will be given for how to react to the player. The more com-
plex the conditions are, the more intelligent the AI will appear. This AI does not
have learning capabilities that is commonly thought of. For an AI to be truly intel-
ligent, to be capable of learning, then it has to be able to have thoughts and behav-
iors that have not been developed for it to do. This kind of AI is similar to human
mind or cognizance simulations that are being developed by scientists however even
though these simulations show signs of learning, it is not the same as human learn-
ing capabilities.
In video games, AI is developed for one purpose only, to function in certain ways
that the developers have given in specific situations of the game. There are however
various types of AI in video games. One type is called the rubber band AI. This AI
is developed to match the skill level of the player to tailor the game experience to the
player and offer more challenges by constantly observing the player’s performance
in the game. This is called Dynamic Game Difficulty Balancing technique. It can
be used in sports games to transform the difficulty of the opposing team, or it can
be used in horror and action games to change the level of enemy aggression and so
forth. This use of AI increases immersion in the game.
It is not necessary for AI to be a truly learning intelligence in video games. If the
video game is developed and designed with lots of detail, AI can be used as addi-
tional art that escalates the overall quality, feel, and art of the game.
Kok Wai Wonge has introduced papers that contribute to the various topics
regarding the use of AI in video games as well as their current state in academia
(El Rhalibi, Wong and Price, 2009). The paper “Performance simulations of moving
target search algorithms” by Peter Kok Keong Loh et al. discusses the use of Moving
Target Search (MTS) when developing bots for computer games (Loh and Prakash,
2009). MTS algorithms raise significant obstacles since they have to meet stringent
specifications involving efficiency in performance and combinative computation. In
this article, the authors analyze the efficiency and actions of current MTS algorithms
when applied to search-and-capture gaming scenarios. As part of the study, they are
also implementing a new algorithm known as MTS abstraction.
The second paper, entitled “A shortest-path lyapunov approach for forward deci-
sion processes” by Julio Clempner, is discussed where the author presents a com-
prehensive model for the representation of the problem of the shortest-path decision
process (Clempner, 2009). Consideration is given to dynamic systems controlled
by normal calculations of difference represented by Petri nets. The paper explains
that by using a discrete Lyapunov-like function, the path over the net is measured
forward. For the deterministic shortest path problem, natural assumptions of the
default results are proven. In this context, the authors adjust the conventional cost
function into a trajectory-tracking function, which is also an efficient cost-to-target
network tracking function. This significantly contributes to the conceptual frame-
work of the problem domain. The Lyapunov method presents a new principle of
balance and consistency to the shortest path decision-making process.
Coleman Ron, in his paper titled “Fractal study of stealthy pathfinding aesthetics”
uses a fractal framework to examine aesthetic qualities for a new category of stealth-
based pathfinding that seeks to prevent detection in video games (Coleman, 2009).
This study is interesting because the research on AI has provided comparatively
limited attention to aesthetic findings in pathfinding. As per the fractal framework,
the data published indicates that stealthy paths are distinct in their aesthetic value in
comparison to control path. The author also demonstrates that paths created by vari-
ous stealth rules are also unique according to statistical results.
Frank Dignum et al. discusses the research on multiagent systems and their
potential promises in developing cognitively intelligent NPC. However, due to com-
patibility differences and issues, the technology is not easily implemented in game
engines (Dignum et al., 2009). Game engines have dynamic and instantaneous fea-
tures which contribute to an increase in centralized control and efficiency, however
multiagent platforms focus on the independence of agents. The use of multiagent sys-
tems to create a more independence and intelligence will help advance and improve
gameplay.
In a paper titled A multiagent potential fields-based bot for real-time strategy
games, Johan Hagelback et al. discusses the use of AI in real-time strategy games
where bots are given field-based systems that allows them to plan attacks, find
enemies, avoid finding each other, and explore their environment (Hagelbäck and
Johansson, 2009).
A paper titled “Combining artificial intelligence methods for learning bots in a
real time strategy game,” written by Robin Baumgarten et al., discusses how AI is
used in strategy games to simulate human centric gameplay that can plan attacks and
movement according to decision tree base learning, case-based reasoning (CBR),
and annealing (Baumgarten, Colton and Morris, 2009).
Fabio Aiolli et al. in their paper titled “Enhancing artificial intelligence on a real
mobile game” discusses the technical issues that encountered during game develop-
ment such as creating a complex and engaging AI to play against. However, such
complexity is difficult for mobile game development (Palazzi and Aiolli, 2009). The
author suggests the use of a ML algorithm that solves this issue by adapting and
predicting human strategies within the game.
In the paper “Breeding terrains with genetic terrain programming—the evolu-
tion of terrain generators,” by Miguel Frade et al. discusses the role of AI in gen-
erating terrains for level design. This allows level designers to create more diverse
and advanced terrain types with better features and aesthetics (Frade, Fernandez
De Vega and Cotta, 2009). The paper conducted a study on the various terrains
created by AI, their characteristics, and resolution. The results have shown that the
use of AI in terrain generation can reserve detailed features without compromising
on resolution.
“Fine-tuning parameters for emergent environments in games using artificial intel-

ligence” by Vishnu Kotrajaras et al. discusses a tool for adjusting properties of develop-
ing environmental maps according to a specific situation (Kotrajaras and Kumnoonsate,
2009). The paper shows that using AI reduces time consuming and intensive work.
3.2.2 AI for Adaptive Computer Games

In this chapter, issues and challenges that are facing AI in video games have been
discussed. These challenges include difficulty in developing complex decision spaces
that do not get repetitive, difficulty in engineering all the behaviors needed for the
game, not enough support for game story authors who are not expert at the program-
ming needed for the complex behaviors, difficulty in anticipating all possible situ-
ations that are likely to be happen during gameplay, including adaptive responses
to player’s game preferences and choices, and lack of variations in game strategies
(González-Calero and Gómez-Martín, 2011).
This chapter proposes the use and development artificial intelligent techniques
that allow video games to be adaptive, smart, and responsive to the player. This can
be done by learning about the player’s choices, strategies, and preferences during
gameplay which in turn will enable the system to adapt and perform behaviors that
are more complex than what the programmers and developers have planned. This
will allow the player to have a more immersive and rich experience. The author
describes such a system as adaptive games.
This chapter presents three approaches for developing adaptive games through
the use of CBR which includes: automatic modification of actions for creating realis-
tic characters; managing drama and game narrative and user simulation for immer-
sive stories; and instantaneous strategy game behavioral planning.
3.2.3 AI in Video Games: Toward a Unified Framework

This chapter discusses and proposes the use of a unified conceptual framework for
developing AI in video games. The authors propose this framework as an approach
to interpret and conduit between human behavior and AI (Safadi, Fonteneau and
Ernst, 2015). Humans are capable to draw connection and similarities between dif-
ferent games which allows them to relate that knowledge while playing other games
through conceptualization. The authors explain that by using a unified framework
that enables developers to conceptualize their video games will contribute to solving
conceptual issues found in video games.
The benefits of developing such a framework will result in reducing the redun-
dancy and increase the robustness of AI in video games through enabling the design
of AI to gain independence from individual game projects and become a central
engine that can be applied to various different games which can be modified by the
developers. This will help game developers in spending less time and resources in
creating complex AI for each individual video game and instead modify an existing
system that is robust for conceptual design.
This conceptual framework however does not apply to AI used in creating story
and environment in video games.
3.2.4 Narrative in Video Games

Video games are distinct from other forms of media because of their interactivity
which has an impact on storytelling. The relationship between games and story-
telling as well as the role of narrative on overall gameplay has formed opposing
views between game developers, scholars, and game players alike which resulted in
two different philosophies to emerge in the field of game studies; Narratology and
Ludology (Aarseth, 2012).
Narratology views that stories are a crucial part of game design and that it enables
the player to experience events and emotions more deeply than any other form of
storytelling medium. They stress that game mechanics should be designed around
the narrative. While Ludology views that interactivity does not coexist with story
and narrative, as well as stress that more focus should be given to mechanics and
gameplay (Carlquist, 2002).
Narrative architecture is the author’s model that is resulted from the division
between Narratology and Ludology. He argued that the debate is based on only one
form of storytelling which is the classic linear method while neglecting other meth-
ods of storytelling such as modernist and postmodernist. Instead of arguing whether
or not games are a form of narrative, there should be an understanding that games
are a form of narrative architecture that has no power on what the player can do once
in the game world but has control on how the story gets shaped according to the pos-
sible actions or choices presented (Jenkins, 2004).
The author argues that viewing games as narrative architecture instead of focus-
ing on Narratology or Ludology, allows for using environmental storytelling or as the
author named it, “Spatial design.” This design model allows for increased immersion
by allowing the elements in the game design to show and embed information about
the narrative. The author has identified four types of spatial design or environmental
storytelling.
The first is Evoked Narrative. This spatial design provides familiar elements
that are well known to the player but offers a new perspective on it. An example of
this can be found in the action-adventure game American McGee’s Alice where the
game developers offer a familiar environment based on the famous story of Alice’s
Adventures in Wonder Land but adds horror elements to it changing the narrative
perspective where the player starts to doubt the sanity of the player character and
immersing the player in a new world that is interesting.
The second is Enacted Narrative. This spatial design provides conflicts and goals
within the game that depends on the player’s movement. This design can be found in
open world games such as Skyrim where the player can freely explore the environ-
ment and find various quests that contain elements the feed the game narrative. Some
quests especially those that contain major game narrative or plot points might require
the player to proceed in the game in a specific order or to finish certain quests first.
The third is Embedded Narrative. This spatial design provides distribution of
narrative information across the game world in order to control the story progres-
sion inside the game as well as challenge the player into piecing together fragments
of information in order to solve puzzles and construct plot points. In this model, the
game world is designed to serve as a memory palace for the player.
The fourth and final spatial design is Emergent Narrative. This provides the
player with an abundant source of world building to allow the player to design their
own narrative. The narrative is not pre-constructed by the developers, it is designed
to be chaotic similar to the real world. The author has used The Sims games as
example where the player can make decisions to interact with other characters, have
desires, and where their decisions in the game have consequences making the player
immersed and engaged while maintaining freedom in their own narrative.
3.2.5 Narrative Game Mechanics

This chapter argues that game mechanics can be the most powerful tools for telling
narrative in video games. Game developers usually use well-known mechanics and
storytelling techniques such as cut-scenes, environmental storytelling, and dialogues
to add narrative in their games however, this limits the narrative experiences that the
developers could offer to the player because these techniques only have room to show
one type of storytelling to the audience. The author believes that using narrative game
mechanics can effectively tell a better story in a video game (Dubbelman, 2016).
The game mechanics and rules developed by game designers have the power to
affect the player’s actions within the game world without completely controlling the
player. This leaves room for player freedom as well as affect the way the narrative
and events take place in the game. The paper mentions the importance of creating
a connection between the player and the character for this leads to immersion and
investment in the game narrative. This can be done by adding conflict within the
character and the environment as well as choices that evoke moral dilemmas for the
player. It is important that these elements all feed the narrative of the game and not
be designed in isolation.
In order to create effective narrative game mechanics, game developers must find
ways to have the player involved in the construction or reconstruction of the game
narrative through the actions that will be designed to be performed by the player.
There must be a link between the player’s actions within the game world, the charac-
ter’s motivation in the game narrative, and the story being told in the game. Unlike
traditional storytelling methods where the creator has control over the behavior of
the characters within the story through mentioning these behaviors, Interactive sto-
rytelling that is supported by games has an indirect control over the characters since
the player has control inside the game world through the created game mechanics.
The author supports narrative mechanics in unfolding meaningful and interactive
stories in games as opposed to focusing solely on mechanics and gameplay or view-
ing games as a traditional form of narrative.
An article on Gamasutra explained the rise of narrative mechanics in Indie
games. Bycer J. has defined narrative mechanics as “A mechanic or set of mechan-
ics, which directly influence the story through player action.” (Bycer, 2012). He
explained further by giving three different examples based on indie games where
narrative mechanics were used. The first being the use of choices that affect the
gameplay completely, the second is the use of a narrator that reacts to the player’s
actions within the game world, and the third is having the game track the actions of
the player and providing an ending that depends on the player’s actions as well as
explaining how those actions led to the specific ending the player got by the end of
the game.
Bycer wrote that there are endless possibilities for how narrative mechanics can
evolve and change gameplay to tell better stories that are more immersive and inter-
esting. He hoped that AAA game developing companies would start implementing
them in their new games and he also recommended the use of narrative mechan-
ics that have not been developed for games yet such as changing the behavior and
appearance of NPC’s according to player’s behavior as well as altering the player
character’s senses within the games based on the rate of success or failure at solving
puzzles in the game.
3.2.6 Interactive Narrative
Interactive narrative is a type of interactive entertainment in video games where the
player can affect the story through their actions during gameplay (Riedl, 2012). Narrative
in video games provide meaning and context to game events and actions, provide moti-
vation for player’s actions in game, and acts as a link to transition the player through the
various events and tasks within the game. However, the player has limited control over
the events of the storyline within the game and can only have a small influence through
choices which allows the game to have multiple branching storylines. Branching story-
lines are not often developed due to technical difficulties but AI can help create interac-
tive narrative by creating multiple storyline branches easily and efficiently.
There are two different approaches in using AI in interactive narrative; the first
is emergent narrative and the second is drama management. Emergent narrative is
where the AI simulates realistic and independent characters within the game, while
drama management is where the AI creates and drives the storyline of the game
according to the player’s actions and preferences. Interactive narrative does not just
give the player control over the story but also the illusion of a realistic world full of
choices and possibilities for the player to explore.
3.2.7 Adventure Games and Puzzle Design

The genre of adventure games allows interactive storytelling in video games through
the use of puzzles and detailed game worlds. There are various sub-genres in adven-
ture games that get classified according to the gameplay as well as themes. They
follow the same steps where the player controls a character and follows a quest which
requires solving puzzles to proceed with the game. Usually, the puzzles have one
solution only (Afram, 2013).
There are many different puzzles in adventure games. Mostly they are about
unlocking barriers such as doors or containers such as boxes. However, no matter
which form the puzzle takes, it is always about coming up with a solution to remove
an obstacle. There are inventory-based puzzles which requires the player to collect
items and combine them. There are environmental based puzzles which requires the
player to examine the environment and solve problems by interacting with it. Finally,
there are puzzles that are based on NPC dialogue which requires players to solve
problems by having dialogues with the NPCs.
Puzzles in video games have a strong impact where they can affect the player’s
actions and progress the story. They give meaning to in-game actions such as inspect-
ing items to find clues and search for a solution. They can also have consequences by
giving punishments or rewards that affects the overall story and progression. Puzzles
must be designed to allow the player to progress in the story and game goals where
they are will not be perceived as meaningless obstacles but meaningful and impact-
ful choices that makes the game progress.
The thesis discusses ten principles for designing puzzles in video games. The first is
to make them easy to understand so that the players do not feel confused. The second
is to make them easy to begin solving them so that the players do not feel intimidated
and try to solve them. The third is showing progress where the player can feel that they
are accomplishing something by interacting and manipulating the objects available.
The fourth is to show that the puzzle can be solved by giving the player cues of prog-
ress. The fifth is making the puzzle interesting by adding more difficult challenges.
The sixth is giving the player multiple options to progress the game. Scattering dif-
ferent puzzles that the player can choose from to avoid frustration resulted from lack
of progression. The seventh is to show the player that all elements of the puzzles are
connected which helps solve more difficult challenges. The eighth is to offer clues that
can help the player in solving the puzzle. The ninth is showing the answer to the puzzle
indirectly so that when the player finds the solution, they gain a sense of accomplish-
ment. The final principle is to be careful of using optical illusions in puzzle design
which might make the player feel frustrated.
3.2.8 The Horror Genre and Video Games

This book section discusses how the Horror genre is one of the most ideal genres to
be adapted and used in video games going back to the genre’s roots which can be
traced back to the text interactive game Zork created in 1980. This game marked the
beginning of the shift from Adventure games to Horror by introducing the concept of
death to the player for the first time. The Horror genre started to evolve and become
more popular with many games being introduced in the market such as Mystery
House, Maniac Mansion, Silent Hill, and many others that followed (Rouse, 2009).
Video games will always have constraints in creating a realistic world. This, how-
ever, is one of the reasons why Horror games are one of the ideal genres for video
games because although Horror games take place in a familiar place that should be
realistic, it is designed with the purpose of making it seem different and out of the
ordinary while offering the player justification for the strange environment, abilities,
and events. The familiar environment also grounds the player and makes the experi-
ence more interesting unlike fantasy and science fiction games, where the unique
characteristics of the game is a result of a foreign world.
The author argues that the video games and the Horror genre share overlapping
goals which has been a driving reason for the success of the genre as well as one of
the reasons why Horror games have potential to grow and evolve.
The Horror genre is one of the most popular video game genres with the youth
(Krzywinska, 2002). It is a unique genre because unlike movies and books, it makes
the player feels scared and immersed in the environment due to its interactivity.
Horror helps young adults confront and master their fears in a safe space (Jamie
Madigan, 2015). This helps releasing feelings of anxiety caused by real world fears
and worries and replacing it with feelings of relief and exhilaration as well as a sense
of reward from conquering their own fear (Elio Martino, 2019).
Experiencing fear in a fictional environment causes a rush of emotions that other
video game genres are unable to provide. This is a result of the excitation trans-
fer theory where fear transfers feelings of pleasure similar to that of riding a roller
coaster knowing that it is safe (Nicolas Brown, 2020).
This article discusses the various issues modern Horror games face and how they
are different to older games of the same genre. The author argues that horror games
have fallen out of fashion and are no longer popular in the game industry. AAA game
development companies have moved away from the genre despite the enormous suc-
cesses of games such as Dead Space 3 and Resident Evil 6. This is due to the indus-
try’s focus on combat which has resulted in the rise of Action-Horror games rather
than focusing on puzzle-solving, narrative, and adventure.
Meanwhile, the independent game development companies such as Frictional
Games, started to stay away from action-horror and implement new game mechan-
ics (such as stealth and sanity) to help the player focus on story while also increasing
fear elements in the game. The success of the game has resulted in a popularity with
indie horror games with many games being released in the market using similar
mechanics of stealth as well as removing combat (Bycer, 2019).
However, many of these games lack an understanding of psychology behind what
evokes tension and fear in the player as well as what immerses the player into the
world and the narrative of the game. This includes providing the player with details
about the game world and narrative as well as game mechanic and loops that keeps
the player immersed in the story. The difficulties of designing these elements are
one of the reasons that independent game developers have switched to a different
approach; designing games for an audience instead of players. With a rise of a trend
called “Let’s Play” has started gaining popularity on YouTube in 2010. With many
YouTubers creating videos of themselves playing and reacting to these games.
This article is significant to the proposed project because it explains the current
issues in Modern Horror game design and shows how they are different from the
classic Horror games.
In this chapter, a tool has been designed to help game developers in creating a
horror game. The author points out the difficulty in developing horror games due
to the different interpretation the nature of fear. This chapter offers a solution by
categorizing fear on a scale tool that measures the level of fear in game events. This
tool helps the developer in understanding how certain game mechanics, lighting, and
enemy design should be like in order to achieve the desired emotional response from
the player as well as plan the intensity of certain game events along with the pacing.
The tool also acts as a bridge between understanding the psychology behind fear and
enjoyment in horror video games (Ntokos, 2018).
It is crucial to understand the emotions that are desired to be provoked in the
player and understand how these emotions contribute to the player’s enjoyment and
immersion. The level of fear tool contains ten levels of fear which explain the inten-
sity of the emotion.
The author advices game developer to use the scale shown in Table 3.1 to plan the
game events and develop accordingly. The scale tool can be applied to be used on
atmosphere, audio, and Enemy AI.
Atmosphere is crucial in horror games where it controls the player’s feel in the
environment. Atmosphere is the element that sets the feel, style, and tone in the
environment. It contains elements such as sound, lighting, and gameplay experi-
ence. Darkness is suitable for creating a scary atmosphere for the player where the
darkness pushes the player to imagine their surroundings and increase tension. Any
element that blocks player’s visibility will create a sense of unease.
Audio has a strong impact on player experience and game atmosphere. In
Horror games, Auditory Hallucinations affect the player’s experience by increas-
ing tension and fear through hearing sounds that did not happen (Demarque and
TABLE 3.1
Levels of Fear (see Ntokos, 2018)
Level Emotion Meaning Example

1 Calmness Peaceful state of mind. No fear. Beginning of the game where the
player is exploring the environment
and learning the mechanics. It could
also be used in safe rooms or guarded
areas.
2 Nervousness Player gains information of strange Non-direct events that the player hears
knowledge that causes interest and or reads about without enough
doubts. explanation.
3 Vigilance Player starts to feel fear and pays Direct evidence of the strange events
attention to the surroundings. the player encountered before.
4 Restlessness The player’s fear starts to hinder Sound effects and strange changes in
level exploration. the environment.
5 Tense Player feels fear from any minor Hearing voices, sounds in the
change in the environment without environment like doors and wood
direct evidence. creaking.
6 Distress Players starts to have survival Witness a distressing event through
instincts instead of thinking clearly. visuals or audio cues.
7 Fright Players starts to have quick reflexes as An enemy getting nearby and the
a reaction to the survival instincts. events of subtle Jump scares.
8 Dread The player’s fear and survival An enemy patrolling near the player.
instincts cause immersion and
focus in the game to avoid getting
detected by the enemy.
9 Panic The player must make quick Combat, enemy chase, managing
actions to survive. resources to survive.
10 Terror Player losses control to fight or Player death, near death by an enemy,
flight instinct. Player might jump surprise attack.
out of the seat or scream.
Relief and enjoyment due to rush of adrenalin
Lima, 2013). They have various types and forms such as strange noises of ghosts
or crying.
This chapter conducted an experiment where people got to play two different
versions of the same game. One has Auditory Hallucination and one without. The
results have shown that people who played the game with Auditory Hallucination felt
more scared playing the game.
The Survival Horror genre has many elements that makes it perfect for story-
telling and narration and trying to understand the genre from the narrative per-
spective helps shed light to many aspects in Horror games that makes it unique
from other genres as well as better at implementing effective narrative structures
in the game. The player character is always put in situations where (s)he takes the
role of a detective who is exploring and investigating their surrounding environ-
ment. This makes the player engaged in the game and trying to reconstruct the
narrative of the game while also being unsure about the strange and paranormal
elements happening in the game world and threatening the character. The player’s
goal is not just to discover the story, but to find a solution and escape at the same
time. This structure is one of the reasons that characters in Horror games usu-
ally suffer from memory loss. This enables the player to share the experience
of uncovering the past with the character creating a connection and increasing
immersion.
Environmental storytelling is one of the powerful mechanics used in horror
games where narrative is being unfolded to the player through the game world and
architecture. The buildings, objects such as toys, furniture, or notes are used to tell
stories from the past. The level design becomes a crucial part of the game narrative
and almost treated as one of the characters in the game (Kirkland, 2009).
Survival Horror games show that there are ways to represent narrative in an
interactive game experience through spatial environmental design even though
the player’s pathway is designed in the game world through quests and tasks.
Storytelling in Survival Horror is also designed to feed the player the sense of not
being in control in an interactive narrative experience. However, in order to achieve
this, storytelling usually takes a linear design in the game giving the player the
illusion of being in control over the story events while providing the player with
interactivity. This causes a sense of fear and tension in the game through showing
that there are more powerful dark forces that control the narrative and manipulate
the player in the game world.
3.3 PLAYER EXPERIENCE

A framework has been presented in a paper published by Hunicke, Leblanc and
Zubek (2004) that has been developed for the Game Developers Conference. The
MDA framework which stands for Mechanics, Dynamics, and Aesthetics is a formal
method to understand games and game design. It is also developed to link between
game development, criticism, and research creating a unified and easy approach for
studying and designing games that can be used by both developers and scholars alike.
Design methodologies are crucial for creating any artifact. They are the processes
that guarantee the quality of a project. Using an iterative, quantitative, and qualitative
methods help the game designer in analyzing both the end results to improve project
implementation, and the implementation process to improve the final result. This
method of approaching the project from both the viewpoints, allow various possibili-
ties and dependencies when creating systems and subsystems in a game that need to
interact with each other dynamic and complex behaviors inside the game.
The MDA framework is designed to help designers and game developers to
closely examine interdependencies in a game in order to create the desired end result
while scholars and researchers need to study and understand them before making
any conclusions on the game experience resulted by them.
Players view the game as a set of rules which need to be followed, these rules
build the system that is being played and playing the system results in the player
having fun. This view can be formalized using the MDA model to be used by the
developers as follows; the game contains mechanics which are the basic structure
of the game and the code that forms the rules, Dynamics are the result of the inter-
dependencies of the mechanics, how they affect the player’s input and the game’s
output to the player, and finally the Aesthetics are the player’s emotional responses
and experiences in the game.
This framework is built on a fundamental idea that video games should be
described as artifacts and not as media. The author suggests that video games identi-
fied as artifacts mean that the behavior of a game is formed by its content and their
interactions together forming a system and not the media elements that are presented
to the player.
Each element of the MDA framework can be used to view separate parts of the
game that can be linked together to provide the full experience desired by the devel-
opers as shown in Figure 3.1. The designer views the mechanics as the elements
that creates the dynamic behaviors in the system, these dynamics create the desired
aesthetic experiences to the player. However, the player views the aesthetics as the
main element that sets the mood of the game which is a result of the dynamics and
lastly, the underlying mechanics.
The author urges developers and researchers to think about both the views of the
designer and the player in order to understand how game elements interdepend and
interact with one another as well as create a game that focuses on player experience
as opposed to a game that focuses on the given features and mechanics.
FIGURE 3.1 The MDA framework (see Hunicke, Leblanc and Zubek, 2004).
In a thesis done by Tanskanen (2018), on the effects of immersion on player expe-

rience explains that immersion is created from different parts; the first is narra-
tive which contains the story and plot in the game. The second is spatial which
includes the player exploring the environment. The third is effective which includes
the player’s emotional involvement in the game. The fourth is shared which is about
the interaction between the various elements of the game. The fifth is kinesthetic
which includes controlling the player character, and finally the sixth and last is ludic
which includes players in game choices and their consequences.
The thesis explains how narrative and spatial involvement the biggest impact on
player immersion can have where the player moves in the environment and interacts
with the narrative forming emotional involvement. This is why the author suggested
the importance of creating a detailed document for the game environment in order to
plan and list the aspects of immersion to make the game believable and immersive.
In a paper by Boonen and Mieritz (2018), methods for manipulating the player’s
agency in horror games have been discussed. Agency is defined as the player’s reac-
tion when a wanted action is supported by the game. The paper has designed a model
for Agency that explains the relationships between the various elements that can be
used to manipulate the player’s agency. Those are parameters of player character
which contain the aspects that concern the player character, the second is parameters
of system which contains the aspects of game and game world limitations, param-
eters of player.
The parameters of player character contain two sub-parameters; the first is physi-
cal parameters which are the physical limitations of the player character. And the
second is psychological parameters which are the limitations of the player charac-
ter’s psyche such as having a sanity system.
The parameters of the system also contain two sub-parameters; the first is about
the obstacles in the environment such as enemies, building, and resources. The sec-
ond is about the limitations created by game mechanics such as locks. Finally, the
parameters of the player are about the limitations and skills of the player in the game.
It is crucial to plan and consider the movement of the player character within the
environment as well as the space and placement of the objects within it. The level
design should be realistic following the rules of physics. This is important to con-
sider because it plays a role in player immersion.
There are two different types of level structures; linear and non-linear. Linear
level design is suitable for story-focused games where the player must follow a
certain path to reveal the plot. This limits the freedom of the player withing the
level. Meanwhile, non-linear level design is flexible allowing the player freedom to
explore without locking certain game areas or objectives due to following a specific
story. However, it is argued that non-linear games are considered linear because the
player’s movement is planned.
A research done by Milam and El Nasr (2010) in the different types in level design
and how they affect the player character based on their patterns. The paper discuses
five different push and pull game mechanics as approaches for game designers to
control and manipulate players movement in the game environment.
The first pattern is path of movement and resistance which is used with narrative
purposes in game such as player objectives which sets the player on a path to reach a
certain goal while resistance is the challenges that stand in front of the player block-
ing the player from achieving their objectives. This kind of pattern leads the player
to find a solution to remove the resistance and continue with the objective.
The second pattern is based on AI movement where the player moves in the envi-
ronment responding to the movements and actions of AI characters in game which
can be NPCs or enemies.
The third pattern is based on a target path where the player needs to reach a cer-
tain visible target which directs and moves the character to that location.
The fourth pattern is based on collection paths where the player moves in the
environment motivated by collecting rewards or resources such as battery pickups
or coins.
The fifth and final pattern is player vulnerability path which forces the player to
adapt to the environment due to vulnerability in a given scenario such as needing to
hide to avoid enemies.
3.4 METHODOLOGIES USED IN GAME DEVELOPMENT

The Agile methodology is one of the most simple and efficient methods that a
developer can rely on to design and develop a software. It is a method that uses
continuous planning, learning, collaboration, and improvement to deliver early
and offer flexibility for changes due to client requirements or emerging new
technologies.
The Agile methodology focuses on how projects should be worked on differ-
ently depending on their type and needs. It combines both iterative and incremental
processes with a focus on having a highly adaptive approach and focus on client
satisfaction through rapid development. It breaks down the project tasks into incre-
mental parts or time boxes that are developed in a series of short iterations. The
project is not designed in one sequential order from beginning to end, but develop-
ing small features and functionalities in small time periods. Each iteration is like
a mini version of the project where the results are used to review and adjust the
project plan.
These iterations are like a short time frame where every iteration lasts around one
to four weeks. Each iteration goes through the project life cycle including planning,
requirements gathering, design, coding, and testing. They include certain features in
the project where a working product is shown to the client by the end of each iteration
and at the last iteration, the final project includes all the features and requirements
that were defined at the start of the development process.
The initial requirements and scope of the research is decided and defined by the
start of the development phase including the number of needed iterations and the
overall duration of the project which avoids the risk of scope creep. The Agile meth-
odology also focuses on developing a project in a swift manner through iterations
and increments which helps reduce the risk in the project as well as reducing the
time for working on tasks and requirements before a functional product is shown to
the client.
A functional prototype has been designed and implemented inside Unreal Engine
using Unreal Engine Blueprints in a duration of four weeks. During this time, various
FIGURE 3.2 Proposed functionality diagram.
games have been observed to gather references and understand their requirements
and functionalities until initial features for the proposed project have been identified.
Various sources and tutorials have been studied to learn how to implement the initial
identified features inside Unreal Engine as shown in Figure 3.2.
Game development projects share many similarities with software engineering
projects (Kortmann and Harteveld, 2009). Agile methodologies are suitable for com-
plex projects where requirements needed are not clear at the start of the project.
Game development projects are complex with uncertain requirements that might
need to change during development due to its multidisciplinary nature where vari-
ous skills and tasks are needed from art design, sound, gameplay, and programming
(Petrillo and Pimenta, 2010). This causes difficulties for clients to have a clear pic-
ture of the needed outcomes at the start of project development. Agile methodology
allows developers to gain better understanding to the needs of the clients as well as
provide suitable solutions to improve the quality of the project as well as add effi-
ciency in development by spending more time adding improvements to the project.
Therefore, many game developments companies have applied the use of Agile meth-
odologies (Stacey and Nandhakumar, 2008).
The functional prototype is implemented in the third person design template of
Unreal Engine. This template includes a third person character and camera where the
camera follows the player character in the three-dimensional (3D) space as shown in
Figure 3.3. The first-person mode has been designed by moving the camera to the
head of the player and linking it to the head mesh so that it moves with the movement
of the player character. A spotlight has been added in front of the head mesh to cre-
ate the light for the flashlight. The arrow attached to the character mesh points at the
direction which the player character faces when spawned in the level.
FIGURE 3.3 Player character class in UE4.
The player can walk and look around in all directions inside the 3D space by
using the mouse and keyboard to navigate in the x, y, and z axes. When the player
presses the w, s, a, and d keyboard buttons, the coordinate in the background gets
updated accordingly.
Screenshot of the player’s screen in the functional prototype. The player has a
user interface where it shows the player’s hunger and stamina levels. These bars
decrease by time where hunger can get regenerated through picking up consumable
items (minor food pickup and full food pickup), the stamina gets regenerated by time
however the player character will not be able to sprint as long as the stamina is get-
ting regenerating.
In the bottom, there is the flashlight battery that shows the player how much bat-
tery is left as shown in Figures 3.4, 3.5 and 3.6. The player will not be able to turn on
the flashlight if the battery is empty and must pick up batteries to refill.
FIGURE 3.4 First person game view.

FIGURE 3.5 Inspect item prompt.
There is an objectives icon that is dynamic meaning it changes and get updated
according to the player’s actions. If the player collides with a new objective, it will be
added and shown on the UI. If the player completes an objective, objective complete
message will be shown.
The screenshot shows the use of the interaction and inspection system in the pro-
totype. This is implemented by creating a Line Trace that reads if there are any
inspectable items for the player to pick. A message is shows to the player if the player
is looking at an inspectable item.
Items are created as child to a parent item having similar main characteristics includ-
ing being picked up and inspected by the player. In the screenshot, the player is looking
at the parent item where a message shows the player that (s)he can inspect the item.
The player picks up the inspectable item where a message is shown explaining
that using left click allows item inspection while right click drops the item on the
floor being impacted by physics.
When the player left clicks to inspect the item, (s)he can rotate the item by left
clicking and moving the mouse. There is also an option to zoom by scrolling the
FIGURE 3.6 Player picks item.

FIGURE 3.7 Item inspection.
mouse well. Item description can be visible on the side. The description is dynamic
and changes according to the item being held. This is scripted inside the script graph
of the inspection UI. The player can exit inspection mode by right clicking twice;
first to hold the item and second time to drop the item.
The design of the user interface has been designed inside Unreal Engine wid-
gets as shown in Figure 3.8.
Figure 3.6 shows the blueprints for an AI going toward the player. This is done
by creating a trigger box and placing it in the scene at the desired location. A trigger
box is a box collision that is invisible to the player in game. It is used to trigger events
when something overlaps with it. In Figure 3.7, the box is triggered by the player
character by casting or calling the player’s class and from this class, the location of
the player is stored in the blueprint and used to for playing a sound effect. This is
done by getting a play sound at location node and connecting the player’s location
FIGURE 3.8 User interface.

FIGURE 3.9 Main menu.
to the location input of the node. Then the AI actor is stored in the blueprint by get-
ting all the actors from the AI class. This enables the developer to access any actor
in the specified class as shown in Figures 3.9 and 3.10. The actors are then stored
as a copy and the AI actor set to visible. An AI move to node is added. This node
enables the AI to move toward the desired location. The pawn is connected to the
object reference, specifying which actor or which AI should be moved. The location
input is a vector variable representing the desired destination the AI should move to.
The target actor input is also an object or actor reference however, this input is for
specifying an actor whom the AI is targeted to move toward. Acceptance radius is
a float variable that specifies the value on which the success of the event is based on
upon reaching the desired destination. Finally in the input is stop on overlap which
is a Boolean variable that specifies whether or not the AI will stop moving when
overlapping or colliding with something. In the output there are three different exits;
if the first is connected, then any nodes or events that follow will happen no matter if
the AI move node succeeded in going to the desired location or not. The second exit
FIGURE 3.10 Journal blueprint.

FIGURE 3.11 Player hunger and stamina.
FIGURE 3.12 Blueprint of AI based functionality.
is for the success of the node, and the third is for the failure. Finally in the output is
the movement result which stores the result of the path followed for later use.
After the success of the AI move node, the AI actor is destroyed as well as the
trigger box to ensure that the player does not trigger the event again and avoid any
bugs. Overall blueprint of the game is shown in Figures 3.11 and 3.12.
3.5 AI ELEMENTS IN THE PROPOSED

NARRATIVE GAMING MODEL
There are multiple types of ML algorithms that could be applied for the gaming.
However, supervised ML is not appropriate for narrative gaming. Therefore, it is
recommended to either adopt unsupervised or reinforced algorithms. Although the
gaming is done in controlled environment, but the patterns are being read, monitored
and recorded for the pattern detection.
This pattern paves way for the next set of player experience when the player finds
a totally new experience based on its previous input pattern (Comi, 2018). The use
of AI in video games is not compatible despite the widespread belief. AI is often
mistaken for artificial behavior of agents in a video game. Agents in video games
are required to offer enough engagement for the player to enjoy a challenging but fun
experience. Agents should not outsmart the player resulting in a frustrating experi-
ence, but imitate human behavior in having weaknesses and imperfections.
Video games can be viewed as more than just a medium of fun and enjoy-
ment, they can be viewed as a medium for studying and experimenting the dif-
ferent developments that can be done in a variety of fields related to AI. Google
DeepMind worked on developing an agent capable of outsmarting players in their
game AlphaGo which led to beating the best Go player and score a goal that was
not considered possible.
The article written by Comi discusses how an artificial intelligent agent can be
developed to play a game such as the “Snake” game. To achieve this, Comi explained,
a deep reinforcement learning algorithm should be implemented using Keras on top
of Tensorflow. This is a method where the system parameters are given a reward;
either positive or negative, based on its actions. The agent is not given any informa-
tion about the game rules, leaving it to figure out what needs to be done by itself and
create an appropriate strategy to gain the positive reward by scoring the highest in
the game.
In this way, a video game can show how an AI agent with the help of Deep
Q-learning, can learn to play the game snake, and come up with suitable gaming
FIGURE 3.13 Overview of the scene.

FIGURE 3.14 User interface of game design platform – Unreal Engine.
strategies in order to gain the highest scores. The author has conducted study where a
Deep Q-learning algorithm has scored 50 points after just five minutes of training on
playing the game Snake. The Snake game was developed using Python and Pygame,
which is a library that allows developers to create simple games using Python. At
first the AI agent did not know how to play the game as it was not trained yet, how-
ever, after 100 iterations which is equivalent to five minutes of playing, the agent
started to score points and after 200 iterations, the agent scored 83 points.
The method of reinforcement learning is based on the Markov Decision Process.
This process provides a mathematical framework that creates random decision-
making outcomes. The author has used Deep Q-learning instead of ML because in
ML, the agent is trained with a targeted input and output. The output is the correct
answer where they are required to be predicted according to the inputs provided.
This method is not effective because it does not show which action is the best to be
taken in the game to score the highest score.
When using reinforcement learning, two components are present, those are, the
game which represents the environment and the snake which represents the AI agent
that has the deep neural network creating the decisions of the snake in game. When
the agents performs and action, a positive or negative reward is then provided by
the environment which depends on how that specific action taken by the agent was
beneficial to that game state. The agent then has an objective to learn the best strat-
egy of actions that provide the best rewards in each game state. The game state is
the observations received by the agent throughout different iterations in the game
environment. It can vary between variables such as position and speed. The agent’s
decision-making process to determine the best strategy of actions in the game is
called a policy according to the reinforcement learning method.
The AI agent makes decisions according to the Q-table which is a matrix that
shows the association between all the possible actions that can be taken by the agent
according to its state in the environment. Values are given in the table according
to the probability of each action’s success and their rewards. However, the author
argues that there is a problem with the policy in reinforcement learning because it is
represented in a table. This constrains the space for the states in the environment and
makes it difficult to have a large number of different states. Therefore, using Deep
Q-learning is preferred because it represents the policy in a deep neural network
where the values are changed by applying the Bellman equation.
The algorithm for the snake agent is explained as follows: Firstly, the game
starts and a Q-value is given randomly. Secondly, the state is recognized by the
system. Thirdly, an action is taken according to the recognized state. The action
taken can be random at first to allow exploration and maximize learning so that
later the agent can depend on its neural network. Finally, the agent gets rewarded
from the environment according to the action taken and how it affects the current
state. Then the Q-value is updated by using the Bellman equation which results in
generating a new state accordingly. The data of the original state, action, reward,
and the updated states are all stored to be used in training the neural network in
a method called Replay Memory. The operations repeat until the game ends or
another condition is met.
A state is explained as representing the agent’s situation as well as the neural net-
work input. In the study done by the author, an array of eleven Boolean variables was
given for the states in the snake game. Those states contained Booleans for whether
or not the agent is close to danger, the direction of the agent’s movement, and the
position of food.
The agent should try to get the maximum number of positive rewards and get
the minimum number of negative rewards. In the snake game, positive rewards are
given by adding ten points to the score every time the snake (AI agent) eats a fruit,
while negative rewards are given by removing 10 points from the score every time
the snake (AI agent) hits itself or a wall. Additional points can be added to the score
when for each move the snake takes while avoiding death, this affects the decision-
making process for the AI were moving in a in a certain way can get additional
rewards. This shows that by applying reinforced learning, an AI can outsmart the
human players by using flaws in opponents’ strategies to come up with action that
are unanticipated.
The study done by the author shows that a simple AI agent can learn how to under-
stand the mechanics of an environment, create decisions to gain positive rewards and
avoid negative rewards without being informed of the rules, and respond in unantici-
pated way by using flaws in opponents’ strategies.
3.6 Q-ALGORITHM FOR AI IN GAMING

Using Q-algorithm for reinforcement learning for AI has various advantages such
as being simple and flexible to use. Q-learning is divided into two stages: one for
training and the other for exploitation. In the training stage, the AI agent explores the
environment at random and storing values in the Q-table of all the positive and nega-
tive rewards it has received during exploration. These values are later used to predict
which actions would be beneficial to get positive rewards and which actions would
FIGURE 3.15 AI behavior tree for the two stages of Q-learning.
lead to negative rewards or no rewards at all in the future. This process is the process
behind the AI agent decision making during the second stage which is exploitation
(sacredgames, 2017).
The author has given the equation below for Q-learning:
Q ( state, action ) = R ( state, action ) + Gamma * Max [Q ( next state, all actions )]
An example has been done in Unreal Engine to show the use of Q-learning on an
AI agent where the agent or none player character starts the training stage by mov-
ing at random between four locations in the environment. The four locations include
three bowls containing food and a switch. Each location has their own values associ-
ated to them and the AI agent learns their values as well as associated behavior (how
to get food from the bowls and how the light switch can affect the availability of
food) during the training stage.
The desired outcome is for the agent to start making decisions and behave with
intention according to the values and behaviors found and observed during the train-
ing stage. The agent should be able to identify that food can be obtained from the
three bowls only if the light switch is turned on therefore the AI agent is expected to
be able to go to the light switch first then go to the food bowls.
Q learning algorithm provides an AI agent with realistic lifelike and dynamic
behavior by imitating the drive behind the carrot and stick example. The agent’s
decisions are based on the rewards available in the environment which is stored in a
table containing all the possible rewards both positive and negative. It can be used for
creating various intentional behaviors to enhance gameplay by having more lifelike
non player characters.
3.7 CONCLUSION
AI has become an integral part in game development. This chapter focuses the use
of AI in 3D narrative games for enhancing the player’s experience and increasing
immersion. This is proposed by using AI or ML to create immersive, complex,
and dynamic game narratives, which changes according to the player’s play-pattern
and allows the player to have more freedom and make interesting choices. Various
approaches have been studied in this chapter including approaches for narrative
game design, enhancing the player’s experience through the use of the MDA model,
and the use of AI in creating dynamic video games. It has been found that there
are various types of AI that is used in game development which are usually cre-
ated on a rule-based system which conditions the AI to make certain behaviors
as a response to different scenarios but sometimes ML is used to allow the AI to
learn and adapt to the player’s patterns. However, supervised ML is not appropriate
for narrative games. Therefore, this study recommends the use of unsupervised or
reinforced algorithms in developing narrative games. This allows the player to be
in a controlled environment but also allows the AI to be able to learn and record
the player’s patterns and responds accordingly creating an enjoyable and immersive
playing experience.
REFERENCES
Aarseth, E. (2012) ‘A narrative theory of games’, in In Proceedings of the international confer-
ence on the foundations of digital games, pp. 129–133. doi: 10.1145/2282338.2282365.
Afram, Rabi. “Puzzle Design in Adventure Games.” (2013),Degree Project in Game Design,
15 ECTS Credits Game Design and Programming, Spring 2013.
Anderton Kevin (2018) The Impact Of Gaming: A Benefit To Society [Infographic], Forbes.
Baumgarten, R., Colton, S. and Morris, M. (2009) ‘Combining AI methods for learning bots
in a real-time strategy game’, International Journal of Computer Games Technology,
2009, pp. 1–10. doi: 10.1155/2009/129075.
Boonen, C. S. and Mieritz, D. (2018) ‘Paralysing fear: player agency parameters in Horror
games’, in Proceedings of Nordic DiGRA 2018.
Brown, A. D., Stacey, P., & Nandhakumar, J. (2008). Making sense of sensemaking narra-
tives. Human Relations, 61(8), 1035–1062. doi:10.1177/0018726708094858
Butler, M. (2016) Why People Play Horror Games, iNews.
Bycer, J. (2012) Extreme Storytelling: The Use of Narrative Mechanics, Gamasutra.
Bycer, J. (2019) The Problem of Modern Horror Game Design - Game Wisdom, Game
Wisdom.
Carlquist, J. (2002) ‘Playing the Story Computer Games as a Narrative Genre’, Human IT:
Journal for Information Technology Studies as a Human Science, 6(3), pp. 7–53.
Clempner, J. B. (2009) ‘A Shortest-path Lyapunov Approach for Forward Decision Processes’,
International Journal of Computer Games Technology, Volume 2009, pp. 1–12. doi:
10.1155/2009/162450.
Coleman, R. (2009) ‘Fractal Analysis of Stealthy Pathfinding Aesthetics’, International
Journal of Computer Games Technology, Volume 2009, pp. 1–7. doi: 10.1155/
2009/670459.
Comi, M. (2018) ‘How to Teach AI to Play Games: Deep Reinforcement Learning’, Towards
Data Science. Available at: https://towardsdatascience.com/how-to-teach-an-ai-to-
play-games-deep-reinforcement-learning-28f9b920440a (Accessed: 7 March 2021).
Demarque, T. C. and Lima, E. S. De (2013) ‘Auditory Hallucination: Audiological Perspective

for Horror Games’, Sbgames.Org, Volume 2013, pp. 19–26.
Dignum, F. et al. (2009) ‘Games and Agents: Designing Intelligent Gameplay’,
International Journal of Computer Games Technology, Volume 2009, pp. 1–18. doi:
10.1155/2009/837095.
Dubbelman, T. (2016) ‘Narrative game mechanics’, in International Conference on Interactive
Digital Storytelling. Springer Verlag, pp. 39–50. doi: 10.1007/978-3-319-48279-8_4.
Elio Martino (2019) Why We Love Survival Horror, Psychology Today.
Ferguson, C. J., & Colwell, J. (2020). Sexualised video games, sexist attitudes and empathy
towards victims of rape: Correlational evidence for a relationship is minimal in an
online study. Criminal Behaviour and Mental Health. doi:10.1002/cbm.2138
Frade, M., Fernandez De Vega, F. and Cotta, C. (2009) ‘Breeding Terrains with Genetic
Terrain Programming: The Evolution of Terrain Generators’, International Journal of
Computer Games Technology, Volume 2009, pp. 1–13. doi: 10.1155/2009/125714.
Gameranx (2016) How Does Video Game AI Work?, YouTube.
González-Calero, P. A. and Gómez-Martín, M. A. (2011) ‘Artificial Intelligence for
Computer Games’, Artificial Intelligence for Computer Games, (Ccl), pp. 1–200. doi:
10.1007/978-1-4419-8188-2.
Hagelbäck, J. and Johansson, S. J. (2009) ‘A Multiagent Potential Field-Based Bot for Real-
Time Strategy Games’, International Journal of Computer Games Technology, Volume
2009, pp. 1–10. doi: 10.1155/2009/910819.
Hunicke, R., Leblanc, M. and Zubek, R. (2004) ‘MDA: A formal Approach to Game Design
and Game Research’, AAAI Workshop - Technical Report, WS-04-04, pp. 1–5.
Madigan, J. (2015) The Psychology of Horror Games, The Psychology of Video Games.
Jenkins, H. (2004) ‘Game Design as Narrative Architecture’, in Harrigan, P. and Wardrip-
Fruin, N. (eds.) First Person: New Media As Story, Performance, and Game. US,
Cambridge, pp. 118–129.
Kirkland, E. (2009) ‘Storytelling in survival horror video games’, in Perron, B. (ed.) Horror
Video Games: Essays on the Fusion of Fear and Play. Jefferson, North Carolina:
McFarland & Company, Inc., pp. 62–78.
Kortmann, Rens & Harteveld, Casper. (2009). Agile game development: Lessons learned
from software engineering.
Kotrajaras, V. and Kumnoonsate, T. (2009) ‘Fine-Tuning Parameters for Emergent
Environments in Games Using Artificial Intelligence’, International Journal of
Computer Games Technology, Volume 2009, pp. 1–9. doi: 10.1155/2009/436732.
Krzywinska, T. (2002) ‘Hands-On Horror’, ScreenPlay. Cinema/Videogames/Interfaces.,
2(Fall), pp. 12–23.
Loh, P. K. K. and Prakash, E. C. (2009) ‘Performance Simulations of Moving Target Search
Algorithms’, International Journal of Computer Games Technology, Volume 2009,
pp. 1–6. doi: 10.1155/2009/745219.
Milam, D. and El Nasr, M. S. (2010) ‘Design patterns to guide player movement in 3D games’,
in Proceedings - Sandbox 2010: 5th ACM SIGGRAPH Symposium on Video Games,
pp. 37–41. doi: 10.1145/1836135.1836141.
Nicolas Brown (2020) Why Do We Enjoy Horror Games?, TechRaptor.
Ntokos, K. (2018) ‘“Level of Fear”: Analysis of Fear Spectrum into a Tool to Support Horror
Game Design for Immersion and Fear’, An International Journal (CGDEIJ), 1(1),
pp. 33–43.
Palazzi, C. E. and Aiolli, F. (2009) ‘Enhancing Artificial Intelligence on a Real Mobile
Game’, International Journal of Computer Games Technology, Volume 2009, pp. 1–9.
doi: 10.1155/2009/456169.
Petrillo, F., & Pimenta, M. (2010). Is agility out there? Proceedings of the 28th ACM International
Conference on Design of Communication - SIGDOC ’10. doi:10.1145/1878450.1878453.
El Rhalibi, A., Wong, K. W. and Price, M. (2009) ‘Artificial intelligence for computer games’,
International Journal of Computer Games Technology, Volume 2009, pp. 1–3. doi: 10.1155/
2009/251652.
Riedl, M. O. (2012) ‘Interactive Narrative: A Novel Application of Artificial Intelligence
for Computer Games Artificial Intelligence in Computer Games’, in Proceedings of
the Twenty-Sixth AAAI Conference on Artificial Intelligence. Atlanta, Georgia, USA:
Georgia Institute of Technology, pp. 2160, 2165.
Rouse, R. (2009) ‘Match Made in Hell: The Inevitable Success of the Horror Genre in Video
Games’, in Perron, B. (ed.) Horror Video Games Essays on the Fusion of Fear and
Play. Jefferson, North Carolina: McFarland & Company, Inc., pp. 15–25.
sacredgames, (2017) “Reinforcement Learning: Q-Algorithm in a Match to Sample Task –
Machine Learning In Unreal Engine,” sacredgames, December 19. https://unrealai.
wordpress.com/2017/12/19/q-learning/ (accessed Mar. 16, 2021)
Safadi, F., Fonteneau, R. and Ernst, D. (2015) ‘Artificial Intelligence in Video Games: Towards
a Unified Framework’, International Journal of Computer Games Technology, 2015,
pp. 1–30. doi: 10.1155/2015/271296.
Sikali, K. (2020) ‘The Dangers of Social Distancing: How COVID-19 can Reshape our Social
Experience’, Journal of Community Psychology, 48(8), pp. 2435–2438. doi: 10.1002/
jcop.22430.
Tanskanen, S. (2018) ‘Player Immersion in Video Games Designing an Immersive Game
Project’, p. 75.
4 Review on Using
Artificial Intelligence
Related Deep Learning
Techniques in Gaming
and Recent Networks
Mujahid Tabassum, Sundresan Perumal,
Hadi Nabipour Afrouzi, Saad Bin Abdul
Kashem, and Waqar Hassan
CONTENTS
4.1 Introduction.....................................................................................................66
4.2 Internet of Things (IoT)................................................................................... 67
4.3 IoT Infrastructure............................................................................................. 68
4.3.1 Components......................................................................................... 68
4.3.1.1 Identification Block............................................................... 68
4.3.1.2 Sensing Block........................................................................ 68
4.3.1.3 Communication Block.......................................................... 68
4.3.1.4 Computation Block............................................................... 69
4.3.1.5 Service Block and Semantics Block..................................... 69
4.3.2 Protocols.............................................................................................. 69
4.3.3 Applications......................................................................................... 73
4.3.3.1 Home Automation................................................................. 74
4.3.3.2 Smart Agriculture................................................................. 74
4.3.3.3 eHealth.................................................................................. 74
4.3.3.4 Logistics................................................................................ 75
4.4 Deep Learning in Gaming and Animation..................................................... 75
4.4.1 MotionScan Technology...................................................................... 77
4.4.2 Framework (Architectural Model)....................................................... 77
4.4.3 Appearance Model............................................................................... 78
4.5 IoT and 5G Technology................................................................................... 78
4.6 Artificial Intelligence....................................................................................... 79
4.6.1 Supervised Learning............................................................................80
4.6.1.1 Regression............................................................................. 81
4.6.1.2 Classification......................................................................... 81
DOI: 10.1201/9781003231530-4 65
4.7 Discussion on Machine Learning and Data Analytics in IoT Networks......... 83

4.8 Trends and Challenges..................................................................................... 87
4.9 Conclusion....................................................................................................... 88
References................................................................................................................. 88
4.1 INTRODUCTION
Thanks to Internet technology which has brought various new concepts and opportu-
nities to life. Today the Internet of Things (IoT) concept has become a fast-growing
industry in developed and developing countries. Due to Internet availability, we can
access instant real-world devices and digital information anywhere at any time. With
the Internet’s help, most of our daily usage things are connected us while we are
away from home; for example, watches, home TVs, kids monitoring, house moni-
toring, air conditioners, car and many more. Therefore, IoT industry’s rapid growth
needs reliable IoT applications, protocols, and platforms to support future industries
such as smart cities, smart homes, smart appliances, gaming, e-health, logistics, and
intelligent forming requirements.
IoT can be recognized as a network of neighboring things that connect via Internet
to share collected information without requiring human-to-human or human-to-
computer interaction. It is a network combination of many tiny sensor devices which
sense and collect various information depending on their characteristics and trans-
mit to the respective user via a wireless medium. The environmental and physical
parameters are sensed and monitored to take pre-defined actions. Users can access
live data, instantly notified to take appropriate measures.
IoT networks offer several benefits to various industries. According to an IDC
forecast, there will be 41.6 billion IoT devices connected over the Internet by 2025,
which will generate approximately 79.4 zettabytes (ZB) amount of data [1].
IoT systems include numerous business applications to offer many benefits to
humankind. The new Industrial IoT and the fourth digital revolution (Industry 4.0)
provide designers and implementers’ versatility and increase their decision-making
abilities using IoT and machine learning (ML) developments. Besides ML, cloud
computing services provide more benefit to companies and individual in term of
usability and productivity.
Artificial intelligence (AI) and ML techniques facilitate the communication
between systems, allowing them to make their own autonomous choices [2]. A sim-
ple IoT Network is used to offer self-optimized networks service. The individually
designed network is expected to be configured for massive data transfer and recep-
tion between a various autonomous device with time and channel free freedom. The
connected devices will determine and calculate the shortest flow path for success-
ful transmission. AI make these devices more efficient and communicate with each
other effectively.
In the field of gaming, opponents of human players often have advanced AI
involving genetic algorithms (GA). Strategies used in the past are being programmed
using AI to ensure that AI can learn and develop from past experiences. Learning
techniques allow AI to repeat past mistakes, enhance the game. This allows for a
more realistic experience for human players as they need to change their strategy
Review on Using AI Related Deep Learning Techniques 67
over time. It also helps to avoid situations where the human player finds a sequence
of moves that ultimately leads to success, meaning that the game no longer becomes
a challenge. GA an instance is needed to illustrate the challenges in the solution
situation and the optimal process of determining the quality of the example. The fit-
ness function first recognizes the variability of a unit and determines its properties.
In addition, these special functions are tailored to the problem domain. The fitness
function can only be a system timing function in most cases, especially code optimi-
zation. As genetic representation and fitness functions are determined, GA create an
initial candidate sample that will then apply multiple repeat operators with options,
crossovers, and options to increase the value of the candidate’s fitness.
This chapter has discussed the deep learning (DL) involvement in IoT networks,
components, application, protocols, gaming and expectation from 5G network. We
have discussed the various ML; DA algorithms and concepts enhance the IoT net-
work productivity and efficiency. In the last part, we have discussed the future trend
and challenges.
4.2 INTERNET OF THINGS (IoT)

In the last two eras, IoT-based networks have become famous and utilized in many
industries such as agriculture, medical, manufacture, and so on due to their usabil-
ity and performance. Wireless Sensor Network (WSN) is one example of IoT base
network that can be utilized in many areas. Throughout the years, WSNs have been
implemented for smart agriculture, emphasizing environmental tracking, precision
cultivation, system, process management automation and traceability. WSN’s offers
self-manage, self-configure, self-diagnosis, and self-heal rendered as intelligent agri-
culture monitoring network. The WSN is a device made up of transceivers, cameras,
microcontrollers, and radio frequency (RF) sources. However, when IoT became a
significant force in smart agriculture, it has shifted the model from using the WSN
for smart agriculture to IoT. The IoT integrates other established innovations, includ-
ing WSN, RF recognition, cloud storage, middleware platforms, and end-user soft-
ware [3]. The goal of IoT in agriculture and other industries is to empower farmers
with decision-making tools and automation technology that seamlessly combine
resources, information, and services to increase production, efficiency, and revenue.
Latest studies have centered on obstacles and restrictions faced by IoT and WSN
industries to manage and handle large amounts of data for a long time that required
new innovative models, analysis techniques, data protection and safety, and proper
data management approaches to take current obstacles. The usage of WSNs has
become a key component in several industries [3].
Multi-player online skill games are popularly played for money and offer a per-
fect form of fun and leisure. To ensure ethically responsible gaming, service provid-
ers that have interactive environments for playing social games for real money must
contribute to detecting and avoiding potentially dangerous addictive behaviors. To
put it another way, when playing cash, use care and decision [4].
Recent IoT solutions offer market share and provide competitive advantages and
usability for consumers to use it and promote preference. It offers a range of resources
and technologies, such as data collection and processing, system management,
convergence, surveillance, user visibility into processes, and capacity to recognize

and handle these apps. There are various IoT types, such as on-site applications,
which run on the same premises and networks as off-site systems that usually utilize
cloud storage. Companies also pick based on these templates.
4.3 IoT INFRASTRUCTURE
4.3.1 Components
Usually, IoT associated networks are a combination of the large number of devices
connected worldwide. IoT technology connects smart devices, gateways, data net-
works via cloud computing and applications. These intelligent devices are mostly
at various distance from each other under multiple scenarios and controlled by the
centralized management system to process and save the data or information. Entire
IoT infrastructure is made of various elements, blocks, components, and protocols.
IoT components consist of a sensing unit, a communication unit, computation, and
Internet (connectivity) unit and appropriate protocols and services. In addition to
featured features and application capabilities, IoT communication protocols, compu-
tational processing speed, and cloud services define IoT platforms, strengths, and lim-
itations [5]. The growing demand and involvement of IoT in recent industries signify
the interoperability improvement needs between various applications and services
as per user requirements. IoT infrastructure model consists of six blocks such as [5]:
4.3.1.1 Identification Block

Each IoT object must have a unique ID and connected to the IoT network infrastruc-
ture properly.
4.3.1.2 Sensing Block

Includes sensors to sense and observe the physical environment parameters. IoT
network’s main objective is to keep and access the physical environmental param-
eters in any industry mainly completed by sensor nodes deployed in vast or small
areas. These sensor nodes sense physical environmental values such as temperature,
humidity, air pressure, etc., and convert them into electrical signals and transmit. The
sensing system collects the sensed data and forwards to a connected network, such as
the data management system. Other algorithms are then applied for data validation
and visualization. Mostly information is transmitted via Internet and saved via cloud
computing infrastructure. Sensing unit also contains actuators which are connected
by wires or wireless medium as part of network objects. Actuators are mechanical
hardware devices such as switches, gateway, etc. that used in IoT platforms. These
devices get the user’s instruction or the sensors and perform appropriate actions such
as open/close windows, doors, curtain, light on/off, etc. Actuator devices convert
electrical signals into physical measures.
4.3.1.3 Communication Block

To transmit the data over a wireless or wired medium. The communication unit is
another aspect of IoT network. All connected sensor nodes should communicate
with each other especially with the upper system that handles the collected data. A
gateway or bridge is used in a condition when connected devices cannot adequately
communicate with other systems via a specific protocol. Therefore, a gateway is
used to communicate among various network via communication protocols. The IoT
devices used several communication technologies such as WiFi, ZigBee, Near Field
Communication (NFC), Bluetooth Low-Energy (BLE), Long Term Evolution (LTE),
LoRa, SigFox, NarrowBand (NB)-IoT, etc., to connect with the Internet. Hence, in
IoT networks identification methods are considered better than authentication in
terms of performance gains. For different recognition technologies, each entity is
unique through its recognition [6]. There are several communication protocols used
in IoT networks such as Constrained Application Protocol (CoAP) and Message
Queue Telemetry Transport (MQTT) to connect IoT objects with the data manage-
ment system.
4.3.1.4 Computation Block

Responsible for computing and processing data. In the IoT network, mostly data
is stored on the cloud network to access it from anywhere and anytime. There are
various protocols, technologies used for data management services. Sensing devices
send the data to the respective data server mostly stored on the cloud and perform
computational tasks.
4.3.1.5 Service Block and Semantics Block

Offers Internet connectivity and data analytics and ML techniques to get relevant
results from the collected data. These techniques help the user to visualize, manage
and secure the appropriate data. As a service, networking is offered by the free use
of apps. Data obtained by the sensors and devices can be used for analytical tools in
application creation.
4.3.2 Protocols
IoT protocol can be divided into two basic types such as IoT Network Protocol
related to managing the network traffic and IoT Data Protocols related to managing
data availability. In the IoT network, the connected devices are not entirely machine-
to-machine M2M systems because M2M communication devices directly commu-
nicate. In M2M communication, devices such as sensor, actuators, and embedded
systems capture or sense the data and share with other connected devices. For exam-
ple, they controlled electrical applications like bulbs, air conditioner, fridge and fan
RF or Bluetooth technology using smartphones or remotes. Therefore, electrical
appliance and smartphones considered as two different machines which are interact-
ing with each other. However, in IoT networks, physical devices are embedded with
sensors, software, and electronics to communicate using a cloud base or Internet-
related network. IoT networks are about sensor automation and Internet platform [7].
Figure 4.1 shows the concept difference of M2M and IoT.
In IoT network, most of the system parts need to be configured, maintained, and
monitored to offer expected services and data management [7].
FIGURE 4.1 Difference of M2M and IoT [7].
CoAP is a new communication protocol mainly designed for IoT networks, low
power devices. It allows connected devices to communicate among themselves
and with the Internet using similar protocols. Its lightweight generator traffic and
intended to use for resource-constrained Internet devices such as sensor nodes. It is a
service layer protocol that offers one to one communication like Hypertext Transfer
Protocol (HTTP) and does not support TCP/IP. It uses User Datagram Protocol
(UDP) over IP to offer efficiency as compared to HTTP. It uses fewer resources than
HTTP and provides more observation, execution, discovery, reading, and writing.
CoAP is the best choice for web services based [8].
MQTT is another communication protocol that is used in IoT networks. It was
developed in 1999 by Arlen Nipper (Arcom) and Andy Stanford-Clark (IBM) to
collect data from various electrical devices. It is mostly used for monitoring from
a remote area in IoT networks [8]. It is implemented over TCP/IP as a lightweight
communication protocol and work on a hub-and-spoke architecture concept. The
correspondence between devices uses a message brokers’ server that does not make
an M2M communication platform. It utilized three elements such as subscriber, pub-
lisher, and a broker. MQTT support Secure Sockets Layer (SSL) and Transport Layer
Security (TLS) for security services. In WAN IoT base networks, MQTT is consid-
ered better to use due to its broker concept. The broker is a center point of contact
between sensor devices. MQTT protocol is a favorite choice for all IoT based devices
because it provides ample routing information functions to cheap, memory-intensive,
and small devices on low and poor bandwidth-based networks. It is useful for poor
bandwidth networks, especially in numerous remote locations. Therefore, Amazon
and Microsoft Azure are using the MQTT protocol for their services.
Bluetooth is one of the most used short-distance wireless technologies used in
IoT networks. User can get Bluetooth applications that offer wearable technology
to connect smart devices quickly. One of the newly developed Bluetooth protocol
for IoT networks is known as Bluetooth Low-Energy (BLE) protocol as shown in
Figure 4.2. It offers the same services as Bluetooth with lower power consumption.
However, BLE is not designed to transfer large files and is mainly preferable for
small file size [9].
FIGURE 4.2 Bluetooth Communication.
WiFi is another famous and favored IoT network-related protocol and accepted
by many industries because it offers fast data transfer speeds and many data. The
large WiFi 802.11 standard allows user to transmit hundreds of megabits in a
second only. However, WiFi technology consumes a considerable amount of bat-
tery that is one of its drawbacks. It operates on 2.4 GHz and 5 GHz frequencies
band [9, 10].
ZigBee is another protocol which is mainly designed for IoT related industries as
compared to the consumers. It operates mostly on the 2.4 GHz frequency band with
a smaller data transmission rate [9, 10].
The Data Distribution Service (DDS) is another standard protocol choice for
the high-efficiency, expandable and efficient machine-to-machine communication.
It is developed by Objective Management Group. Users can transfer data on both
low-scale devices and cloud services using DDS. It used two important layers such
as Data-Centric Publish-Subscribe (DCPS) and Data Local Reconstruction Layer
(DLRL). The DCPS used to deliver information to the subscribers, and DLRL offer
an interface to DCPS [9].
NFC is another IoT related technology that is mostly used in the latest mobile
phones. It allows users to connect with electronic devices to use digital contents to
perform contactless payment transactions. It mainly emphasis on contactless com-
munication between electronic devices. It has a minimal distance range of 4 cm
between both electronic devices [9].
Many IoT applications are available to call for service over a longer distance. The
use of cellular networking technologies such as 4G/5G is possible in such IoT appli-
cations. Cellular is an IoT Communication Protocol that can send or transmit a large
quantity of data over the longer distance. However, users must understand the cost
of the cellular network, which might be quite expensive and consume more power.
Therefore, SparqEE has introduced a cellular kit with a name CELLv1.0 which
can be used in Arduino and Raspberry Pi platforms to get the cellular technology
benefits in IoT related networks. CELLv1.0 used a 2G+3G chipset to communicate

worldwide with longer distance and lower battery consumption [11].
Another recent IoT-related protocol is Advanced Message Queuing Protocol
(AMQP) known as application layer protocol and designed for middleware envi-
ronment. It is a message-oriented protocol and has three main components such as
Exchange, Message Queue and Binding. The Exchange segment operates by arrang-
ing the message in the queues. The Message Queue is responsible for processing the
directive until the client’s app safely creates the message. The linking element of the
exchange component indicates the relation to the message queue portion [9].
LoRaWAN is a wide area network protocol which is also used in IoT network.
It is mainly developed for low power devices to serve many node networks primar-
ily used in heavy networks such as smart cities or similar concepts. LoRaWAN is
also recognized in many industries for protecting bi-directional communication.
LoRaWAN frequency may vary between networks with various data rates of 0.3 –
50 kbps in IoT networks. Data transmission range is between 2 km and 5 km in the
urban areas approximately 15 km in the suburban areas [9].
The Radio Frequency Identification (RFID) is another wireless technology used
in low power IoT short-range networks. RFID used the Active Reader Active Tag
system that does not require any power. This technology used an electromagnetic
field to identify the smart device or objects and the Radiofrequency communication
range is around 10 cm to 200 mm [9].
Z-Wave IoT networks Protocol communicate on low-energy radiofrequency and
mainly used in home automation systems. This low latency technology also has more
wireless security features and is used for low power RF devices such as sensors,
lamp controllers, etc. Z-Wave used 900 MHz frequency and have a 30–100 m range
with 40–100 kbit/s data rates [9].
On top of all, SigFox is recognized as one of the best solutions with cellular
and WiFi characteristics. It is designed for M2M applications to transmit low-level
data. It can handle 10–1000 bits per second transfer speed using Ultra Narrow Band
(UNB) and consume only 50 microwatts of the power. It also runs on 900 MHz fre-
quency and has cloud access. SigFox ranges are between 30–50 km in rural areas and
3–10 km in the urban areas.
Thread is another new IoT security protocol created by Thread Group and getting
famous in-home automation applications. It is IPv6 network protocol based on the
6LowPAN. It was mainly designed to complement indoor Wi-Fi networks with the ben-
efit of free protocol. It supports mess topology structure and uses IEEE802.15.4. It runs
on 2.4 GHz (ISM) frequency band ad cover 10–20 m distance for data transmission. It
can handle approximately 250 nodes by having encryption and authentication features.
EnOcean takes an innovative spin under the IoT Protocols. It is ideal for design-
ing networks that need quick response in various circumstances such as tempera-
ture changes, humidity, lighting, and others. It runs on multiple frequencies such
as 315 MHz, 868 MHz, and 902 MHz and covers 300 m communication range for
outdoor and 30 m range for indoors networks. This protocol is currently being used
in various industries such as transportation, home and industrial automation, logis-
tics, and others [9]. In Figure 4.3, we can observe the operating layer model of IoT
protocols and their services [12].
FIGURE 4.3 IoT operating layer model.
4.3.3 Applications
IoT field is still emerging technology, but it helps many industries play important roles
[13]. While IoT is an evolving area, it has helped to enhance numerous applications that
have changed our lives in many ways. The IoT-I project 2010 survey described 65 IoT
scenarios covering 12 areas: travel, smart home, smart city, lifestyle, gaming, agricul-
ture, supply chains, emergency services, public health care, user engagement, culture
and tourism, climate, and environment [14]. The IoT analytical web portal results show
the growing number of IoT projects in various worldwide regions [15]. In Figure 4.4,
we can observe the ever-increasing number of IoT involved in multiple industries.
Smart City and Industries are on top with the heavy usage of IoT-related networks.
FIGURE 4.4 Top 10 IoT involved industries.

4.3.3.1 Home Automation

Using IoT network in the home environment helps manage and control user life
and enhance daily life product productivity. Individual users and companies can
get any data that could be used for many useful purposes such as understanding
the user health, daily activities, electricity usage, kids monitoring, monitoring com-
mon house problems, etc. Furthermore, the bedrooms’ heating can be adjusted to
the needs and conditions of the environment using IoT networks. Rooms lighting
and air-conditioner usage can be controlled and monitored on a timely basis. IoT
devices will automatically off/on and maintain electrical appliances such as fridges
and air-conditioners to save electric power. The washing machine consisting of mas-
ter RFID readers and slave readers will recognize dirty clothes and wash them auto-
matically using RFID tags to collect resentful information on color, contents, and
a suitable washing program. In kitchens for the cooking equipment’s, to automate
similar cooking technologies can be used. IoT network also can boost home security
to a large extent. High-level security devices such as smoke alarm, camera, and an
infrared detector, can be used with the latest security dynamics. The user can view
the home anytime from anywhere using an Internet connection [13].
4.3.3.2 Smart Agriculture

Today, the agriculture industry is also expanding and taking advantage of new indus-
try standards and platforms with the Internet and technology growth. The develop-
ment of IoT networks has reformed the entire agriculture industry into a new era.
In a traditional agricultural concept, crops are planted into a large paddy field area,
and farmers used to inspect regularly crops parameters such as growth, diseases
attacks, water level, humidity, temperature and so on, by visiting the paddy field
themselves which was a solemn and hectic task especially in a bad weather condi-
tion [13]. However, now with IoT networks and WSN, these monitoring tasks have
become easy for farmers. They do not need to regularly visit the paddy field and
periodically monitor the data from home using IoT-based networks. Furthermore,
using AI and Data Analytical knowledge, the deployed networks can predict future
data trend base on the collected data figures [16]. The weather, humidity, wind, pres-
sure, and other details can be obtained in real-time with accuracy using WSN. This
information could use to improve agriculture crop production and quality control.
For example, by understanding future weather information, farmers can choose
ways and precautions to enhance and control the fruits and vegetables. Besides, by
monitoring soil moisture, humidity, and related information, farmers can control the
amount of sugar in grapes plants and grapevine health. Various companies produce
several kinds of industrial hardware, systems, and protocols implemented for agri-
culture monitoring [17, 18].
4.3.3.3 eHealth
Health industry also getting optimal benefits from the usability of the IoT network.
Due to the ubiquitous IoT network capacities, all devices connected can be tracked
and monitored to collect respective and live data [19]. Using WSNs, patients’ live
medical data can be collected from anywhere and anytime, saving lives and under-
standing the variable parameters. Several AI-based smart applications such as live
heath monitoring or fall detection, diet monitoring, body mass index (BMI) monitor-
ing, blood pressure and heartbeat monitoring helps older people and disabled people
to live independently. With the help of these smart devices and applications, doctors
can continuously and effectively monitor patients’ health over the longer distance.
The growing rate of ageing has created many healthcare problems in today era. For
instance, some countries have old-houses and rehabilitation centers that manage old
age people or sick people and offer health services. IoT networks help in these areas
to continuously control and monitor patients’ health parameters [13].
Google Health application was released in 2008 as a personal health monitor
application. The application allows users to share their personal health information
with health service providers voluntarily. The Google Health application gets input
from users or monitors the user’s health and generates a complete report on its health
record. Later, the application was upgraded, and the user’s records are synchronized
on the cloud for further analysis. In the current era, various patient monitoring and
hospital-orientated systems are designed, developed, and significant changes have
been made to incorporate multidisciplinary information fields [20].
4.3.3.4 Logistics
Like other fields, transport and logistics industries are also taking advantages of
IoT systems. IoT helps the transportation industry improve transport efficiency and
precision across the entire supply chain, real-time monitoring and object movements
tracking from source to destination using attached RFID tags. In addition, IoT pro-
vides promising solutions for transforming transport and can predict vehicles moving
in parking or on the road. For example, BMW, Honda, and other vehicle manufac-
ture companies used several sensors and tags to monitor the environment to provide
drivers with driving direction with an intelligent computer system. Moreover, other
transportation and logistics operations such as routes control, warning emission,
track monitoring, etc., can also be enhanced by IoT [13]. Overall, IoT applications
can be grouped into 54 primary application containing 12 main domains due to their
usability and involvement in many industries [15].
4.4 DEEP LEARNING IN GAMING AND ANIMATION

Researchers from the University of Edinburgh and Adobe Research have created an
AI that video game characters can use to communicate more naturally with their sur-
roundings. This method employs a “Neural State Machine” (NSM) which is a deep
neural network that predicts a character’s movement in each scene. Character ani-
mation was very simple in the days of 8- and 16-bit video games, with most games
featuring a stable world and dialogue. As a result, the avatar action does not neces-
sitate a vast number of separate animations. The challenge of making animations
became more difficult with the introduction of 3D games. Now gaming technology
has grown to include a wide-open world and the potential to communicate with it,
animating a character requires hundreds or thousands of different movement abili-
ties as shown in Figure 4.5. Using motion capture (mo-cap) to digitize the actor’s
gestures into character-appropriate animation is one way to lighten and speed up the
saturation of the animation process. As a result, the game’s action seems to be more
FIGURE 4.5 IoT application and industries.
believable. However, understanding the single way the player communicates with the
world is almost impossible.
Furthermore, the switch between animations can look clunky and canned.
Changes between moves are usually done by reusable algorithms that work in the
same way every time. Consider how a character could sit in a chair or place a box.
This is made more difficult as the objects are in varying sizes. Picking up things
in various shapes and sizes or resting your arms on seats of various sizes may be
uncomfortable. A variety of moves should be considered and animated when moving
an object, including beginning to walk, slowing down, turning appropriately when
positioning the foot, and engaging with the object. At the same time, our.ng produces
a wide range of high-quality moves and actions from a single network. NSM has
been taught how to transition from one movement to the next in a normal manner.
Based on the previous poses and scenario geometry, the network advances the next
character.
Mo-cap is the process of recording people’s real-life movements with a cam-
era aimed at capturing those precise movements in a scene created by a computer
camera. As someone who is fascinated using this technology in game development

to create animations, it was great to see the major improvement achieved with DL. I
was delighted to see the great improvements that have been made to this technology
with the help of DL [21].
4.4.1 MotionScan Technology
That was back in 2011 when L.A. Noir really brought a great life-like face animation
that looks ahead of every other game. Now, almost a decade later, we still have not seen
many other games close to the level when it comes to displaying facial expressions [22].
L.A. in 2011 to create life-like face animations. MotionScan technology used by
Rockstar Studios in Noir. This is because the face scanning technology used in the
development of this game, called MotionScan, is very expensive and the file size of
the captured animation is very large, which is why most publishers use this tech-
nology for their games. However, this may soon change due to recent advances in
mo-cap driven by in-depth training.
In the following research work, the authors introduce a DL Framework for creat-
ing animations from the source image of a face, following the movements of another
face in a moving video like MotionScan technology. They propose a self-monitor-
ing training method that can use labelled video data without a specific category to
study the significant dynamics that define movement. Then show how those motion
dynamics can be combined with static imagery to create motion videos [23].
4.4.2 Framework (Architectural Model)

DL Framework architect template in Figure 4.6 shows a motion module and a visual
module [23]. Drive is the input of the video motion module and the source image is
our target object which is the input of the visual module.
FIGURE 4.6 First order model framework.

Motion module has an encoder that must study the incredible presentation which
contains important key points that are very important in relation to the motion of
the object, which is the face in this view. The motion of these key points in different
frames of the driving video creates an area of motion, driven by the function that we
want to study our model. The author uses the Taylor expansion to predict this process
from the first order that builds this plane of motion. According to the authors, this is
the first time a first-order approach has been used to model movement. In addition,
a solid motion field is created by combining the studied Eiffen transformations of
these keys. A solid motion plane predicts the speed of each pixel of distance, not
just focusing on the main points of the sparse motion plane. In addition, the Motion
Module also generates an acceleration map, highlighting the frame pixels that need
to be painted [24].
4.4.3 Appearance Model
Visual Module uses an encoder to encode the source image, which is combined with
a motion field and acceleration map to animate the source image. Generator models
are used for this purpose. During the self-examined learning process, a style frame
from the driving video is used as the source image and the studied motion plane is
used to scent the source image. Since the actual video frame serves as the truth for
the movement of the results, this is self-monitoring training. During the testing esti-
mate phase, this source image can be replaced with another image of the same object
category, and it does not need to come from the driving video [24].
There are several DL applications in the current market such as Self Driving
Cars, Entertainment, Visual Recognition, Virtual Assistants, and Natural Language
Processing.
4.5 IoT AND 5G TECHNOLOGY

The current 4G networks are continuously evolving by connecting various devices
and application. IoT was also utilizing the benefits of 4G networks for communica-
tion. It raises the demand for 5G networks with better speed and efficiency. Current
IoT solutions face various challenges, such as large number of connection nodes, lack
of security and standards that lead to the Internet’s future. The incoming 5G net-
work evolution was bringing significant growth in IoT applications. According to the
IDC study, 70% of companies worldwide would invest approximately $1.2 billion on
connectivity-management related solutions due to 5G networks in coming time. Thus,
due to emergence of IoT new technologies and potential business models needs new
performance standards and criteria, such as heavy convergence stability, security,
throughput, more extended cellular communications coverage, incredibly low latency,
and more efficiency. To fulfill new IoT standards, the 5G technologies are anticipated
to offer unique connectivity interfaces. The development of incoming 5G is still early,
aiming to run on new radio access technology (RAT) along with antenna improve-
ment, high frequencies utilization, and networks re-architecting [25].
Gartner expresses [25] that 20.4 billion IoT devices are expected to be connected
online by the end of 2020. The 5G evolution will bring more devices online to meet
FIGURE 4.7 Wireless network growth.
the market demand of IoT networks in the coming time, to stimulate new social devel-
opment and economics needs. The emerging conditions of potential IoT networks and
the advancement of 5G wireless technologies are two significant developments that
propel the 5G IoT [25]. At present, the IoT networks use 3G and 4G networks exten-
sively but not entirely optimized by IoT applications. The 4G generation refers to LTE
with significant capabilities to offer various Internet base services for IoT networks.
4G network are more reliable, fast, and provide consistent services to users than
other technologies such as BLE, WiMaxB, ZigBee, SigFox, LoRa, and others [20].
However, with the proliferation of IoT networks and applications, 5G networks are
expected to offer fast and sustainable Internet services. Figure 4.7 shows the mobile
network evolution from 3G to coming 5G enables IoT generation [25].
The new 5G networks are expected to operate exclusively on 4G LTE core net-
work to offer, data and Internet services with the speed of 10Gbps along connect-
ing thousands of devices. A lot of academic and industrial research are going on
to understand various aspects of IoT and 5G networks. The main objective is to
enhance the productivity of the IoT network using 5G technology. In a collabora-
tion project involving CISCO, Intel, and Verizon, 5G has unveiled a novel series
of “neuroscience-based algorithms” that adaptive video quality to the human eye’s
demand, indicating that the wireless networks are integrated into human intelligence.
The 5G generation will contribute to IoT success by connecting thousands of intel-
ligent devices and interacting without human participation. The 5G-IoT networks
are expected to offer real-time, on-demand, re-configurable services that required
the 5G-IoT architecture to provide smart operations on each phase. The 5G-IoT net-
works are expected to offer [25]:
• Offer a logically independent network as per application requirements.

• Provide a large connection of various standards and implement the on-
demand deployment of RAN by using cloud base radio access network.
• Simplify core network architecture to implement the on-demand configura-
tion of network functions.
4.6 ARTIFICIAL INTELLIGENCE

By considering industries and user expectation from the IoT, scientists and data engi-
neer brought AI concept to IoT [26]. AI is the science of teaching knowledge into the
machine to execute tasks that the human mind traditionally demanded and need time
FIGURE 4.8 Overview of AI.
to complete. Due to heavy demand for AI applications and equipment, implementa-

tions, configurations, operational speed, and functionality of AI-based systems are
developing rapidly. A variety of disciplines have been integrated to improve AI’s
interdisciplinary aspect, such as economics, computational sciences, engineering,
statistics, genetics, physics, sociology, psychology, and several more. Intelligence is
obtained from all the data produced in these areas. To understand the values behind
this data, proper interpretation is necessary.
AI strongly depends on data science, ML, and data analytic field. In simply, data
science is the study of designing techniques and methods to interpret and extract
knowledge from vast data quantities. Discipline is also a blend of several other fields of
study. Software developed concepts come predominantly from computer science, pri-
marily concerned with efficient algorithmic performance and data management scal-
ability. AI concepts come from far more different fields for research. AI techniques
used as a methodology for optimally utilizing data can be accessed by the individuals
and should be able to update and modify to describe in a meaningful manner [5].
Architecture and technology for digital devices built with Internet connectivity
in keeping with the IoT definition will be highly scalable. The network must be
supplied, keeping in mind the complexity of technologies and the constant need for
seamless connectivity. Just as in the IoT model interpretation, it is not obvious how
to construct its architecture. Many general IoT network solutions exist, but not all of
them are robust enough [5].
4.6.1 Supervised Learning

Supervised learning is one of ML most effective method of data analysis. The user
provides the set of inputs and outputs to find the relation between them through the
system training. Later, the user can get a function from an input x with the best esti-
mation of output y such as (f: x → y). The significant task of supervised learning algo-
rithms is to construct a model that describes relationship and dependency connections
between inputs and forecast objective outputs. It is used in many applications to solve
various problems such as in WSN. It helps to solve routing, localization, fault detec-
tion, data aggregation, congestion and energy harvesting related issues. It can be dif-
ferent categories into two sub-categories, such as Classification and Regression [26].
4.6.1.1 Regression
This method predicts some value (Y) based on a given set of parameters (X) along
with continuous or quantitative variables. It is a simple ML method used to predict
accurate results with fewer errors.
Linear Regression: The purpose of linear regression is to learn a function f that is
mapping with Y value. It can be represented with the following mathematical model.
In the equation, Y refers to the dependent variable (output), x represents independent
variable (input), f refers to a function that makes the relation between x and Y, and ε
deal with an expected random error.
Y = f ( x ) + ε
Support Vector Regression (SVR): The Support Vector Machine (SVM) method is
used to solve regression problems along a process known as SVR.
4.6.1.2 Classification
Classification can be further divided into various fields.
• Decision Trees (DT): DTs comes under supervised ML technique of clas-

sification using if-then rules, and it is used to improve the data readability.
A DL consist of two nodes known as leaf and decision nodes. The DT is
used to predict a class or outcome by creating a training model based on
decision rules inferred from training data. These techniques are more trans-
parent, reduce ambiguity in decision-making and offer a comprehensive
analysis. DTs are used to solve connectivity, anomaly detection, data aggre-
gation, mobile sink path selection related problems in WSNs. An example
of DT is shown in Figure 4.9.
FIGURE 4.9 Machine learning algorithm.

FIGURE 4.10 Decision tree example.
• Random Forest (RF): RF techniques are supervised ML methods that

used trees collection, and each tree represents a classification or value. This
algorithm works in two steps, the creation of RF classifier and prediction of
the results. These algorithms mainly used effectively in larger datasets and
heterogeneous datasets. These algorithms primarily used to solve coverage
and MAC protocol-related problems in WSNs.
• Artificial Neural Networks (ANN): ANN is another supervised ML tech-
nique using human neuron model to classify the data. It contains many
neurons to process information and produce accurate results. It used three
layers that relate to nodes, and each node is associated with an active func-
tion. Each ANN connected with three layers such as input layer, hidden
layer, and output layer. ANN easily can classify complex and non-liner data
sets without any input instructions. ANN techniques are used to enhance
the efficiency of WSN is several areas such as localization, node failure
deduction, congestion, routing, and data aggregation.
• Deep Learning (DL): DL is another ML technique used for classification and
comes under ANN. DL use multi-layers’ data representation and composed
with a simple non-linear module. It can work with or without labels to extract
high-level objectives from the data. DL methodology is used in various domain
such as Business Intelligence, Social Network Analysis, Biotechnology,
Medical and many more. It also being used in WSNs to solve routing, data
quality estimation, energy distribution, and fault detection related problems.
• Bayesian: Bayesian is another supervised ML technique that used a statisti-
cal learning method to find the dataset’s relationships by learning the con-
ditional independent using various statistical methods such as Chi-square
test. It takes the set of inputs X1, X2, X3, X4,……, Xn and returns a label
θ, the probability P(θ | X1, X2, X3, X4,……, Xn) to be maximize. It uses
various probability functions for different class nodes variables. It has been
used in WSNs to solve event detection, routing, data aggregation, tracking,
path selection and synchronization aspects.
• Support Vector Machine (SVM): SVM is also supervised ML techniques
that search an optimal hyperplane to classify the data. It is a non-probabilistic,
binary classifier which mainly use to find the dividing hyperplane that sep-
arates both training classes with maximum margin. SVM is suitable for
interacting with learning tasks in which the number of training instances is
large. SVM techniques are used in WSNs to solve localization, fault detec-
tion, congestion, routing, and localization related problems.
TABLE 4.1
Overview of ML Algorithm used in IoT Networks
Overview of ML Algorithm used in IoT Networks
ML Algorithm IoT Network Use Case Data Set to Optimize

Classification Smart Traffic Traffic Prediction
Clustering Smart Traffic, Smart Health Traffic Prediction
Anomaly Detection Smart Traffic, Smart Environment Traffic Prediction
Support Vector Regression Smart Weather Prediction Forecasting
Linear Regression and Regression Economics, Market Analysis, Smart Real time Prediction
Trees Citizen
SVM Various Data Real time Prediction
k-NN Smart Citizen Passengers Travel Pattern
Naïve Bayes Smart Citizen, Agriculture Food Safety, Passengers
Travel Pattern
• k-Nearest Neighbor (k-NN): It is one of the straightforward laziest instance-

based learning techniques in regression and classification. In this algorithm,
the K nearest training set is considered input from the feature space and clas-
sifies based on the distance between specified training samples and sample
tests. It utilized different distance functions such as Canberra distance,
Euclidean distance, Manhattan distance, Hamming distance, and many
more. This method’s complexity depends on the size and optimal output of
the input dataset for the same data scale. This method defines function space
for potential missed values and therefore eliminates dimensionality. It mainly
used to optimize fault detection and data aggregation problem in WSNs.
Table 4.1 shows the overview of ML algorithm used in IoT Network [27, 28].
4.7 DISCUSSION ON MACHINE LEARNING AND

DATA ANALYTICS IN IoT NETWORKS
ML and Data Analytics are two most prominent techniques that have been used in all
industries to optimize the outcome and future predictions. Both methods are vastly
used in various sectors, especially have become common in IoT networks. In co-oper-
ative ML and IoT, better connectivity and computational efficiency, controllability
and decision-taking can be achieved. IoT has enormous potential to improve the qual-
ity of human life and the potential industrial application productivity, with the help
of advanced monitoring from thousands to billions of ubiquitous sensor devices and
improved communication capabilities. Through converging ML, AI, and data analyt-
ics, IoT ability has been dramatically increased. Advanced technology in machine
intelligence has enabled a large volume of IoT sensory data to be better understood and
take critical operational decisions on a range of real-world problems. To resolve world
complex problems and satisfy the computation and communication needs, IoT and
ML must balance each other and effectively fulfill them. The popularity and demand
for data analytics also have increased in recent years, along with the ML algorithm.
ML integration in the IoT field has opened new productivity, efficiency, economical,
and accuracy IoT solutions. ML usability in IoT applications improves communica-
tion and computational performance, better controllability, and enhances decision-
making skills. IoT networks consist of thousands to billion ubiquitous sensing devices
to offer reliable and consistent communication to improve and secure society lifestyle
and industry workflow. Recently, IoT application has significantly improved with the
convergence of ML and AI techniques. Latest ML and AI techniques allow user
easily to observe the collected data and make critical operational decisions rapidly.
Therefore, in many industries, ML and AI techniques are being used to make the
rapid and successful decision based on collected data. Today, Data Analytics has got
significant attention in IoT network due to the following reasons [29]:
• Massive Volume of Data Generated from IoT devices: Large number
of devices are connected within a one IoT network. They are producing
high volume data that make hard for the user to maintain and understand.
Therefore, it is essential to use intelligent data analytic techniques to extract
relevant data rapidly.
• Variability in Huge Volume Data Collected from Heterogeneous
Devices: In IoT network, many devices are connected and processing the
information among themselves. Due to IoT network heterogeneity, data qual-
ity, processing, and storage have been considered challenging tasks to handle.
• Data Uncertainty: In IoT networks, devices can change their characteris-
tics due to certain limitations such as battery drain, network failure, path
loss, interference, and others. Uncertainty is considered standard in practi-
cal data analysis. Data loss is always present in IoT, which needs efficient
analytics techniques to pre-process the data. It is essential to use proper
analytical techniques and assessment models to enhance decision-making
accuracy based on collected data.
• Balance Scalability with the Efficiency: Normally, IoT networks store
the collected data on the cloud and perform analytics to make decisions.
Moving data from smart devices to the cloud could cause a delay in trans-
ferring between different networks and speed. This could be a challenging
task for time-critical application, especially in many IoT connected devices.
Therefore, it is essential to manage and balance data analytics techniques
accuracy and speed in many IoT networks.
Data Analytics as shown below in Figure 4.11 can be classified as following [29]:
• Descriptive Analytics: Normally, IoT networks collect a large amount of
continuous data from a large, deployed area via smart devices. This informa-
tion is stored on the cloud for user interaction. The user can generate detailed
observation and opinions based on historical data collected from smart
devices using appropriate ML algorithms. Descriptive analysis is defined
as the method used to studies and interprets the raw data into meaningful
information. Data aggregation, data mining, mathematical, logical opera-
tions, data summarization, and others are examples of descriptive analytics.
FIGURE 4.11 Data Analytics.
• Predictive Analytics: On the other hand, Predictive analytics utilizes his-

torical data to model actions or patterns by utilizing predictive statistics
or ML techniques to forecast the probability of possible future events or
anomalies in data. In simple, it predicts future data by observing past data
patterns and data correlations. It has been used in various applications to
forecasts the future. Two major predictive methods, such as classification-
driven models, conduct prediction studies through class membership and
regression base models, are used to predict a number based on past obser-
vations and patterns. Furthermore, predictive modelling techniques include
statistical regression-based models, Bayesian Analysis, DTs, and Deep
Neural Network-based models.
• Prescriptive Analytics: This type of analysis briefs users on responding to
future events based on past data analysis. It forecasts future outcomes with
proper justifications. It is one of the production techniques that leverages
the benefits of both descriptive and predictive analytics. These analyses are
mainly used to optimize business outcomes by looking at the past and pre-
dicting future events. These techniques help the Industrial IoT setup, where
business intelligence-related decisions are utilized with big data analytics
and ML. Table 4.2 shows the overview of various applications using proto-
cols, ML techniques, and analyzed data types [27, 29].
• Adaptive Analytics: The findings obtained from predictive analytics must
be adjusted for real-time data during actual implementation. Therefore,
adaptive analyses are used to adapt or optimize processes based on the
latest data history and by looking at their interactions. Adaptive analysis,
especially in the IoT paradigm, can be useful for processing data in real-
time streams and optimizing model efficiency and minimizing errors.
To arrive at the correct decisions for intelligent data analysis, it is essential to deter-
mine the process needs to be completed out of framework exploration, locate odd
data points, predict values, predict categories, or extract features. Clustering algo-
rithms can provide the most appropriate tools for discovering the structure of unla-
beled data. K-means is the most popular and often used clustering algorithm and
can handle a wide range of data types with a large volume of data. Furthermore,
to identify patterns and odd data points in smart data, two algorithms can be used,
such as SVM and principal component analysis (PCA). Both algorithms used to
offer effective results in anomalies and noisy data sets. The linear and SVR methods
TABLE 4.2
Overview of IoT Applications, Protocols and Algorithm
Application Protocol ML IoT Data

Predicting energy consumption MQTT, CoAP RF Temperature, energy
through machine learning using a consumption,
Smart-Metering Architecture humidity, wind
speed – with
15 min interval
Smart City Traffic Monitoring MQTT Azure ML Light, temperature,
System based on 5G Cellular fuel consumption,
Network, RFID and ML. magnetic, time to
reach destination
Inferring occupancy from building MQTT RF CO2, HVAC
sensors using ML temperature, air
volume
Early detection system for gas MQTT Classification and Temperature, smoke,
leakage and fire in Smart home Regression Trees gas, CO2 and CO
using ML flame, humidity
Vibration condition monitoring CoAP Neural network Motor vibration level
using ML
Fall detection system for elderly MQTT, CoAP, RF Gyroscope data,
eople using IoT and ensemble ML HTTP Accelerometer data
Delay-Sensitive Health Care Data CoAP Deeping Learning Oxygen level, plus
analytics at the Network Edge rate, respiratory
based on DL rate, sleep
condition, gait
tracking, washroom
visit frequency
Quantifying colorimetric tests AMQP Support Vector Blue, average, red,
using a smartphone app based on Machines green values
ML classifiers
IoT and distributed ML powered MQTT Kalman filter Heartbeat, location,
optimal state recommender solution acceleration
An IoT-based solution for MQTT, Rule engine Animal posture
intelligent farming AMQP
An IoT system to estimate personal AMQP Support Vector Heart rate, humidity,
thermal comfort Machines air speed, skin &
room Temperature,
Integrating IoT, streaming analytics AMQP Azure ML Blood sugar data
and ML to improve intelligent
diabetes management system
Application of ML to an Internet of MQTT RF, k-NN Swaps per hour,
Cattle lying time, step
count
Cloud computing based smart MQTT Decision forest Garbage level and
garbage monitoring system regression times
are used to predict future values and to classify analyzed data. The classification
and regression tree used to classify intelligent citizens’ behavior is a further rapid
training algorithm. Neural networks are appropriate learning models for function
approximation problems to predict the categories of data [30].
4.8 TRENDS AND CHALLENGES

IoT implementations depend not only on their versatility in determining the extent
of connectivity steps but also on their quick reaction to environmental change. The
scheduling and predictability of IoT implementations with time restrictions may
therefore play an important role mainly. There are a variety of IoT applications
which must be reliable in addition to reliability. Appropriate Healthcare and avionics
facilities are sources for these systems. The term “predictable” applies to the mecha-
nisms that maintain a particular pause for the production values.
The development of IoT and wide-ranging data analytics has generated massive
opportunities. AI-based IoT network has created an enormous demand in society
and industries to solve several complex problems. The IoT system operating with
ML algorithms have better monitoring, analyzing, and controlling data capabilities.
Furthermore, it helps businesses and industries better understand upcoming chal-
lenges and make the right decisions. Although AI and IoT integration have opened
several potentials, many obstacles have constrained their growth due to many reasons
such as security, many ubiquitous devices, network speeds, protocols and many other.
To enhance productivity, improve decision-making, and perform future fore-
cast analysis, the AI algorithm mainly relies on the IoT networks data. It has been
observed that the application layer and transportation layer could be suspected of
cyber-attacks vulnerabilities in IoT networks perception layer. Therefore, it is nec-
essary to maintain a robust risk protection mechanism to protect the IoT network
against such cyber-threats. IoT equipment requires itself attention owing to the
absence of sufficient protection on certain apps. Security process in IoT networks
is different as compared to traditional networks. IoT networks combine tiny size
and ubiquitous devices with additional security measurement limitations due to low
memory and ubiquity [27]. Therefore, the compromise between lower capability and
higher security is often challenging to achieve in IoT networks. IoT network consists
of a wide variety of heterogeneous nature devices that increase security risk due to
mixing various vendors, types, and technological devices [31, 32].
To fulfill the Low-Power Wide-Area (LPWA) specifications, different systems
such as ultra-narrowband channels and transmission bandwidth are used. As more
approaches to LPWA problems are being built, innovations are needed to help higher
bandwidth, long-distance coverage, a high-path error connectivity budget and longer
battery life. Many NB-IoT cellular networks also allow frequency bands that run
duplex dividing modes. In bands running the duplex time division modes, further
research is required. These LPWA devise are required to have a robust energy-saving
algorithm to achieve a longer lifetime.
Quality of Service (QoS) is expected in all IoT network levels. The opportunity
to allow IoT usage in all connectivity systems through a system that wants to trans-
mit vital data remains an active field of study [26]. Like LoRa, the use of NB-IoT
networking technologies is promising robust QoS. Further research is required in the

implementation of a QoS system through multiple IoT network layers.
Most robust and strong ML and DL algorithms are required to read images and
extract the problem from them in the gaming and recent networks field to help the
related users. Many IoT devices are connected, advanced compression and multiplex-
ing techniques are required to share hundreds to thousands of data sizes. It is particu-
larly needed when NB-IoT cell connectivity is used for the transmitting of images and
video data. Furthermore, multiplexing allows for combining data from various sources.
In term of protocol comparison, MQTT has more reliable features as compared to
CoAP. For IoT computer systems, MQTT has greater efficiency and predictability;
on average, it needs 364uW higher energy consumption than CoAP to provide the
same data volume. In comparison with CoAP, 4.99x is also required when using
MQTT as a robust application layer protocol. Overall, the MQTT protocol is suitable
for IoT systems where power and efficiency problems are not considered. CoAP has
fewer reliability mechanisms and is an appropriate choice for real-time IoT systems
with high power and performance constraints.
4.9 CONCLUSION
IoT consists of a wide variety of linked appliances and shares enormous data vol-
umes. The IoT platform has been a part of our daily lives. However, IoT equipment
is constrained in computational and connectivity capabilities that are bottlenecks
in creating flexible, smart solutions using ML techniques. Although advancements
in technology and software upgrades pave the way toward a future consisting of
accelerated IoT expansion, device delivery and precise IoT data high Volume review,
we also concluded that it has proved challenging to combine intelligent technologies
from various fields. ML techniques play an important role in IoT networks success
and assist in product enhancement. Due to IoT networks continuous growth, many
improvements are required in security, data handling, processing, speed, protocols,
and platform connectivity. The optimized power consumption and reduction of sen-
sor size are expected to minimize the deployment cost. Future work is going on to
develop poor infrastructure and operating cost with better performance.
A few years ago, we would never have imagined an in-depth learning app that
includes driverless cars and Alexa, Siri and Google Assistant virtual assistants. But
today, these creations are part of our daily lives. DL is captivating us with endless
possibilities like fraud detection and pixel recovery.
REFERENCES
1. N.a, Help Net Security, 41.6 Billion IoT Devices will be Generating 79.4 zettabytes of
Data in 2025, 2019. Access date: 10 July 2020, Access Link: https://www.helpnetsecu-
rity.com/2019/06/21/connected-iot-devices-forecast/
2. Tabassum, M., & Mathew, K., A Genetic Algorithm Analysis towards Optimization
Solutions. Int. J. Digit. Inf. Wireless Commun. (IJDIWC), 4(1), 124–142, 2014.
3. Elijah, O., Rahman, T.A., Orikumhi, I., Leow, C.Y., & Hindia, M.N., An Overview of
Internet of Things (IoT) and Data Analytics in Agriculture: Benefits and Challenges.
IEEE Internet Things J., 5(5), 3758–3773, 2018.
4. Seth, D., Eswaran, S., Mukherjee, T., & Sachdeva, M. (2020). A Deep Learning
Framework for Ensuring Responsible Play in Skill-based Cash Gaming. In 19th IEEE
International Conference on Machine Learning and Applications (ICMLA), IEEE,
454–459.
5. Hejazi, H., Rajab, H., Cinkler, T., & Lengyel, L., Survey of Platforms for Massive
IoT. In 2018 IEEE International Conference on Future IoT Technologies (Future IoT)
(pp. 1–8). IEEE, January, 2018.
6. Tabassum, M., Perumal, S., Mohanan, S., Suresh, P., Cheriyan, S., & Hassan, W., IoT, IR
4.0, and AI Technology Usability and Future Trend Demands: Multi-Criteria Decision-
Making for Technology Evaluation. In P. Suresh (ed.), Design Methodologies and Tools
for 5G Network Development and Application (pp. 109–144). IGI Global, Hershey,
Pennsylvania, USA, 2021.
7. B. Mishra and A. Kertesz, “The Use of MQTT in M2M and IoT Systems: A Survey,” in
IEEE Access, vol. 8, pp. 201071–201086, 2020, DOI: 10.1109/ACCESS.2020.3035849.
8. AVSYSTEM, Internet of Things, “IoT vs M2M - What Is the Difference?”, 2019.
Access date: 10 July 2020, Access Link: https://www.avsystem.com/blog/iot-and-m2m-
what-is-the-difference/.
9. Fries, J. IoT Agenda, “Why are IoT Developers Confused by MQTT and CoAP?”, 2017.
Access Date: 19 July 2020, Access Link: https://internetofthingsagenda.techtarget.com/blog/
IoT-Agenda/Why-are-IoT-developers-confused-by-MQTT-and-CoAP#:~:text=CoAP%20
was%20started%20by%20a,it%20is%20in%20the%20lead.
10. Ubuntupit, M. Top 15 Standard IoT Protocols That You Must Know about, Access
Data: 25 June 2020. Access Link: https://www.ubuntupit.com/top-15-standard-iot-
protocols-that-you-must-know-about/
11. SpareqEE, “SparqEE CELLv1.0”. Access Date: 10 July 2020, Access Link: http://www.
sparqee.com/portfolio/sparqee-cell/.
12. Tabassum, M., & Zen, K., Performance Evaluation of ZigBee in Indoor and Outdoor
Environment. In 2015 9th International Conference on IT in Asia (CITA) (pp. 1–7).
IEEE, August, 2015.
13. Safaei, B., Monazzah, A.M.H., Bafroei, M.B., & Ejlali, A., Reliability Side-effects
in Internet of Things Application Layer Protocols. In 2017 2nd International
Conference on System Reliability and Safety (ICSRS) (pp. 207–212). IEEE,
December, 2017.
14. Sharma, T. & Tabassum, M., Enhanced Algorithm to Optimize QoS and Security
Parameters in Ad hoc Networks. In P. Suresh (ed.), Design Methodologies and Tools
for 5G Network Development and Application (pp. 1–27). IGI Global, Hershey,
Pennsylvania, USA, 2021.
15. IoT-I, Internet of Things Initiative, FP7 EU project, FP7-ICT-2009-5-257565.
16. Scully, P. IOT ANALYTICS, “The Top 10 IoT Segments in 2018 - based on 1,600
real IoT projects”, 2018. Access Date: 28 June 2020, Access Link: https://iot-analyt-
ics.com/top-10-iot-segments-2018-real-iot-projects/?_scpsug=crawled_5484401_
4ca49840-17a8-11e8-854f-f01fafd7b417.
17. Tabassum, M. & Zen, K., Signal Interference Evaluation of Eko Wireless Sensor
Network. In 19th International Conference on Transformative Research in Science and
Engineering, Business and Social Innovation (SDPS 2014), 2014.
18. Tabassum, M. & Zen, K., Evaluation and Improvement of Data Availability in WSNs
Cluster Base Routing Protocol. J. Telecommun. Electron. Comput. Eng. (JTEC), 9(2–9),
111–116, 2017.
19. Tabassum, M., Perumal, S., & Ab Halim, A.H., Review and Evaluation of Data
Availability and Network Consistency in Wireless Sensor Networks. Malaysian J. Sci.
Health Technol., 4(Special), 56–64, 2019.
20. Brendard, S., Tabassum, M., & Chua, H., Wireless Body Area Networks Channel
Decongestion Algorithm. In 2015 9th International Conference on IT in Asia (CITA)
(pp. 1–6). IEEE, August, 2015.
21. Chen, S., Gao, X., Wang, J., Xiao, Y., Zhang, Y., & Xu, G., Brand-new Speech
Animation Technology based on First Order Motion Model and MelGAN-VC. J. Phys.
Conf. Ser., 1828(1), 012029, 2021.
22. Saxena, S. MotionScan: Towards Brain Concussion Detection with a Mobile Tablet
Device. 2016.
23. Siarohin, A., Lathuilière, S., Tulyakov, S., Ricci, E., & Sebe, N. First Order Motion
Model for Image Animation, 2020. arXiv preprint arXiv:2003.00196.
24. Xue, W., Madonski, R., Lakomy, K., Gao, Z., & Huang, Y. Add-on Module of Active
Disturbance Rejection for Set-Point Tracking of Motion Control Systems. IEEE Trans.
Ind. Appl., 53, 4028–4040, 2017.
25. Yuehong, Y.I.N., Zeng, Y., Chen, X., & Fan, Y. The Internet of Things in Healthcare:
An Overview. J. Indus. Inform. Integ., 1, 3–13, 2016.
26. Li, S., Da Xu, L., & Zhao, S. 5G Internet of Things: A survey. J. Indus. Inform. Integ.,
10, 1–9, 2018.
27. Kumar, D.P., Amgoth, T., & Annavarapu, C.S.R. Machine Learning Algorithms for
Wireless Sensor Networks: A Survey. Inf. Fusion, 49, 1–25, 2019.
28. Mahdavinejad, M.S., Rezvan, M., Barekatain, M., Adibi, P., Barnaghi, P., & Sheth,
A.P. Machine learning for Internet of Things data analysis: A survey. Digit. Commun.
Networks, 4(3), 161–175, 2018.
29. Tabassum, M. & Mathew, K. A Genetic Algorithm Analysis Towards Optimization
Solutions. Int. J. Digit. Inform. Wireless Commun. (IJDIWC), 4(1), 124–142, 2014.
30. Chang, J.H., Tabassum, M., Qidwai, U., Kashem, S.B.A., Suresh, P., & Saravanakumar,
U., Design and Evaluate Low-Cost Wireless Sensor Network Infrastructure to Monitor
the Jetty Docking Area in Rural Areas. In Advances in Smart System Technologies (pp.
689–700). Springer, Singapore, 2020.
31. Liang, C.B., Tabassum, M., Kashem, S.B.A., Zama, Z., Suresh, P., & Saravanakumar,
U., Smart Home Security System Based on Zigbee. In Advances in Smart System
Technologies (pp. 827–836). Springer, Singapore, 2020.
32. Ali, A.B., Tabassum, M., & Mathew, K. A Comparative Study of IGP and EGP Routing
Protocols, Performance Evaluation Along Load Balancing and Redundancy Across
Different AS. In Proceedings of the International MultiConference of Engineers and
Computer Scientists (Vol. 2, pp. 487–967), 2016, March.
5 A Review on Deep
Learning Algorithms
for Image Processing in
Gaming and Animations
Sugandha Chakraverti, Ashish Kumar Chakraverti,
Piyush Bhushan Singh, and Rakesh Ranjan
CONTENTS
5.1 Introduction..................................................................................................... 91
5.2 Machine Learning...........................................................................................92
5.3 Deep Learning................................................................................................. 93
5.3.1 Important Architectures in Deep Learning.........................................94
5.4 Content-Based Gaming Image Retrieval (CBGIR).........................................94
5.5 Image Classification......................................................................................... 95
5.5.1 Adding Deep Learning to Neural Networks....................................... 95
5.5.2 Benefits of Image Classification.......................................................... 95
5.5.3 The Image Classification Theory........................................................96
5.6 Image Processing.............................................................................................97
5.6.1 Point Operations..................................................................................97
5.7 Image Processing using Deep Learning.......................................................... 98
5.8 Conclusion.......................................................................................................99
References............................................................................................................... 100
5.1 INTRODUCTION
Limited information content or any information that is unilateral yet associated with
the most of the images due to the differentiation in focal points when objects appear
to be blurred in any of the one image. Most of the cases, images lack valid informa-
tion or say, they are unilateral with the flow of the content. Multi-modal images are
categorized into their several types determined with austere as they reflect diverg-
ing information as per the types of images. In case of doctors, for an instance, at
times, they find the information too much disintegrated that it results in their mis-
understandings. Not only doctors, even several researchers associated in the field of
image processing, utilization, and gaming diagnosis found the same problem as well.
Throughout the recent years, several researches are continuously assembling with
the image processing faculties related to gaming and animation as it still continues
DOI: 10.1201/9781003231530-5 91
to prove its effectiveness. The image fusion solution automatically detects different
types of images along with proper integration of clear images. In short, the image
fusion is precisely structured algorithm that combines more than two images to cre-
ate a new one.
Image fusion is also known for producing multi-modal gaming images that
grabbed the attention of a wide range of associated audience and researchers in
quiet recent years later categorizing its domain into three potential classes such as
DL algorithm, transform domain algorithm, and algorithms for spatial domains.
Recently, the focus is majorly on the methods and techniques of DL encouraging
scholars from across the globe to conduct DL research even from their comfort zones
resulting in image processing and alternatives. Some of the recent analysis on such
DL include pixel-level image fusions (steps conclude-training, classification, weight,
and fusion), convolutional neural networks (CNN), convolutional sparse representa-
tion, stacked auto-encoders (SAEs), and deep Boltzmann machine.
Humans use their vision to collect information that mostly and significantly
depends on the image quality. Hence, Spatial resolution such as High-Resolution
(HR) is considered much effective for measuring this important attribute because
of their plenty delegation, prime features of shooting, and their artificial factors. It
is at times quite difficult for the imaging system to receive an information without
any interruption, deformity or changes. Therefore, in practical sense, it still remains
strictly limited.
Improving the quality of the images is not an easy or inexpensive task. It is quite
trending in research and engineering fields related to image sciences. The expensive-
ness in the field lies in the hardware device practically while approaching toward
increasing the superlative image resolutions. In this chapter, we therefore shall also go
through the image super-resolution technology used for improving the image quality.
Image super-resolution technology [1] is an image processing approach toward
acquiring clear and high-quality images that is believed to be replicated just from the
image of a single-frame that is actually structured low resolution (LR). The visual
effects of the image get on improving the image super-resolution that is also ben-
eficial during the extraction, recognizing of proper data, traffic, as well as security
monitoring.
The deficiency of conventional schematic learning and their over-simplification
intensify the image processing information with a rate of higher frequency so that
the comparison between the images shall be regenerated using the sporadic prac-
tices. This results in DL approaches for melding with the sublimed with resolutions
to coherently meliorate the interpretation for reconstructing.
5.2 MACHINE LEARNING

Machine learning is a necessary attribute during the current researches, discoveries,
and investigation processes. They are basically used in identifying hidden patterns
from the datasets, acquiring of quick and accurate predictions and results, and many
other advanced level research programs, just as it goes with the customers. Using
machine learning, their preferences can be collected from the pre-existing records
equivalent to near interests of other customers [2].
Review on Deep Learning Algorithms for Image Processing 93
Major studies such as artificial learning, DL, and machine learning has their posi-
tions in the areas of computing, real-time imagery, finance, medical science and
health, etc. Their urgeness to level up their performance can be broadly seen in open-
sourced frameworks during the research methodologies of practical and advanced
technologies, with a hope to touch the $24 trillion world market by 2024. Since
they find themselves available as well as a necessity in every often field these days,
these advanced learning and research are trying their best to reach out to the lowest
factor of errors. Therefore, government is discussing broadly to invest more on such
advanced learnings so as to be quick with the competitive flow on developing more
upon the data-set sectors and research-based dimensions [3].
5.3 DEEP LEARNING

After machine learning, deep learning (DL) technology is used massively in recog-
nizing the patterns and in the areas where there is a conduction of huge image pro-
cessing. A multi-mode gaming image fusion with DL will be proposed, according
to the characters of multi-modal gaming image, gaming diagnostic technology and
practical implementation, according to the practical needs for gaming diagnosis. The
usage cannot be only seen in MRIs, CT, and SPECT image fusions but also help-
ful during the issues related to multi-modal gaming image fusions. In addition, the
shows the sign of effectiveness in overcoming the limitations of one-page image pro-
cessing. In order to analyze the performance and stability of the multi-modal gaming
images, several experiments are conducted continuously based on visual quality and
the various criteria of their quantitative evaluation [4].
Throughout this review, the concentration is greatly upon the applications of DL
models that can be used for image processing, but it is quite hard to say how much
is the utilization made for the multi-modal batch processing. When compared to the
gaming images, the same results may show richness in their clarity. The data quan-
tity of a single image is maximized by the image fusions. Therefore, in this chapter
we tried to demonstrate the practical gaming problems, and resolving them, with a
tentative research using regulated DL methods.
The introduction of DL model into this chapter is primarily intended so as to
develop a new idea of image fusion. Using the practical methodology of DL, we can
acquire a new structure that shall become suitable for the fusion for advancing the
efficiencies and accuracies resulted from an image processing [5].
DL and neural networks (NN) are the two types of fundamental techniques that
are expansively used during classification as well as during prediction processes such
as the identification of images and objects, smart homes, automated machines, self-
driven cars, etc. Their popularity has risen vastly because they can easily adapt to
multiple data types over the domain variants.
DL generally acts as a human brain where it keeps on training a system that refines
the data inputs with the help of classification and prediction process, as described
above. This filtration process is basically a structure of different active layers where
each layer tries to act as a feedback for the adjacent or succeeding layer. There is a
continuation in the feedback cycle unless an accurate output i.e., the information, is
obtain from the process.
These days, DL techniques are commonly seen taking immeasurable help from
artificial neural network during image processing and extraction. Just like DL, NN
takes help of CNN for image processing. This CNN is automated in nature and
works similar to recurrent neural network (RNN), only exceptional with the fact is
the computational language is basically used by the RNN. This RNN therefore uses
the concept of feedback loops where the output of one layer is fed as the input of the
next layer.
There is hardly any visible limit to the discussions upon DL. DL is used so much
these days that it is even used in data processing units. Their active potentiality
encourages them to generalize huge data of the smart and metropolitan cities and
evaluate them in the areas of gaming and animation. DL in particular aids the faster
analysis for rendering an accurate diagnosis. Not only this, the complexity of the
health sector is very well managed using DL. This also diagnoses diseases, reduces
labor-force, etc. The following section provides an overview of DL applications for
gaming image processing [5].
5.3.1 Important Architectures in Deep Learning

With the developments of the previous section, much progress was made toward
improving the signals, images, videos, and audio processing. Notwithstanding to
astray from the topic of discussion, only minute detailing has been added to each
section. For more sources, refer to [6–8] and references wherein. Our discussions
with be advanced yet short inclusive of some network architectures that we believe
had, or will have, an impact on gaming image processing and animation.
Auto encoders use a contracting and an expanding branch for finding the input
of a lower dimensionality [9]. They do not require annotations, as the network is
trained to predict the original input using loss functions such as L(θ)=|| f ˆ(x)-x||22. DL
variants use CNN [10], for adding noise to the input [11], or aiming to find the sparse
presentations [12].
5.4 CONTENT-BASED GAMING IMAGE RETRIEVAL (CBGIR)

Content-Based Gaming Image Retrieval (CBGIR) is one of the most challenging and
ambiguous tasks used to minimize the semantic gap between images and human
queries in datasets with rich information content, and to recognize images accu-
rately. The consequences derived from the review goes vivid with ingression for
producing convenient yet large balanced datasets for the images. Possibly, there is a
shortfall of such datasets in the areas of gaming and animation, even today.
The imbalance gaming image datasets are likely presented by minimizing the
semantic gaps that probably evolve between the subordinate features and distinctive
semantics, directed by the four-step effective hash code generation techniques. The
first stage justifies CNN to be the potent with features so as to withdraw discrimina-
tive features such as Fully Connected Layer (FCL) from the images automatically.
There is a reduction in the imbalanced dataset within the second stage that makes use
of Synthetic Minority Over-sampling Technique (SMOTE). As per the third phase
is concerned, we can now clearly see the balanced features practically transmuted
to a 13-charactered symbolized code structured from deep SAE. Some of the other
relatable dataset experiments are the retrieval performances of the proposed method,
error parameter, and the classification performance [13].
Gaming image retrieval research is an aid to radiologists for being beneficial to
the usage of gaming and animated images for a long run. This brought advance-
ments to the imaging techniques and automatic diagnosed systems in the medi-
cal sector. The gaming images in the modern hospitals store Digital Imaging and
Communications in Medicine on text format, for retrieving repositories of the
gaming images. Therefore, CBGIR system is actually reliable based on visual-
izations that always gets itself ready to retrieve continuous and effective gaming
images [14].
5.5 IMAGE CLASSIFICATION

Image classification in a computer helps in analyzing the “class” of the images that
is exemplified often as an “animal,” “plant,” “vehicle,” etc. With not a big deal, the
perfect example fitting to this term is that of Moravec’s paradox when it comes to
understanding the classification within the machines or various artificial intelligent
based learnings.
Image classification in the preliminary stage depended on raw pixel data that
broke down the image into individual pixels. As a matter of fact, two different
images need not necessarily be alike, even though both of them belong to the same
category, because of their feature variants contrast to each other. This, thus, brought
challenges to the computers for effective visualization and categorization of the
images [15].
5.5.1 Adding Deep Learning to Neural Networks

DL is a type of machine learning as well as a segment to the artificial intelligence
(AI). Such advanced learning is altogether advantageous with NN for allowing
the machines to read the given data. While advanced with the NN, data is input
through the node layers for communicating the results to the next node layer. This
process is continued until the machine reaches the output layer, and provides a
concrete answer.
There are various types of NN among which the CNN are mostly associated with
DL during the process of image classification. In CNNs, the nodes in the hidden
layers don’t always share their output with every node in the next layer (known as
convolutional layers) (Figure 5.1). DL allows machines to identify and extract fea-
tures from images, furthermore, aiding programmers to hardly enter into the filters
manually.
5.5.2 Benefits of Image Classification

Image classification are potentially active in vast yet reliable growths. For instance,
the self-driving cars (Figure 5.2) make use of the image classification for identifying
their surrounding objects such as trees, people, traffic lights, etc.
FIGURE 5.1 Computer “vision” via data. (Source: KDnuggets.)
FIGURE 5.2 Classification of object detection, here, the self-driving car.
Image classification also helps in analyzing the gaming images. It is also used
in suggesting and depicting any traces of illness in the human body. Other usages
include proper organization of several types of photo collections.
5.5.3 The Image Classification Theory

Not only humans and creatures, computers have visions too. A vision that can clas-
sify any object just upon seeing a particular image. Today, the pervasiveness of image
classification is in a good shape because of the utilization of DL in the curriculum of
image processing. This advancement also resulted in the involvement of the CNNs
coinciding with strong exploration.
Image classification is a definitive learning for targeting set of classes and rec-
ognizing them with proper set of exemplified images. In such classification process,
CNN regards the raw pixel data as an input for extracting the necessary features
from a particular image such as textures and shapes. This way the CNNs extract the
contents from the image successfully. The input image of a CNN has dimensions
W×H×C, where W and H are the width and height of the image in pixels, respec-
tively, and C is the number of images color channels.
Generally, CNN constitutes of stack modules known to perform three operations.
They are the Convolution, Rectified Linear Unit (ReLU), and Pooling. Convolution
helps CNN in creating a filter map for the input image. ReLU brings non-linearity
into the model whereas Pooling acts helpful in the reduction of the dimensions of the
feature map. When the obtained feature map is shaped once again into a long vector,
many connected layers are used in the place for completing the final task with the
help of Softmax activation function. The output derived from this function is valued
between 0 and 1. In this review, any image processing method that takes the Softmax
activation function and modifies it compared to baseline methods is considered as a
“post-processing” technique [16].
It is basically the accurate calculations that measures the performance of deep
networks in the subject of image classification. The calculation is such that the ratio
between the correctly classified images and the aggregate quantity of images defines
how metric they are in solving the tasks of CNN in gaming imaging such as patch-
based classification processes [17].
5.6 IMAGE PROCESSING

Image processing is an operation that begins with one grey scale or colored image
(images with an array of pixels) and ends or returns with the other one. Throughout
this chapter, we are explaining certain techniques useful for operating the grey-scale
images in order to locate their features edges. The derived result is such that there is
an equal quantity of pixels as well as grey levels. The resulting image with this simi-
lar size is presented with modified brightness with the help of the spatial relationship
to many other pixels. This thus, produces a new image. Depending on what is consid-
ered to be of an absolute interest, the image is processed and modified intentionally
to emphasize the original one. With the resulted and desired features such as objects,
edges, boundaries, etc., the image later goes through under operation for extract the
background (i.e., uninteresting) data from that particular image for isolating other
additional information.
5.6.1 Point Operations
Each pixel value when replaced by a new one that is wholly and solely dependable
on the original value gives an idea of a simple methodology. Thus, this is definitely
known as a “point” operation keeping it distanced from other operations such as
“neighborhood” and “global.” In particular, if we carefully examine the original
brightness that is covered with a small fraction of the total range, new values are seen
increasing the contrast of the image. This is called “Point Transformation.” The rela-
tionship between original and replacement brightness values is often called a transfer
function, and can alter the appearance of an image as in Figure 5.3.
FIGURE 5.3 Examples of different transfer functions applied to an image: (a) original,
(b) logarithmic, (c) histogram equalization, and (d) negative.
The given functions don’t necessarily need to be a linear. Logarithmic function

is a type of function that imitates the works of photography. This thus gives the
function an inclined slop for semi-log plotting which is also known as the “gamma”
value. Its positive values try to constrict the brightness displayed on the end range
whenever it expands to the darker segments. Negative transfer functions are also
often used for reversing the contrast of the image. Histogram equalization method is
a type of non-linear transfer function useful for particular images where pixels of a
proportionate value consist of particular grey shades.
Transfer function derives its shape from the brightness histogram of the origi-
nal image in higher gradients so that the details are much clear to observe. It is
mostly used in the display devices and are often monotonic. It also grants permission
for inserting grey level variants for deriving good output. This causes the values to
become unconvinced and cynical about either accepting the minimized or the maxi-
mized ones. Brightness gradients are visualized in hegemony settling the brightest
pixels into extreme bright or dark values. The result is to superimpose brightness
contour lines on the image. No other techniques are so impactful upon transforming
the original image. Indeed, with many modern systems it is not even necessary to
operate on the image itself to produce the transformation.
Look Up Tables or LUTs are non-destructive and quick-witted methods for exe-
cuting the transfer functions. With the help of this method, the function is known to
produce the display and keep the video output under-control. Thus, it is very conve-
nient to change the façade of the display.
5.7 IMAGE PROCESSING USING DEEP LEARNING

Image processing plays an important role in the fields like image mining, gaming
imaging, gaming image processing, web mining etc. It is also the Content-Based
Gaming Image Retrieval (CBGIR) system that is quick enough for retrieving the
gaming images. This chapter on image processing and segmentation is deep with AI
type DL. Its recapitulation presents several possible advantages for using DL meth-
ods into gaming and image mining. It also throws an insight of content-based image
retrieval and identifying a flaw in the gaming image.
Besides optical sensing and image capturing systems [18], process images are
the new standpoints for process monitoring. When these process images are put to
use, the Deep Belief Network (DBN) are directly manipulated into the existing net-
works for extracting the features from those images and detect the scheduled faults.
Meanwhile, the sub-networks are operated to extracting the local features from the
sub-images. The global network exceptionally extracts and improves the training
efficiency without deteriorating the fault detection accuracy. On release, a new sta-
tistic is further developed in the course of the framework customized for DL [19].
As much as the current AI techniques are concerned, DL techniques are the most
stabilized variables in the health, infrastructure, research, gaming, and UI/UX sec-
tors. Their available is widespread rapidly each passing day because of its effective
and accurate results. They also aid in the implicit abilities of engineering workshops,
integrational ability for embedding words, and again an ability for dealing with com-
plex and unstructured dataset. Beyond this, the concept is also related digital texts
in electronic health records (EHRs), clinical texts on social media, text in electronic
gaming reports and gaming images. On the growing popularity of DL and its algo-
rithms, the responsibility toward several domains have risen as well. It is very well
observed in the 2017–2020 reports [20], where the 2019 report stated the publication
up to 4 times, that is more than the year 2018. Growths are seen increasing rapidly
after the learning acted even in the gaming techniques and animating applications
for image processing [20].
5.8 CONCLUSION
The technical translation of DL methods in image processing is still in the juvenile
phase. These techniques have grown vigorously useful in tackling with the tasks
conducting image processing and yet stay aback at times due to certain limitations.
Since there was a deficiency in the exclusiveness of the annotated experimental
data, the researchers adjusted to the utilization of simulated data only, validating
hardly any proposed techniques to run it on a large-scale. This chapter discusses
the challenges for clinical translations used for DL methods in image processing by
concluding and summarizing the major key findings on processing images in dif-
ferent aspects. Wilson et al. [21] examined the challenges and issues faced during
the clinical translation using the techniques of spectroscopic optical imaging. This
prepared a broader space for us to focus more on the upcoming hurdles impending
while translating imaging modalities. Therefore, this thesis particularly concentrates
on the facilitations made by the DL application while integrating the approaches of
image processing.
At last, reviewing this chapter shows that the methods used for DL is prominently
utilized in the field of personal activity intelligence (PAI), CBGIR, etc., for sim-
plifying extensive number of translations that would go on a long term. Therefore,
with the help of several studies, research, and investigation, we analyzed that even
these minimalistic studies on image processing are some of the little concerns for
contributing to the bigger picture challenges to what we call it as the future open
challenges associated to the applications on DL such as detecting problems in optical
inversion, image post-processing, and annotation on semantic imaging.
REFERENCES
1. Zhang J., Shao, M., Yu, L., and Li, Y., Image Super-Resolution Reconstruction Based
on Sparse Representation and Deep Learning, Signal Process. Image Commun., 87,
115925, September 2020.
2. Arunakranthi, G., Rajkumar, B., Chandra, V., Rao, S., and Harshvardhan, A., Advanced
Patterns of Predictions and Cavernous Data Analytics Using Quantum Machine
Learning, Mater. Today Proc., 2021.
3. Raju, B., and Bonagiri, R., A Cavernous Analytics Using Advanced Machine Learning
for Real World Datasets in Research Implementations, Mater. Today Proc., 2021.
4. Li, Y., Zhao, J., and Zhihan, Lv., Li, J., Medical Image Fusion Method by Deep
Learning, Int. J. Cogn. Comput. Eng., 2, June 2021, 21-29.
5. Bhattacharyaa, S., Reddy Maddikuntaa, P.K., VietPhamb, Q., Gadekallu, T.R.,
Krishnan, S.R., Chowdhary, C.L., and MamounAlazab, Md. J.P., Deep Learning and
Medical Image Processing for Coronavirus (COVID-19) Pandemic: A Survey, Sustain.
Cities Soc., 65, 102589, February 2021.
6. Benjio, Y., Courville, A., and Vincent, P., Representation Learning: A Review and New
Perspectives, IEEE Trans. Pattern Anal. Mach. Intell., 35, 1798–1828, 2013.
7. Goodfellow, I., Bengio, Y., Courville, A., and Bengio, Y., Deep Learning, vol. 1, MIT
Press, Cambridge, 2016.
8. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., et al.,
A Survey on Deep Learning in Medical Image Analysis, Med. Image Anal., 42, 60–88,
2017.
9. Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P.-A., Extracting and Composing
Robust Features with Denoising Autoencoders, Proc. 25th Int. Conf. Machine Learning,
ACM, 1096–1103, 2008.
10. Holden, D., Saito, J., Komura, T., and Joyce, T., Learning Motion Manifolds with
Convolutional Autoencoders, SIGGRAPH Asia 2015 Technical Briefs, ACM, 1–4.
November 2015 Article No.: 18. https://doi.org/10.1145/2820903.2820918.
11. Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., and Manzagol, P.-A., Stacked
Denoising Autoencoders: Learning Useful Representations in a Deep Network with a
Local Denoising Criterion, J. Mach. Learn. Res., 11, 3371–3408, 2010.
12. Huang, F.J., Boureau, Y.-L., LeCun, Y., Huang, F.J., Boureau, Y.-L., LeCun, Y., et al.,
Unsupervised Learning of Invariant Feature Hierarchies with Applications to Object
Recognition, IEEE Conf. Comput. Vision Pattern Recog., 2007. CVPR’07, IEEE, 1–8,
2007.
13. Ozturk, S., Stacked Auto-Encoder Based Tagging with Deep Features for Content-
Based Medical Image Retrieval, Expert Systems Appl., 161, 113693, 2020.
14. Shamna, P., Govindan, V.K., and Abdul Nazeer, K.A., Content-Based Medical Image
Retrieval by Spatial Matching of Visual Words, J. King Saud Univ. Comput. Inf. Sci.,
2018. https://doi.org/10.1016/j.jksuci.2018.10.002
15. https://www.thinkautomation.com/eli5/eli5-what-is-image-classification-in-deep-
learning/
16. Salvi, M.,Acharya, U.R., Molinari, F., and Meiburger, K.M., The Impact of Pre- and
Post-Image Processing Techniques on Deep Learning Frameworks: A Comprehensive
Review for Digital Pathology Image Analysis, Comput. Biol. Medicine, 128, 104129,
January 2021.
17. Y. LeChun, Y., Bengio, Y., and Hinton, G., Deep Learning, Nature, 521, 436–444,
2015.
18. Karanam, S.R., Shrinivas, Y., and Vamshi Krishna, M., Study on Image Processing
Using Deep Learning Techniques, Mater. Today Proc., 2020.
19. Lyu, Y., Chen, J., and Song, Y., Image-Based Process Monitoring Using Deep Learning
Framework, Chemom. Intell. Lab. Syst., 189, 8–17, 2019.
20. Pandey, B., Pandey, D.K., Mishra, B.P., and Rhmann, W., A Comprehensive Survey
of Deep Learning in the field of Medical Imaging and Medical Natural Language
Processing: Challenges and Research Directions, J. King Saud. Univ. Comput. Inf. Sci.,
29 January, 2021.
21. Wilson, B.C., Jermyn, M., and Leblond, F., Challenges and Opportunities in Clinical
Translation of Biomedical Optical Spectroscopy and Imaging, J. Biomed. Opt., 23(3),1–
13, 2018. doi: 10.1117/1.JBO.23.3.030901. PMID: 29512358; PMCID: PMC5838403.
6 Artificial Intelligence
in Games
Transforming the
Gaming Skills
Abhisht Joshi, Moolchand Sharma, and
Jafar Al Zubi
CONTENTS
6.1 Introduction................................................................................................... 103
6.1.2 How will AI Reinvent the Experience of Gaming?........................... 106
6.1.2.1 Advantages of AI in Gaming.............................................. 106
6.2 Gaming Experience....................................................................................... 107
6.2.1 Power of Voice in Gaming................................................................. 107
6.2.2 Real Gaming Experience to Players.................................................. 108
6.2.2.1 Three-Dimensional Visualization Techniques................... 108
6.2.2.2 Simulation Based on Physics.............................................. 108
6.2.2.3 Virtual Reality.................................................................... 109
6.2.2.4 Augmented Reality............................................................. 110
6.2.2.5 Extended Reality................................................................. 110
6.2.3 Necessity of RL for Adaptability in Games...................................... 112
6.2.4 In-Game Support to Players by AI Powered Chatbots...................... 113
6.2.5 Gives an Overall Ultimate Gaming Experience................................ 114
6.3 Early Game AI vs Complex Game AI........................................................... 115
6.3.1 Early Game AI................................................................................... 115
6.3.2 Complex Game AI............................................................................. 115
6.3.3 Video Game AI.................................................................................. 116
6.3.3.1 Finite State Machines......................................................... 116
6.3.3.2 Path-Finding....................................................................... 117
6.3.3.3 Real-Time Play Complex AI............................................... 118
6.4 Conclusion and Future Scope........................................................................ 119
References............................................................................................................... 120
6.1 INTRODUCTION
It is important to consider the past and present before embarking on future
game production. In October 1958, the game design and production began.
William Higinbotham, an American physicist, created the first video game. It was
DOI: 10.1201/9781003231530-6 103
an old tennis game known as “Pong” that was very common at the time. Since then,
the number of games, video game platforms, and game types has increased dra-
matically. As a result, many people showed a willingness to play, and many people
took it out on them. The gaming industry has evolved rapidly because of signifi-
cant advancements in hardware and design and development strategies. Artificial
Intelligence (AI) has aided the expansion by improving the gaming experience. In
addition, it has piqued players’ attention by exceeding their standards for the game.
AI software improves creation processes such as image production, animation gen-
eration, improving the performance of non-player characters (NPCs), story layers,
characterization, and graphics authenticity.
When you talk about AI, you are talking about computers that mimic the human
mind. The primary goal of AI in gaming is to make games smarter. It gives the games
a natural system to support NPCs’ intelligent actions. AI offers intelligent game con-
trols that allow for clever character communication and movement. To put it another
way, AI brings the game closer to reality. It investigates the dynamic relationships
between agents and game environments in general. Various games provide agents
with fascinating and complex problems to solve, making video games ideal for AI
study and studying how to use AI technologies to achieve human-level success.
In contrast, playing games is known as AI in the gaming experience. It investi-
gates the dynamic relationships between agents and game environments in general.
Various games provide agents with fascinating and complex problems to solve, mak-
ing video games ideal for AI study. These virtual worlds are safe and manageable.
Furthermore, these game environments have an endless supply of data for machine
learning (ML) algorithms and are much faster than in real-time. Because of these
characteristics, games are a particular and common domain for AI study. On the
other hand, AI has aided games in being better in terms of how we play, understand,
and style them [1].
In general, game AI is concerned with perception and decision-making in virtual
environments. There are some significant challenges and solutions associated with
these components. The primary challenge is that, particularly in strategic games,
the state space of the sport is extremely broad. The entire framework has effectively
modeled large-scale state-space with deep neural networks as representation learn-
ing has increased. The second issue is that it is difficult to learn proper policies
for making decisions in a dynamic, unknown environment. Data-driven approaches
such as supervised learning and reinforcement learning (RL) are viable solutions for
this issue. The third issue is that the vast majority of game AI is created in a con-
trolled virtual world. Therefore, the ability of AI to be transferred between games
could be a major challenge. To address this issue, a more generalized learning sys-
tem is also required. Hence, if we talk about AI in gaming detail, there are two types
of game AI techniques: deterministic and non-deterministic [2, 3].
i. Deterministic: Deterministic performance or action is predetermined and

predictable. There is not any ambiguity. A simple chasing algorithm is an
example of deterministic action. By advancing along the x and y coordinate
axes until the character’s x and y coordinates coincide with the target posi-
tion, you can specifically code an NPC to travel toward a target point.
Artificial Intelligence in Games 105
ii. Nondeterministic: The opposite of deterministic behavior is non-deter-

ministic behavior. There is a degree of ambiguity in behavior, and it is
very unpredictable (the degree of uncertainty depends on the AI method
employed and how well that method is understood). An NPC learning to
adapt to a player’s combat strategies is an example of non-deterministic
actions. A neural network, a Bayesian method, or a genetic algorithm may
be used in this form of learning.
The bread and butter of game AI are deterministic AI techniques. These tech-
niques are easy to implement, understand, test, and debug because they are predict-
able, fast, and simple. Despite their many advantages, deterministic methods predict
all scenarios and code all actions on the developers’ shoulders. Furthermore, deter-
ministic approaches obstruct learning and evolution. Deterministic habits begin to
become predictable after a little practice. This, in a sense, shortens a game’s lifespan.
On the other hand, learning and volatile gameplay are made easier with non-
deterministic methods. Furthermore, developers are not required to code all actions
directly in advance of all potential scenarios. Non-deterministic approaches can also
learn and extrapolate independently and foster emergent behavior, which occurs
without clear instructions.
However, if we considered ML methods, RL has been commonly used for a
long time to solve these problems. In recent years, deep learning (DL) has achieved
impressive success in computer vision and natural language processing (NLP) [4]. As
a result, many video games with deep reinforcement learning (DRL) have achieved
performance beyond human ability. At the same time, there are still some challenges
in this domain that need to be tackled.
The gaming industry worldwide has values of around $109+ billion, making this
industry one of the central points to perform research and thus improve its per-
formance by integrating it with AI, specifically DL. Figure 6.1 shows the gaming
industry size as compared to other industries. Most gamers do not consider gaming
as a sector, but it is becoming such a large and influential entertainment industry that
it attracts an increasing number of professionals. We must first dismantle the gaming
value chain. Game engines, which game engine owners normally license, are used to
FIGURE 6.1 Comparison of gaming industry size with other industries.

create games. These game creators then depend on publishers to bring their creations
to life [4].
6.1.2 How will AI Reinvent the Experience of Gaming?
With AI changing the landscape of gaming, we are revolutionizing the gaming industry
with captivating gaming experiences that cannot be achieved on any other platforms.
Jefferson Valadares, CEO, Doppio Games
AI has shown positive results in the gaming industry since its inception. In every
aspect of the gaming industry, AI has always been the same. It not only learned to
play chess, but it also defeated some of the best players in the world. Furthermore,
gaming companies are now employing AI in a more dynamic setback process, such
as determining the main AI benefits in games.
6.1.2.1 Advantages of AI in Gaming

The major advantages of AI in gaming are as follows:
• Creating a real gaming experience: A player will normally play a game if

they can relate to it; the more authentic the experience, the more the player
will find it enjoyable and worthwhile to play. As ML and DL become more
mainstream, more progress is made in these fields. As a result, the gaming
experience matters a lot and AI has emerged as one of the major factors
leading to a better gaming experience.
• Improves the experience of audio and video: There are two ways a game
communicates with the player: audio and video; if both are not of high
quality, the player is more likely to dislike the game. Thus, image and audio
auto-regeneration are promising ML technology for real-time communica-
tions. Furthermore, the algorithms can produce higher-quality images from
lower-quality images. ML algorithms can also detect and suppress loud
noises produced in the background that are unintelligible and unimportant
during your conversation.
• Identification of bugs: The ubiquitous restricted random/coverage-
oriented verification technology, or structured verification technology, is
used in today’s bug-finding approaches. However, as designs become more
complex, new methodologies and methods are needed to ensure long-term
quality. This can range from AI and ML to upgrading methodologies to
provide more automation and less error tracking in a notebook or spread-
sheet. If the AI model is correctly and appropriately trained, it will assist in
identifying bugs and removing them immediately. This area is growing in
popularity because it saves developers from having to hardcode bug fixes
every time.
• Enhancing the gaming experience and many more: Though AI is widely
used, it is focused on a single target in the gaming industry: to have a better
gaming experience. While there is still much improvement in AI in gaming,
there is much substantial advancement going on right now. Day by day,

progress is made in this area, and improvisation and technology updates
occur.
6.2 GAMING EXPERIENCE

As the name suggests, the gaming experience means how the player feels the game
every time they play it. Therefore, the gaming experience depends upon the follow-
ing factors:
• Interaction with the player: Certain games necessitate live interaction

with a live player, while others necessitate interaction with an in-game
character; the more interaction, the better the gaming experience.
• Audio capability and quality of graphics: Audio and graphic quality con-
nect with the gamer and provide the actual gaming experience. This is the
most important factor of the gaming experience as the more it is better, the
user will find it real, and the gamer will be able to believe that he is actually
in the game.
• Degree of difficulty: Even though every game gives the player the option
to choose the difficulty level before starting the game, developers must bear
in mind that the game should not be too difficult for the player to leave, nor
should it be too straightforward for the game to end in a few hours.
• The level of intelligence in-game characters: Every player’s intelligence
is important because they are the ones who will lead the game to its conclu-
sion; if they are not properly qualified, the game will be stuck indefinitely.
NPC intelligence is also essential, as NPCs cannot simply be programmed
to do anything they want, as this would seem strange. As a result, the NPC’s
intelligence level is also significant.
AI is focusing on the variables mentioned above to have a superior gaming experi-

ence. In addition, AI assists game developers in improving the gaming experience on
mobile devices and other platforms. The five ways to use AI to redefine the ultimate
gaming experience are as follows:
1. Power of voice in gaming

2. Real gaming experiences to players
3. The necessity of RL for adaptability in games
4. In-game support to players by AI-powered chatbots
5. The overall ultimate experience of gaming
6.2.1 Power of Voice in Gaming

The way games are played changing thanks to AI-based voice intelligence. The term
“AI” sounds more rational now than at any other time in gaming history. This also
contributed to a more realistic gaming experience. Additionally, voice-controlled
games allow players to use voice commands to control the game. This results in a
more enjoyable gaming experience. When playing these games, players will simply
tell their characters what to do. For example, gamers may use simple verbal com-
mands to tell their game characters to run, swim, jump, hide, open the door, change
weapons, and do something else. Since voice commands save them from remember-
ing various keys and their variations, they can provide a more smooth and realistic
gaming experience. In contrast to buttons or keys, using voice commands is a more
natural process.
The following are some examples of voice-controlled AI games:
a. Bot Colony
b. Broken Seal
c. The 3% Challenge
d. Westworld: The Maze
e. LEGO: Duplo Stories
The accent problem is a major issue in this type of game; various world regions
have different accents, which confuse AI games because the user has to repeat the
command repeatedly, resulting in poor game results. Since AI-created voice sounds
natural, and the voice generated also plays a crucial role in making it more realistic,
voice AI can make the game more interactive. AI provides digitally created voices,
and its ability to initiate interactions through dialogues aids players in completing
several tasks and more. Voice AI has only recently started to reshape the gam-
ing industry. Voice AI has limitless potential, but it also creates a customized and
dynamic gaming experience for gamers. There is still much progress in voice AI,
and definitive research is being conducted to improve the current voice AI. Scientists
and game developers are collaborating to improve its accuracy [5, 6].
6.2.2 Real Gaming Experience to Players

The hyper-realistic gaming experience is another advantage of AI in gaming. AI
brought the following changes in creating realistic gaming experiences.
6.2.2.1 Three-Dimensional Visualization Techniques

The realism of three-dimensional (3D) visualization in simulation software is getting
better. Although most simulation packages had two-dimensional (2D) visualization
featured ten years ago, most now have 3D visualization features. The simulation
techniques used, however, do not provide a realistic representation of the virtual
world. To improve the quality of 3D visualizations, we should look to the video
game industry for research solutions and avoid hardcoding wherever possible.
Furthermore, since these games’ common special effect techniques are implemented
in various open-source and commercial game engines, they can be modified and
reused in other areas [7].
6.2.2.2 Simulation Based on Physics

Since a game created with AI will be trained on many data, it will have a greater
understanding of physics as it will be fed with the data following physics laws,
FIGURE 6.2 Making AI model to adapt the physics laws.
making the AI model adapt to physics law, making it to be real so the player can
relate game with real-life making the gaming experience more real as shown in
Figure 6.2.
6.2.2.3 Virtual Reality

Virtual reality, also known as VR, is a computer-generated simulation. A person can
interact with an artificial 3D world using electronic devices such as special goggles
with a screen or sensors-equipped gloves [8]. Anyone with these devices will enjoy
the unique VR experience depicted in Figure 6.3.
FIGURE 6.3 Unique experience of VR.

FIGURE 6.4 AR used in Pokémon Go.
6.2.2.4 Augmented Reality

Augmented reality (AR) is an immersive experience of a real-world environment.
Computer-generated perceptual knowledge is used to improve the objects in the
real world, often through several sensory modalities such as visual, auditory, haptic,
somatosensory, and olfactory. Figure 6.4 depicts AR in the famous game Pokémon
Go, while Figure 6.5 depicts the distinction between AR and VR.
6.2.2.5 Extended Reality

Extended reality (XR) refers to all real-and-virtual collaborative environments and
human-machine interactions generated by computer technology and wearables. The
FIGURE 6.5 AR vs. VR.

FIGURE 6.6 Extend reality with AR, MR, and VR.
“X” represents a variable for any current or future spatial computing technologies.
Figure 6.6 depicts a clearer picture of the XR.
Games today are intelligent and visually pleasing in ways that were unimaginable
10–15 years ago. Figure 6.7 shows the evolution of game characters, and Figure 6.8
shows an example of the evolution of graphics in games over the years.
The definition of gaming has changed thanks to AR. The VR game “Pokémon
Go” is a well-known example. Another example is “Until you fall,” a VR sword
combat game depicted. One of the most common problems with video games is
that items look fine from afar, but as you get closer to take a closer look, objects
FIGURE 6.7 Evolution of game characters.

FIGURE 6.8 Evolution of graphics in Need for Speed.
make poorly and become pixelated, resulting in bad game quality. NVidia and
Microsoft are cooperating to solve this issue by investing in DL to improve render-
ing. Computer vision and DL algorithms can help with the dynamic representation
of finer details, a problem in the past. Algorithms acting as NPCs are another signifi-
cant factor contributing to improving the real-world gaming experience. AI-based
NPCs allow you to play against less predictable opponents. These adversaries may
also change their difficulty level based on the user. As a player learns how to play the
game, the enemies may become smarter and react to the user in specific ways that the
user has never seen before based on the user’s behavior. Companies are developing
NPCs by using data from top players as training data, allowing for better and quicker
reinforcement training.
The way players communicate with friendly NPCs is another big challenge in
creating a realistic virtual environment. To complete your goals in many games,
you must speak with scripted characters. These interactions, on the other hand, are
normally brief and obey on-screen prompts. NLP could allow you to speak to in-
game characters aloud and receive accurate responses, similar to how Siri, Alexa,
and Google Assistant work. Furthermore, games with VR haptics or imaging of the
player may allow computer vision algorithms to detect body language and intentions,
enhancing the experience of interacting with NPCs even more [9, 10].
6.2.3 Necessity of RL for Adaptability in Games

RL is based on the reward and punishment theory. Software is trained to reward and
punish in RL, a form of ML, as shown in Figure 6.9. When software gets it right, it
gets praised, and when it gets it wrong, it gets punished. This decreases the number
of incorrect moves while increasing the number of right moves.
RL’s hyper customization will revolutionize the gaming experience. For example,
RL determines the player’s skill level, and the strategy or strategies are adjusted
accordingly [11]. To better grasp the application of AI in the gaming room, consider
FIGURE 6.9 Reinforcement learning.
Deep Mind’s Alpha Zero. Alpha Zero is a program that aims to understand the recur-
ring patterns and properties of games, such as board symmetry.
This is possible, thanks to a new type of RL in which AlphaGo Zero serves as its
instructor. The software begins with an untrained artificial neural network that has
no prior knowledge of Go. Then, combining this neural network with a strong search
algorithm, the program plays the game against itself. Finally, the neural network is
trained and tuned to anticipate m as it plays. Figure 6.10 shows the Alpha Go Zero
beating Ke Jei.
6.2.4 In-Game Support to Players by AI Powered Chatbots

AI-powered chatbots benefit both game developers and players by offering the best
in-game support. These chatbots can apply ML algorithms to implement human-level
FIGURE 6.10 AlphaGo Zero beating Ke Jei.

reasoning and logic for efficient gameplay. VR and XR are also driving up demand
for AI-powered chatbots like never before. Chatbots operated by AI contribute to the
improvement of gaming experiences in the following ways:
• Enhancement to players’ interactions: With the aid of chatbots, real-time

interactions with players make the gaming experience more interactive.
Conversational AI is used to empower these chatbots.
• Recognition of voice-based commands: In contrast to text-based or key-
based commands, voice-based commands are the more natural way to
communicate. NLP and speech-to-text recognition technologies are used
to allow these commands.
• Enhancement to customer support: The new generation is enthralled
with video games, with gamers playing them at all hours of the day and
night. As a result of this zeal, assistance is needed immediately, regardless
of the time. AI-powered chatbots improve customer service by offering in-
game support 24 hours a day, seven days a week, at lower operating costs.
6.2.5 Gives an Overall Ultimate Gaming Experience

When playing a game, gamers pay close attention to every detail. Gamers evaluate
the gaming experience not only in terms of visual appeal or graphic quality but also
in terms of how realistic and related the game is. The game is evaluated in every way
possible. AI plays a critical role in elevating the gaming experience. The in-game
experience of a player is first decoded in order to improve it. Developers use ML
algorithms such as SVM or neural networks to decode the data. As in the AlphaGo
Zero game, models of game player experience are developed. For data preparation,
the game’s specific pattern and details of game-player interaction are needed. Data
may be collected using a variety of methods, such as a questionnaire or a survey.
Since each player reacts differently to various scenarios, hardcoding a common algo-
rithm is difficult; this is where ML and DL come into play [2, 12].
Several other tech firms, including Facebook, have been focusing on AI in gam-
ing. For example, Facebook’s new AI will remove characters from YouTube vid-
eos for a better gaming experience. In addition, the Facebook Research team has
developed Vid2Game, an AI that will allow playable characters to be extracted from
real people’s images. It is a more advanced variant of the 1980s’ full-motion video
(FMV) games. The Facebook Game Development team used two neural networks,
Pose2Pose and Pose2Frame. The video was fed into the Pose2Pose neural network in
the first level, designed for specific actions such as dancing, tennis, and fencing. The
system recognizes and separates the individual from their surroundings, isolating
him and his pose. Pose2Frame then inserts the individual and their movements and
whatever object they are carrying into a new scene with small artifacts in the second
stage. The character’s action is then controlled using a joystick or keyboard.
A few short videos were used in the system’s training to analyze events such as
fencing, dancing, and tennis. It was able to screen out unwanted people and integrate
various camera angles. This research is like Adobe Premiere Pro’s “content-aware
fill” function, which uses AI to remove different elements from video. NVIDIA and
a few other companies have also worked on AI to convert real-world videos into
virtual worlds, making it ideal for video games [13].
6.3 EARLY GAME AI VS COMPLEX GAME AI

AI has been used in a variety of games in recent years, ranging from modern real-
time strategy games like StarCraft II to classic games like checkers. As a result,
AI and ML have a long background in game evolution. The development of virtual
worlds in video games has proven to be a useful test-bed for AI algorithms. Using AI
in games is also a good idea.
6.3.1 Early Game AI
The first known example of AI in gaming was inspired as a demonstration of com-
puting at the Festival of Britain in 1950. In 1951, Ferranti introduced a custom-built
Nimrod machine, which used the Nim game to demonstrate its mathematical poten-
tial. Nim is a two-player game where players take turns by removing one to three
items from the pile. On the contrary, the player who re-enters the game is the one
who re-enters the game. Although Nimrod was created to play Nim, Ferranti thought
that creating a computer that would play a complex game would also solve complex
problems. As a result, Nimrod had no program. Instead, it had a set of complex hard-
wired collections of logic to follow. Thus, Nimrod did not have traditional AI, but its
ability to play Nim made it competitive, surprising, and intimidating players alike.
6.3.2 Complex Game AI
Arthur Samuel of the tech giant IBM introduced one of the first applications to learn
about conventional ML in sports (who coined the term “machine learning”). Samuel
accepted the challenge of the game checkers in 1956, which required both easy play
and complex strategy. His work on the first commercial computer, the scientific IBM
701, resulted in the development of two key principles in ML. The invention of alpha-
beta pruning was the first.
“Do not take the time to see how bad an idea is if it is undeniably awful,” Patrick
Winston said of the basic definition of alpha-beta. The Minimax algorithm is used
to reduce the number of nodes that must be evaluated inside a tree—in this case,
for two-player games—and alpha-beta pruning is a tree-search optimization tech-
nique. In a nutshell, when the machine chooses a move, it is represented as a game
state space tree, with each level representing the computer or player’s next move.
The Minimax algorithm aims to optimize the smallest benefit possible. Alpha-beta
prunes parts of the tree that don’t improve a player’s reward out of the evaluation;
only the middle move can help player X win (assuming that both players will be play-
ing optimally), shown in Figure 6.11.
Samuel’s second central idea was the combination of self-play and rote learning.
Samuel pitted his software against itself to enhance its efficiency, as well as remem-
bering each move it had experienced (dutifully recorded on magnetic tape), as well
as its terminal “reward” value. His software was a checkers game for amateurs, but
it incorporated ideas that are still used today. For example, the alpha-beta search
FIGURE 6.11 Minimax tree for Tic-Tac-Toe.
method was revived in the game of chess 40 years later. Garry Kasparov, a chess
grandmaster, was defeated by IBM’s Deep Blue in 1996. Deep Blue ran simultane-
ous game-state tree search, pruned with alpha-beta pruning, to speed up the process
of calculating the computer’s next step (by searching 200 million positions per sec-
ond and decide up to 20 moves for the future). In 2015, Google used DL (neural net-
works) for move selection and Monte Carlo search for applying previously learned
moves to the game of Go. In 2015, AlphaGo defeated a professional human player,
and in 2017, it defeated the world’s top player. AlphaGo reapplied Self-play (which
was developed by Samuel for checkers approximately 60 years ago) to improve its
play and store the moves that were previously known.
6.3.3 Video Game AI
Early video games had no idea that instead of relying on state machines to predict
movement, their rivals used AI (such as in Space Invaders). Pac-Man raised the dif-
ficulty of the enemy in 1980 to aid the player in path-finding (or away from the player
in the case of escape). Furthermore, each opponent had a distinct personality, mak-
ing the game more unpredictable.
Let us look at some of the current methods in video games.
6.3.3.1 Finite State Machines

Though not technically AI, finite state machines (FSMs) have long been a part of
games, providing the foundation for what appears to be intelligent actions. An FSM is
a computation model that allows states to change in response to external input. FSMs
have differing behaviors for game entities in the context of games. For example,
FIGURE 6.12 The finite state machine for a real time strategy worker NPC.
Figure 6.12 depicts an NPC to whom the player can delegate a role. Once delegated,
the NPC performs the task based on the various behaviors required to complete it.
The FSM in Figure 6.12 depicts a worker NPC who is tasked with collecting
resources by the player. Following its assignment, the NPC locates the desired
resource, collects it, and transports it to a collection point. This cycle repeats until
all resources have been depleted, at which point the NPC will become idle.
6.3.3.2 Path-Finding
In games, the ability to navigate in their surroundings by the NPCs is a common
feature. Path-finding is a term used to describe an ability that can be learned in
several ways. One of the most well-known is A. The A algorithm is a variant of
Dijkstra’s shortest path algorithm, in which NPC destinations are represented as the
nodes of a graph (see Figure 6.13). There are two types of lists that A* can deal with:
FIGURE 6.13 Implementation of an A* graph to find the best way.

open and closed lists. The empty list contains all the nodes that non-visited nodes,
while the closed list contains all the visited nodes. The algorithm starts with a node
available in the list and looks for nodes accessible from it, adding them to the empty
list (if they are not already added in the closed list). Next, the current node’s score
is determined using the node’s value (some of the nodes might be more expensive
than others to pass through) and a heuristic from the current node (the shortest dis-
tance between a node and the goal). The process is carried out repeatedly until the
desired result is obtained. Then, the algorithm follows the node’s decreasing score
back again to the starting node to determine the shortest path.
There are several path-finding algorithms, including A*. An original path-finding
algorithm can cluster the universe into larger nodes in applications with a large set of
maps and small-sized CPUs. On the other hand, hierarchical path-finding generated
a smaller number of nodes, decreasing the amount of time spent looking.
6.3.3.3 Real-Time Play Complex AI

The growth of ML players in Dota 2 is the final example we will look at. Dota 2 is
a real-time strategy game in which two teams of five players compete to demolish
an ancient structure. Each of the team members works together to kill the ancients
of the other while defending their own. This game features a variety of characters
with varying playstyles (carry vs. support). Characters can increase their level to
improve their ultimate skills and earn gold to buy products with unique properties.
Dota 2 has an interesting feature in that it takes a lot of coordination and collabora-
tion to win.
In April 2019, OpenAI Five released an advanced Dota 2 AI that beat the
Dota 2 e-sports world champions. With no previous knowledge of the game or
heuristics to guide its behavior, the AI was designed from the ground up. It began
by using a random weight initialization technique for its neural networks and then
self-taught itself to play the game. In 80% of the games, the eligible AIs (5 for each
character) competed against each other, and 20% against previous AIs. The AIs
used 128,000 CPU cores and 256 GPUs to play 180 years’ worth of games every
day. Each game’s AI makes about 20,000 moves, meaning a long time horizon for
data collection to match an AI scenario. Thanks to a team spirit hyperparameter,
OpenAI Five consists of five different neural networks that work together. As a
result, individual AIs can focus on either their own reward (which involves killing
other characters and gathering their treasure) or the team’s overall reward (such as
collecting treasure).
The three parts of AI are perception, processing, and action selection (see
Figure 6.14). Different types of environmental inputs are fed into an embedding
max-pooled as well as into the perception processing core. The long short-term
memory (LSTM) is a Turing complete recurrent neural network (RNN) that can
perform all of the computations of a general-purpose machine. Furthermore,
LSTMs excel at learning sequences, making them ideal for this use. After that,
the LSTM sends its production to a group of action networks, which decide what
action to take (alongside the other pertinent information, like target). Furthermore,
this takes into account the decisions that the character is free to take as well as the
character’s attention.
FIGURE 6.14 Dota 2 hero AI for long-horizon learning.
The agent’s perception space is very large since the network’s inputs are made
up of around 20,000 real values. With about 170,000 total unique actions to choose
from, the number of actions available is also very large (about 1,000 actions that
are typically possible among them as they are reduced due to cooldowns, items not
present, etc.). The OpenAI Five used proximal policy optimization for learning,
which is simple to use and performs well in real-world scenarios. Both how we play
games and how they are made are being transformed by AI. While the methods
used have evolved over time, some ideas have remained unchanged, such as Arthur
Samuel’s principle of self-play in the creation of learning agents. From checkers
and the IBM 701 to complex real-time games trained with the aid of distributed
networks of CPUs and GPUs, ML is an integral part of games and a test-bed for
potential ML methods [14].
6.4 CONCLUSION AND FUTURE SCOPE

AI is a powerful technology that can be used at any point of the game, from game
design to customer loyalty and a realistic gaming experience. As a result, investing in
ML solutions is important, as the advantages of AI in gaming cannot be overlooked.
Furthermore, game developers can use all game development methods, track and
evaluate player results, provide round-the-clock customer service, and improve their
gaming experience with AI and its subset of understanding, ML. AI’s advantages
will continue to evolve and improve as games become more interactive, realistic,
and lifelike. Intelligent actions in games will provide attractive features in virtual
movement and interaction between game characters and game players. The game’s
protagonist can use sensors and actuators to interact with his surroundings and react
appropriately. This character’s action is edited by the coder, who then develops the
character through observation and communication. Because of the existence and
incorporation of IoT and AI, all of this is possible. AI is paving the way for a future
centered on custom-made games.
Learning is the next big thing in game AI. Instead of making all NPC actions pre-
determined when a game is published, it should evolve, learn, and adapt as it is played.
As a result, as the player progresses, the game becomes more difficult to foresee,
extending the game’s lifespan. Since learning and changing games are inherently
unpredictable, AI developers have historically approached learning strategies with
caution. Non-deterministic AI, which has its own set of problems, encompasses the
techniques for understanding and reacting to character behavior. However, building
and evaluating non-deterministic learning AI techniques takes longer. In addition,
debugging becomes more difficult as it becomes more difficult to comprehend what
the AI is doing fully. These obstacles have proven to be major impediments to AI’s
widespread adoption. However, a lot of this is changing. Several famous games,
including Creatures, Black & White, Battlecruiser 3000AD, Dirt Track Racing,
Fields of Battle, and Heavy Gear, used non-deterministic AI. Their achievements
sparked renewed interest in AI techniques such as decision trees, neural networks,
genetic algorithms, and probabilistic methods. Non-deterministic methods are used
in combination with more conventional deterministic methods in these success-
ful games. They are used only when they are required and for problems for which
they are ideally suited. A neural network is not a magic pill that will solve all AI
problems in a game; however, it can be used with excellent results for specific AI
tasks. When using non-deterministic methods, we suggest taking this approach.
This way, you can isolate the parts of your AI system that are unpredictable and
difficult to create, test, and debug while maintaining the majority of your AI system
in its traditional form.
REFERENCES
1. Tamburrini, G. & Altiero, F. (2021). Research Programs Based on Machine
Intelligence Games. In Chiodo, S. & Schiaffonati, V. (Eds.), Italian Philosophy of
Technology. Philosophy of Engineering and Technology, vol 35. Springer, Cham. DOI:
10.1007/978-3-030-54522-2_11
2. Singh, T. & Mishra, J. (2021). Learning With Artificial Intelligence Systems:
Application, Challenges, and Opportunities. In Verma, S. & Tomar, P. (Eds.), Impact
of AI Technologies on Teaching, Learning, and Research in Higher Education
(pp. 236–253). IGI Global, Hershey, Pennsylvania, 2021. DOI: 10.4018/978-1-7998-
4763-2.ch015
3. Westera, W., Prada, R., Mascarenhas, S. et al. (2020). Artificial Intelligence Moving
Serious Gaming: Presenting Reusable Game AI Components. Education and
Information Technologies, 25, 351–380. DOI: 10.1007/s10639-019-09968-2
4. Ho, M. (2021, February 2). Video Games and Scientific Research. DOI: 10.31219/osf.io/
cj593
5. Vamsidhar, E., Kanagachidambaresan, G.R., & Prakash, K.B. (2021) Application of
Machine Learning and Deep Learning. In Prakash, K.B. & Kanagachidambaresan, G.R.
(Eds.), Programming With TensorFlow. EAI/Springer Innovations in Communication
and Computing. Springer, Cham. DOI: 10.1007/978-3-030-57077-4_8
6. Ahmad, F., Ahmed, Z., & Muneeb, S. (2021). Effect of Gaming Mode Upon the
Players’ Cognitive Performance During Brain Games Plays: An Exploratory
Research. International Journal of Game-Based Learning, 11(1), 67–76. DOI: 10.4018/
IJGBL.2021010105
7. Spronck, P., Liu, J., Schaul, T., & Togelius, J. (Eds.) (2020). Artificial and Computational
Intelligence in Games: Revolutions in Computational Game AI: Report from Dagstuhl
Seminar 19511. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, Dagstuhl. https://
drops.dagstuhl.de/opus/volltexte/2020/12011/
8. Yannakakis, G.N. & Togelius, J. (2015, December). A Panorama of Artificial

and Computational Intelligence in Games. IEEE Transactions on Computational
Intelligence and AI in Games, 7(4), 317–335. DOI: 10.1109/TCIAIG.2014.2339221
9. Yannakakis, G.N. & Togelius, J. (2018). Game AI Panorama. Artificial Intelligence and
Games. Springer, Cham. DOI: 10.1007/978-3-319-63519-4_6
10. Yannakakis, G.N. & Togelius, J. (2018). Frontiers of Game AI Research. In Artificial
Intelligence and Games. Springer, Cham. DOI: 10.1007/978-3-319-63519-4_7
11. Shao, K., Zhentao, T., Zhu, Y., Li, N., & Zhao, D. (2019). A Survey of Deep
Reinforcement Learning in Video Games. DOI: https://arxiv.org/pdf/1912.10944.pdf
12. Bijl, J.L. & Boer, C.A. (2011). Advanced 3D Visualization for Simulation Using Game
Technology. Proceedings of the 2011 Winter Simulation Conference (WSC), pp. 2810–
2821. DOI: 10.1109/WSC.2011.6147985
13. Su, Y. (2018). The Application of 3D Technology in Video Games. Journal of Physics:
Conference Series, 1087, 062024. DOI:10.1088/1742-6596/1087/6/062024
14. Khakpour, A., & Colomo-Palacios, R. (2020). Convergence of Gamification and
Machine Learning: A Systematic Literature Review. Technology, Knowledge and
Learning. DOI: 10.1007/s10758-020-09456-4
7 A Framework for
Estimation of Generative
Models Through an
Adversarial Process for
Production of Animated
Gaming Characters
Saad Bin Khalid and Bramah Hazela
CONTENTS
7.1 Introduction................................................................................................... 123
7.2 Literature Survey........................................................................................... 125
7.2.1 Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks.............................. 126
7.2.2 Conditional Generative Adversarial Networks.................................. 126
7.2.3 Image Generation and Recognition................................................... 127
7.2.4 Progressive Growing of GANs for Improved Quality, Stability,
and Variation..................................................................................... 129
7.2.5 Diverse Image Generation via Self-Conditioned GANs.................... 129
7.2.6 Deformable GANs for Pose-Based Human Image Generation......... 130
7.2.7 Systematic Analysis of Image Generation Using GANs................... 131
7.3 Methodology.................................................................................................. 131
7.3.1 Working Intuition.............................................................................. 132
7.3.2 Architecture....................................................................................... 132
7.3.3 Training............................................................................................. 133
7.4 Results............................................................................................................ 134
7.5 Conclusion..................................................................................................... 135
References............................................................................................................... 135
7.1 INTRODUCTION
Artificial intelligence has long been associated with gaming. Deep learning (DL)
research has improved the consistency and innovations of games significantly. DL
has proven to be a highly beneficial improvisation in game development. DL is used
DOI: 10.1201/9781003231530-7 123
to power various algorithms in gaming, including multiplayer, character control,

voice recognition, avatar characteristics, and other areas. Generative Adversarial
Networks (GANs), which Ian Goodfellow presented in 2014, have proven to be a
significant improvement in the field of forgery (legally speaking) by automation.
GANs are used in this chapter to create animated characters that can be used as
avatars in games and generate various types of characters based on the data they
were trained on.
It was decided to use Convolutional Neural Networks (CNNs) and transpose con-
volutions. The CGANs, or conditional GANs, are an update to the DCGANs, or deep
convolutional GANs, which were an update to the standard GANs in the article. A
Generator (a network of transpose convolutions for generating data from the vectors
it receives) and a Discriminator are both used in GANs (a classification model that
determines whether the images are authentic or generated by the Generator). Both of
these networks are pitted against each other. The Generator’s job is to convince the
Discriminator that the fake images it produces are accurate. The Discriminator’s job
is to determine if the image is actual or generated by the Generator. Both Generator
and Discriminator start dumb and are conditioned to become more competent at the
same time when working against each other. We do not use any pre-trained models
because an untrained model will struggle to compete with its much brighter coun-
terpart. Adversary, resulting in the generation of pointless noise and, as a result, two
useless models competing for no purpose [1].
DCGANs are an advancement over GANs in that they produce and distinguish
between images using deeper networks. There are three significant steps in the train-
ing of both models in this phase. The all-convolutional net uses stride convolutions
to replace deterministic spatial pooling functions like max-pooling, allowing the
network to learn spatial down-sampling on its own. Both the Discriminator and the
Generator have used the technique above, allowing the Generator to learn spatial
up-sampling on its own. The second step is to remove the dense layers of convo-
lutional features that are entirely connected on top. Global average pooling, which
existing state-of-the-art image classification models use, is the best example that
supports the method. Direct connections of the highest convolutional features to
the Generator and Discriminator’s input and output proved to be a good compro-
mise. The GAN’s first layer, which takes a uniform noise distribution as input, is
completely connected since it is just matrix multiplication. However, the output is a
reshaped four-dimensional (4D) tensor used as the convolution stack’s first layer. The
last convolutional layer in the Discriminator is flattened and fed into a single sigmoid
output. The third step is Batch Normalization, which stabilizes learning by normal-
izing the input provided to each unit to achieve a zero mean and unit variance. This
aids in resolving training issues that may occur as a result of poor initialization, and
it also aids gradient flow in models with greater depth. This proved crucial for deep
Generators to learn while avoiding the typical failure mode in GANs, where the
Generator collapses all samples into a single stage. To prevent sample oscillation and
model instability, batch normalization was not applied directly to the Generator’s
output layer and the Discriminator’s input layer. Except for the Generator’s output
layer, which uses the Tanh function, the Generator uses the Leaky Rectified Linear
Unit (ReLU) activation [2].
A Framework for Estimation of Generative Models 125
To train both models, CGANs are a better option. It works by giving both the
networks, the Discriminator and the Generator, preconditioning and training them
in a controlled manner. These make use of a separate Generator and Discriminator.
If the Generator and Discriminator are initially preconditioned on some additional
details, such as Y, the GANs can be upgraded to a conditional model. Y may be
anything, like class codes, data from other modalities, or some other type of auxil-
iary data. If both networks are fed with Y as an extra layer of input, the condition-
ing can be done. The prior input noise and the information Y are combined in a
joint hidden representation in the Generator. The adversarial training frame allows
for considerable flexibility in how this hidden representation is composed. Y and
the Generator’s output are provided as inputs to a discriminative function in the
Discriminator [3].
7.2 LITERATURE SURVEY

Ian Goodfellow proposes a novel framework for estimating generative models
via an adversarial process in which two models are simultaneously trained: a
generative model, the Generator, that receives the data distribution and builds an
image using the received vectors, and a discriminative model, the Discriminator,
that classifies whether a data sample belongs to the training data rather than hav-
ing been produced by the Generator and estimates the probability of it being real
or fake. The training technique for Generator is to maximize the classification
error induced by the Discriminator. This structure is comparable to a two-player
Minimax game. There is a particular solution in the space of arbitrary functions
Generator and Discriminator, in which Generator recovers the Discriminator’s
training data distribution and updates the weights. When Discriminator and
Generator are made up of multilayer perceptrons, backpropagation can train the
entire machine for both Discriminator and Generator. Neither the training nor
the generation of samples requires unrolled approximate inference networks or
Markov chains [1].
The Generator is pitted against a rival in the proposed nets framework: a
Discriminator, a network equipped to determine if the sample data obtained is from
the Generator (or fake) or belongs to the actual data distribution. According to the
person behind this model, the Generator can be thought of as a team of forgers
attempting to counterfeit currency. In contrast, the Discriminator can be thought
of as a team of detectives attempting to distinguish between fake and authentic
currency. Both the forgers and the detectives improve their skills until they forged
money that is almost indistinguishable from the real thing.
This model can be used to generate training algorithms for a variety of models
and optimization algorithms. In this post, we will look at a unique situation. The
Generator produces samples from random noise using a multilayer perceptron, while
the Discriminator classifies the data as false or true using its multilayer perceptron.
Both models can be trained on the dataset and samples from the Generator using
commonly used and highly efficient backpropagation algorithms. There is no need
for Markov chains or approximate inference.
7.2.1 Unsupervised Representation Learning with Deep

Convolutional Generative Adversarial Networks
This research aims to close the gap between the effective use of CNNs for supervised
and unsupervised learning. This chapter introduces DCGANs, which are superior
to conventional GANs for image generation using CNNs. The architecture has been
trained on various datasets, and the results show that the deep convolutional adver-
sarial pair in both Discriminator and Generator has learned a hierarchy of represen-
tations. These newly acquired features are then used in more advanced tasks, such as
demonstrating their utility in general image representation.
The authors’ approach is based on adopting and modifying three recently dem-
onstrated improvements to CNN architectures. First, the all-convolutional network
replaces deterministic spatial pooling functions (like max-pooling) with stride con-
volutions, allowing the network to learn spatial down-sampling on its own. Both
the Discriminator and the Generator have used the technique above, allowing the
Generator to learn its spatial up-sampling.
On top of convolutional features, the removal of completely connected dense lay-
ers is the second step. Global average pooling, which existing state-of-the-art image
classification models use, is the best example that supports the method. The direct
relation of the highest convolutional features to the Generator and Discriminator’s
input and output proved a good compromise. The GAN’s first layer, which takes a
uniform noise distribution as input, is completely connected since it is just matrix
multiplication. However, the output is a reshaped 4D tensor used as the convolution
stack’s first layer. The last convolutional layer in the Discriminator is flattened and
feeds its data into a single sigmoid output.
Finally, Batch Normalization is a method of learning stabilization that involves
normalizing the input provided to each unit to achieve a zero mean and unit variance.
This aids in resolving training issues that may occur as a result of poor initialization,
and it also aids gradient flow in models with greater depth. This proved crucial for
deep Generators to learn while avoiding the typical failure mode in GANs, where the
Generator collapses all samples into a single stage. To prevent sample oscillation and
model instability, batch normalization was not applied directly to the Generator’s
output layer or the Discriminator’s input layer. Except for the Generator’s output
layer, which uses the Tanh function, the Generator uses the Leaky ReLU activation.
The model’s training progressed to the point that it saturated and filled the entire
color space of the training data distribution. It was discovered that the Leaky recti-
fied activation works well, particularly for higher resolution modeling. This was in
contrast to the standard GAN’s max-out activation previously used [2].
7.2.2 Conditional Generative Adversarial Networks

In this chapter, we propose a conditional version of GANs, in which both the
Generator and the Discriminator can be conditioned on data by simply feeding it into
both networks. The model has been learned to generate Modified National Institute
of Standards and Technology (MNIST) datasets, which contain binary images of
handwritten digits based on class labels, according to the authors. The work also
shows how the proposed model can be used to learn a multimodal model and gives
examples of how it can be applied to image tagging, demonstrating how the method
can generate descriptive tags that are not part of the training labels.
Using GANs to train generative models has proved to be a promising alternative.
They are getting around the problem of approximating a large number of intractable
probabilistic computations. Adversarial networks have the advantage of not requir-
ing Markov chains or probabilistic intervention because gradients are obtained using
backpropagation, and a wide range of incorporations and factors can be easily incor-
porated into the model. Furthermore, the networks will produce realistic samples
and log-likelihood estimates that are up to date. The modes of data generation are
totally out of balance in a generative model that has not been conditioned. However,
using some additional information to condition the model, the data generation pro-
cess can be guided. For inpainting, the conditioning could use data from a particular
modality, class marks, or a portion of the data. This chapter is a demonstration of
how to go about building CGANs. For analytical findings, two sets of experiments
have been demonstrated. The model was trained on the MNIST digit dataset condi-
tioned on class labels in one experiment and on the MIR Flickr 20.000 dataset for
multimodal learning in the other.
The GANs can be upgraded to a conditional model if the two networks, the
Generator and the Discriminator, are preconditioned on some additional details, say
Y, before training. Y may be information from other modalities, class marks, or
some other type of auxiliary data. As an extra layer of data, Y can be fed into both
the Generator and the Discriminator to perform the conditioning. In the Generator,
the obtained input noise and Y are combined in joint hidden representations. The
adversarial training frame aids in imparting considerable versatility in how this hid-
den representation is implemented. Y and the output from Generator are provided as
inputs to a discriminative function in the Discriminator [3].
7.2.3 Image Generation and Recognition

This study takes a deep dive into recent GAN research, summarizing the various
architectures and their implementations in various fields and methods for training
GANs for optimal results and dealing with latent space. Following that, this chapter
addresses possible future GAN research areas, including improving GAN assess-
ment, exploring different techniques for GAN training, and gaining a deeper overall
understanding of GANs. The report’s second section examines the compiled picture
dataset describing each of the seven fundamental human emotions and the experi-
ments conducted when conditioning a StarGAN on this dataset and the FER2013
dataset [4].
This chapter talks about different GAN architectures that have been proposed
since the idea of GANs was introduced in 2014 and discussed several techniques for
training the GANs. This chapter talks about the following architectures:
i. CGANs: CGANs were introduced into the field to condition the GAN model
and give it the ability to guide what it produces. This is accomplished by
using the data as an additional input layer to condition both the Generator
and the Discriminator models. One of the many benefits of CGAN is that
it allows for better one-to-many mapping representation, which means that
conditioning on a single class (e.g., cat) will synthesize a range of cats with
different colors and features [3].
ii. Stack GANs: Stack GAN is a two-stage GAN that uses text descriptions
to create photo-realistic images (proposed in 2016 by Zhang et al.). The
problem is broken down into two simpler “stages,” significantly improving
on previous approaches. The first stage (Stage-I GAN) summarizes what
is mentioned in the text before adding the primary context and object
colors. The second stage (Stage-II GAN) takes both the text description
and the low-resolution image generated in stage 1 as input. It performs
image detailing to make it more photo-realistic and accurate to the text
description [5].
iii. Cyclical GANs: Cyclical GANs function by using a cycle continuity loss,
which allows information to be transferred from one domain to another and
then back to the previous domain without loss. While cycle consistency loss
had previously been used in other areas, the CycleGAN was the first time
it was applied to GANs. The CycleGAN enables translation between two
separate image domains by training on two unordered image sets, one for
each domain. This is possible because the CycleGAN assumes an underly-
ing relationship between the two domains, allowing for training without
paired training data [4].
iv. Self-attention GANs (SAGANs): Long-range dependence modeling
and self-attention are combined in the SAGAN to generate images of
scenarios and artifacts linked in a way that is compatible with realistic
images (proposed in 2018 by Zhang et al.). Self-attention looks for con-
nections between different sections of a series in order to represent and
recognize each one. The SAGAN is the first paradigm to incorporate
self-awareness into GANs. In general, GAN models employ convolu-
tional layers in their architecture since they are effective at modeling
local dependencies, but CNN’s fall short when it comes to long-range
modeling dependencies. By applying self-attention to both networks, the
SAGAN effectively models both local and global dependencies in an
image, ensuring consistency in the highly detailed features in different
portions of the created image [6].
The different training techniques this chapter discusses are:
i. Feature matching: Making the Generator’s goal fit the expected value of
the Discriminator’s intermediate layers.
ii. Minibatch discrimination: Deals with mode collapse by making the
Discriminator examine a small number of examples rather than a single
example, allowing it to determine if the Generator is generating the same
outputs. Feature matching was found to be inferior to this form.
iii. Historical averaging: By keeping track of the historical average, it penal-
izes significantly different criteria from the average.
iv. One-sided label smoothing: This approach lowers the Discriminator’s

goal to 0.9 from 1, preventing it from reaching overconfidence and, as a
result, providing weak gradients.
v. Virtual batch normalization: Takes advantage of batch normalization,
but instead of using it, it normalizes each sample using a reference batch set
before the training begins. Since the Generator network is computationally
intensive, this is only applied to it [4].
7.2.4 Progressive Growing of GANs for Improved

Quality, Stability, and Variation
The study describes a new GAN training method. The strategy is to improve both
the Discriminator and the Generator gradually. New layers are introduced as the
training progresses, starting with a low-resolution image, modeling increasingly fine
details. This stabilizes and speeds up the preparation, enabling us to produce images
of unrivaled quality, such as CelebFaces Attribute (CELEBA) images at 10242. The
research also proposes a simple method for increasing image variation and achieves
an unsupervised CIFAR10 inception score of 8.80, a new high. The authors also go
through some implementation specifics that are important to prevent unhealthy com-
petition between Discriminator and Generator. Finally, the authors propose a metric
for evaluating GAN performance based on the shaped image quality and variation in
the images generated. Besides, the authors create a higher-resolution version of the
CELEBA dataset [7].
However, this chapter’s main contribution is a training method for GANs that
requires starting with low-resolution images and gradually increasing the resolution
by adding layers to the networks. Instead of trying to get into the finer details simul-
taneously, this increasingly finer image update helps the network learn the overall
broader scope of the image delivery first and then turn its focus to more delicate
tweaks that need to be done. A Discriminator and a Generator, which are opposites
of each other and often evolve synchronously competing with each other, are used to
learn and update their weights. Both layers in both networks are trainable during the
training period. Any layers added to the networks are faded in gradually, preventing
abrupt shocks to the layers that have already been trained for a lower resolution. As
seen in the work, Progressive training has many advantages, such as the development
of smaller images becoming more stable early on due to fewer modes and fewer class
details. Compared to the final goal of discovering a mapping from latent vectors to
10242 pictures, increasing the resolution incrementally makes weight changes much
more accessible. This method is conceptually similar to Chen and Koltun’s work,
which successfully stabilizes the training for sound megapixel image synthesis using
WGAN-GP loss and LSGAN loss.
7.2.5 Diverse Image Generation via Self-Conditioned GANs

This chapter aims to introduce a simple but successful unsupervised technique for
creating realistic and diverse photographs. Instead of training on manually annotated
class labels, the model is trained and conditioned on labels derived automatically
from clustering in the Discriminator’s feature space. The approach’s clustering

phase automatically detects variable modes, which the Generator must specifically
cover. Experiments on traditional mode collapse benchmarks show that the proposed
technique outperforms most competing approaches to addressing mode collapse.
Compared to previous approaches, the proposed methodology performs well on
large-scale datasets like ImageNet and Places365, enhancing both standard quality
metrics and image diversity.
The Generator collapses, generating minimal varieties of samples, resulting in the
Generator’s output distribution being much smaller than the support of actual data
distribution. This is a crucial problem faced in generating a wide diversity of out-
puts in such a high-dimensional space as images. A class-CGAN, which specifically
penalizes the Generator for not supporting each class, is one way to reduce mode
collapse empirically. The authors propose to use this class-conditional architecture,
but instead of accessing actual class labels, they synthesize labels in an unsupervised
manner. The technique dynamically divides the real data space into several clusters
at its most basic level, which are then used to train a class-CGAN. Since genera-
tion is conditioned on a cluster, index optimization is done about the corresponding
conditioned Discriminator, with each Discriminator allocated to each cluster of the
actual distribution. The proposed method enlivens the Generator’s output data dis-
tribution, allowing it to cover any partition of the actual distribution. The data set is
then partitioned into k clusters 1,…, k that are calculated during the training to train
the GNN to synthesize the target distribution with perfect imitation following the
actual distribution. There are no ground-truth labels used; instead, the samples are
randomly clustered in the Discriminator feature space, and the clusters are modified
regularly. A class-conditioned GAN architecture is used to separate the Generator
and the Discriminator [8].
7.2.6 Deformable GANs for Pose-Based Human Image Generation

This chapter discusses the problem of generating a human image that resembles
a specific pose that the image must be produced in. To be more specific, given an
image of a person in a particular pose, an image of the same person in a different,
alternative pose must be synthesized. The solution uses deformable skip connections
in the network’s Generator to deal with pixel-to-pixel misalignment caused by pose
differences. Furthermore, a proposal for a nearest-neighbor loss is made in place of
the usual L1 and L2 losses for matching the information of the rest of the created
image with the target image. Compared to previous work in this field, the method
used photographs of people in various poses, and the model outperforms the compe-
tition in two benchmarks. On the condition that the central point detector can extract
the articulated object’s pose, the proposed method can be used in the broader field of
deformable object synthesis.
Local information can be transferred back and forth between the encoder and
decoder parts of the Generator using the deformable skip connections. The local
data that must be transferred is typically represented by a tensor, representing the
feature map activations of the encoder’s given convolutional layer. However, unlike
before, the details to be shuttled must be chosen when taking into account object-
shape deformation, as summarized by the discrepancy between P(xa) and P(xb) (xb).
To do so, the global deformation is broken down into a series of local affine trans-
formations, which are characterized by subsets of joints in P(xa) and P(xb) (xb). The
content of F is deformed using these local masks and affine transformations built
using the relevant joints, and then skip connections are used to copy the transformed
tensor, which is then concatenated with the corresponding tensor in the destination
layer [9].
7.2.7 Systematic Analysis of Image Generation Using GANs

In recent years, GNNs have played a critical role in the advancement and growth
of unsupervised learning. The GANs, which are the epitome of image synthesis
from text or other images, have outperformed conventional approaches in terms of
efficiency. These Adversarial training networks aim to estimate the expected data
distribution from the actual distribution and use that estimate as input in synthetic
data generation. This basic concept can create a variety of frameworks that are
eidolon implementations in various domains of real-life applications, such as high-
resolution output generation, art synthesis, and image synthesis from human-drawn
sketches, to name a few. Though GANs are an improvement over conventional
approaches in principle, providing better outcomes in various areas, implementing
these architectures for real-world critical applications remains a challenge. This
research delves into and investigates these structures, including a glossary of terms
and their applications in text-to-image and image-to-image synthesis. The GNNs
are subjected to a series of crucial examinations, beginning with the basic archi-
tecture and progressing through the numerous changes and upgrades made to the
frameworks. The advantages and disadvantages of using GNNs over other tradi-
tional methods have been examined and discussed and the industries’ reliance on
GNNs for betterment [8].
7.3 METHODOLOGY
a. Convolutional layer: For generating a feature map, this layer adds a feature
detector, which is a 3 × 3 matrix of zeros and ones in general, to the image
matrix, which is also binary. The feature detector strides through the image
matrix, traversing it entirely, ANDed its inputs with the image matrix’s
inputs, performing the addition of the resulting matrix’s components, and
storing it in the feature map. After that, the resulting function map is a
much smaller picture with missing data but highlighted essential features
that can be used in a more streamlined manner later [10].
b. Pooling layer: This layer removes the important features even further and
gives the neural network spatial variance so that the essential features can
be skewed, turned, down, near, or lateral, and the model will still be able
to classify them. In this case, a feature detector strides over the feature map
from the convolutional layer without any feedback and generates a pooled
feature map that is even smaller in scale. After that, the production is placed
in the flattened sheet (a linear of features) [10].
c. Kernel size: The size of the kernel defines the field of view that comes
under that convolution. A general choice for 2D is three—that is 3 × 3
pixels.
d. Padding: The padding specifies how a sample’s border is treated. If the
kernel is more significant than 1, a (half) padded convolution will hold the
spatial output dimensions equal to the input. In contrast, unpadded convo-
lutions will crop away some of the boundaries, reducing the output size.
Padding aids in the management of production size.
e. Stride: The stride specifies the kernel’s step size when traversing the image.
The performance decreases as the stride lengthens. In the same way, that
max-pooling uses a stride of 1, we can use a stride of 2 to down-sample a
picture.
f. Transpose convolution: A transpose convolution is used to reconstruct the
spatial resolution of the size of the output image and perform convolution
resulting in a larger size image. It will be used in the construction of the
Generator [10].
g. Dense layer: This is the fully connected layer that takes the inputs, assigns
them with the wait, and is responsible for the activation caused by the
features.
7.3.1 Working Intuition
GANs are trained using two models, a Discriminator and a Generator, that com-
pete and work against each other while learning and updating their weights. The
Generator’s task is to produce images that resemble the images we are trying to
generate to trick the Discriminator into thinking they are real-world images rather
than created non-real images. The Discriminator’s goal is to determine whether the
images are natural or artificial. The Discriminator has been trained on both authentic
and generated images (initially random noise). The parameters in both the Generator
and the Discriminator are modified after each iteration. The Generator is given the
parameters individually to correct its output rather than adding up the values and
calculating the cost function as a whole [11].
Both the Generator and the Discriminator are trained at the same time to prevent
one of the networks (the Discriminator) from being too bright for the other, result-
ing in random outputs because the Generator would not know which parameters to
tweak to deceive the Discriminator.
7.3.2 Architecture
We start by creating a standalone Discriminator model with a 28 × 28 input
image. A leakyReLU layer with an alpha of 0.2 is accompanied by a convolu-
tional layer of 128 units with a scale of 3 × 3, a stride of 2, and padding of the
same. This is accompanied by a leakyReLU layer with an alpha of 0.2 and a
convolutional layer of 128 units with a scale of 3 × 3, a stride of 2, and padding

as before. Finally, we flatten the layer and use the sigmoid function to complete
the relation [11].
We also describe a standalone Generator that takes a 7 × 7 image as input and pro-
cesses it through the dense layer before applying the leakyReLU layer with an alpha
of 0.2. After that, a transpose convolution with a kernel size of 4 × 4 and a stride of
2 is used for up-sampling the 7 × 7 image from the previous layer to a 14 × 14 image.
The 14 × 14 image obtained as input from the previous layer is then up-sampled to
28 × 28 using a transpose convolution with a kernel size of 4 × 4 and a stride of 2.
Finally, the output is obtained using the Tanh layer [11].
After both the Discriminator and the Generator have been defined, we define the
combined model comprising both the Generator and the Discriminator for updating
the Generator. We make the Discriminator not trainable, get noise and input labels
in the Generator and receive image outputs from it. Image output and label input
from Generator are connected as inputs to Discriminator. GAN model is defined
as taking in noise and label and outputting a classification. Finally, the model is
compiled [12].
We first load the dataset, then pick authentic images, break the dataset into images
and labels, and finally create class labels. Points in latent space must now be created
as input for the Generator, which will then be used to construct fake examples with
class labels.
Now that both the Generator and the Discriminator are ready, the next and final
step is to train them together.
7.3.3 Training
We first need to create two objects of the Discriminator class, one that receives the
actual images. The Discriminator should learn to compute the high values, meaning
the images are accurate for the Discriminator. Another that should learn to compute
the low values, meaning the images are fake for the Discriminator. To accomplish
this, we use the binary cross-entropy function [12].
The Generator tries to achieve the opposite goal to make the Generator assign
high values to the fake images it creates.
Regularization is needed on both the Generator and the Discriminator, for which
two different optimizers are created. It is critical to specify the variables these opti-
mizers should change; otherwise, the Generator’s optimizer might screw up the
Discriminator’s variables and vice versa.
Optimizers modify the weight parameters in order to minimize the loss function.
The loss function serves as a reference to the terrain, indicating if the optimizer is
heading in the right direction to meet the valley’s bottom, the global minimum. The
Adam optimizer is used in this situation.
The Adam optimizer follows this equation:
α.mt
θt + 1 = θt −
vˆ + E
where,
t = mt
m
1 − β1t
vt
vt =
1 − βt2
mt = (1 − β1 ) gt + β1mt −1 and

vt = (1 − β 2 ) gt + β 2 vt −1

where,
α is the step size,
β1 and β2 are decay rates,
mt is the first-moment vector,
vt is the second-moment vector, and
t in the subscript is the timestep.
Initially, the timestep is initialized as zero.

The first-moment vector and the second-moment vector are also initialized as
zero for the timestep equal to zero.
β1 has the value of 0.9, β2 holds the value of 0.999, and as the timestep increases,
both are raised to the powers of the timestep.
Adam algorithm first updates the exponential moving averages of the gradient
(mt) and the squared gradient (vt), the first and second-moment estimates.
Hyper-parameters β1, β2 ∈ [0, 1) control the exponential decay rates of these
moving averages.
Moving averages are initialized as 0, leading to moment estimates biased around
0, especially during the initial timesteps. This initialization bias can be easily coun-
teracted, resulting in bias-corrected estimates.
Finally, the parameters are updated.
Through balancing their losses, we also ensure that neither the Discriminator nor
the Generator is too high. Otherwise, this would inhibit the learning of either part
and could even obviate the network from learning anything at all. That is the reason
that they trained at the same time.
The approach is to grow both Discriminator and the Generator progressively.
Starting from a low resolution, new layers are added that model increasingly fine
details as training progresses. This significantly stabilizes the training while also
speeding it up, allowing use in producing images of much better quality [12].
7.4 RESULTS
The battle in GANs is not about having the Discriminator’s highest accuracy and
lowest loss. This would have the exact opposite result to what the model is aim-
ing for. Instead, we want to train the Discriminator as much as possible to classify
FIGURE 7.1 Graph of loss of Generator and Discriminator in CGAN.
accurate data (from the data distribution) and false data (Generator by the Generator),
and then train the Generator to reduce the loss as much as possible. Both models are
trained in parallel, with the idea that if one of them is too intelligent for the other, we
will end up with two dumb models that cannot learn from each other.
The loss graph Generator in the training process has been shown in Figure 7.1.
The graph in Figure 7.1 has been attained after 100 epochs with a batch size of
32 images per batch. The model shows a decent performance in creating images
since the graphics are not too demanding in the creation of animated images.
7.5 CONCLUSION
In terms of image creation, the model performs admirably. The Generator’s batch
normalization was not applied at first, resulting in a reduction in its efficiency.
The learning significantly improved after batch normalization layers were applied.
The dense layer had to be a balanced job in terms of size because a too large layer
resulted in the model generating too similar images. Simultaneously, a too-small
layer resulted in Generator failing to learn and update the weights effectively, result-
ing in distorted images. It must be ensured that the Generator does not overfit or
underfit, and tuning must be meticulous.
REFERENCES
1. Generative Adversarial Nets by Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza,
Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengi. https://
arxiv.org/pdf/1406.2661.pdf
2. Unsupervised Representation Learning With Deep Convolutional Generative
Adversarial Networks. Alec Radford & Luke Metz, Soumith Chintala. https://arxiv.org/
pdf/1511.06434v2.pdf
3. Conditional Generative Adversarial Nets by Mehdi Mirza, Simon Osindero. https://
arxiv.org/pdf/1411.1784.pdf
4. CSGAN: Cyclic-Synthesized Generative Adversarial Networks for Image-to-Image

Transformation by Kishan Babu Kancharagunta, Shiv Ram Dubey. https://arxiv.org/
abs/1901.03554
5. A Style-Based Generator Architecture for Generative Adversarial Networks by Tero
Karras, Samuli Laine, Timo Aila. https://arxiv.org/abs/1812.04948v3
6. Self-Attention Generative Adversarial Networks by Han Zhang, Ian Goodfellow,
Dimitris Metaxas, Augustus Odena. https://arxiv.org/abs/1805.08318v2
7. Progressive Growing of GANS for Improved Quality, Stability, and Variation by Tero
Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen. https://arxiv.org/pdf/1710.10196.pdf
8. Systematic Analysis of Image Generation using GANs by Rohan Akut, Sumukh
Marathe, Rucha Apte, Ishan Joshi, Siddhivinayak Kulkarni. https://arxiv.org/abs/
1908.11863
9. Deformable GANs for Pose-based Human Image Generation by Aliaksandr Siarohin,
Enver Sangineto, Stephane Lathuili, and Nicu Sebe. https://openaccess.thecvf.com/
content_cvpr_2018/papers/Siarohin_Deformable_GANs_for_CVPR_2018_paper.pdf
10. Convolutions: Transposed and Deconvolution. https://medium.com/@marsxiang/
convolutions-transposed-and-deconvolution-6430c358a5b6
11. Conditional and Controllable Generative Adversarial Networks. https://towardsdatasci-
ence.com/conditional-and-controllable-generative-adversarial-networks-a149691ddae6
12. Implementing a Generative Adversarial Network (GAN/DCGAN) to Draw Human Faces.
https://towardsdatascience.com/implementing-a-generative-adversarial-network-gan-
dcgan-to-draw-human-faces-8291616904a
8 Generative Adversarial
Networks Based
PCG for Games
A Comprehensive Study
Nimisha Mittal, Priyanjali Pratap
Singh, and Prerna Sharma
CONTENTS
8.1 Introduction................................................................................................... 138
8.2 Background.................................................................................................... 140
8.2.1 Procedural Content Generation (PCG).............................................. 140
8.2.1.1 Game Bits............................................................................ 140
8.2.1.2 Game Space........................................................................ 140
8.2.1.3 Game Systems..................................................................... 140
8.2.1.4 Game Scenarios.................................................................. 141
8.2.1.5 Game Design....................................................................... 141
8.2.1.6 Derived Content.................................................................. 141
8.2.2 Procedural Content Generation Using Machine Learning
(PCGML)........................................................................................... 141
8.2.3 Deep Learning in PCG...................................................................... 142
8.2.4 Generative Adversarial Networks...................................................... 142
8.3 Overview of GAN in Video Games............................................................... 142
8.3.1 Level/Map Generation....................................................................... 143
8.3.2 Height Map Generation..................................................................... 144
8.3.3 Texture Synthesis in Games.............................................................. 145
8.3.4 Characters or Face Generation.......................................................... 145
8.4 Overview of Datasets and Games.................................................................. 146
8.4.1 Popular Games................................................................................... 146
8.4.1.1 Super Mario Bros (1985)..................................................... 146
8.4.1.2 The Legend of Zelda (1986)................................................ 148
8.4.1.3 DOOM (1993)..................................................................... 148
8.4.1.4 Pac-man (1980)................................................................... 148
8.4.1.5 StarCraft (1998).................................................................. 149
8.4.2 Popular Datasets................................................................................ 149
8.4.2.1 Video Game Level Corpus (VGLC)................................... 149
8.4.2.2 Idgames Archive................................................................. 149
DOI: 10.1201/9781003231530-8 137

8.4.2.3 General Video Game AI (GVG-AI)................................... 150

8.4.2.4 Sarah and Lucy Character.................................................. 150
8.4.2.5 NASA SRTM30.................................................................. 150
8.5 Future Advancement...................................................................................... 151
8.6 Conclusion..................................................................................................... 152
References............................................................................................................... 153
8.1 INTRODUCTION
Procedural Content Generation (PCG) is the automation of content production
through algorithmic means. The content can be anything that requires the involve-
ment of humans, from the creation of images to videos, poetry to music, paintings
to architectural designs. Owing to its capability to augment human creativity, with
limited or no human contribution, PCG has been an integral aspect of game devel-
opment and technical games research for years now. The increasing prominence of
PCG in games is due to its promising potential to escalate the replay values, reduce
development costs and effort, optimize storage space, or simply improvise the aes-
thetics. With games as the prime focus, PCG refers to the creation of game contents,
such as textures, maps, levels, quests, characters, stories, ecosystems, sound effects,
weapons, or even game mechanics, player interaction, and rules.
PCG for games is broadly divided into two categories, namely functional and cos-
metic. The functional PCG covers aspects like game space, systems, and scenarios
[1], which includes maps, ecosystems, levels, and stories, etc. This majorly focuses
upon the player interaction and game mechanics to enable state-of-the-art game
experiences and enable adaptivity in games as per the player. The cosmetic PCG,
on the other hand, focuses upon graphics and visualization, vegetation, architecture,
textures, and sounds. It also includes the set of rules and mechanics of the game that
constitute the game design [1].
The most sought-after and elusive goals of PCGs are to achieve the ability to
control the generated content (controllability), express a variety of unrepetitive con-
tent (expressivity), and generate believable content (believability) [2]. The techniques
used in applications of PCG in games to achieve these goals are broadly catego-
rized as constructive or traditional and artificial intelligence-based (AI-based). The
simpler traditional approaches have been to use grammars, noise-based algorithms,
fractals, Pseudo-random number Generators. The AI-based techniques are further
classified into search-based, solver-based, machine learning-based (ML-based),
and subsequently, deep learning-based (DL-based). The search-based methods [3]
employed evolutionary algorithms and stochastic and metaheuristic search/optimi-
zation techniques, while solver-based methods [4] use the design space approach to
function. The need to be fine-tuned or even explicitly be designed for specific gen-
erations is the primary flaw in standard/constructive PCG that gave rise to the devel-
opment and advancement of ML-based PCG (MLPCG). MLPCG aims to design
general algorithms that can create vast amounts of content using reasonable amounts
of data. ML-based methods have a wide category of algorithms to choose from, such
as n-grams, Markov models, Recurrent neural networks, autoencoders to name a
few. Some of the most famous algorithms deployed under DL methods are neural
Generative Adversarial Networks Based PCG for Games 139
networks, Generative adversarial networks (GANs) [5], deep variational autoencod-

ers (VAEs) [6], and long short-term memory (LSTM) [7, 8]. The dependency of
most algorithms on initial designs from humans for future automatic developments
opens up the gateway for research on DL-based game mechanics, capable of craft-
ing algorithms, parameters, constraints, and objectives for games without human
intervention.
The method under consideration in this chapter is GANs, a two-deep neural
networks-based ML algorithm. The most common issues observed in Procedural
Content Generation using machine learning (PCGML) are ill-detailed content gener-
ation, unguaranteed playability, unavailability of training data, and poor creation of
personalized content [9]. The two networks in GANs, a Discriminator and Generator,
that work in tandem and in opposition to develop content, show the potential to pro-
duce high-quality content as well as appropriate identification of unplayable games—
two of the most grieve problems. Being statistically and computationally efficient,
GANs have emerged as one of the most successful frameworks that are based on
unsupervised generative modeling [10]. Depending upon the dataset on which the
model is trained, the content can be partial or complete, autonomous, interactive, or
guided [9], stochastic or deterministic, necessary or optional [2].
Following the boom in the field of AI, there has been considerable exploration of
AI implementation in games. As opposed to highly recognized Atari 2600 games
that were used as the benchmark for visual learning algorithms [11], attention has
shifted toward AI-backed games to provide a more realistic gaming experience.
First-person shooting games have been the first choice to test new algorithms by a
fairly large number of researchers for their ease of utilization and examination. There
have been steady-state genetic algorithms [12], rule base evolution-based algorithms
such as REALM [13], Hybrid Fuzzy ANN Systems [14], navigation by imitation
via Playback of human movement [15] among many others. The convolutional deep
neural network-based VizDoom [11], a highly customizable lightweight and fast soft-
ware that is based on the classical first-person shooter (FPS) video game, Doom,
garnered attention from the industry for its rudimentary approach to visual rein-
forcement learning in three-dimensional (3D). Perhaps the most popular milestone
in this endeavor are the uses of Deep Convolutional Generative Adversarial Network
(DCGAN) [16], Token-based One-shot Arbitrary Dimension GANs (TOAD-GANs)
[17], and LSTMs [18] in the game Mario to evolve levels. Furthermore, to create
photorealistic co-creative level designs, VAEs [19] generate level segments that
can combine properties of levels from two different games—Kid Icarus and Super
Mario. The antagonistic nature of the two neural nets used in GANs presents signifi-
cant similarity with games and hence it has been the area of interest to many devel-
opers. A noteworthy development is the creation of GameGAN [20], a regenerative
deterministic and stochastic modeling algorithm. The most successful implementa-
tion of GameGAN has been on Pac-Man, where visual imitation of the desired game
is learned by the model to create newer aspects for the game, by enhancing the inter-
pretability and providing high visual consistency.
The review of the application of GANs in the field of gameplay content genera-
tion is divided into two parts. Initially, the research based on PCG is explored to
gain a broad yet well-elucidated understanding of PCG and its classification. This
is followed by shifting the focus on GAN and in particular to their application in

games PCG and a compendious overview of games based on GANs to study their
compelling nature.
Following the introduction, this chapter is further segregated into different sec-
tions that include: Section 8.2, which gives a brief overview of earlier research
done in the field of PCG, and Section 8.3 presents our review regarding these algo-
rithms. Section 8.4 describes the various games along with datasets utilized in them.
Section 8.5 provides a brief overview of potential future direction of research for
GANs in video games. Finally, Section 8.6 includes the conclusion.
8.2 BACKGROUND
Further, this section presents reviews of PCG along with its implementation via ML
and DL methodologies for several video games. Further, the procedural generation
of game contents (six classes) using different approaches is analyzed.
The comprehensive exploration of these algorithms highlighted the recent explo-
ration of GAN and its implementation by several scholars. However, the litera-
ture available is extremely limited and has not been structured properly for future
research in the same domain.
8.2.1 Procedural Content Generation (PCG)

With the propitious stature of PCG, it is natural for it to be able to create countless
aspects of games and therefore it is of the obvious need to classify the aspects into
recognizable categories. Following are the six categories [1, 2] under which the pro-
cedurally generated content can be classified into.
8.2.1.1 Game Bits

Game bits are considered the primary elements of any game content that can be
either essential or unimportant for the game. This categorization depends on whether
the game bits are concrete or abstract. However, it is notable that these components
independently cannot be sufficient for user interactions. Vegetation and buildings are
examples of the concrete domain that could be interacted directly within the game
environment while other abstract units such as texture and background music need
to be combined with other components for creating a concrete unit.
8.2.1.2 Game Space

Game space is referred to as the surrounding environment of the game that a player
interacts with where several game bits. Similar to game bits, game spaces can also be
categorized as either concrete or abstract. Game space is considered the most crucial part
of a game as it is responsible for making the gameplay interesting for the players. Indoor
maps, outdoor maps, and bodies of water are the three main types of this game content.
8.2.1.3 Game Systems

It generates and simulates complex surrounding in the game that makes the game-
play more realistic and appealing for the players. These systems include urban
environment, road network, and ecosystems along with complex player interaction
with the virtual world that makes these games more real-life.
8.2.1.4 Game Scenarios

The sequence in which the events of the game will be used to make player interac-
tion is depicted by game scenarios. These consistently keep the users engrossed in
the games by challenging the players with different types of problems to accom-
plish. Puzzles, stories, storyboards are some of the common types of scenarios that
are generally present in any game. Additionally, the concept of level has always
been a key element in the creation of games that keeps players motivated to reach
their goals.
8.2.1.5 Game Design

Game designs are another important element of gameplay that is in charge of main-
taining a balance between all the players. Specifically, these designs generate several
game rules to be followed while playing and the goals that have to be achieved by
the players.
8.2.1.6 Derived Content

This module makes players be more immersed in the gameplay by providing them
with tools to access and review their experiences in-game. News or broadcasts and
leaderboards are two main examples of these types of content that are used in almost
every arcade game.
8.2.2 Procedural Content Generation Using

Machine Learning (PCGML)
PCGML [9] concepts is based on creating novel content for arcades by initially train-
ing the model of existing content of the gameplay available. ML has been used for
several applications when integrated with PCG such as in autonomous content gen-
eration, creating co-creative and mixed-initiative design where AI assists humans for
gameplay content creation. PCGML due to its approach to use training on existing
data has been able to identify repair areas (not playable) and analyze the generated
components. Moreover, initially, PCG was introduced to compress game data due to
a lack of disk space.
There are several areas in game content where PCGML has been applied, like
real-time-based strategy levels, collectible card generation in arcade games, and
integration of interactive fiction. However, it has been evident from the research in
this domain that the major focus has always been on level generation problems. For
instance, the n-grams method was used by Dahlskog et al. [21] on original Super
Mario Bros. gameplay to create new levels. Further, this domain and Super Mario
data were explored by Jain et al. [22] by applying the concept of autoencoders.
Snodgrass and Ontan integrated Multi-Dimensional Markov chains with PCG[23].
They have surveyed Markov chains in-depth to ensure usable and higher levels for
gameplay. Furthermore, Bayes Net and Principal Component Analysis (PCA) [24]
have been implemented on Legend of Zelda for level and room creation. Researchers
have also worked on predicting resource locations on StarCraft II maps using neural
networks [25].
8.2.3 Deep Learning in PCG

Recently, DL is having a great impact in the field of PCG [26] by integrating several
techniques including GAN [5], reinforcement learning such as LSTM [7, 8], VAEs
[6], and evolutionary computing. These methods have generally been employed to
generate creative game content, however, sometimes these generated results require
certain adaptations for creating content that is practical to use. Of these, reinforce-
ment learning is a recent domain being explored in the field of arcades. Similar to
PCGML, most researches in DLPCG to date focus majorly on generating playable
levels [26]. Another interesting area in DLPCG is the implementation of GAN in
gameplay content generation that will be discussed in detail in Section 8.3.
8.2.4 Generative Adversarial Networks

GANs [5] are one of the most prominent deep generative models that are based
on adversarial training techniques. It can be assumed to be an adversarial game
between two players, Generator (G) and Discriminator (D) whose objective is to
maintain the Nash Equilibrium (NE). Hence, the basic principle behind adversarial
networks is to map a random noise vector through a Generator to generate a sample.
This resultant is further analyzed by the Discriminator to classify the sample as real
or fake. Training these adversaries ensures the Discriminator achieves a maximum
state of correctly distinguishing the generated samples.
At present, GANs are being extensively employed in different fields such as
computer vision and natural language processing for several tasks. In vision-related
systems, GANs are being applied for image translation, improving the resolution
of images, texture synthesis, face synthesis, and image colorization. While in the
domain of natural language processing (NLP), music generation, and dialogue gen-
eration. Moreover, GANs have also been applied to the medical domain for solv-
ing different problems. Furthermore, adversarial learning is being employed in the
area of video games for various generation tasks that will be elaborated in the next
section.
8.3 OVERVIEW OF GAN IN VIDEO GAMES

This section gives a brief overview regarding different approaches been employed in
gameplay content generation using the concept of adversarial learning. GANs and
their variants are the most popular model being explored recently for several tasks.
In the area of games, to date, they have been mostly used for either 2D or 3D level/
map generation, along with certain research in creating face/characters for game
and texture synthesis for game modules. Figure 8.1 represents the rate of evolution
of GAN in the area of arcade games. We will discuss all the research in this field in
detail in subsequent subsections.
FIGURE 8.1 Evolution of GAN in video games.
8.3.1 Level/Map Generation
Level generation is the process that is responsible for creating a playable and interest-
ing environment or scenarios for gameplay and is the most important aspect of any
video game. Fabricating levels require designers to have both artistic and technical
acumen. Generating the arcade maps requires extensive playtesting, hence, from the
past few years, designers have inclined toward exploiting the domain of ML for mod-
eling-level designs. Video games are either generate 2D levels or convert these to 3D
maps for creating advanced games. There has been the implementation of several
techniques to create levels using PCG. However, recently GANs [5] have emerged to
be the most popular deep generative method for arcade content generation.
In 2018, Giacomello et al. [27] applies the concept of GAN for generating new
and unique levels on DOOM [28] video game series. Their approach included imple-
menting Wasserstein GAN with Gradient Penalty (WGAN-GP) in two ways. The
first one only had images and noise vector as the input to the model and was named
as unconditional WGAN-GP. On the other hand, in conditional WGAN-GP there is
an additional input of extracted features from existing DOOM levels that showed
better results in comparison to the unconditional model. Volz et al. [16] processed
another popular game by Nintendo, Super Mario Bros [29] for stage creations. In
this research, the main intent was to generalize the generation process for different
games, hence, the authors divided the technique into two phases. In the first stage, an
unsupervised method is used to train the GAN model which is followed by identify-
ing ideal input vectors from the latent space. However, to avoid random sample gen-
eration evolutionary control in the form of Covariance Matrix Adaptation Evolution
Strategy (CMA-ES) [30] is applied while exploring the inputs followed by fitness
function evaluation. This procedure is extended by Giacomello et al. [31] where they
used the CMA-ES technique on DOOM to generate novel levels according to the
features required in the new gameplay stage. Prominently three types of levels were
considered for this experiment, arena level, labyrinth level, and complex level.
The education domain is considered challenging for creating video games by
applying PCG due to its requirement of fulfilling the learning objectives through it.
Therefore, there has been limited research in this field. However, Park et al. [32] pro-
posed a multistep DCGAN (deep convolutional GAN) for generating novel and solv-
able levels. The model comprises two Generators, where the first one is responsible
for creating a large dataset of synthetic levels from a small set of example stages. The
generated data is further fed to the second Generator to produce levels with increased
solvability. Furthermore, Torrado et al. [33] proposed conditional embedding self-
attention generative adversarial network (CESAGAN) where the bootstrapping tech-
nique was integrated to efficiently train Generators and Discriminators. The model
is based on combining self-attention GAN (SAGAN) [34] with a conditional vector
to improve the diversity, uniqueness, and playability of the generated levels.
Gutierrez and Schrum [35] integrated both GAN and graph grammar to generate
dungeon rooms for The Legend of Zelda [36] video game. In their initiated work, the
model employs both GAN and graph grammar equally that resulted in the generation
of an interesting and playable dungeon layout with the placement of obstacles and items
in it. Followed by this research, Schrum et al. [37] combined GAN and Compositional
Pattern Producing Network (CPPN) [38] to tackle the issue of arranging structured lev-
els using segments. The CPPN exploited the latent vectors that are directly associated
with game segments that resulted in the generation of complete levels.
Furthermore, Awiszus et al. [17] introduced TOAD-GAN, a unique methodology
to generate new Mario levels by considering the scarcity of data. The proposed tech-
nique is based on SinGAN architecture [39] and employs a single training level to
produce tile-based game stages. Bontrager and Togelius [40] have recently proposed
a notable methodology on a 2D dungeon crawling game by implementing Generative
Playing Networks (GPN). The model comprises of Generator and agent that have
direct communication for generation of gameplay levels that results in the require-
ment of less training data. Moreover, Constrained Adversarial Networks (CANs) by
Di Liello et al. [41] penalizes the GAN network for inappropriate and invalid structure
generation. The application of CAN has produced efficient results without affect-
ing the run-time of the model. Moreover, GameGAN, the recent development by
NVIDIA [20] has integrated deterministic and modeling algorithms on Pac-Man to
generate novel aspects related to the gameplay and providing high visual consistency.
Apart from specific models, there have been research on developing a general
method for game content generation. The first one is introduced by Irfan et al. [42]
where the model consists of DCGAN being train on three different games includ-
ing Colourescape, Zelda, and Freeway. It was observed through this research that
the generated stages were playable and a large amount of data assisted DCGAN to
capture required data in the levels. Another approach was initiated by Kumaran et al.
[43], in this approach branched GAN was trained that can generate levels of four
different arcades using a single random vector as the input to the architecture and
captures the variation present in the data while training.
8.3.2 Height Map Generation

In computer graphics, height maps are the 2D projected image of points for dis-
placing the 2D mesh to create the 3D terrain. Creating interactive 3D landscapes is
essential for user interactions and to maintain the engagement of players. Moreover,
as modeling of level maps might require a height map, researchers have also explored
the domain of generating height maps from 2D images. As compared to the domain
of level generation, research resources are limited in the area of height map genera-
tion. Broadly there have been four works on producing height maps in video games
using GANS.
Beckham and Pal [44] in their research identified the first technique to generate
height and texture maps using GANs. Initially, they trained their data on DCGAN to
produce height maps that were further processed using pix2pix GAN for generating
their corresponding textures. As shown in that research, they rendered height and
texture maps to generate 3D terrain for video games. Furthermore, Wulff-Jensen
et al. [45] in 2018 worked on Digital Elevation Maps (DEM) for alps [46] to gener-
ate 3D landscapes for enhancing gameplay interactivity. They trained DCGAN on
Elevation Maps and then converted the same to 3D maps with Unity3D. The resultant
maps were realistic and could be deployed while creating different arcade games.
In 2019, Spick and Walker [47] worked on creating a model that could generate
several variants of a specific landscape region. For this purpose, they trained Spatial
GAN (SPA-GAN) [48] with non-spatial feature learning. The height and texture
maps were rendered through a 3D game engine to produce realistic and endless vari-
ants of any specific area.
Recently, there has been a novel approach to generate terrain for gameplays using
rough sketches. Wang and Kurabayashi [49] proposed a unique methodology involv-
ing a generative model of two phases where the conditional generative adversarial
network (cGAN) is responsible for generating different variants of elevation bitmap
of the sketch. The model in this phase generates maps corresponding to the terrain
data that has been used for training. In the second phase, the deterministic algorithm
produces the actual asset for the terrain by interpreting the details on the elevation
bitmap. Hence, this approach will help various designers to create interactive game
maps by simply converting their rough sketches.
8.3.3 Texture Synthesis in Games

Texture synthesis is an essential component for enhancing realism in game con-
tent and makes the environment more aesthetic for game players. This enhancement
eventually increases the interactivity of the arcade with the users. Moreover, it is
evident from earlier research work that texturing is either applied through supervised
or automated methods. However, the research in the field of synthesizing textures for
gameplay content is a limited single experiment by integrating it with GANs.
Fadaeddini et al. [50], in their research, proposed a novel technique to apply tex-
ture synthesis for PCG in video games. They constructed a simple adversarial net-
work where the network is trained on sample images of textures collected and then
generated game levels using those textures.
8.3.4 Characters or Face Generation

Sprites are 2D animations or bitmap images that are integrated into the video
game. These gameplay elements are not static and perform animation by moving
independently. In games, they represent the users in the gameplay by acting as an

interface between game players and the game environment. There have been several
studies on generating sprites through PCG. Nevertheless, in this subsection, we will
discuss their generation with the help of GANs.
Horsley and Perez-Liebana [51] introduced the technique of integrating DCGAN
to generate sprites in video games. They considered three categories for analysis
including Faces, Creatures, and Human-like sprite generation. Each of these cat-
egories produced satisfactory results even with a small dataset, however, at certain
points the model needs to have a higher epoch count for better results.
Another interesting research by Hong et al. [52] proposes Multi-Discriminator
GAN (MDGAN) that the unique sprites could be generated by only considering
three image-related parameters; color, shape, and animation involved. The approach
involves encoder networks for shape and color extraction while the decoder pro-
cesses each encoder’s inputs and applies a Discriminator network for these sprites.
This methodology performs image-to-image translation for the 2D production of
sprites.
In 2019, Serpa and Rodrigues [53] performed research on generating enhanced
gameplay sprites using GANs. Hence, they designed a variant of the pix2pix model
that comprises two parts. In the first part, there is a line-to-gray algorithm that con-
verts the initial sprite lines to grayscale which is followed by conversion to colored
sprites using a gray-to-color algorithm. The results obtained from this model depict
enhancement in the quality of sprites as they involve the shading effect.
Table 8.1 depicts the main contributions of each GAN in each subdomain of
gameplay content generation along with their short description.
8.4 OVERVIEW OF DATASETS AND GAMES

This section gives a brief elucidation on different games that are being used for the
intent of research in the domain of PCG using GAN following their corresponding
datasets from which they are extracted.
8.4.1 Popular Games
Games have always been favored by people from all generations and hence this
area has innately advanced with time. Earlier the games used to be created manu-
ally, after which methods for automatic creation were examined, leading the inter-
est toward PCG algorithms for the same. However, in recent years researchers are
becoming interested in integrating them with advanced techniques like GANs for
richer and enhanced content. However, throughout this period, certain games have
been admired by researchers for experiments. Therefore, in this subsection, we will
discuss some of those favorable games.
8.4.1.1 Super Mario Bros (1985)

Super Mario Bros [29] is a classical side-scrolling platform game, developed by
Nintendo. With different power-up options, the evaluation for the games starts with
Mario in a normal state, only capable of left-right scrolling and jumping. The game
TABLE 8.1
Main Contributions of the Above Explored Algorithms
Game Content Type Variant Author Short Description

Map/Level WGAN-GP for Giacomello Implemented WGAN with gradient penalty
Generation DOOM Levels [7] et al. in conditional and unconditional way
Mario Levels with Volz et al. Integrated the concept of latent space and
DCGAN [33] CMA-ES to generate Mario Levels
WGAN-GP with Giacomello Extended WGAN-GP by employing latent
Latent Space [34] et al. space search
Multistep DCGAN Park et al. Developed educational game based on
[35] DCGAN
CESAGAN [36] Torrado et al. Based on SAGAN and bootstrapping
technique for improved performance
GAN Rooms for Gutierrez and Incorporated GAN and graph grammar to
Zelda [37] Schrum generate rooms
CPPN2GAN [38] Schrum et al. Utilized CPPN to generate complete levels
with segments
TOAD-GAN [17] Awiszus et al. Based on implementation of SinGAN and
single level training data on Mario dataset
Generative Paying Bontrager and Architecture similar to GAN with
Networks [40] Togelius Discriminator replaced to agent for direct
communication
Constrained Di Liello et al. Novel method where GAN penalizes
Adversarial Network network for invalid generated content
[41]
Game Gan [20] Kim et al. Integrated deterministic and modelling
algorithms on Pac-Man
DCGAN based Irfan et al. Generated playable levels by training the
General PCG [42] GAN architecture on 3 datasets
Branched GAN based Kumaran et al. Based on concept on random vector and
General PCG [43] variation training
Height Map Terrain Generation Beckham and Generation of 3d height maps using
Generation using DCGAN [44] Pal DCGAN and pix2pix GAN
Height Map using Wulff-Jensen Enhanced gameplay interactivity using
DEM [45] et al. digital elevation maps of alps
Height and Texture Spick and Trained Spatial-GAN on non-spatial
Map using Walker features to produce endless variants
SPA-GAN [47]
Sketch2Map [49] Wang and Utilized conditional GAN and deterministic
Kurabayashi algorithm
Texture Synthesis Texture Synthesis in Fadaeddini Applied texture synthesis of custom created
Video Games [50] et al. dataset on game images
Character or Face Sprite Generator with Horsley and Integrated DCGAN for three categories,
Generation (Sprites) DCGAN [51] Perez-Liebana human-characters, faces and creatures
Game Sprite using Hong et al. Implemented multi-Discriminator GAN and
MDGAN [52] image-to-image translation
Pixel Art for games Serpa and Employed variant of pix2pix with
[53] Rodrigues line-to-gray and gray-to-color algorithm
consists of eight “worlds,” each world with its own set of four sub-levels called
“stages.”
Each game starts with players having a certain number of lives and which may be
increased by gaining additional lives by defeating several enemies in a row with a
Koopa shell, picking up certain power-ups, or by collecting a certain number of coins
or bouncing on enemies successively without getting in contact with the ground.
Contact with enemies converts him into big mode and subsequent contact with ene-
mies converts him into the regular state, instead of dying. The lives count decreases
if Mario takes damage while small, falls in a bottomless pit, or exceeds the time
limit. Once the lives count strikes zero, Mario dies and the evaluation ends.
8.4.1.2 The Legend of Zelda (1986)

The Legend of Zelda [36] is an action-adventure dungeon crawler, from Nintendo.
The major genres explored in the game are puzzles, action, adventure/battle game-
play, and exploration itself. The three principal areas in the games are overworld,
interconnection of all the areas; areas of interaction with other characters; and dun-
geons, the areas of labyrinthine layout wherein most of the gameplay occurs. The
main character, Link, explores the overworld to find and explore several maze-like
dungeons full of enemies, traps, and puzzles and takes aid to complete his missions
by gaining special items or advice from other characters, or purchase equipment, or
even complete side quests in the interaction areas.
The evaluation progresses as the player completes the assigned tasks without
depleting the life meter. The life meter drops with combats with dungeon bosses and
can be increased beyond just three hearts by attaining “Heart Containers.” At the
start of the evaluation, the player is given the extendable string of three hearts, rep-
resenting the number of hits that can be withstood before the end of the evaluation.
8.4.1.3 DOOM (1993)

Developed by id Software, Doom [28] is a classical early 3D graphics-based FPS
game, for MS-DOS. The game has a series of levels, with increasing difficulties, and
several distinct levels constitute episodes. The game features three episodes: “Knee-
Deep in the Dead,” “The Shores of Hell,” and “Inferno.” The continuous succession
in completion of levels in an episode progresses to entry in the next episode.
Evaluation requires exploration of some military base, in each level, in search
of exits by the player as the protagonist character, Doomguy. While traversing the
base, the player is required to defeat enemies in each level while managing supplies
of ammunition, health status, and armor points. With each combat, the health points,
armor points, and ammunition of the player decreases owing to the damage incurred
during the combat. The player can replenish the depleted points by collecting power-
ups, weapons, and ammunition throughout the levels and from defeated enemies.
The evaluation ends as and when the health points cease to zero.
8.4.1.4 Pac-man (1980)

Developed by Namco, Pac-Man [54] is a maze chase arcade video game. The game-
play requires controlling an eponymous character, Pac-man, through an enclosed
maze, populated with four varieties of enemies, that are presented as ghosts-Inky
(cyan), Clyde (orange), Pinky (pink), and Blinky (red). Each ghost is given its own
unique and distinct “personality,” and operates differently, albeit with the same goal
to catch the protagonist. The aim is to eat all the dots are in the maze, consistently
avoiding the ghosts. Successful consumption of all dots in a level advances the player
to the next, more challenging level.
The evaluation in this game begins with a certain number of lives allotted to the
player and each contact with ghost results in loss of life. The seemingly routine play
is coupled with energizers, warp tunnels, and bonus items. The evaluation ends when
the lives count is exhausted.
8.4.1.5 StarCraft (1998)

StarCraft [55] is a military science fiction-based real-time strategy game developed
by Blizzard Entertainment. Instead of a single protagonist, the game features three
different intelligent races—Terrans, Zerg, and Protoss—each fighting to establish
dominance over others. Each species is granted its own unique set of strengths, pow-
ers, and abilities. The game ensures that no race has an innate advantage over any
other and that there is uniqueness in the composition of each race.
The primary focus of the game is on strengthening the capabilities of the cho-
sen race by collecting resources, building bases, constitute armies, and finally wage
wars against the other race. Though each race relay on two key resources, minerals
and vespene gas, and supply mechanic operates in the same fashion for each race,
the resource management for each race is unique as the nature of the supply differs
between the races. On a very similar tone, base constructions for different species all
differ and provides different possibilities as well challenges. Given the competitive
nature of the game, it has been widely credited with revolutionizing the real-time
strategy genre and rightly so, considering the scaling AI that has been used in it.
8.4.2 Popular Datasets
Data is the primary and essential element for research in any domain. In-game con-
tent creation, some popular corpora are employed frequently when PCG is integrated
with GANs. Hence, in the following subsections, a brief overview of these datasets
is mentioned.
8.4.2.1 Video Game Level Corpus (VGLC)

VGLC [56] is the most famous repository used by researchers for accessing data to
be used for training purposes in the field of PCG for gameplay. The corpus comprises
12 games that include some of the prominent games in PCG such as Super Mario
Bros, The Legend of Zelda, DOOM, and many more. The data has been annotated as
Graph, Tiles, and Vectors. Moreover, most of the data is currently being employed in
research for arcade level generations.
8.4.2.2 Idgames Archive

Idgames archive [57, 58] is the primary source for Doom engine-related tools, levels,
and modifications maintained by the community. This repository has been one of the
largest open-source archives for all the Doom data, and although it started in 1994,
it still receives weekly entries for Doom games that further assists in the research
domain. The database consists of several different types of training data for Doom
including, levels, textures or skins, music-related, and many more.
8.4.2.3 General Video Game AI (GVG-AI)

General Video Game AI (GVGAI) [59, 60] was initially introduced by researchers by
encouraging novel research in the domain of gameplay to generate game agents with
unknown testing environments. This competition framework comprises mainly two
PCG generation objectives namely, level generation and rule generation. Moreover,
recently this platform has been used by researchers as small dataset for different
PCG implementations in generating gameplay content.
8.4.2.4 Sarah and Lucy Character

This dataset comprises two characters, Sarah and Lucy [53], selected from Trajes
Fatais arcade that is integrated for gameplay character generation or sprite production.
From the Trajes Fatais game, we selected the Sarah and Lucy characters as data-
sets to evaluate the usefulness of the Pix2Pix architecture. The Sarah character has
only 87 completed sprites while 207 uncompleted ones. It’s also a character with a
relatively fluid personality, with many smooth and complex areas. Lucy, on the other
hand, is completed, so it has 530 completely drawn sprites and is relatively simple to
draw, with mainly smooth features.
8.4.2.5 NASA SRTM30

The NASA SRTM dataset [61] contains the elevation map data of different regions
of the planet. This corpus is used to create a height map using the points from the
Shuttle Radar Topography Mission (SRTM) dataset. Moreover, the data comprises
of single texture where white depict high values in comparison to black regions
depicting low height values.
Apart from the above-mentioned datasets, there were other corpora used as data
for producing arcade game content that was manually created by the researchers
for their specific purpose. Moreover, of these five training datasets, the Sarah and
Lucy characters and NASA SRTM30 have been comparatively less popular among
the researchers. Table 8.2 depicts details regarding the three prominent datasets for
research.
TABLE 8.2
Popular Datasets in Video Game Content Generation
Number of
Dataset Games Available Most Prominent Games Dataset Composition
VGLC [56] 12 Super Mario Bros, Doom, Tiles, Graph, Vector
The Legend of Zelda
Idgames Archive [57, 58] 1 Doom Levels, Skins, Music, etc.
GVGAI [59, 60] No specific count Random generation of data Two categories: level and
rule generation
8.5 FUTURE ADVANCEMENT

A noteworthy pressing issue commonly encountered in GANs specialized for games
is the scarcity of the research done for aspects such as character generation, terrain
generation, and texture synthesis for games. While there has been extensive research
for level map generation for video games, other functional aspects have hardly seen
any exploration. Even cosmetic aspects such as texture generation, vegetation gen-
eration, or utilization of different architecture do not have any substantial automa-
tion development. Some potential promising architectures for developing GANs for
games are cycle consistent adversarial networks, that are capable of translating the
styles and characteristics in one sample into another; and conditional GANs, capable
of investigating the incorporation of additional constraint checkers into the solver
and explore their impact and how the generated content address specific learning
objectives. Another possibility for future scope is to explore the self-attention GAN
model which can better track patterns with long-range dependencies.
A noticeably rarely researched area is the two-stage GAN framework for games,
due to the obvious extensive computational requirements. There have been notable
developments in deep generative modeling in recent years. However, none of the
models, new or old, has shown to satisfactorily generate content on its own. An
obvious step in this regard would be the development of more sophisticated models.
Another not-so-obvious yet potentially more practical approach would be to use a
mixture of existing models. The Generator and Discriminator both could be based
on different models or two sets of GANs could be trained on different models and the
final results could be combined to generate more realistic content. This idea opens
up a plethora of possibilities in games as this lifts the burden of single-handedly
creating rich, believable content from a single GAN. A game-based GAN that can
perfectly create 3D meshes can be coupled with another GAN that adds texture to
the mesh, resulting in a richer experience. It also holds the possibility to be further
developed into a full-fledged tool to aid creators and producers in generating high-
quality content. This, of course, comes with its computational limitations, and exten-
sive research is required. Also, more advanced GAN-specific loss reductions need
to be explored to ensure minimization of reconstruction losses, especially in height
maps and prevent mode collapse and partial mode coverage.
Using GANs for games is essentially optimization of a Minimax objective [5],
that is optimizing multiple objectives jointly. Theoretically, this is a Nash equilib-
rium (NE) problem; but given the NP-hard nature of finding a global NE, gradient-
based algorithms, especially gradient descent ascent (GDA), are more commonly
used. But this approach too comes at a cost. The GDA has its inherent flaws when
dealing with joint optimization. There is a potential for the existence of cycles,
and therefore there are no convergence guarantees [62]. Next, in case the gradient
descent does converge, there is a high probability that the rate is too slow in practice
as recurrent dynamics require extremely small learning rates [62, 63]. The existence
of multiple objectives leaves no room for progress monitor [63]. From the game-
theoretic aspect, the duality gap (DG) [64] seems to be a natural consequence and
can be used to evaluate the performance of GANs. Interestingly the evolution of the
DG tracks the convergence of the algorithm to an optimum. Even so, only the issue
of progress monitoring, using DG, has been satisfactorily catered to. There has been
the utilization of Parallel Nash Memory (PNM) or the addition of fake uniform data
to act as a guiding component for adversarial training [10]. Another suggestion has
been to optimize games centric GANs by using mixing strategies, such as NE itself
with gradient descent. But so far this is only of theoretical interest as there is no
known DL (i.e., gradient-based) method yet to optimize these mixed strategies that
can stand the large support sizes needed in games.
Yet another unexplored area in game-centric GAN development is disentangle-
ment. Disentanglement is learning to distinguish between distinct, informative fac-
tors of variations of data. Unlike Bayesian generative models with their probabilistic
framework, in GANs, there is the perpetual absence of sample likelihood and poste-
rior inference for latent variables. Therefore, learning a factorized representation or
especially disentangling the interclass variation in GANs poses an obvious obstruc-
tion in its advancement. Improvisation in disentanglement for GANs specifically
utilized for games is essential to attain the goals of believability and expressivity that
PCG demands. There are limited disentanglement methods for GANs and even more
scarce in respect to gaming requirements. They can be classified into supervised
ones and unsupervised ones, however, the theoretical impossibility of unsupervised
disentanglement learning without inductive biases [65] has been a topic of debate and
has led to a piqued interest in contrastive learning [66–68], another barely explored
area, that needs little supervision, in contrast, to complete unsupervised alternatives.
8.6 CONCLUSION
GANs-based PCG for games has an excitingly promising potential to significantly
reduce costs and development time of games, improvise replay values by producing
high-quality fresh content with incredible aesthetics, while also managing the opti-
mization of memory. A fairly modern research topic, it is the descendent from ML
(DL) based PCG and is yet to be fully explored and the therefore little intersection
of academic research and industrial development is found. This survey attempts to
provide a comprehensive flow of how constructive PCG paved the path for AI-based
PCG and consequently MLPCG and eventually to PCG-GAN. It lists the various
challenges faced and potentials of former PCG models—constructive, search-based,
solver-based, simple CNN-based, and other regenerative-based models, among oth-
ers. This chapter also explores how PCG-GAN can prove to be a futuristic attempt
at providing a rich, photorealistic gaming experience and provide playable, authen-
tic gaming content superior to prior generated content and comparable to human
augmentations-based creations. A multitude of PCG models is examined, each with
a detailed description so that this survey can act as a single point stop for a com-
prehensive overview of all the work done in this field. Since the major point under
consideration is to demonstrate the potential of GAN on games, a variety of GAN-
based games are presented along with the datasets they utilize. This section also
sheds light on some of the ready-to-use off-the-shelf tools that have already been
and can further be utilized for advancement. Different models based on which vari-
ous researchers have trained the GANs for games have also been elaborated on to
understand and build upon.
Beyond the typical problem of hardware and GPU-related issues faced in most
GAN applications, the specific problems associated with the application of GANs
for games, such as the inability for disentanglement, lack of exploration of aspect
creations beyond level generations, overpowering of networks, etc. have also been
briefly discussed. Based on the problems faced an enormous amount of scope for
future work, this chapter concludes with the elucidation of some potential future
works that can be carried out for further advancements in the application of GANs
for games.
REFERENCES
1. Hendrikx, M., Meijer, S., Velden, J.V.D., & Iosup, A. Procedural Content Generation
for Games: A Survey. ACM Transactions on Multimedia Computing, Communications
and Applications (ACM TOMCCAP), 9, 1–22 (2013).
2. Barriga, N.A. A Short Introduction to Procedural Content Generation Algorithms for
Videogames. International Journal on Artificial Intelligence Tools, 28(11), 1–12 (2019),
1930001.
3. Togelius, J., Yannakakis, G.N., Stanley, K., & Browne, C. Search-Based Procedural
Content Generation: A Taxonomy and Survey. IEEE Transactions on Computational
Intelligence and AI in Games, 3, 172–186 (2011).
4. Smith, A.M., & Mateas, M. Answer Set Programming for Procedural Content Generation:
A Design Space Approach. IEEE Transactions on Computational Intelligence and AI in
Games, 3, 187–200 (2011).
5. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
Courville, A.C., & Bengio, Y. Generative Adversarial Nets, NIPS, 2672–2680 (2014).
6. Kingma, D.P., & Welling, M. Auto-Encoding Variational Bayes. CoRR. abs/1312.6114
(2014).
7. Greff, K., Srivastava, R., Koutník, J., Steunebrink, B., & Schmidhuber, J. LSTM: A
Search Space Odyssey. IEEE Transactions on Neural Networks and Learning Systems,
28, 2222–2232 (2017).
8. Hochreiter, S., & Schmidhuber, J. LSTM can Solve Hard Long Time Lag Problems.
NIPS (1996).
9. Summerville, A., Snodgrass, S., Guzdial, M., Holmgård, C., Hoover, A.K., Isaksen,
A., Nealen, A., & Togelius, J. Procedural Content Generation via Machine Learning
(PCGML). IEEE Transactions on Games, 10, 257–270 (2018).
10. Oliehoek, F.A., Savani, R., Gallego-Posada, J., Pol, E.V., Jong, E.D., & Groß, R.
GANGs: Generative Adversarial Network Games. ArXiv. abs/1712.00679 (2017).
11. Kempka, M., Wydmuch, M., Runc, G., Toczek, J., & Jaśkowski, W. ViZDoom: A Doom-
based AI research platform for visual reinforcement learning. 2016 IEEE Conference
on Computational Intelligence and Games (CIG), 1–8 (2016).
12. Esparcia-Alcázar, A., García, A.M., Guervós, J.J., & García-Sánchez, P. Controlling
bots in a First-Person Shooter game using genetic algorithms. IEEE Congress on
Evolutionary Computation, 1–8 (2010).
13. Bojarski, S., & Congdon, C. REALM: A Rule-Based Evolutionary Computation Agent
that Learns to Play Mario. Proceedings of the 2010 IEEE Conference on Computational
Intelligence and Games, 83–90 (2010).
14. Rhalibi, A., & Merabti, M. A Hybrid Fuzzy ANN System for Agent Adaptation in
a First-Person Shooter. International Journal of Computer Games Technology, 2008,
432365:1–432365:18 (2008).
15. Karpov, I.V., Schrum, J., & Miikkulainen, R. Believable Bot Navigation via Playback
of Human Traces. Believable Bots (2012).
16. Volz, V., Schrum, J., Liu, J., Lucas, S., Smith, A., & Risi, S. Evolving Mario Levels in the
Latent Space of a Deep Convolutional Generative Adversarial Network. Proceedings of
the Genetic and Evolutionary Computation Conference (2018).
17. Awiszus, M., Schubert, F., & Rosenhahn, B. TOAD-GAN: Coherent Style Level
Generation from a Single Example. Proceedings of the AAAI Conference on Artificial
Intelligence and Interactive Digital Entertainment, 16(1), 10–16. ArXiv. abs/2008.01531
(2020).
18. Summerville, A., & Mateas, M. Super Mario as a String: Platformer Level Generation
via LSTMs. ArXiv. abs/1603.00930 (2016).
19. Sarkar, A., Yang, Z., & Cooper, S. Controllable Level Blending between Games using
Variational Autoencoders. ArXiv. abs/2002.11869 (2020).
20. Kim, S.W., Zhou, Y., Philion, J., Torralba, A., & Fidler, S. Learning to Simulate
Dynamic Environments With GameGAN. 2020 IEEE/CVF Conference on Computer
Vision and Pattern Recognition (CVPR), 1228–1237 (2020).
21. Dahlskog, S., Togelius, J., & Nelson, M.J. Linear Levels Through n-Grams. MindTrek,
200–206 (2014).
22. Jain, R., Isaksen, A., Holmgard, C., & Togelius, J. Autoencoders for level generation,
repair, and recognition. In Proceedings of the ICCC Workshop on Computational
Creativity and Games (2016).
23. Snodgrass, S., & Ontañón, S. Experiments in Map Generation Using Markov Chains.
FDG (2014).
24. Summerville, A., & Mateas, M. Sampling Hyrule: Multi-Technique Probabilistic Level
Generation for Action Role Playing Games. AIIDE (2015).
25. Lee, S., Isaksen, A., Holmgård, C., & Togelius, J. Predicting Resource Locations in
Game Maps Using Deep Convolutional Neural Networks. AAAI (2016).
26. Liu, J., Snodgrass, S., Khalifa, A., Risi, S., Yannakakis, G.N., & Togelius, J. Deep
Learning for Procedural Content Generation. Neural Computing and Applications. 33,
19–37 (2021).
27. Giacomello, E., Lanzi, P.L., & Loiacono, D. DOOM Level Generation Using Generative
Adversarial Networks. 2018 IEEE Games, Entertainment, Media Conference (GEM),
316–323 (2018).
28. Doom (franchise). This page was last edited on 13 September 2021, at 01:55 (UTC)
https://en.wikipedia.org/wiki/Doom (franchise).
29. Super Mario Bros. This page was last edited on 13 September 2021, at 17:08 (UTC)
https://en.wikipedia.org/wiki/Super_Mario_Bros.
30. Hansen, N., Müller, S., & Koumoutsakos, P. Reducing the Time Complexity of the
Derandomized Evolution Strategy with Covariance Matrix Adaptation (CMA-ES).
Evolutionary Computation, 11, 1–18 (2003).
31. Giacomello, E., Lanzi, P.L., & Loiacono, D. Searching the Latent Space of a Generative
Adversarial Network to Generate DOOM Levels. 2019 IEEE Conference on Games
(CoG), 1–8 (2019).
32. Park, K., Mott, B., Min, W., Boyer, K., Wiebe, E., & Lester, J.C. Generating Educational
Game Levels with Multistep Deep Convolutional Generative Adversarial Networks.
2019 IEEE Conference on Games (CoG), 1–8 (2019).
33. Torrado, R., Khalifa, A., Green, M.C., Justesen, N., Risi, S., & Togelius, J. Bootstrapping
Conditional GANs for Video Game Level Generation. 2020 IEEE Conference on
Games (CoG), 41–48 (2020).
34. Zhang, H., Goodfellow, I.J., Metaxas, D.N., & Odena, A. Self-Attention Generative
Adversarial Networks. ICML (2019).
35. Gutierrez, J., & Schrum, J. Generative Adversarial Network Rooms in Generative Graph
Grammar Dungeons for The Legend of Zelda. 2020 IEEE Congress on Evolutionary
Computation (CEC), 1–8 (2020).
36. The Legend of Zelda. This page was last edited on 12 September 2021, at 15:53 (UTC)
https://en.wikipedia.org/wiki/The_Legend_of_Zelda.
37. Schrum, J., Volz, V., & Risi, S. CPPN2GAN: Combining Compositional Pattern
Producing Networks and GANs for Large-Scale Pattern Generation. Proceedings of
the 2020 Genetic and Evolutionary Computation Conference (2020).
38. Stanley, K. Compositional Pattern Producing Networks: A Novel Abstraction of
Development. Genetic Programming and Evolvable Machines, 8, 131–162 (2007).
39. Shaham, T.R., Dekel, T., & Michaeli, T. SinGAN: Learning a Generative Model From
a Single Natural Image. 2019 IEEE/CVF International Conference on Computer Vision
(ICCV), 4569–4579 (2019).
40. Bontrager, P., & Togelius, J. Fully Differentiable Procedural Content Generation
through Generative Playing Networks. ArXiv. abs/2002.05259 (2020).
41. Gobbi, J., Liello, L.D., Ardino, P., Morettin, P., Teso, S., & Passerini, A. Efficient
Generation of Structured Objects with Constrained Adversarial Networks. ArXiv.
abs/2007.13197 (2020).
42. Irfan, A., Zafar, A., & Hassan, S. Evolving Levels for General Games Using Deep
Convolutional Generative Adversarial Networks. 2019 11th Computer Science and
Electronic Engineering (CEEC), 96–101 (2019).
43. Kumaran, V., Mott, B., & Lester, J.C. Generating Game Levels for Multiple Distinct
Games With a Common Latent Space. AAAI, 16(1), 109–115 (2020).
44. Beckham, C., & Pal, C. A Step Towards Procedural Terrain Generation With GANs.
ArXiv. abs/1707.03383 (2017).
45. Wulff-Jensen, A., Rant, N.N., Møller, T.N., & Billeskov, J. Deep Convolutional
Generative Adversarial Network for Procedural 3D Landscape Generation Based on
DEM. ArtsIT/DLI, 85–94 (2017).
46. Ferranti, Jonathan de. Viewfinder Panoramas (2012).
47. Spick, R.R., & Walker, J. Realistic and Textured Terrain Generation using GANs.
European Conference on Visual Media Production (2019).
48. Jetchev, N., Bergmann, U., & Vollgraf, R. Texture Synthesis With Spatial Generative
Adversarial Networks. ArXiv. abs/1611.08207 (2016).
49. Wang, T., & Kurabayashi, S. Sketch2Map: A Game Map Design Support System
Allowing Quick Hand Sketch Prototyping. 2020 IEEE Conference on Games (CoG),
596–599 (2020).
50. Fadaeddini, A., Majidi, B., & Eshghi, M. A Case Study of Generative Adversarial
Networks for Procedural Synthesis of Original Textures in Video Games. 2018
2nd National and 1st International Digital Games Research Conference: Trends,
Technologies, and Applications (DGRC), 118–122 (2018).
51. Horsley, L., & Liebana, D.P. Building an Automatic Sprite Generator with Deep
Convolutional Generative Adversarial Networks. 2017 IEEE Conference on
Computational Intelligence and Games (CIG), 134–141 (2017).
52. Hong, S., Kim, S., & Kang, S. Game Sprite Generator Using a Multi Discriminator
GAN. KSII Transactions on Internet and Information Systems, 13, 4255–4269 (2019).
53. Serpa, Y.R., & Rodrigues, M.A. Towards Machine-Learning Assisted Asset Generation
for Games: A Study on Pixel Art Sprite Sheets. 2019 18th Brazilian Symposium on
Computer Games and Digital Entertainment (SBGames), 182–191 (2019).
54. Pac-Man. This page was last edited on 14 September 2021, at 00:21 (UTC) https://
en.wikipedia.org/wiki/Pac-Man.
55. Blizzard Entertainment. StarCraft (1998). https://starcraft2.com/en-us/.
56. Summerville, A., Snodgrass, S., Mateas, M., & Ontañón, S. The VGLC: The Video
Game Level Corpus. ArXiv. abs/1606.07487 (2016).
57. https://www.doomworld.com/idgames/.
58. idgames archive | Doom Wiki | Fandom, https://doom.fandom.com/wiki/Idgames_archive.
59. http://www.gvgai.net/.
60. Khalifa, A., Liebana, D.P., Lucas, S., & Togelius, J. General Video Game Level
Generation. Proceedings of the Genetic and Evolutionary Computation Conference
2016 (2016).
61. https://topex.ucsd.edu/WWW_html/srtm30_plus.html.
62. Mescheder, L.M., Geiger, A., & Nowozin, S. Which Training Methods for GANs do
actually Converge? ICML (2018).
63. Gidel, G., Hemmat, R.A., Pezeshki, M., Huang, G., Priol, R.L., Lacoste-Julien,
S., & Mitliagkas, I. Negative Momentum for Improved Game Dynamics. ArXiv.
abs/1807.04740 (2019).
64. Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., & Abbeel, P. InfoGAN:
Interpretable Representation Learning by Information Maximizing Generative
Adversarial Nets. NIPS (2016).
65. Locatello, F., Bauer, S., Lucic, M., Gelly, S., Schölkopf, B., & Bachem, O. Challenging
Common Assumptions in the Unsupervised Learning of Disentangled Representations.
ArXiv. abs/1811.12359 (2019).
66. He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R.B. Momentum Contrast for Unsupervised
Visual Representation Learning. 2020 IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 9726–9735 (2020).
67. Wu, Z., Xiong, Y., Yu, S., & Lin, D. Unsupervised Feature Learning via Non-parametric
Instance Discrimination. 2018 IEEE/CVF Conference on Computer Vision and Pattern
Recognition, 3733–3742 (2018).
68. Chen, T., Kornblith, S., Norouzi, M., & Hinton, G.E. A Simple Framework for
Contrastive Learning of Visual Representations. ArXiv. abs/2002.05709 (2020).
Index
A G
Adventure Games and Puzzle Design Game Theory 3, 6
37, 44 Gaming Experience 103, 107
AI for Adaptive Computer Games 37, 41 Generative Adversarial Networks 123, 126
AI for Computer Games 37, 38
AI in Gaming 37, 60 H
AI in Video Games: Toward a Unified
Framework 37, 41 Heuristic Function 3, 4, 5
AI Latest Techniques in Animation 19, 23 Home Automation 65, 74
AR Technology 19, 26 Horror Genre and Video Games 37, 45
VR Technology 19, 28 Hybrid Fuzzy ANN Systems 139
AI Latest Techniques in Animation
19, 23 I
AI Reinvent 103, 106
The Image Classification Theory 91, 96
AI’S Role in Animation 19, 21
Image Generation and Recognition 123, 127
How AI Replaces Animation 19, 21
Image Processing Using Deep Learning 91, 98
Various Agents in AI 19, 22
Important Architectures in Deep Learning
Alpha-Beta Pruning 8
91, 94
Augmented Reality 110
Interactive Narrative 37, 44
Internet of Things (IoT) 65, 67
C Communication Block 65, 68
Protocols 65, 69
CBGIR 91, 94 Sensing Block 65, 68
CGANs 127 IoT and 5G Technology 65, 78
Classification 65, 81 IoT Networks 83
Complex Game AI 103, 115
Cyclical GANs 128
L
D Logistics 65, 75
Deep Learning 75, 82, 93 M

Deep Learning in PCG 137, 142
Deformable Gans for Pose-Based Human Image Machine Learning 91, 92
Generation 123, 130 Minimax Approach 4, 8
Deterministic 104 MINIMAX Vs APLHA-BETA PRUNING 8
Diverse Image Generation via Self-Conditioned Motionscan Technology 65, 77
GANs 123, 129 MQTT 69, 70
Multi-Discriminator GAN (MDGAN) 146
E N
Early Game 103, 115 Narrative Game Mechanics 37, 43
eHealth 65, 74 Narrative in Video Games 37, 42
Extended Reality 103, 110 Nondeterministic 105
F P
Finite State Machines 103, 116 PCGML 137, 141
Future Aspects of Animation with AI Player Experience 37, 48
19, 33 Procedural Content Generation (PCG)
Futuristic Approach 37 137, 140
157
158 Index
Q T
Q-Algorithm for AI in Gaming 37, Three-Dimensional Visualization Techniques
60 103, 108
TOAD-GANs 139
The Traditional and Modern Animation
R 19, 30
Regression 65, 81
U
S Utility and Application 14, 15
SAGANs 128 V
Search Tree 2, 3, 8
Simulation Based on Physics 103, 108 Virtual Reality 103, 109
Smart Agriculture 65, 74 Voice in Gaming 103, 107
Stack GANs 128
Supervised Learning 65, 80 Z
Systematic Analysis of Image Generation Using
GANs 123, 131 Zero-Sum Game 4, 6

Deep Learning in Gaming and Animations

Uploaded by

Copyright:

Available Formats

Deep Learning in Gaming and Animations

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deep Learning in Gaming and Animations

Uploaded by

Copyright:

Available Formats

Deep Learning in Gaming

Explainable AI (XAI) has developed as a subfield of artificial intelligence, focusing

Deep Learning in Gaming and Animations: Principles and Applications

For more information about this series, please visit: https://www.routledge.com/

and by CRC Press

First edition published by CRC Press 2022

CRC Press is an imprint of Taylor & Francis Group, LLC

Library of Congress Cataloging‑in‑Publication Data

Names: Chaudhary, Vikas, editor. | Sharma, Moolchand, editor. | Sharma, Prerna,

ISBN: 9781032126098 (hbk)

Typeset in Times LT Std

Chapter 2 The Future of Automatically Generated Animation with AI.............. 19

Chapter 3 Artificial Intelligence as Futuristic Approach for Narrative Gaming..... 37

Chapter 4 Review on Using Artificial Intelligence Related Deep Learning

Chapter 5 A Review on Deep Learning Algorithms for Image Processing

Chapter 6 Artificial Intelligence in Games........................................................ 103

Chapter 7 A Framework for Estimation of Generative Models Through

Chapter 8 Generative Adversarial Networks Based PCG for Games................ 137

Figure 3.3 Player character class in UE4............................................................ 53

Figure 6.3 Unique experience of VR................................................................ 109

Moolchand Sharma is an Assistant Professor in the Department of Computer Science

Prerna Sharma is an Assistant Professor in the Department of Computer Science

Deevyankar Agarwal works as a lecturer in the Engineering Department–EEE

Aparna Agarwal Bramah Hazela

Ashish Kumar Chakraverti Saad Bin Abdul Kashem

Sugandha Chakraverti Saad Bin Khalid

Priyanshi Gupta Preety Khatri

Vikas Rao Naidu Prerna Sharma

Game Using Game

1.2 RELATED WORK

1.3.1 Game Theory

1.3.2 Zero-Sum Game

FIGURE 1.1 An example of payoff matrix for a zero-sum two-player game.

FIGURE 1.2 Code implementation of the heuristic function.

1.3.4 Search Tree

FIGURE 1.3 Binary Search tree example.

FIGURE 1.4 A Minimax tree.

In this project, the Minimax [10] algorithm is used as a recursive algorithm to

FIGURE 1.5 Algorithmic representation of a Minimax tree example.

1.3.7 Minimax vs Alpha-Beta Pruning

FIGURE 1.6 Alpha-Beta pruning illustration.

FIGURE 1.7 Alpha-Beta pruning illustration.

1.4.1 Game Algorithm

FIGURE 1.8 Alpha-Beta pruning illustration.

FIGURE 1.9 Flowchart depicting the implementation algorithm.

1.4.2 Graphical User Interface

FIGURE 1.10 Before single capture of a piece.

FIGURE 1.11 After single capture of a piece.

FIGURE 1.12 Before multiple capture of pieces.

FIGURE 1.13 After multiple capture of pieces.

FIGURE 1.15 Displaying the result of the game.

1.5 UTILITY AND APPLICATION

FIGURE 1.16 Use case flowchart.

• It develops concentration skills and promotes confident decision-making.

1.5.1 System Environment

FIGURE 1.17 System environment diagram.

1.6 CONCLUSION AND FUTURE SCOPE

1.6.2 Future Scope

FIGURE 2.1 Setting up keyframes for timing the bouncing balls.

2.2 AI’S ROLE IN ANIMATION

Figure 3.3 Player character class in UE4............................................................ 53

Figure 6.3 Unique experience of VR................................................................ 109

3.5 AI ELEMENTS IN THE PROPOSED