CiF-Bannerlord, a Social Artificial Intelligence System for Mount and

CiF-Bannerlord, a Social Artiﬁcial Intelligence System for

Mount and Blade II: Bannerlord

Filipe Miguel Dias da Silveira

Thesis to obtain the Master of Science Degree in

Information Systems and Computer Engineering

Supervisors: Prof. Pedro Alexandre Sim

oes dos Santos

Prof. Jo

ao Miguel de Sousa de Assis Dias

Examination Committee

Chairperson: Prof. David Manuel Martins de Matos

Supervisor: Prof. Pedro Alexandre Sim

oes dos Santos

Member of the Committee: Dr. Joana Carvalho Filipe de Campos

November 2021

This work was created using L

X typesetting language

in the Overleaf environment (www.overleaf.com).

Acknowledgments

I would to start by expressing my appreciation to my supervisors, professor Pedro Santos and pro-

fessor Jo

ao Dias. Thank you for all the guidance and advice throughout the whole project. Thank you to

my colleagues, Joaquim Quadrado, Marco Cabral, Marcos P

ego and David Ricardo, that accompanied

and helped me through this journey. Thank you to my friends for providing me with the encouragement

to complete my studies. A special thanks to my girlfriend Cristiana Rodrigues, for always helping me

with everything I needed. Thank you for the endless support and for always believing in me. Last but

not least I would like to thank my family. To my grandparents for raising me and helping me grow into

the man I am today. To my parents for giving me their unwavering support, love and care throughout my

whole life and providing me with everything I needed. And to my brother for being my best friend and

always being there for me.

Abstract

Mount and Blade II: Bannerlord is an intricate and complex game that aims to simulate a realistic me-

dieval world. Although Bannerlord offers an immersive world that allows for engaging role play, there

are some aspects that can be improved in order to provide the full experience, such as character re-

lationships and social interactions. Developing a Social Artiﬁcial Intelligence System system for a AAA

game improves the player’s immersion and overall gameplay experience. The purpose of this thesis is

to create a set of engaging Non Playable Characters (NPCs) with unique personalities, beliefs, desires

and struggles, by applying and improving existing social models. In order to achieve this, the Comme il

Faut (CiF) model was considered and developed for the game Mount and Blade II: Bannerlord.

To test the success of the implemented model, User tests were performed. These measured the

users Immersion and their NPC Believability when playing the game with the CiF-Bannerlord changes.

The tests acquired positives results, demonstrating an improvement of the determined metrics.

Keywords

Non-Playable Character; Social Artiﬁcial Intelligence; Mount and Blade II: Bannerlord;

iii

Resumo

Mount and Blade II: Bannerlord

e um jogo complexo que procura simular um mundo medieval e real-

istico. Este jogo oferece um mundo extenso e imersivo que permite interpretar personagens criadas

pelo jogador e conversar ou conviver com personagens virtuais. Apesar disto existe certos aspectos

que podem ser melhorados de modo a oferecer uma experi

encia mais completa atrav

es das relac¸

oes

e interac¸

oes com estas personagens virtuais. Este projeto procura desenvolver um sistema de In-

telig

encia Artiﬁcial Social para um jogo AAA de modo a oferecer um mundo mais imersivo. Este sistema

consiste em criar um grupo de personagens (Non Playable Characters (NPCs)) com personalidades,

crenc¸as unicas e com vontades e problemas pessoais. Foi desenvolvido e aplicado um modelo social

existente chamado Comme il Faut (CiF) (Comme il Faut). Para testar o sucesso das mudanc¸as trazidas

pelo modelo CiF, foram feitos tests com utilizadores. Houve um total de trinta e um participantes nos

testes. Com base nestes testes foi poss

ıvel deduzir o sucesso positivo que a implementac¸

ao deste

modelo social revelou no jogo Mount and Blade II: Bannerlord.

Palavras Chave

Personagens; Intelig

encia Artiﬁcial Social; Mount and Blade II: Bannerlord;

Contents

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Document Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Mount and Blade II:Bannerlord 7

2.1 Campaign Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Non-Playable Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.3 Code Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3.1 Module Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3.2 Important Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.4 Bannerlord Modding Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Related Work 15

3.1 Social AI Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1.1 CiF (Comme il faut) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1.2 Ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.1.3 CiF-CK (Comme il faut - Creation Kit) . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.1.4 CiF-Ex (Comme il faut - Exiles) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.1.5 FAtiMA Toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2.1 Games with Social AI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2.1.1 The Sims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2.1.2 Fac¸ade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.2.1.3 Prom Week . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.2.1.4 Persona 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2.2 Current Bannerlord Mods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

vii

4 Comme il Faut: Bannerlord 33

4.1 CiF-Bannerlord Additions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.1.1 CiF-Bannerlord Removed Features . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.2 CiF Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.2.1 Desire Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.2.2 Intent Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.2.2.1 Intent based on Personality . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.2.2.2 Intent based on Social Feelings . . . . . . . . . . . . . . . . . . . . . . . 41

4.2.2.3 Determine desired Social Interaction . . . . . . . . . . . . . . . . . . . . 42

4.2.3 Exchange Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.2.4 Exchange Fallout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.3 CiF-Character . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.3.1 CiF Player Character . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.4 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.5 Social Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.5.1 Social Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.5.2 Social Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.6 Rumor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.6.1 Creating Rumors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.6.2 Spreading Rumors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.7 Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.7.1 CiF-Bannerlord Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5 Implementation 57

5.1 Behaviours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.1.1 SubModule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.1.2 NPC Dialog Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.1.3 Player Dialog Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.1.4 Rumor Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.1.4.1 Saving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.1.5 Microtheories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.2 Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5.2.1 Character Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5.2.1.1 Saving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.2.2 Rule Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.2.2.1 Condition Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

viii

5.2.2.2 Inﬂuence Rule Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.2.2.3 TriggerRuleManager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.2.2.4 Social Interaction Manager . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.2.2.5 Dialog Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6 Evaluation 69

6.1 New Lords Mod . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

6.1.1 Online Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.2 User Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.2.1 Instruments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

6.2.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6.2.3 Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.2.4 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.2.4.1 Demographic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.2.4.2 Player Immersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6.2.4.3 NPC Believability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

6.2.4.4 Interaction Enjoyment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.2.5 Participant Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

6.3 Results Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

7 Conclusion 89

7.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Bibliography 93

A Appendix 95

List of Figures

2.1 Campaign Screen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Screenshot of a conversation in a keep . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.3 Visual representation of the module system of Bannerlord . . . . . . . . . . . . . . . . . . 12

3.1 CiF Architecture Diagram [11] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.2 Data ﬂow of schema package components [17] . . . . . . . . . . . . . . . . . . . . . . . . 20

3.3 CiF-Ex Structure [15] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.4 The Sims 4 Screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.5 Screenshot Fac¸ade’s gameplay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.6 Screenshot of Prom Week’s gameplay [12] . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.7 Persona 5 conversation screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.1 CiF models and their addition to the architecture . . . . . . . . . . . . . . . . . . . . . . . 35

4.2 Visual representation of the CiF cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3 Visual representation of the thresholds and outcome determination . . . . . . . . . . . . . 43

4.4 Example graph for the token based dialog system in Bannerlord . . . . . . . . . . . . . . 53

5.1 Visual representation of the manager system in CiF-Bannerlord . . . . . . . . . . . . . . . 64

6.1 New Lords mod data as of 20th of October 2021 . . . . . . . . . . . . . . . . . . . . . . . 71

6.2 Questions present in the questionnaire and their respective metrics . . . . . . . . . . . . . 75

6.3 Graph detailing the procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.4 Playtime in hours of gaming per week intervals and their count . . . . . . . . . . . . . . . 78

6.5 Usual emotional attachment to characters in-game . . . . . . . . . . . . . . . . . . . . . . 78

6.6 Likelihood to skip dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6.7 Boxplot illustrating immersion question comparison based on version . . . . . . . . . . . . 80

6.8 Boxplot illustrating believability question comparison based on version . . . . . . . . . . . 82

6.9 Count of the Nice question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

6.10 Count of the Annoying question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6.11 Count of the Friend question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6.12 Count of the Enemy question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.13 Boxplot with enjoyment difference regarding version . . . . . . . . . . . . . . . . . . . . . 86

xii

List of Tables

3.1 Conversion of social exchange into quest [3] . . . . . . . . . . . . . . . . . . . . . . . . . 22

4.1 CiF-Character Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.2 Personality Traits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.3 Social Interaction Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.4 Social Exchange Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.5 Base game Dialog Lines Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.6 CiF-Bannerlord Dialog Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.1 Character Manager Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6.1 Metrics and their respective cronbach’s alpha . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.2 Wilcoxon signed-rank test with their respective Z-values, p-values and r-values . . . . . . 80

6.3 Wilcoxon signed-rank test with their respective Z-values, p-values and r-values . . . . . . 82

xiii

xiv

Acronyms

NPC Non Playable Character

RPG Role Playing Game

MMORPG Massive Multiplayer Online Role Playing Game

CiF Comme il Faut

SI Social Interaction

SE Social Exchange

AI Artiﬁcial Intelligence

xvi

Introduction

Contents

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Document Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

The gaming industry is currently one of the most considerable markets in the entertainment industry,

even surpassing movie and music industries, while still growing with a swift rhythm [6]. Thousands

of games are released every year, adding new entries to every genre in the market. By improving

mechanics, features and content, players all around the world look forward to additions to their favorite

genres and series of games. One of the most important features of games is immersion, since it is a

distinguishing features when compared to other entertainment mediums. Immersion is often viewed as

critical element to game enjoyment and refers to the outcome of a good gaming experience [5].

Games have evolved over the years, improving their immersion either by graphic ﬁdelity, with a re-

alistic capture of the world or by creating interesting and engaging narratives that the players can be a

part of. Developers create speciﬁc worlds to envelop their story and current games are even capable

of generating new random worlds, different each time the game starts. One of the struggles of game

development is adding content or mechanics while saving time and work. All games have some sort of

reuse on content but there is an important balance to look for in order to not seem repetitive.

Mount and Blade II Bannerlord is a free roam single-player campaign Role Playing Game (RPG)

where players can engage in battles, economics and politics. In Bannerlord, there are several of unique

villages, differing from culture to culture. However two villages from the same culture are equal, therefore

it can get repetitive while battling in an area with the same culture. Even though TaleWorlds announced

to work on this problem, it still proves as a big authoring burden. A reliable solution is procedurally

generated content, creating different level layouts and items automatically. This provides a signiﬁcant

boost to replayability while saving developers time and effort to work on other aspects of the game.

1.1 Motivation

Non Playable Characters (NPCs) have the same authoring issue: creating features by the developers

can become a bit too linear and predictable. Nowadays, most games have two sets of characters,

a small group loaded with dialog and personalization and a large group of really simple dialog with

repetitive lines. This second group that consists of the majority of NPCs encountered in the game,

effectively breaks the immersion since players feel they are not speaking to some other person or entity,

but are instead reading repetitive dialog. The NPCs are devoid of personality and players are never

compelled to explore the world in search of characters and develop relationships with these due to the

lack of complexity in the social interactions. The game would be more believable and immersive if the

characters and their dialog lines were procedurally generated to ﬁt in the environment and the story.

Another potential issue related to NPCs is the lack of impact in the player or the story, which boils

down to the choices a player can make throughout the game. Most games have a linear story and

progression and the characters do not have much impact. The player can usually do whatever they

desire, such as breaking everything inside a store, and the NPC will, regardless of what happened,

maintain the same dialog. If a player is rude or disrespectful towards certain characters, they shouldn’t

feel welcome around them and the NPCs should convey this clearly. As a result, the game would be

more immersive with a lively reactive world, where the players can feel the consequences to their actions

and observe the consequences of the NPCs actions as well.

1.2 Goals

To improve the overall gameplay experience, we decided to tackle player immersion and NPC believ-

ability. For this purpose we explored Social Artiﬁcial Intelligence (AI) models, like Comme il Faut [14],

to improve and develop emotional non-playable characters with satisﬁable interactions. The ultimate

success of these characters is that the player will not only search for the items, locations and points of

interest but will also search for new relations with the residents of the world, to try to make friends or en-

emies and get to know new characters because they are engaging enough and can provide meaningful

social interactions. Thus this system requires that:

NPCs should be believable, one of the most important things for people to understand each other

is their emotions. In order to feel empathy and form a relation people try to perceive other’s desires and

beliefs. Therefore, to create believable virtual characters, they should have beliefs and desires in order

to transmit emotions to the player. A believable character is deﬁned as the one who seems life-like,

whose actions make sense, who allows for suspension of disbelief [2].

NPCs should be reactive, to world events, other characters and the player’s actions should affect

the respective NPCs, depending on what they do, how they speak and are treated, characters should

behave towards others respectively. NPCs should have a notion of others based on their interactions

and should respond to events like war.

Social environments should be immersive, with a set of believable characters, virtual social en-

vironments should replicate real social environments making the player perceive other NPCs with their

own worries and interests, by talking to each other. A group of unique personalities allows for the player

to choose the characters to align and speak to further improve the player’s bonds with them and raise

their perceived immersion and emotional connections.

The ultimate goal of this project is to create a believable set of Non-Playable Characters in Mount

and Blade II: Bannerlord in order to improve the immersion and provide satisfactory social interactions

for the player.

1.3 Document Outline

The document is divided into seven chapters, including the ﬁrst which is the introduction. The second

chapter is the description of Mount and Blade II: Bannerlord, how the game works, important concepts,

how it will be to work with and why it is a good choice for modding and developing Social AI. The third

chapter is about Related Works, where we will discuss in the ﬁrst section, current Social AI architectures

and in the second section, current mods for Bannerlord and games that implement Social AI, some of

which are previously discussed architectures. The fourth chapter is about CiF-Bannerlord, the main

chapter, where we will explore the Comme il Faut (CiF) architecture developed within the game Mount

and Blade II: Bannerlord. The ﬁfth chapter explains the implementation and design choices of CiF-

Bannerlord. The sixth chapter is the study about the results from the modiﬁcations brought by the

CiF-Bannerlord model and the results from the user tests performed. The seventh and last chapter is

the summary of this project as well as considerations of future work to further improve it.

Mount and Blade II:Bannerlord

Contents

2.1 Campaign Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Non-Playable Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.3 Code Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.4 Bannerlord Modding Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Mount and Blade II: Bannerlord is the second game on the Mount and Blade series, developed by

TaleWorlds Entertainment. The ﬁrst is Mount and Blade, with a few expansions released from 2008

to 2011. Bannerlord is a very recent game, released as Early Access on March 30th, 2020. The

game is still in development so there are currently a lot of patches and updates to improve and add

features. Bannerlord is also a game with a booming modding community, due to modders coming from

the previous installment or because the game has a robust engine that supports all kinds of mods.

The developers also made the game not only easy to mod but also, easy to add existing mods. It has

both multiplayer and singleplayer modes, the latter being a campaign where players can role-play and

discover the world of Calradia.

2.1 Campaign Description

Bannerlord’s campaign is a RPG (Role Playing Game), where the player assumes the role of a character

inside the virtual world. RPGs are a very broad genre that a lot of people enjoy. It is the oldest game

genre, which has started as written stories the players could partake in, for example, Dungeons and

Dragons

. Usually, the RPG worlds are ﬁlled with characters and locations for the player to explore,

thus making these interactions an integral part of the game. There are several sub-genres in RPGs,

one of which is Massive Multiplayer Online Role Playing Games (MMORPGs), that have both NPCs and

real players in a massive world, appealing a little more to players looking for social interactions [20]. In

the Bannerlord’s campaign, the player starts with the character creation, they pick from several options

describing the character’s early years which determines the ﬁrst traits and skills unlocked. Afterwards,

the player enters the world in a speciﬁc location and chooses the course of action, where to move and

what to do.

Figure 2.1: Campaign Screen

Dungeons and Dragons, Gary Gygax, Dave Arneson, 1974

The main objective is to grow the character’s clan and eventually, transform it into a kingdom.

Through quests, battles and tournaments, the player is awarded points. Quests provide relationship

points, to signify how friendly or not the player is with each character. Both tournaments and battles

award inﬂuence points and renown if they are won. The bigger and more impressive the battle, the more

points it awards. Inﬂuence points are the relative power the player has inside the kingdom they belong.

Renown points are more important, as they determine the player’s clan level, serving as one of the main

objectives of the game.

Combat is a pretty important part of the game, although simple it can be very satisfying and addictive.

In battle, the player only accounts for one soldier and it leads the army from within the battleﬁeld. The

main restriction is money, usually, a bigger army is better, however, some balance is required, regarding

the troop costs, funds, food and available party size.

The player is the one who decides what is the best course of action, though. They can become a

merchant, buying and reselling items in towns, a war commander, to ﬁght battles and conquer land, a

bandit, to steal and raid villages or be roaming in the world, helping the NPCs through quests. There

is also a story related to the player’s character and a story related to the world recent events for the

player to ﬁnd out. These have no time constraints nor obligation to complete, therefore, they can be

done whenever, or even if, the player feels like it.

2.2 Non-Playable Characters

In Bannerlord’s campaign the player can interact with several NPCs on different occasions. There are

some types of NPCs in the game such as:

Figure 2.2: Screenshot of a conversation in a keep

Townsmen have limited interactions and no possibility of relationships. In this group of characters,

there are merchants, that sell products and services to the player, guards in prisons, castles and keeps

and ﬁller NPCs to make the villages and taverns look more crowded.

Important Town folk have more complex interactions since they can form a unique friendship re-

lation with the player. This group includes quest givers present in towns that allow the player to recruit

units. This allows the player to recruit more or less based on the relation value. Currently, this relation

is described by a number from -100 to 100, positive actions to the character, for example, helping it

with quests, raises this number while negative actions, such as raiding the village, will lower it. Another

kind of NPCs that have a friendship relation with the player are companions. Unfortunately, most of the

interaction between the player and companions is at the tavern, where they are met and recruited. The

overall interaction is lackluster, as some of these often accompany the player in their party while still not

interacting much.

Lords are from noble families, which comprise of the leader of the clan and their family members.

They can command armies and ﬁght either alongside or against the player. They also have a friendship

relation that, in extreme cases, can cause them to declare war. The player can interact with lords in

castles, or the player can approach moving parties and armies commanded by these lords to interact with

them as well. These groups of NPCs are important because the player can form romantic relationships

with them. They can court lords and ladies with a simple “mini-game” where they try to persuade

the receiver into liking the player. This interaction is luck-based and it is affected by both the player

character’s and the receiver character’s traits. These romantic relationships can develop into marriage

and eventually provide children. This mechanic is surprisingly important since the player can continue

the game through children characters when their main character dies, therefore, having children can

prolong the player’s “legacy” in the world.

2.3 Code Architecture

The previous installment of Mount and Blade had a strong modding community and TaleWorlds wanted

to maintain this community. Therefore, Bannerlord was developed to be very extensible and easily

moddable. Even the ofﬁcial game launcher allows for the selection of several mods.

2.3.1 Module Organization

The access to the game code was enabled by using a tool to decompress the binary ﬁles called dnSpy.

This tool is very helpful because, not only to decompress and access the code without having the source,

but also to debug and edit ﬁles. The code obtained from this tool is not the original one, meaning there

are no comments nor documentation hampering the understanding of the architecture.

Figure 2.3: Visual representation of the module system of Bannerlord

The code is separated into two parts, the base and submodules. There are ﬁve submodules, how-

ever, the game considers these as mandatory mods. Mods are added to the game as modules, and

any number of them can be added at the same time. In the ﬁgure 2.3 we can observe how the module

system works, considering Module 1 and Module 2 as mods. The main reason behind this organization

is because there is a load order to the submodules that can be modiﬁed for custom mods to replace

certain assets or code present in the base game’s submodules. The most important submodule for the

campaign is SandBox, which contains most of the logic behind the campaign system. There is also

SandBoxCore which contains campaign assets and StoryMode, only responsible for the story’s logic

and assets. This architecture can be extremely effective to work within mods. Creating a submodule is

very simple in Bannerlord since the game already loads a few as part of the base game.

2.3.2 Important Classes

As mentioned before, the developers made the code very extensible, with several interfaces to facilitate

modding. There are two prevalent classes that help to easily work on features, the base for Submodules,

which gets automatically loaded and the base for Behaviours, which function as a simple and effective

way to add new classes.

The characters in the code are represented by two classes, Agent and Hero. The agent class deals

with most of the logic regarding characters, especially in battle. The Hero class is responsible for the

character logic regarding the campaign system, while Agent contains most of the general character

logic like moving and data. There is also an interface named AgentComponent which can be used to

add new features or roles to the Agent in an easy and elegant way. The Hero and Agent classes are

rather detailed and complicated.

Fortunately, in the architecture, there are managers enveloping a lot of the logic, for example, the

character relation manager, that stores and updates all relations between characters. Modding these

speciﬁc classes, besides being more simple to understand, are a better way to to implement desired

modiﬁcations.

In chapter 5, we will review how these classes are connected to the project developed.

2.4 Bannerlord Modding Tools

On the 1st of October, 2020, TaleWorlds released an ofﬁcial beta version of Modding Tools aiming to im-

prove the quality of modding development. However, these Modding Tools are simple and currently more

focused on visual and asset based mods. These tools just enable the user to change and customize the

map and other assets in the game.

Related Work

Contents

3.1 Social AI Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

In the ﬁrst section of this chapter, we will discuss what are some important concepts in AI, existing

social AI models, how they work, their weaknesses and strengths as well as their uses. Furthermore,

in the second section, it will be discussed games that implement social AI and both popular and social

mods existing for Bannerlord.

3.1 Social AI Models

Games are being released every day, with current the huge search and demand in the industry, it is not

uncommon for people to look forward to a game they desire at any given moment. Most people look

forward to AAA games, which have a massive budget and a great deal of work put into from sizeable

game companies. AAA games are very balanced, focusing on most aspects of the game, graphics, story

and gameplay, trying to please most people.

They are also, however, very undeveloped on AI, in particular, on social AI. Most people can un-

derstand how frustrating escort missions with slow or buggy companions can be. Games praised for

their AI are only praised for their effective decision making, like The Last of Us

and F.E.A.R

or praised

for tactical prowess like the Total War games

. However, there are very few AAA games distinguished

for their social AI, engaging characters without scripted dialog. The standard is just dialog choices or

a relationship value to receive bonuses from the character (like the case with Bannerlord). Games with

social AI (like Prom Week [14] or Fac¸ade

are all designed around this feature, lacking in aspects AAA

excel in and that the majority of players enjoy. Thus, adding this speciﬁc characteristic to an AAA game

would complement it.

3.1.1 CiF (Comme il faut)

CiF, Comme il faut, is a french expression and it translates to “as the standard” or “as it should be”.

The main objective of the CiF Social Model is to reduce the authoring work load developers have on

working with NPCs and their narrative. In order to create a dynamic social environment, CiF empowers

the NPCs with a set of characteristics that allow complex reasoning and pursue individual gain.

CiF is rule based system. CiF manages to lower the authoring burden and allow developers the

need to only develop rules for the environment. One example of a rule in natural language might be:

a character who is vengeful will more likely do something mean to someone who has recently done

something mean to them [10]. The interaction between characters in CiF is called a social exchange.

It encapsulates a social action a character takes with another character (with the intention of changing

The Last of Us, Naughty Dog and Sony Computer Entertainment, 2013

F.E.A.R. - First Encounter Assault Recon, Monolith Productions, Vivendi Universal Games, Warner Bros. Games, 2006

Total War series, Creative Assembly and Feral Interactive, 2000-2020

Fac¸ade, Procedural Arts, 2005

their relationship) as well as how the other character responds [13]. Social exchanges have an initiator

and its intent, the responder and its response, as well as possibly a third element. An example of social

exchange is, John wants to be friends with Mary and so does Mary, therefore, John asks Mary to be his

friend and she accepts.

Figure 3.1: CiF Architecture Diagram [11]

In CiF each character contains:

• Desires, characters are motivated by theirs desires, they will try to achieve these through social

interactions and serve as the primary reasoning for their actions. As an example, John, who wants

to be friends with Mary will try to be nice around her.

• Basic Traits determine the character’s response to the situations they encounter. Even though

Mary wants to befriend John, she is too shy (a trait) to start a conversation (social exchange).

• Temporary Statuses are the character’s internal private feelings that have an impact on their

actions. John after successfully befriending Mary is happy (happiness is the temporary status).

• Social Network is the collection of feelings the character has regarding the others. All the opinions

the character has on others, for example, John liking Mary as a friend.

There is also a public database that all NPCs can have access to. It contains:

• Relationships in CiF are the known relations between two characters, all the NPCs have access

to this. These can change at different rates depending on the character’s traits, desires and tem-

porary statuses. After John and Mary become friends, every agent in CiF receives that knowledge.

• Social Fact Database is the collection of the social exchanges that occurred. For example, when

John initiated a conversation with Mary to become her friend.

• Cultural Knowledge Base is the storage for the general opinion on the objects present in the

world as well as the individual opinions on that speciﬁc object. For example, in a certain setting,

movies are considered cool, however, John does not like movies. Both these opinions are stored

in the Cultural Knowledge Base.

CiF works in a set of simple processes on each social state.

1. The ﬁrst process is desire formation, this determines the character’s inspiration to execute a social

exchange with another. Every character determines its volition on all its available social exchanges

with every other character.

2. Then the character, which is now the initiator chooses an exchange to perform with another.

3. The next step is to determine the outcome of said social exchange. Similar to the volition in the

desire formation, the responder calculates a value based on its rule set and this determines if it

accepts or rejects the social move.

4. Since the aftermath of the social exchange is dependent on the social state and its outcome, the

system creates something called effects. Then the system will choose the situation with most

effects and instantiate it.

5. An instantiation is a set of template-based dialogue acts and associated animations. After the

instantiation is realized, the social state change associated with the chosen effect is applied [10].

6. Lastly, a set of trigger rules over the new social state is run. These rules account for the changes

performed in the previous step, updating the characters’ statuses and relationships.

CiF is based on a rule system and these can be added manually or dynamically at runtime. Predi-

cates are a very important characteristic of these rules, as they function as representational primitives

that can be evaluated for truth in a social state [11]. The main ruleset in CiF is inﬂuence rules, these

alter the desires of the character to engage and to respond to social exchanges. Inﬂuence rules have a

social condition and a weight associated with them. For example, since John is friends with Mar, he will

be more likely to want to do something nice to her. Another set of rules is trigger rules. These are very

similar to the effect changes associated with instantiations, however, unlike the effect changes, these

rules actually change the social state if the condition holds true and they are not associated with the

instantiation, meaning they can run independently and in a more broad manner.

However, as the system grows in complexity, a few problems arise. The authoring problem is still

troublesome, although less pronounced, as well as the problem to keep track of the existing rules since

they can become repetitive or contradictory. To help address this problem, CiF breaks rules into sets

called microtheories, uniﬁed by a precondition [11]. Microtheories contain a group of less general rules

that apply to one predicate on the social state. There are a set of rules that only apply to friends, for

example, not lying to friends. Since John and Mary are friends, they will not lie to each other.

3.1.2 Ensemble

Ensemble was developed as an upgrade of CiF, as such, it is also a rule based social physics model

with the aim to transform CiF into a more expressible and accessible system, providing a family of AI

techniques to developers and promoting the spread of social physics. Ensemble expands on a few key

features of CiF, namely the social exchange, now called social practice, and instantiations are now a di-

alog set between characters instead of scripted dialog and animations resulting from a social exchange.

Ensemble also upgraded the rule system to be capable of describing and referencing complicated social

situations involving any number of characters [17].

Figure 3.2: Data ﬂow of schema package components [17]

Ensemble added an important feature called schema, which is a simple way of developers to cate-

gorize the state of the system. CiF’s social state was not a reliable solution, since it was hardcoded and

inefﬁcient. Ensemble tried to make a more generalized version. Schemas are the proposed solution

for the authoring burden present in CiF. Developers design a schema package that deﬁnes the system’s

properties and contains rule deﬁnitions. These schemas are then validated in order to create the world’s

backstory, rule sets and more importantly, Categories, a generalization of the CiF’s traits, relationships

and temporary statuses. Categories have a direction type, which means what the category affects if it

is between one, two or several characters. This change transforms the character’s traits and statuses

into “undirected” categories, relationships into “reciprocal” categories and turning the social network into

“directed” categories. These upgrade the engine to be more customizable since terms such as “rela-

tionships” and “statuses” can be named [17]. Categories are very broad and have several attributes to

help deﬁne all kinds of categories, such as duration, an interval, a string to declare a type, or a boolean

value.

A social practice is deﬁned as a normative pattern of social interaction that captures the nuances

that result from the individual agent’s situation in the social state [19]. The dialog between characters

is now interactive, being generated after each action taken. Social practices are made up of a series of

linked stages with an active character that is alternated by the two participants. This stage is selected

based on the social state as well as the previous actions and microtheories. The Stage is comprised of

a group of actions that contain a character performance and updates to the social record. For example,

on a stage “reciprocate compliment” could contain several actions where participants reciprocate the

compliment in different manners. There are a few special stages like event stages that are triggered on

speciﬁc circumstances and entry stages that are used to start the social practice.

Actions express the character’s intent and performance regarding it is calculated, the highest score

will be the chosen action. Like CiF, the action is calculated based on microtheories, inﬂuence rules

and new in Ensemble, practice-speciﬁc bias. This bias is a weight on the equation to account for the

most common path taken through a practice, that is independent of the characters’ relation or emotions.

For example, a practice-speciﬁc bias may be introduced in a “How’re you doing?” social practice where

agents almost always respond with “Good! How’re you doing?” [19].

3.1.3 CiF-CK (Comme il faut - Creation Kit)

CiF-CK is a CiF implementation in a RPG environment. Similar to Ensemble, it is an improvement

on CiF, with a focus on generalizing its characteristics, however, CiF-CK also has an emphasis on

quests, since it is dedicated to RPGs. In CiF the player has access to control all the NPCs, which

is not the norm in games. In RPGs, the player plays with a single character and CiF-CK developed

the model accordingly. This model was implemented in the game The Elder Scrolls V: Skyrim

, which

was developed in Bethesda’s engine, the Creation Engine, like a few other games. Therefore, it is

immediately extendable to these games, the lastest being Fallout 76

. Skyrim has a modding tool called

Creation Kit, that helps developers mod the game and CiF-CK was implemented with this. Although

Mount and Blade II: Bannerlord does not have modding tools this useful, they would not be required

since the game is already very accessible as mentioned in section 2.3. CiF-CK has an architecture

very similar to CiF’s, with the main characteristics being social exchanges, characters, social state and,

added in CiF-CK, beliefs.

Social Exchanges: In Skyrim everything revolves around quests [4]. Both the main storyline and

small sidequests are quests, but they can also be used to store variables, dialogue, actors, performances

or even locations. The adaptation of CiF into Skyrim transformed the social exchanges into quests. An

adaptation to the characters was also made in CiF-CK. In CiF agents have a collection of their desired

social exchanges, this was converted into a collection of quests with targets.

The Elder Scrolls V: Skyrim, Bethesda Game Studios and Bethesda Softworks, 2011

Fallout 76, Bethesda Game Studios and Bethesda Softworks, 2018

CiF Function CiF-CK

Name Unique Identiﬁer Quest ID

Intent Pretended social state

change

Used once the quest has ﬁn-

ished

Pre Conditions Conditions for it to happen Quest start conditions

Initiator Inﬂuence

Rules

Character’s desire to initiate

the exchange

Set of rules

Responder Inﬂu-

ence Rules

Target’s reaction to the ex-

change

Set of rules

Effects The resulting consequences

of the exchange

Applied in the ﬁnal stage of a

quest

Instantiations Performance of the social ex-

change

Dialogue between NPCs

Table 3.1: Conversion of social exchange into quest [3]

The social state is one of the most important characteristics of CiF. This is a representation of the

state of the world, however, the world in CiF is relatively small compared to RPGs, like Skyrim. There

are public attributes like relationships and social networks in CiF, that all agents have access to, which

is possible in small scenarios, in CiF-CK that is not possible though. To tackle this problem, instead of

loading all values and attributes, CiF-CK only loads the ones close to the player, like other things in the

game such as assets and scripts. Relationships, Cultural Knowledge Base and Social Facts Database

are all the same as CiF. Social Networks, which are the feelings each character has for another were

converted from public access to private.

The last main feature of CiF-CK, which was an addition instead of a transformation, is beliefs. This

feature comes in tandem with the fact that the feelings the characters have are now private (Social

Network), since other characters no longer have access to these private feelings, they can only speculate

on each private social network. For example, John likes Mary, which is documented in his social network,

but, since Mary doesn’t have access to this knowledge, she might think John does not like her, that is

her belief. This creates more realistic scenarios where agents do not know the other agent’s intention,

feelings or reasoning.

Social NPCs was the mod release with the implementation of CiF-CK and it was a big success within

the community, proving as a useful improvement to the game. Using CiF-CK adds something new to the

NPCs present in Skyrim and that is something players clearly want [3].

3.1.4 CiF-Ex (Comme il faut - Exiles)

Comme il Faut - Exiles (CiF-Ex) is another extension of the CiF model, using CiF-CK as a baseline, it

set out to improve even more the model. It extended CiF-CK by exploring more the characters’ emotions,

extending the belief system, improving the temporary status of the characters, adding another outcome

for the social exchanges and adding the ability for the player to respond to social exchanges. CiF-Ex

was implemented as a mod for Conan Exiles, a game developed by Funcom and released on May 8,

2018. It was developed in the Unreal Engine 4 and was made available for mods with some tools called

blueprints. However, these tools were not the most adequate and brought some problems by lowering

the access and freedom modders had for development.

Figure 3.3: CiF-Ex Structure [15]

Although Social Exchanges had no major changes compared to CiF-CK, the characters did. They

can now be busy for social interactions and they also have a more detailed version of temporary statuses.

To avoid excessive social exchange initiations, characters in CiF-Ex have a cooldown, that is reset

when it initiates the exchange. The attribute is called On Cooldown and it is inﬂuenced by the character’s

traits, such as shy, where the cooldown is much higher than an extrovert. Agents can also be occupied,

they then have a ﬂag called Busy. The character can be targeted for a social exchange, marking them

as busy and forbidding other characters from interacting with them, however, the player can bypass this

ﬂag, basically being able to interrupt any existing exchange.

Statuses were improved on the previous iterations by grouping them in categories, Emotion Status,

what the character is feeling, e.g. happy, Social Emotion Status, which are the character’s feelings

towards others, e.g. worried and Physical Status, that are the physical needs, like tired or hungry.

Social Exchange Memory stores representations of the social exchanges that happened in the

vicinity of the character [15]. This virtually replaces CiF’s relationships, by having these private and in

each character’s memory. Instead of having access to all relationships existing in the world, the entries

only include the agent’s relationships and the ones it witnessed. Social State is categorized by the

same four attributes as CiF, Social Networks, Social Networks Beliefs, Relationships and Social Facts

Knowledge Base.

• Social Networks are the character’s private feelings. It has a reference to how it feels towards the

others, with a current value, the goal of the character on the relationship and the belief, what the

character believes the other feels about itself. In John’s case, regarding Mary, he likes her as a

friend, his goal is to continue as friends and, since they are friends, he believes Mary likes him as

well.

• Social Network Beliefs are similar to the beliefs on the Social Network, though, it stores external

relations as well, what some character thinks of another one. As an example, John has seen

another character, Ana, being friendly with Mary, consequently, he believes Ana likes Mary.

• Relationships in CiF-Ex are no longer public knowledge and can be inferred from social network

beliefs.

• Social Facts Knowledge Base is the archive of all social exchanges that occurred.

Although Cif-Ex, similarly to CiF, has microtheories and trigger rules, these have not been improved

in this iteration.

3.1.5 FAtiMA Toolkit

Fearnot AffecTIve Mind Architecture (FAtiMA) is an open source model designed to use emotions as

the primary input to decide the agent’s behaviour. It was initially created for autonomous 3D characters

in a serious game about bullying (FearNot!). It was later developed as a toolkit for socio-emotional inter-

actions. It is based on appraisal theory, which is the claim that emotions are elicited and differentiated

on the basis of a person’s subjective evaluation of the personal signiﬁcance of a situation, object, or

event [18].

Even though it was not from the start, FAtiMA’s architecture is modular, which means it has several

functional independent components that come together to build FAtiMA. This approach makes FAtiMA

easier to work in as well being possible to choose the desired modules. The main module is FAtiMA

Core, which is the one that binds all the components as well as most of the base architecture.

Events are the atomic element in FAtiMA, they can be properties or actions. These are represented

as Well Formed Names (WFNs) through strings that follow a set of syntactic rules. Events contain

symbols and variables to express constant entities and entities not speciﬁed, respectively. There is also

composed names that represent relations between two or more symbols. Event example: Event(Action,

John, Eat, [object]), where John and Eat are symbols and object is a variable.

Knowledge Base is capable of storing beliefs about the beliefs of other agents, also known as

Theory of Mind. It is done with an additional ﬁeld on each belief called perspective, the name of the

agent’s perspective and, by default, it uses its own.

Emotional State is the structure that contains the agent’s current emotions. These emotions are

generated by the Emotional Appraisal library. It is capable of generating 8 different types of emotions

with a total of 22 different emotions. The Emotional Appraisal library uses 5 different appraisal variables

to generate emotions, Desirability, desirability of the consequences, Desirability for others, desirable

consequences for others, Praiseworthiness, how praiseworthy is the action, Goal Probability, how

probable is the goal and Like, how appealing it is. Goals, FAtiMA generates emotions based on the

goals and their probability. By storing the previous goal probability, it compares it when updated gen-

erating a goal based on the comparison. If it is less likely to achieve the goal it will generate Fear or

Disappointment, if it is more probable, it will generate Hope or Relief.

Decision Making in FAtiMA is achieved by keeping a list of speciﬁed decision rules. Each decision

rule contains a set of attributes: action, target, conditions, priority and layer. The actions that can be

performed by the agent, constricted by conditions as well as a priority. The priority is calculated based

on beliefs and meta-beliefs. The layer is similar to type of action, a way to group actions together.

An example of a chosen action could be, running. Where the action is run, the target is self, with no

conditions, a high enough priority to surpass others and with reactive as the layer.

With the modular design on the FAtiMA architecture it is possible to add Reasoner components.

These components add meta-beliefs that can be used in logical conditions [7]. An example for reasoner

compatible with FAtiMA is CiF-CK, a social AI model already studied in section 3.1.3.

The dialog produced by agents operating with FAtiMA is a hybrid solution between dialog trees, that

game developers are well used to, and character-centric design, which means it is the character picking

dialog. A set of dialog lines are prepared for each state and the character can choose what can convey

what it is trying to say best. However, the dialog progress is done through a state tree that is updated

accordingly. A conversation usually starts with greetings, the initial state. The agent can then choose

from a set of lines compatible with the conversation start state, like “Hello” or “How are you?”. The

conversation then progresses to a new state where the agent has a new set of dialog lines to choose

from.

3.2 State of the Art

In this subsection we will review some games with social AI, their impact and their methods for creating a

social environment. It will also be discussed current Mount and Blade II: Bannerlord mods, both popular

and more socially focused implementations.

3.2.1 Games with Social AI

3.2.1.1 The Sims

The Sims is a series of AAA games developed by Maxis and published by Electronic Arts and the ﬁrst

came out in 2000. It has been an extremely successful series, even being the best selling PC game ever

by 2002. The Sims is a series of life simulation and sandbox games. The main gameplay loop consists

of creating characters and their houses as well as accompanying these through their daily lives. The

player mainly babysits the sims so that they can live the best possible way.

Figure 3.4: The Sims 4 Screenshot

Although the player never directly interacts with sims, they have desires and personalities. They can

communicate with each other making the player manipulate the way the interaction progresses. Sims

have a relationship value similar to Bannerlord’s, where there is a value that determines the type of

relationship ranging from -100 to 100. If it is lower than a threshold sims consider each other enemies,

while high, there is the possibility to start new interactions, for example, a romantic relationship. In Sims

4, they added a different value for romantic ones, existing now two different relationships, similar to

CiF. The player can control the interactions the sims have with each other, raising or decreasing their

relationship value, depending on the traits present in each sim.

While this method of quantifying the relationship with a value seems simple, it proves to be quite

successful as we can see in this franchise. The fact that the sims interact with each other, while the

player controls them from outside, removes the immersion from these social interactions. This effectively

makes the player view them with a more rational approach, instead of the usual emotional decision

making present in relations.

3.2.1.2 Fac¸ade

Fac¸ade is a game that was initially released in 2005, winning the 2006 Slamdance Independent Game

Festival and developed by Procedural Arts. The game takes place in an apartment owned by a couple,

Trip and Grace, the two NPCs in the game. The player assumes the role of a close friend of the couple,

however, with very little context to it, the player should try to understand what is happening. Afterwards

the player is forced to make choices in the drama that is happening, Trip and Grace will ﬁght to win the

player’s side. The characters are designed to respond robustly to a variety of open-ended dialog from

the player, including questions and provocations. Players are rewarded for being proactive and acting

dramatically. Through dialog, gesture and action, the player has continuous effects on their afﬁnity with

Grace and Trip, the current level of tension, and the speciﬁc information revealed about their marriage [9].

Figure 3.5: Screenshot Fac¸ade’s gameplay

Fac¸ade uses innovative and experimental technology regarding AI. Both characters are very expres-

sive since they have facial expressions and a vast group of gestures and body expressivity. This was

possible using a reactive planning language, ABL (A Behaviour Language), improving on the agent lan-

guage Hap by changing the syntax, generalizing mechanisms, adding multi-agent language support [8].

Fac¸ade also has a rule system capable of understanding and interpreting natural language, allowing

players to be expressive and to explore the game in a unique way. Although the result is an impres-

sive story with an insane level of freedom and expressivity, the game required an enormous authoring

effort [9].

3.2.1.3 Prom Week

Prom Week is a game released in beta in 2011 and featured in the 2012 Independent Games Festival,

where it was nominated as a ﬁnalist in Technical Excellence. It was developed as an implementation

of the CiF architecture, the social AI model seen in section 3.1.1 and developed at the University of

California at Santa Cruz. Prom Week places players in a typical high school, abuzz with excitement

over the upcoming prom. Players indirectly sculpt the social landscape by having these hapless high

schoolers engage in social exchanges with each other [12].

Figure 3.6: Screenshot of Prom Week’s gameplay [12]

Prom Week revolves around the social lives of eighteen characters, each with a story that contains

a collection of levels. Each character has a different set of goals, the gameplay loop is to compose a

plan to complete each goal. The ending of the story, also known as the last level (which is common to

all stories), takes place at the prom and a customized ending, based on the player choices and actions

plays out. As stated previously, Prom Week uses the CiF AI model, which was tailored to the game.

CiF manages to build an environment that is able to make the player really take into consideration the

characters’ feelings and personality traits using the social physics engine to create an engaging and

satisfactory story.

3.2.1.4 Persona 5

Persona 5 is a RPG developed by Atlus. It is the sixth installment in the Persona series, which is part

of a larger franchise, Megami Tensei. This series started as region locked, with the ﬁrst installment

released in 1996, Japan and 2010, worldwide. Persona 5 was released in 2017 and it had a recent

remake (Persona 5 Royal) which came out in March 2020 worldwide. This game was somewhat a

breakthrough in western audiences, being praised for its art, visual presentation, gameplay, music and

most importantly, the story/characters’ impact on the player.

In this RPG, players take the role of a high school student in Tokyo, Japan. The gameplay is sep-

arated into two parts. A realistic world, where the purpose is to meet characters and help solve their

problems. And a different realm, where the player and party transverse and battle through dungeon

crawler missions affected by the overworld activities and interactions. The entire game is played out in a

school year, meaning it ends after a pre-determined amount of in-game time. Persona 5 playthroughs’

goals are to develop relationships and to complete the story, which is heavily dependent on relations,

within this limited amount of time.

Figure 3.7: Persona 5 conversation screenshot

The relationships the player has with the characters are called Conﬁdants. Conﬁdants are catego-

rized by ranks, the higher it is, the stronger the bond. Higher ranks unlock dialog, game events, battle

features and the possibility of dating. There are two aspects to Conﬁdant ranks, relation points and rank

up. The player gains relation points with the conﬁdants through dialogue and social interactions. When

these points reach the limit, it is possible to rank up through a quest unique to the character in question.

These quests are very sensible aspects of each character’s life and, with the player’s help, are able to

complete it and overcome their problems.

Illustrating the characters’ emotions and stories make them very believable. Due to this, Persona

5 manages to successfully create a strong bond between players and characters and fully immerse

players in the world.

3.2.1.5 Conclusions

The Sims have a simple, yet effective relationship system. It allows players to easily understand the

relation and to try to change it for their purpose. However, players are not invested in the relationships,

losing immersion in the social aspect. The similar happens with Bannerlord, which also has a similar

system.

Although, both Prom Week and Fac¸ade are unique phenomenal games, they are exclusively centered

on their social physics, missing out on some of the core characteristics AAA games usually offer and

that most players enjoy, like graphic ﬁdelity and gameplay mechanics.

In Persona 5, characters’ stories and relationships are closely tied together which translates into a

strong immersive connection with the player. Persona 5 sets a great example for social interaction in

AAA games, which is desirable for Bannerlord’s characters, considering the game’s storyline.

3.2.2 Current Bannerlord Mods

Mod is short for modiﬁcation and it means an unofﬁcial alteration of a game. There are a large variety

of mods, ranging from changing the game’s graphical ﬁdelity, gameplay or any content present. A game

renown for its mods is, for example, Minecraft

, a game with so many mods that the players often refer to

the unmodiﬁed game as vanilla. Mods are a great way to ﬁx some problems, to add necessary features

the player base misses or even a great way to extend the game’s longevity by adding new content and

creating an active community [1].

Mount and Blade II: Bannerlord came out on 30th March, 2020, meaning the game is relatively new,

therefore, even though people are substantially more excited to develop mods on a new game, it is

expected that there are few mods, as well as existing being relatively simple or underdeveloped.

Some mod developers are very ambitious, trying to bring universes like Star Wars

or Game of

Thrones

into the Mount and Blade, develop new mechanics, bring new objects like guns. These will

take a lot of effort and time, thus it might be a while until mods of this caliber show up in the scene,

especially considering modders are working in this as a hobby.

Most Bannerlord mod can be found in Nexus Mods

, a website that supports modding for all PC

games. The three most popular mods for Bannerlord on the site are:

• ModLib, that is designed to help and provide some features for modding developers.

• Bannerlord Tweaks, that adds a collection of tweaks and modiﬁcations to improve the quality of

life of certain aspects.

• Fast Dialogue, that makes the dialogue faster by removing loading screens from interactions with

essentially all map interactions.

There are very few mods that work on the characters or their relations and the ones that exist are

fairly simple as well and more oriented towards ﬁxing bugs and improving gameplay. The most popular

of these are:

Minecraft, Markus Persson, Mojang, 2009

Star Wars, George Lucas, Lucasﬁlm, 1977-2020

Game of Thrones, David Benioff and D. B. Weiss, HBO, 2011-2019

Nexus Mods, www.nexusmods.com

• Diplomacy Reworked, that adds a few options on the diplomacy side, basic and advanced politics

and more bargain options.

• True Relations, that improve how the relation between character change, by improving how often

it does from killing bandits, winning or losing battles and tournaments. All characters in a clan have

the same relation towards the player and this mod ﬁxes that by separating each lord relation into

its own. It also upgrades how the trait system works, leveling up the player character as well as

companions and being possible to imprint the players characteristics onto companions.

Since the game came out in an early beta stage, meaning although playable, it is not complete and

is currently under development. There are constant patches, usually once a month, adding missing

features and improving on others. Consequently a lot of mods, in fact, most mods focus on solving

these problems, adding necessary missing features and ﬁnishing uncompleted ones. As we can see

from both existing social mods and from the most popular mods on the site, these are mainly small ﬁxes

to the game. We can conclude the modding scene for Bannerlord is somewhat lacking.

Comme il Faut: Bannerlord

Contents

4.1 CiF-Bannerlord Additions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.2 CiF Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.3 CiF-Character . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.4 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.5 Social Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.6 Rumor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.7 Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

The Comme il Faut (CiF) model seemed the most appropriate to implement in Bannerlord. CiF is a

simple and efﬁcient model, that can be easily extended. Bannerlord is a game where the interaction with

NPCs is performed inside instances ﬁlled with other characters, due to this situation, the model needs

to be computationally light in order to not worsen the efﬁciency of the base game. The CiF model also

shines in crowded situations where characters can converse with each other, making the environment

lively and enjoyable. For these reasons Comme il Faut was the chosen model to be developed in Mount

and Blade II: Bannerlord.

In this chapter, we will discuss the implementation of Comme il Faut: Bannerlord, CiF-Bannerlord. In

the ﬁrst section, we will go through the differences and improvements in CiF-Bannerlord regarding other

CiF models. The second section explains the CiF cycle, the procedure of CiF-Bannerlord occurring

in-game . Afterwards, we will examine the implementation of the different features of CiF-Bannerlord in

the third, fourth and ﬁfth sections, respectively CiF-Character, Social Exchange and Rules. In the sixth

section, we will discuss a new addition from CiF-Bannerlord: Rumors. In the seventh and last section,

we will examine how dialog was established.

4.1 CiF-Bannerlord Additions

Comme il Faut: Bannerlord is an implementation of the Comme il Faut model for the game Mount and

Blade II: Bannerlord. CiF-Bannerlord extends other existing models, Comme il Faut - Exiles (CiF-Ex)

and Comme il Faut - Creation Kit (CiF-CK), by improving some aspects and adding new features. The

ﬁgure below represents the addition of each CiF model to the overall development of this architecture.

Figure 4.1: CiF models and their addition to the architecture

The graph in the ﬁgure 4.1 represents in blue the additions each CiF model added to the overall CiF

architecture. As we can see CiF-Bannerlord adds to Character several features represented in a lighter

blue, like Continuous Trait values. The arrows represent direct additions or changes to speciﬁc features,

like the Power relationship being added as a Relationship.

A – Extended Personality Traits - In order to improve believability in the characters present in

the game, CiF-Bannerlord extends character traits by making the traits continuous instead of a sim-

ple boolean value. This change was aimed at improving character uniqueness, diversity and in turn

boost the ambiguity of NPCs characteristics further improving the complexity of a social environment

and creating more non-deterministic situations.

B – Cultural Differences - Bannerlord offers a world rich with different cultures and countries. Cre-

ating a set of characters with different cultural backgrounds in each social environment reinforces the

enjoyment of the player in discovering cultures or social environments. These environments were aimed

to create biases on the player based on the localization and it raise emotional attachment to groups

or factions. The change would consequently boost player immersion through their attachment to game

entities.

C – Rumors - In order to improve immersion and believability in these social environments, a new

feature was implemented: Rumors. These symbolize gossip existing in usual social environments where

elements talk behind others’ back and events that occurred. Making the characters reactive to events

and to other characters creates a more robust, lively, and immersive social environment that simulates

more accurately the real world.

D – Power Relation - CiF had two relationship types possible between characters, friendship and

romantic. Usually, in a relationship, there are two parts, two friends or two partners, however, the

introduction of rumors, allows for a new relationship to be possible, a power relation. Instead of the other

two, this is a one-sided relationship where one character can have an idea about another one based

on the things they hear from rumors. Commoner people should naturally be careful around powerful

leaders or dangerous warlords.

4.1.1 CiF-Bannerlord Removed Features

The latest implementation of the CiF model was already a relatively complex and complete implemen-

tation, therefore in order to further extend the model, some aspects were scraped and simpliﬁed. This

decision allowed to maintain a simple system while improving the complexity in other aspects.

A – Status Simpliﬁcation - CiF-Ex was a version of CiF that allowed for characters to be especially

expressive. With the addition of Statuses types and with a continuous value for it, characters could be

more expressive since they had a better and more complex representation of the Statuses they could

be feeling. In Conan Exiles, CiF-Ex’s base game characters are constantly present and have game

mechanics such as hunger. Unfortunately, this additional complexity on the Status feature has been

scrapped. In Bannerlord, the player (and characters in general) are not actively together in an instance,

instead most of the gameplay in the normal campaign is just managing the party. The player only enters

a social environment when entering the town center, keep or tavern. Therefore it does not make sense

to work on such a feature for this particular game and instead it was chosen to focus on other features

or additions to the CiF model.

B – Social State - In the CiF model, the concept of Social State consists of a snapshot of the state

of the whole system, how relationships have evolved, the Social Networks of each character (their non-

reciprocal feelings), and the Social Facts Database that contains information about every Social Ex-

change (SE) that occurred.

All of the mechanics featured in the Social State are present in CiF-Bannerlord. However, during the

development of CiF-Bannerlord there has not been a need to link these mechanics in order to form the

concept of Social State. Instead, it has become a decentralized concept being possible to have a notion

of the state of the system through each mechanic. For example, there is not a snapshot of the Social

State, but it is possible to have a snapshot of every current relationship or a snapshot of all SEs that

were performed.

4.2 CiF Cycle

The CiF-Bannerlord procedure has four distinct stages. Desire Formation, Intent Formation, Exchange

Realization and Exchange Fallout.

Figure 4.2: Visual representation of the CiF cycle

4.2.1 Desire Formation

The character’s decision to act depends on its energy attribute. Each character has a different threshold

to decide the fullness of their energy and the time it takes to ﬁll up or down is also unique, based on the

delta energy. The threshold is calculated based on the personality set and based on the equation below.

energy

threshold

= α + (−w

e,annoy

annoy

+ w

e,calc

calc

− w

e,curious

curious

) (4.1)

In the equation 4.1 there are two components, a constant α, and the offset based on the personality

set. The t-values are the personality traits, while the w-values are their respective weight on the formula.

The character energy thresholds have values between 0 and 1, and α was empirically determined to

be 0.5. The chosen traits to calculate the energy threshold were annoying, calculating and curious.

More annoying characters will perform exchanges more frequently and therefore have lower thresholds.

Calculating NPCs should not be so impulsive, but instead, contemplate their actions before acting. This

aspect is modeled by having calculating characters with a high energy threshold. Curious characters,

similarly to annoying, lower the threshold since they do not want to become bored and act sooner. The

w-values are weights (empirically determined) that are assigned to each particular trait to have more or

less impact on the equation.

The energy starts at zero and it is set at zero in every time a Social Exchange is performed. The

energy level of a character increases as time elapses (once each tenth of a second). The increase is

determined by the energy delta, which is also inﬂuenced by a character’s personality.

∆energy = (β + (−w

∆e,stubborn

stubborn

− w

∆e,calc

calc

+ w

∆e,curious

curious

− w

∆e,shy

shy

)) × noise

(4.2)

noise ∼ U (0.9, 1.1) (4.3)

Similar to the threshold, delta energy has a base value, β, and an offset calculated with the person-

ality set. The delta energy has a noise factor with the interval illustrated in equation 4.3. This noise

factor has the purpose to lower predictability, making each NPC not have a constant timer to perform

exchanges. The base delta energy, β, was determined empirically, with a base value of around ﬁfty

times smaller than the base threshold, that was selected in order to have an average of ﬁve seconds

to reach full energy. The traits used to calculate the offset were stubborn, calculating, curious and shy.

The delta energy is lowered by the shyness trait since shy characters ﬁnd it harder to gather energy to

interact and it is the most impactful trait out of the four. The stubborn trait lowers the delta energy since

the characters do not feel as much need to interact and change their situation. Calculating and Curious

traits work similarly to the threshold equation.

When the character reaches full energy it is determined that it desires to perform a Social Exchange

and proceeds to the Intent Formation step.

4.2.2 Intent Formation

After the initiator has decided to start an exchange, it needs to discover the type of the exchange, as

well as the receiver for it. For each character present in the instance, the initiator calculates an intent for

that character as well as the desired interaction.

In CiF-Bannerlord, the intent has ﬁve different values: Neutral, Positive, Negative, Romantic and

Embellish. Neutral does not have an impact, interactions with positive intent are generally compliments

or similar while negative is generally more towards insults. Romantic intent is related to romantic interest

in the receiver. Embellish intent is similar to positive, however, the initiator is trying to be positive to praise

in order to win favor. Embellish is related to the power relation where the NPCs need to act nice around

powerful characters.

There are two distinct ways of determining the Intent that is going to be selected: a stochastic process

based on the character’s personality, and a stochastic process based on the character’s social feelings.

The ﬁrst step is to determine which of these two processes is going to be used for the intention

formation. This is also determined stochastically, as described in Equation 4.4.

personality

= w

annoy

− w

caref ul

− w

calc

(4.4)

feeling

= 1 − p

personality

(4.5)

The probability of using the personality based process is given by three traits, annoying, careful and

calculating. These traits were chosen as the ones that determine the NPC’s unpredictability. The mod-

iﬁer is negative for careful and calculating traits, demonstrating that these characters are less probable

to be unpredictable while annoying ones are considered unpredictable in CiF-Bannerlord. The w-values

are empirically determined weights that are assigned to each trait. There are two methods to decide

intent, either based on personality or feelings as we can see in the equation 4.5.

4.2.2.1 Intent based on Personality

The intent selected based on the personality method is a stochastic process where each intent has

a respective strength value that determines the probability of being the chosen intent. For instance,

equation 4.6. shows how the strength value for the positive intent is calculated.

Str

positiv e

= w

positiv e,kind

.fn(t

kind

)+w

positiv e,helpful

.fn(t

helpf ul

)+w

positiv e,sensitive

.fn(t

sensitiv e

) (4.6)

Traits values were modiﬁed by a normalization function, fn, as we can observe in the equation above.

There are a total of four intent values calculated using similar formulas to equation 4.6, with the difference

being the traits chosen for each.

Positive and negative strength values are calculated using the same three traits, kind, helpful and

sensitive. These traits are negative for the negative strength. Kind characters want to be positive to

others. Helpful NPCs like to help and be friendly and sensitive ones are careful with others emotions

making them want to have positive intent. These traits also apply to negative intent, however with the

opposite reasoning, where unkind characters want to be negative for example.

Romantic strength values are calculated using charm, careful and shy traits. Charming characters

are more drawn to romantic situations. Careful and shy NPCs usually run away from taking risks and

are careful or fearful of ruining conversations, therefore these traits are considered with negative values

for this strength formula.

Embellish strength uses the traits, helpful, honor and liar. Characters wanting to embellish others,

have a friendly intent (helpful), however, it is shallow with a secondary purpose. These types of be-

haviours are not honorable and are associated with lying. The trait honor is considered with negative

values due to this.

The neutral strength is the rest of the probability considering all of the others combined, as seen in

the formula below.

Str

neutral

= 1 −

i6=neutral

Str

(4.7)

The ﬁnal intent is determined as the softmax of the strength of the possible intentions.

4.2.2.2 Intent based on Social Feelings

The intent selected based on the social feelings of the character method, like the other method, is a

stochastic process where each intent value has a respective strength value that determines the proba-

bility of being the chosen intent. This strength, instead of being calculated based on the personality, is

calculated using the character’s social feelings. Further details on feelings are on section 4.3.

The neutral intent chance has an empirically predetermined value 30%. All other strength values are

obtained from the social feelings in the characters, which in total account for 70%.

total = f riendly + romantic + admiration (4.8)

In CiF-Bannerlord, social feelings are represented through a value, an intensity of the feeling. Char-

acters have three different types of social feelings as illustrated by the equation 4.8. This total intensity

is used to calculate the strength of each possible intent.

Str

positiv e

friendly

total

× 0.7 (4.9)

The formula above illustrates how the probability strength of positive intent is calculated. It uses the

friendly intensity in the character’s social feelings and the total intensity obtained from the equation 4.8.

The negative intent probability strength is calculated with the same formula as the positive one. The

difference is the intent in the friendly feelings, which determines which of the two intents is chosen. If

the character has negative friendly feelings for the other, it has a chance to be negative, if it has positive

friendly feelings, it has a chance to be positive.

The probability of each value is then calculated as exempliﬁed in the equation above for the positive

intent. The positive and negative chance depends on the intent on the friendly feelings. The romantic

chance depends on the romantic feelings and the admiration feelings determine the embellish intent

chance.

To exemplify the intent situation, a mean character is more likely to be negative to others, while a

kind character is more likely to be positive. Although personality has some impact on the intent, NPCs

are always much more likely to have neutral intent with strangers. To other characters they know, based

on their feelings, of friendship or romantic, they are more likely to have an intent similar to the speciﬁc

feeling. A character that has negative friendly feelings for another, will be more likely to be negative. A

character that has both positive romantic and friendly feelings, will be more likely to be either romantic

or positive.

4.2.2.3 Determine desired Social Interaction

After determining the intent, the initiator calculates the value of each Social Interaction (SI), using the

intent and receiver as input for this. The initiator calculates the SI’s value as the sum of the values on

each inﬂuence rule applied to the situation. These consist of the initiator inﬂuence rules contained in the

Social Interaction, the inﬂuence rules related to the character’s culture and general inﬂuence rules that

apply to every character.

value

r∈inf l uence rule(SI)

r.value (4.10)

The equation 4.10 determines the value of the Social Interaction (SI), currently being calculated. This

formula is applied to all SIs available to the character. This procedure is ﬁnished by storing the highest

valued SI, together with the respective SI, intent and receiver.

The initiator then has the same procedure for every character available (as receivers), calculating

intent, and highest valued SI. The ﬁnal interaction chosen is the one with the highest value of all the

receivers and after achieving the desired interaction the initiator proceeds to perform the exchange.

4.2.3 Exchange Realization

When the initiator character has decided both the receiver character, their speciﬁc SI, and intent, the

Social Exchange begins. These characters are now performing an exchange, they get a ﬂag signaling

they are busy, to stop other characters from interacting with them and lose the ability to gain energy to

perform other exchanges. First, the initiator moves towards the receiver. After arriving a determined

threshold of distance of the receiver, the initiator chooses a dialog line contained in the SI, further details

on dialog are explained in the section 4.7.

After a few seconds, the receiver responds to the initiator. The receiver calculates the response value,

based on the SI’s receiver Inﬂuence Rules, cultural rules and general response rules. This calculation

is similar to the equation 4.10, with the difference that it uses the receiver inﬂuence rules instead of

initiator.

The sum of all rules combines determines this value, which then is used to compare against the SI

threshold intervals, Upper Threshold and Lower Threshold. A higher value than the Upper Threshold

consists of a Positive outcome, a lower value than the Lower Threshold consists of a Negative outcome

and any in between are Neutral outcomes.

Figure 4.3: Visual representation of the thresholds and outcome determination

With this outcome, the receiver then proceeds to choose a response dialog line contained in the SI,

based on the personality set of the receiver, outcome of the exchange and dialog used by the initiator. It

is possible for the initiator to respond to that response, in some instances. Further details on dialog are

explained in the section 4.7.

At the end of the exchange, a Social Exchange object is ﬁnally created with an initiator, a receiver,

date, outcome and SI and the respective fallout is applied.

4.2.4 Exchange Fallout

When the SE is over, the fallout begins. There are two types of rules being responsible for the fallout,

Instantiation Rules and Trigger Rules. Instantiation Rules are speciﬁc functions to the SI and are exe-

cuted at the end of an exchange. In the Date exchange, for example, the initiator asks the receiver to

start a romantic relationship and if the outcome is positive, there is an Instantiation Rule responsible for

creating a romantic relationship and updating the characters. Trigger Rules are general functions that

are executed at the end of an exchange. They are responsible for updating the feelings and beliefs of

the characters present or creating rumors for example.

The SE is saved in the Social Facts Database and a reference of this exchange is stored in the

memory of the parties involved, initiator and receiver. The characters stop following each other, get their

energy at zero, and the ﬂag signaling they are occupied is turned off, so are now able to start exchanges

with other characters and their energy is able to be raised again.

4.3 CiF-Character

In CiF-Bannerlord, the NPCs are represented as CiF-Characters. The table below represents the com-

ponents of a CiF-Character.

Table 4.1: CiF-Character Attributes

Attribute Description

Personality set of CiF-Character Traits

Status List of statuses currently being felt by the character

Friendly Feelings List of friendly feelings on other characters

Romantic Feelings List of romantic feelings on other character

Admiration Power relations with other character

Beliefs List of feelings the character thinks the other characters feel

Energy energy the character has available to initiate exchanges

Threshold threshold for the energy

Culture Cultural reference of the character

Memory List of exchanges the character took part in

Last Exchange List of the last exchanges the character took part in

A – Personality - is a collection of Traits, unique to each character. Each Trait is a number that can

have a value between -1 and 1. Traits have one bidirectional dimension that represents two comple-

mentary opposites in a spectrum. Values near -1 correspond to a negative Trait and a value near 1

corresponds to the opposite. For instance, a character with -0.5 Kind means it is somewhat malicious,

and a character with 0.9 Kind means it is very kind.

Table 4.2: Personality Traits

Personality Trait Description

Kind Represents kindness and maliciousness

Helpful Represents selﬂessness and selﬁshness

Sensitive Represents sensitivity and insensitivity

Honor Represents honorableness and shallowness

Calculating Represents intelligence

Stubborn Represents stubbornness and gullibility

Liar Represents easiness to lie and invent lies

Curious Represents curiosity

Shy Represents shyness

Careful Represents carefulness

Charm Represents beauty and charm

Annoying Represents annoyance and how it perceives other’s annoyances

The table above illustrates the Traits present in CiF-Bannerlord. There are a total of twelve Traits,

which all are created randomly in the initiation of the CiF-Character. The ﬁrst ﬁve Traits, Kind, Helpful,

Sensitive, Honor, Calculating have an input of the character traits already present in the game. These

CiF-Traits are still created randomly, however they have a higher probability based on the value of these

in-game traits, which are respectively: Generosity, Mercy, Valor, Honor and Calculating. The other seven

traits, respectively, Stubborn, Liar, Curious, Shy, Careful, Charm and Annoying were added based on

features present in CiF-Bannerlord. Stubborn, Liar and Curious are related to rumors and hearing

about other characters, Shy, Careful and Annoying are related to the start and response of exchanges.

Charm is related to the romantic aspect of CiF-Bannerlord, representing beauty and innate charm of the

character.

These CiF-Traits are used for most operations and calculations for each CiF-Character and are the

basis for most rules within CiF-Bannerlord architecture. For simplicity purposes, the personality compo-

nent also stores the sexual preference of each character which is also random. With such a large set of

Traits and respective possible values, each character CiF-Bannerlord creates is substantially unique.

B – Status - is a list that represents the characters current state. Each status present in the Status

list signals the character is feeling that emotion/state. There are ten different Statuses with one default

(Normal): Wounded, Bored, Tired, Angry, Fearful, Sad, Conﬁdent, Happy, Ashamed and Prideful.

C – Social Feelings - CiF-Characters have three different types of social feelings, friendly, romantic

and admiration. These Feelings have a recipient, an intent and an intensity. The recipient is a CiF-

Character and the intensity is a number that starts at zero. Without thresholds, the intensity value can

be raised or lowered through interactions. It has usual values between 0 and 10. The intent is helpful to

determine if the feeling is positive or not. For example a feeling with negative intent and high intensity

means the character dislikes the other in great quantities. However a character that has neutral intent

and low intensity, means the character does not think much of the other. Intent corresponds to some

intervals of intensity when related to friendship, when below zero, intensity is raised again and the intent

changes to negative. Though other feelings (not friendship) can have negative values.

Friendly feelings is a list of social feelings, the character in question likes or dislikes. Romantic

feelings is a list of social feelings that the character has romantic interest in.

D – Admiration/Power Relation - is a list of social feelings related to hearing about other characters.

These social feelings are not affected by interacting directly with the character, but it is instead the notion

the character has from others based on rumors. This feeling has two possible intents, positive and

negative. Admiration with positive intent means the character looks up to the other character. Admiration

with negative intent means the character is fearful of the other character.

For example, the character hears a rumor about a battle nearby, the leader of the victorious party is

going to be admired in those rumors, meaning their admiration feeling is positive and it is going to be

boosted. While the loser is going to look worse in those characters’ eyes, meaning the admiration is still

positive, however, it is going to lose intensity, as characters are losing admiration/respect from losing the

battle. Another example is raids. It is usually very badly seen by other characters, when a party raids

a village, destroying everything. Characters that hear rumors about this become fearful of the character

responsible for the raid, having a high intensity admiration feeling with negative intent.

E – Beliefs - are social feelings the character thinks the other characters feel. Each belief of the

character is basically the social feeling another one feels, based on the exchanges performed by the

other that the character could see. These beliefs do not correspond to the exact social feelings of other

characters, they are based on the perceived SIs that occurred. The value that is added to beliefs is

greatly impacted by the character’s stubborn trait, where it refuses to believe things or it is gullible and

believes in other’s feelings with great intensity.

For example, if a character is seen performing several romantic interactions with another, the belief

of surrounding characters is that the character performing all the interactions has high romantic feelings

for the other.

F – Culture - Culture is fairly important in Bannerlord, characters have different stories and mental-

ities from their culture which ties in with the game lore. CiF-Bannerlord reﬂects this by adding different

features to every character. Each character can perform unique SIs only available to that culture. Or

characters can have access to unique cultural dialog lines, Inﬂuence rules, instantiation rules and trigger

rules on any SI. Culture also has modiﬁers applied to the character’s energy or thresholds. Basically,

culture works as a separation of groups of several CiF-Bannerlord mechanics.

G – Memory - CiF-Bannerlord characters store previous interactions. This attribute refers to interac-

tions or rumors the character has found interesting. This interest is determined by values on the rumor,

further details are explained in section 4.6. Last Interactions is an attribute that accounts for the char-

acter’s short-term memory. It stores the last ﬁve interactions the character performed, in order to avoid

repeating the same few interactions.

4.3.1 CiF Player Character

In the original CiF, there was not a main character or a character the player would role-play as. Instead,

the game was focused on each character’s story and the player could play with any character. That is not

the case with RPG games where the player has a speciﬁc character built and developed by them over the

period playing. In CiF-Bannerlord the main character is a CiF-Character like all the other NPCs, however,

it has a few distinctions. Each character’s set of Traits is created generally randomly, except for the main

character. It starts with all of its Traits as 0 and then is increased or decreased through instantiation

rules and based on the interactions the player decides. For example, if the player repeatedly decides to

be negative with other characters, it will raise its Trait of maliciousness.

The main character also differs in dialog. Characters go through the CiF cycle in order to perform

a SE, however, in CiF-Bannerlord, the player decides several steps of the cycle, like when and who to

speak with the base game built-in mechanics, such as walking towards and starting dialog. The player

also chooses the intent in the middle of dialog, at the start of a CiF exchange. More details on the dialog

will be further explained in the implementation chapter.

4.4 Rules

Rules are the foundation of the CiF architecture. There several types of rules in CiF-Bannerlord, condi-

tions and three different implementations of rules, Inﬂuence Rules, Instantiation Rules and Trigger Rules.

Microtheories are not rules in the CiF-Bannerlord model but are instead functions that occur when an

event is triggered.

A – Conditions - are the simplest rules, they have a name that identiﬁes it, and a validate function

that determines if the condition returns positive or not. In the CiF-Bannerlord everything uses condi-

tions to test its availability and these rules can be inﬂuenced by anything, character attributes, SIs or

exchanges, other rules, or any situation in the base game. An example is ”LordsOnly”, a condition that

returns true when both parties are lords. This condition also requires there to be two participants.

B – Inﬂuence Rules - These rules have a name and a list of conditions that implement a validate

function, in order to determine if it is active or not based on the cumulative result of this list, meaning all

conditions need to be true for the rule to be functional.

Inﬂuence Rules also have a delegate function that is used as a value related to the will of the charac-

ter and the speciﬁc SI containing this rule. Like Conditions these rules are broad and can have an input

of anything, like characters or their attributes, interactions, exchanges, rules, or anything from the base

game as well.

An example is ”LikesThem”, an inﬂuence rule that returns a higher value based on the social feelings

the ﬁrst character has for the second. Only the highest value of all three different social feelings is used

for this rule, where romantic feelings are worth double of friendly feelings.

C – Instantiation and Trigger Rules - Instantiation rules and Trigger rules are very similar, they

both contain a name and a validate function implemented by a list of conditions, like Inﬂuence rules,

to determine if the rule is active in the particular situation or not. The difference between them is that

instantiation rules run their effects in the middle of an interaction, while trigger rules run their effects at

the end of it.

Unlike Inﬂuence Rules, instantiation and trigger rules do not return any value and instead change the

environment or characters directly. An example is ”StartDating”, an instantiation rule that marks on the

character relation manager the two characters as dating.

D – Microtheories - In the original CiF model, microtheories correspond to simple rules that can be

applied to speciﬁc situations. An example of a CiF microtheory is: If the character is friendly with another,

he will be more likely to have positive exchanges. In CiF-Bannerlord, these types of microtheories have

been merged with inﬂuence rules, aiming to simplify the use of rules. If the rule is used to calculate

some result or inﬂuence to determine interactions or situations, it is considered an Inﬂuence Rule. In

CiF-Ex for example, there is a Memory Frequency Microtheory that says, characters will have a debuff

on the volition of the last SEs that have been performed, where in CiF-Bannerlord, this situation was

transformed into an Inﬂuence Rule called Repetition.

Microtheories have been rearranged and in CiF-Bannerlord these are connected with the events

happening in-game. They are speciﬁc functions that are executed when an event is triggered. For

example, on the completion of a quest, it raises the friendly feelings between the characters present.

4.5 Social Exchange

SEs are an integral part of the CiF architecture, in the original Comme il Faut, SEs were events that

happened when the player chooses dialog lines. In CiF-Bannerlord this feature has been separated into

two, SIs, which are the type of interaction available and SEs, that are occurred SIs, with an initiator and

receiver of the exchange as well as a date for when it happened.

4.5.1 Social Interaction

In CiF-Bannerlord, SIs function as the SEs from the CiF model. They have a unique set of Inﬂuence

rules that determine how much a character would like to perform it, or how much the character would like

to respond. They contain the dialog speciﬁc to the SI and contain the necessary thresholds to determine

the outcome of the Exchange.

Table 4.3: Social Interaction Attributes

Attribute Description

Name Identiﬁer of the interaction

Preconditions List of preconditions that determine if it is active

Initiation Rules List of initiator inﬂuence rules

Response Rules List of receiver inﬂuence rules

Instantiation Rules List of instantiation rules

Dialog Rules List of custom Instantiation rules related to dialog

Upper Threshold Threshold that determines positive outcomes

Lower Threshold Threshold that determines negative outcomes

Sentences Matrix of dialog containing the available sentences

A – Preconditions - like rules, SIs have a list of conditions that determine if it is active, being possible

only when all the conditions are considered true.

B – Initiation and Response Rules - In order to determine the best course of action, each SI has

a collection of rules, and the sum of their values determines how likely the character is to choose that

interaction. The SI with the highest value calculated with Initiation Rules will be the chosen one. After the

initiator starts the exchange, the response value is determined by the Response Rules and the outcome

is determined based on the thresholds present in each SI.

C – Thresholds - SIs have two speciﬁc thresholds that determine the outcome of the exchange,

Lower Threshold and Upper Threshold. The total value of the exchange is determined by the Response

Rule list’s value, if it is below the lower threshold, the outcome is determined as negative, if it is over the

upper threshold, the outcome is determined as positive. Values in between are a neutral outcome. The

ﬁgure 4.3 illustrates these thresholds.

D – Sentences - Each SI has a complex collection of possible sentences to choose from. This

collection is a matrix of Dialog, a custom class created in the CiF-Bannerlord system. This matrix is

composed of a list of lists, where each list corresponds to either one entire exchange or if it has only one

element, corresponds to one sentence. If the list containing dialog has several entries, it means it has a

direct conversation, where the second entry is the response to the ﬁrst and so on. With only one entry,

the initiator can choose this sentence, however, the response given by the receiver is a general answer

and not directly related to the ﬁrst sentence. This translates to conversations not becoming repetitive

and having the characters always respond the same, however, it also leaves answers a little too simple

since they need to be broad and apply to several situations. Further detail on dialog is on section 4.7.

E – Instantiation Rules - These run at the end of the exchange. They are the equivalent to Instan-

tiations in the original CiF. When a speciﬁc SI is executed, these rules apply its speciﬁc changes and

fallout. For example, the Gift exchange has several Instantiation Rules, one of which takes the gift out of

the initiator and gives it to the receiver.

F – Dialog Rules - These are a special type of Instantiation Rules, that create dialog lines dynam-

ically. The sentences available to be used can be obtained from one of these two options, either a

sentence in the Sentences attribute or a line created using Dialog Rules.

For example, Relay Information is a SI where the initiator talks about some rumor they heard, and for

these interactions, each sentence needs to be created with the speciﬁc rumor as its base.

Dialog Rules are unique to the SI and construct a different sentence based on the information nec-

essary. These rules require a sizeable amount of authoring for each sentence where it needs to be

constructed around speciﬁc variables. The most simple dialog rule, for example, is used to describe the

status a character is feeling. It merges the status with part of a sentence like ”I am ”. It has some com-

plexity because these parts of sentences can be replaced, for others like ”I am feeling”, or with modiﬁers

like ”Today,” or ”Right now,”. These rules allow for a more expressive set of dialog lines, however, they

require a good deal of work and effort for each one, instead of the simple and less expressive dialog

system implemented. Dialog Rules are also not compatible with the dialog engine established in the

base game since the dialog lines cannot be created dynamically.

4.5.2 Social Exchange

In CiF-Bannerlord SEs are the instantiations of SIs. Each SE has to have necessarily a type (the SI

performed) and the initiator and receiver that took part in it. The exchanges are the objects that are

stored in the memory and Social Facts Database. SIs are a more abstract concept, for example, ﬂirting,

while a respective example of SE would be character B was ﬂirting with character B. In the table below

we can observe the different attributes contained in the SE.

Table 4.4: Social Exchange Attributes

Attribute Description

Initiator CiF-Character that starts the exchange

Receiver CiF-Character that responds to the exchange

Type Type of the exchange is a Social Interaction

Outcome Outcome of the speciﬁc exchange

Date Day and time of the exchange

Rumor Rumor used in the exchange

A – Social Interaction - Each SE is unique in the sense that is referenced to the interaction between

two characters. Type is a reference to the SI that corresponds to the SE.

B – Outcome - Each SE has an outcome, that has three different values, positive, neutral, and

negative. This value is calculated using the SI Thresholds and the Receiver’s calculations based on the

sum of the applicable response rules.

C – Rumor - The SE has a reference to the rumor that has been the focus of the interaction. If the

exchange does not relate to any rumor, this attribute is null.

4.6 Rumor

People are very social creatures, communicating is a part of our daily lives. Rumors, gossip, talking

about others are always popular topics and CiF-Bannerlord reﬂects this. Rumors are one of the main

features of CiF-Bannerlord, introduced due to the nature of the game, of visiting different settlements

and towns, talking with different characters, exploring different cultures and social environments. Parties

and armies from all factions roam the map constantly, being possible to spread rumors from a settlement

to another easily.

Rumors have an interest value, which can have values from 0 to 1, where 1 is the most interesting

and 0 has no interest. Generally, when a rumor is created, this value is 1 and then proceeds to be

lowered when it spreads. Rumors also have a type, that can have values of Economic, Warfare and

Gossip.

Economic rumors are a feature implemented in the base game where townsfolk inform the player

about good trade deals/offers happening in other towns. It relates to a merchant style of gameplay

where the player can buy cheap products in one town and sell at higher prices to a settlement.

Warfare rumors are related to wars and battles. They are created through events happening in-game

such as declarations of war or peace or sizeable battles.

Gossip rumors are the usual rumors, that have a speciﬁc character or characters as the information,

as well as a SI. For example, a character that usually ﬂirts with several others can be the focus of a

rumor.

4.6.1 Creating Rumors

When starting a new campaign a number of rumors are randomly generated to make the campaign

seem lively even before it was created. Rumors are also randomly generated throughout the campaign

gameplay, to simulate once again, the livelihood of the characters even when the player is not present

in the settlement. These randomly generated rumors are Gossips since Economic rumors are created

by the base game system.

Warfare rumors are created when war or peace is declared. These rumors are worldwide, reaching

every settlement. Another type of warfare rumors is based on battles. When a sizeable battle happens

(with a large number of combatants), the rumor reaches nearby settlements. Afterwards with the prop-

agation of rumors, it is possible to reach other settlements, but it is a weaker less interesting version of

the rumor.

Gossip rumors are created when the player is inside an instance, by a repetition of SIs. Either the

player or other characters can repeat the same interaction several times, and other characters pick up

on that. For example, the player is ﬂirting with a character for several exchanges in a row, a Gossip

rumor is created that portraits the player as ﬂirting a lot.

4.6.2 Spreading Rumors

Rumors are talked about in two speciﬁc SIs, Relay Information and Ask For Rumor. The initiator in these

interactions gives information about a rumor and asks for rumor information respectively. When a rumor

is used its intensity is lowered by a small percentage, a value determined empirically, 2%. This makes

the rumor lose interest faster when it is considered fresh, but it stagnates the disinterest as do all rumors.

Since SEs only occur when the player is present, the rumor only loses interest from this method when

the player is present as well.

Characters choose the rumor to spread based on its intensity and based on character preference.

This preference is a rumor type that is calculated based on the character’s personality set. Each rumor

type preference is paired to a trait, Economic with calculating, Warfare with careful and Gossip with

curious. When the character has its curious trait above an empirically determined value the preference

is Gossip. Otherwise, the highest valued trait of the three is chosen as the preference. The rumor’s

interest and preference are used to calculate the most interesting rumor for the character to use on its

SE. The rumors being used in these interactions come from the Rumor Holder containing a list of rumors

available for all characters present to use.

There are two types of Rumor Holders in CiF-Bannerlord, Settlements and Parties. Settlements have

a list of rumors, that instantiated characters (in taverns or town centers) can use. Parties are responsible

for spreading the rumors between settlements and have a smaller list of rumors. When a party arrives

at a settlement, the settlement’s rumors are updated with the party’s, resulting in a list of rumors with the

highest interest values. When the party leaves the settlement, the rumors it takes are a few of the most

interesting to the party leader, based on the rumor’s interest and leader preference.

4.7 Dialog

Dialog in Mount and Blade II Bannerlord has been an unexpected challenge. The ﬁrst implementation

of CiF-Bannerlord had every sentence being calculated dynamically on each SE, however, due to the

way Bannerlord’s engine was built this was not possible. In Bannerlord the dialog engine was built with

several attributes and with a token-based state tree.

Table 4.5: Base game Dialog Lines Attributes

Attribute Description

ID Identiﬁer of the dialog line

Input Token state where this line is inserted

Output Token state for the following dialog line

Text Actual sentence of the dialog line

Condition Function that determines the validity of this line

Consequence Fallout and consequences of the dialog line

Value order value

The dialog system from the base game contains Consequences that are the equivalent of Trigger

rules in the CiF model and Conditions which are used in the CiF model as well. This direct translation

allowed for an easy implementation of the CiF model in this system.

A – Token state tree - In this state dialog tree, dialog sentences are added previously, before the

interaction, typically, when loading the game. These sentences are then grouped by their Input Token

being able to be picked only when it applies to the situation and when their respective Condition returns

true.

Figure 4.4: Example graph for the token based dialog system in Bannerlord

At each state, there are two subgroups, Dialog Lines (to be used by the NPCs) and Player Lines (to

be used by the player). There needs to always be at least one dialog line on each state, where the game

then chooses it based on the available ones, where the highest valued one is picked. If there are any

Player Lines in a determined state, the player has a choice between the available lines ordered by their

value. Every dialog sequence starts with the base token/state which is named “start” and an available

dialog line given the circumstances. If the dialog reaches a state where there are no dialog lines, the

conversation ends.

Dialog lines in CiF-Bannerlord are instantiated in the behaviour Player Dialog Behaviour, further

detail on dialog between the player and NPCs is explained in section 5.1.3.

4.7.1 CiF-Bannerlord Dialog

There is a custom Dialog class in the CiF-Bannerlord engine.

Table 4.6: CiF-Bannerlord Dialog Attributes

Attribute Description

ID Identiﬁer of the dialog

Sentence Actual sentence of this dialog object

Sentence Type Type of sentence regarding the delicacy of the statement

Response Type Type of the response based on the Outcome

Intensity Number containing the intensity of the sentence

A – Response Type - has four values, NotResponse, Normal, Positive and Negative. This value is

used to differentiate response sentences from normal ones as well as the type of response based on

the outcome.

B – Sentence Type - has three distinct values, Normal, Cordial and Rude. These function as another

form of clariﬁcation and distinction between dialog lines. Most dialog lines fall in the normal category,

however, there are Cordial ones, for more wordy and complex sentences and Rude as the opposite.

The sentence type is determined by a value that is calculated randomly and applied an offset based

on the character’s personality. The value is then compared to a scale between 0 and 1 to select the

sentence type. Values below 0.15 are considered Rude, above 0.85 are considered Cordial and any

value in between is considered Normal. The formula 4.11 illustrates how the result value is calculated.

result = outcome + w

calc

.fn(t

calc

) + w

caref ul

.fn(t

caref ul

) + w

honor

.fn(t

honor

) + w

shy

.fn(t

shy

) (4.11)

outcome ∼ U (0, 1) (4.12)

There is a normalization function applied to each trait and w-values are the weights of each trait,

which were determined empirically. The outcome represented in the From the traits observed in the

equation above, the higher these are, the more probable Cordial sentence types are and vice versa.

The outcome represented in the equation 4.12, is the random number used to calculate the sentence

type, with an interval of values between 0 and 1. Calculating, careful, honor and shy were the traits

chosen since these are ones that usually make characters more careful about the things they would say.

C – Intensity - The Dialog class has a number indicator of its value. It is used to distinguish dialog

lines inside the SI, working as some sort of intensity. The higher the value, the less probable the dialog

line is to be chosen since it is more ”intense”.

When a character is deciding a dialog line, ﬁrst calculates the sentence type, and a list contained

all dialog lines with that type, ordered by their intensity is formed. To select a dialog line, a recursive

algorithm is used, that runs through the list. This algorithm has a chance to stop at any dialog line which

then is the selected one. This chance is a probability calculated from the personality set of the character

and the value of the determined SI. The personality set is used due to the character’s predisposition to

be more intense about what they are saying. The SI value is used since the characters can be more

or less sure of the interaction chosen. For example a high value for the ”Compliment” interaction could

be interpreted as very sure of the decision and therefore be more likely to choose a more ”intense”

compliment.

The probability of the algorithm stopping and selecting a dialog line is calculated once, which is then

used on all steps of the algorithm. Considering a list with several dialog lines, the algorithm will run a

random number and if this number is outside the range of the probability to stop, it will continue to the

next dialog line. For example, with a probability to stop of 0.4, the ﬁrst dialog line has a 0.4 chance to be

chosen, the second dialog line has a chance of 0.6 times 0.4, and so on. This probability is represented

by the c-value in the formula 4.13.

c = 1 − (f n(−t

shy

− t

sensitiv e

− t

caref ul

) ∗ f n(value

)) (4.13)

Both traits and the SI value are calculated in the formula 4.13 using a normalization function. The

traits chosen for this formula were, shy, sensitive and careful. Shy, sensitive or careful characters are

less likely to be excessive/intense.

Implementation

Contents

5.1 Behaviours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.2 Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

In this chapter we will examine the overall architecture of the implementation of CiF-Bannerlord as

well as design choices. The ﬁrst section is about the behaviours created and added to the game. In the

second section we will discuss the managers created, the design choices for this architecture and the

instantiation of some of the rules and SIs created in CiF-Bannerlord.

5.1 Behaviours

Mount and Blade II: Bannerlord has a large modding community and the game reﬂects it. The game has

several tools that help programmers add functions, one of these is their implementation of Behaviours.

Behaviour is a campaign speciﬁc abstract class from the base game that is designed for modding. This

class has two methods, Register Events which is used to register events from the game and create cus-

tom functions around it, and SyncData which is used as the logic behind saving and loading. Behaviours

are designed to be added in the game and are stored besides the base game’s behaviours. These can

be easily extended through a subclass and without having to extensively study the base game’s code.

It is easy to implement some changes, although, a more in-depth knowledge of the game’s code is

necessary in order to implement more complex changes.

5.1.1 SubModule

Behaviours are added through the class SubModule, which is basically the ”main” class for mods, in-

cluding CiF-Bannerlord’s mod. It functions as the bridge between CiF-Bannerlord and the base game.

SubModule is responsible for initializing all the behaviours. There are two types of Behaviours, from

campaigns and from mission. Campaign Behaviours are always active and function when the player is

in the overworld/campaign map and also inside a mission/instance. Mission Behaviours are only active

when inside an instance. Dialog Manager is an example of a Mission Behaviour.

SubModule is responsible for storing the highest level of most features in CiF-Bannerlord. One of

these is the Social Facts Database, which the SubModule is responsible for saving as it contains a

large part of the Social State of the world. It is also responsible for creating and storing the Managers

examined in the previous section.

SubModule contains an extension of a class, Character Relation Manager, from the base game.

This class is responsible for storing and the logic behind character relationships in the game, keeping a

value from -100 to 100. This extension is a CiF implementation on top of the character relation manager

also storing relationships, not only friendship but also romantic ones. It does not store Power relations

because these are one sided and only available for each individual character. Romantic and Friendship

relationships are public knowledge, therefore, different from the individual non-reciprocal feelings each

CiF-Character stores.

5.1.2 NPC Dialog Behaviour

NPC Dialog Behaviour is the class accountable for most of the logic behind the dialog between NPCs. It

instantiates Character Manager and works as a bridge between the CiF-Characters, the rest of the CiF

mod and the base game.

NPC Dialog Behaviour is the class that manages the CiF cycle between the characters. It raises

the character’s energy, when the event triggers from Dialog Manager. It asks the initiator for the intent,

receiver and SI.

This behaviour is also responsible for timing the characters responses and their content. When the

character picks a dialog line, this class sends it into the Dialog Manager to be shown and keeps a record

of the time. After a few seconds in-game have passed, the behaviour then calls the for the response.

NPC Dialog Behaviour is in charge of making characters move. When the initiator wants to move

closer to the receiver to start the Social Exchange, the initiator’s agent receives an agent behaviour from

the base game, Agent Follow Behaviour. This class also gives this behaviour to the receiver when they

want to respond.

5.1.3 Player Dialog Behaviour

Player Dialog Behaviour is responsible for the dialog between the player and the other characters. It

is an extensive class that adds a “Converse” option to every character speciﬁc dialog tree. It uses the

dialog system built in the base game. It also has the necessary Condition and Consequence Delegates

to make the new dialog option function.

Unfortunately, because of how the dialog system was implemented, dialog lines have to be added

before the game starts. Each Dialog line corresponds to a CiF-Bannerlord Dialog object. This object has

a ﬂag, active, that determines if the dialog line should be active or not. It is connected to the Condition

Delegate on each dialog line, so that if the ﬂag is active the dialog line should also be active. Dialog

Rules do not work on this dialog system, therefore some interactions have to be disabled since it was

not possible to create dialog dynamically in the middle of a conversation between the player and NPCs.

Player Dialog Behaviour is accountable for following the CiF cycle when an interaction is being per-

formed between the player and a character.

5.1.4 Rumor Behaviour

Rumor Behaviour is the class behind all the logic regarding rumors. The rumor system has a class,

Rumor Holder that holds a list of rumors for the character to choose the one they ﬁnd most interesting.

There are two types of Rumor Holders, Settlements and Parties. Rumor Behaviour has two Dictionaries,

that map the Settlements and Parties to their respective Rumor Holder counterparts.

Rumor Behaviour is responsible for creating and storing all rumors that have been created. It tracks

events such as battles, war and peace declarations, raids and other map events that are used to create

the rumors.

5.1.4.1 Saving

Since the saving system in Bannerlord is not robust for modding, Rumor Behaviour is responsible for

saving the Rumor Holders as well as the rumors they contain. The saving is performed with a recursive

method like the Character Manager. Each Dictionary is transformed into a string by asking each Rumor

Holder to return a string and saving it with the respective Settlement or Party reference. Rumor Holders

then ask each Rumor to return a string and it merges all Rumor’s strings. This is done recursively until

reaching the simplest type of data like strings or numbers.

5.1.5 Microtheories

The class responsible for Microtheories is a behaviour related to events happening in-game as refer-

enced in an earlier section 4.4. There are seven implemented Microtheories.

Three microtheories are related to ﬁrst impressions. The ﬁrst time the player enters an instance with

social characters, such as the tavern or the town center, these microtheories create ﬁrst impressions of

the other characters. There are three different ﬁrst impressions: LoveAtFirst, LooksCool and HateAtFirst.

LoveAtFirst is a ﬁrst impression where the character has a boost in the romantic feelings for the other

character, if it aligns to their preference. LooksCool is a ﬁrst impression where the character has a boost

in friendly feelings for the other. HateAtFirst is the opposite of LooksCool, where the character gains

a negative boost. This boost is calculated based on the personality set of each character, therefore, a

very romantic character will be very passionate and pursue strongly the other, for instance.

Every character is compared to another and randomly chooses either one of the three ﬁrst impres-

sions or no ﬁrst impression at all. This chance of picking one of the ﬁrst impressions is based on the

personality set of each character. For example, a mean character will be more likely to dislike other

characters at ﬁrst glance.

Defeat is a microtheory, which occurs when a sizeable battle is over. As the rumor of the battle

spreads to nearby settlements, the characters admire the victor, raising their notion in their respective

Power relation. This raises particularly in the battle’s losing side, while the winner will have a negative

notion of the loser in their Power relation.

Quest Complete is a microtheory executed after a quest is completed. It raises the friendship values

for each party involved as it happens in the base game. The Power relation is also raised towards the

player since they prove to be reliable, helping those in need.

Raid is a microtheory related to the in-game mechanic of raiding villages. This action is widely

frown upon by the characters affected in the raid. This microtheory raises the Power relation of nearby

characters but with negative intent, to promote fear and dislike of the raiding character and their actions.

Bored is a microtheory related to the status of characters. Each character has a timer that is reset

when starting an exchange. Although separate to the energy system, it complements it. After a certain

amount of time, this microtheory attributes the bored status to the character which boosts the desire to

perform exchanges. The amount of time required is dependant on the personality set.

5.2 Managers

CiF-Bannerlord uses an architecture design with managers in order to initialize the other classes.

5.2.1 Character Manager

The character Manager is responsible for everything related to characters. The main responsibility of

the character manager is mapping Characters from CiF-Bannerlord to their respective counterpart in the

base game, Heroes, Agents and Character Objects.

Table 5.1: Character Manager Attributes

Attribute Class

Characters Dictionary<Character Object,Character >

Agents Dictionary<Character, Tuple<Agent, Settlement>>

Tavern People Dictionary<Settlement,List<Character>>

Main Character Main Character, custom subclass to Character

Character Manager is responsible for creating CiF-Characters. It has two dictionaries that map CiF-

Characters. Characters dictionary has Character Objects as keys, and a CiF-Character as their value,

Agents dictionary has CiF-Characters as keys, certifying these objects as being unique, since every key

is unique in the dictionary and all characters added or created are added to this dictionary. Agents’

values are the respective CiF-Character agent as well as the settlement in the creation of the agent.

Since agents can be null when not in an instance, this implementation keeps track of what agents are

currently initiated or not as well as mapping the CiF-Character to an agent.

CiF-Characters are created when entering a new mission and Character Manager adds any new

CiF-Character to both dictionaries. CiF-Characters already instantiated in the dictionary will have their

corresponding agent updated to the new object. Character Objects are not unique in the Bannerlord

engine, therefore when entering a mission, the Character Object townsman for example, will generate

several different agents. To keep track of this, the Character Manager checks whether the settlement

already has character entries in the Tavern People dictionary, and will attribute the CiF-Characters to

the new agents created. Townsman for example, have their personalities jump from an agent object to

another, based on the same Character Object between instances, but it is not lost.

The Main Character has a special corresponding CiF-Character, a subclass called Main Character.

The creation of the CiF-Character is different, its personality attribute does not create

Character Manager is also responsible for creating and storing the general rules that apply to every

character. There are two groups of general rules, ones for initiations and ones for responses. These are

a set of rules that apply every time in each situation. An example is ”Repetition”, that compares the last

interactions the character performed with the one being calculated, the more repeated entries found,

more negative the return value becomes. This rule is useful to block characters from choosing the same

interaction every time. There is also a ”Repetition” response rule, making the characters tired of hearing

the same interaction several times in a row.

5.2.1.1 Saving

Unfortunately, Bannerlord’s saving system is not robust enough to be used by mods. Character Manager

is responsible for saving every character and their connection to the game counterpart. Saving in this

class is performed in a recursive manner, every element is stored inside a string which then is coded

using Json serialize function. First Character Manager asks each CiF-Character to be returned as a

string, and CiF-Character then transforms into strings each of its attributes which are also complex data

structures which will ask recursively for strings until reaching a simple value.

5.2.2 Rule Managers

Most rules are created through a manager, this includes conditions, inﬂuence rules and trigger rules.

This is a very efﬁcient method of instantiating rules since the number of created instances can get very

large. Managers consist of a dictionary containing their respective instantiations.

Figure 5.1: Visual representation of the manager system in CiF-Bannerlord

As we can see in the example in ﬁgure 5.1, there are three Social Exchanges, which all performed

the same SI, and which itself has three different Inﬂuence Rules. In this system, there is only one of each

inﬂuence rule and only one Social Interaction 2. In a system without references, instead instantiating

every object, the situation represented in the image would have nine different Inﬂuence rule objects.

As the situation becomes more complex, it would grow exponentially in object complexity which would

render an inefﬁcient game and possibly unplayable. It is thereby important to create an efﬁcient method

of instantiating rules, such as the current one present in CiF-Bannerlord.

5.2.2.1 Condition Manager

The Condition Manager instantiates eight different conditions. These do not account for all conditions

created, since many are nameless and created inside another Rule or SI. Four examples of conditions

used in CiF-Bannerlord are:

A – Two Characters - This is a simple condition and applied to most rules. It requires tat both

characters exist.

B – Adults - This condition is applied to ﬂirting and dating SIs, that consists on romantic intent. It

requires to condition above to be true.

C – No Romance - This condition is contained of the Date SI, and it requires that both characters

are not in a relationship already in order to start a new relationship.

D – Is Bored - This is a condition made for a speciﬁc trigger rule. It requires the character to have

the Bored status.

5.2.2.2 Inﬂuence Rule Manager

There are thirty one unique different Inﬂuence Rules created by the Inﬂuence Rule Manager. Like

conditions, there are several nameless rules created inside SIs that are not accounted for in this list.

Four of the main Inﬂuence Rules used in CiF-Bannerlord are:

A – Repetition - This Inﬂuence Rule lowers the value of the interaction based on how many SEs

have been performed by the initiator with the type of SI being calculated, in the last ﬁve interactions.

This rule prevents characters from performing the same interactions repeatedly by lowering their value.

B – Is Liked - This rule uses the belief system on the character to determine if it would have success

in performing the desired SI. It searches for the belief that the initiator has based on the social feeling

entry of the receiver with the initiator as the recipient of the feeling. The higher the social feeling, the

higher the value returned by this rule. These beliefs do not account as the actual social feeling the

receiver has and the initiator could have them wrong.

C – Hurtful - This Inﬂuence Rule is used for malicious SIs, like insults. It represents the desire

to be hurtful against other characters. It is calculated based on the statuses the character has. If it is

unbalanced, in the sense that it has more negative status like Angry, Ashamed or Sad than positive ones

like Happy or Conﬁdent, the rule returns a high value. The traits Honor and Careful from the personality

set also impacts the value returned.

D – Enjoy Listening - This rule is related to how much the character enjoys listening to others. It is

is only used as a Response Inﬂuence Rule in SIs. The value returned depends on the personality set.

The traits used to calculate it, are curious and kind.

5.2.2.3 TriggerRuleManager

There are only four different Trigger Rules.

A – Friendly and Romantic Exchange - Two of the trigger rules are related to the type of the ex-

change. Friendly Exchange updates the feelings and beliefs of each character involved based on the

outcome and intent. Romantic Exchanges is a similar rule, being executed instead of friendly exchange,

when the exchange is romantic.

B – Introduction - is a trigger rule that is only executed in the ﬁrst exchange between two characters.

It requires a condition, NotIntroduced. This rule updates the Introduction Relation Manager and has a

strong impact on each of the characters feelings, since it functions as a ﬁrst impression between the two

characters. This impact is dependant on each personality set.

C – UnBored - is a rule complementary to a microtheory that applies bored statuses to characters.

This rule removes that status from characters after they perform any social exchange, either as an

initiator or as a receiver.

5.2.2.4 Social Interaction Manager

The Social Interaction Manager instantiates twenty different SIs. As seen in a previous section 4.5.1,

these interactions include rules and conditions instantiated inside the other managers. Four of the main

SIs used in CiF-Bannerlord are:

A – Complain - Characters using this SI, are complaining to the other about something. It uses Low

Honor, Likes To Speak and Enjoy Speaking inﬂuence rules for the initiator and Low Honor and Enjoy

Listening for the receiver. The dialog in this interaction is broad, in which he characters complain about

not ordinary things, like the weather.

B – Insult - This SI is used by characters that want to be mean to others. The Inﬂuence Rules used

by this interaction are Hurtful, Lower Relation, Negative and Dislikes Them for the initiator. The receiver

has the same inﬂuence rules plus Annoyed. The dialog consists on general insults in order to make the

other character uncomfortable.

C – Flirt - This SI is performed with the aim to improve romantic relations from the initiator. It has

Romantic, Likes Them and Is Liked as Inﬂuence rules for the initiator, while the receiver has the same

rules plus Charming. The dialog in this SI are simple ﬂirting sentences with the purpose to make the

other one ﬂushed.

D – Bad Mouth - This SI is talking badly about a character not present in the conversation. Char-

acters that perform this usually have low honor. The inﬂuence rules used in this SI are Low Honor, Not

Good and Negative. They talk about a character they do not like on their friendly feelings, which the

dialog lines are constructed with dialog rules on this SI. This interaction is also exclusive to the Empire

culture.

5.2.2.5 Dialog Manager

Dialog is instantiated in behaviours, explained on the next section 5.1.3. Although Dialog Manager does

not instantiate dialog, it is responsible for making dialog appear when inside missions/instances.

Dialog Manager is a subclass of Mission View, a Bannerlord class that has access to some aspects

of the mission/instance the player is currently in. One of these is controlling the time in the mission,

since it extends a function called Mission Tick, that is called every tick (depending on the frame rate

the game is operating on). Through this function, Dialog Manager is responsible for timing and raising

energy. There is a timer that operates each tenth of a second, that returns true to call a function to raise

energy of all available characters.

Dialog Manager has a list of Targets. Target is a Bannerlord class that is part of a feature to ﬁnd

characters inside missions/instances. By pressing the Alt key, it shows on screen the distance and

name of notables as well as exits and objects relevant to the mission. This feature was re-purposed

in CiF-Bannerlord to replace the name above the agent, as their dialog. This was possible since every

character represented is an agent in the game’s engine, not only the notables which could have their

name appear.

Evaluation

Contents

6.1 New Lords Mod . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

6.2 User Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.3 Results Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

In this chapter, we will discuss two evaluation methods regarding CiF-Bannerlord. The ﬁrst method

and section is an online mod created for Bannerlord. The second method and section of this chapter

is related to user tests. The objective is to prove, CiF-Bannerlord improves the experience of the play-

ers, through its characters and social environments. This improvement was measured through player

Immersion and their perception of NPC Believability.

The mod was created and published with the objective to understand the user’s enjoyment and

the impact this mod had on their game experience. This study is helpful since it targets players that

already have the game and have spent some time playing. Users usually search for mods after being

satisﬁed with the core experience the game has to offer. Therefore users trying the mod containing

CiF-Bannerlord are mostly experienced in the game and can return more accurate feedback regarding

the impact and differences in the mod.

Additional user tests were performed with the objective to measure the user’s Immersion and NPC

believability in social environments. These tests are performed in a controlled environment, where users

all perform the same tasks. This situation results in more statistically accurate results since their opinion

will be based on the same environment and time spent performing the tasks. Unfortunately, the user

tests do not give the same insight into the impact of the modiﬁcations, since users are not experienced

in Bannerlord like the mod users.

6.1 New Lords Mod

CiF-Bannerlord modiﬁcations were published as a mod for the game Mount and Blade II: Bannerlord

on the 31st of August of 2021. The Mod was titled ”New Lords”

and it was released on the site Nexus

Mods

. The mod was released while in development, therefore, users were asked to use the mod

Better Exception Window, to provide a clear description of errors that might occur. With this mod, it

was possible to swiftly address the bugs. The mod New Lords had continuous support until the end of

October 2021.

In Nexus Mods there are endorsements which are basically likes people award after downloading the

mod if they enjoyed it. Nexus mods also tracks not only downloads but also unique downloads, being

possible through nexus mods accounts, since there needs to be an account in order to download it.

Figure 6.1: New Lords mod data as of 20th of October 2021

New Lords Mod, www.nexusmods.com/mountandblade2bannerlord/mods/3272

Nexus Mods, www.nexusmods.com

In the ﬁgure above we observe mod data until the 20th of October 2021. It acquired 18 endorsements,

1,199 downloads and 12,105 views on the mod page. From the downloads, 874 are unique downloads,

from different nexus mods users.

6.1.1 Online Feedback

In order to consider the impact of the mod New Lords a questionnaire was created. This questionnaire

aimed to measure NPC believability and player immersion of users with a good understanding of the

game. Though people that search for mods are somewhat biased since they would only look for and in-

stall something they ﬁnd desirable or missing in their game. Mod users are not an accurate portrait of the

seasoned player base. The questionnaire was available in the mod page and several announcements

were made in an attempt to get users to respond to it.

In the time period between the launch of the mod, in the 31st of August, up to the 20th of October,

the questionnaire only obtained ﬁve responses. Meaning the questionnaire did not accrue a reasonable

amount of responses required to perform any statistical evaluation on the data. However, there has been

some feedback in the mod page. In Nexus Mods, it is possible to leave comments/posts and to report

bugs. There have been several posts on the New Lords page, a total of forty three comments and four

formal bug reports. There were a few the posts related to bugs, crashes and errors.

Direct feedback related to the functionality and features of New Lords was overwhelmingly positive.

There was not been a single negative comment, while most of the positive ones conveyed that would

have liked this mod to be a core feature of the game. Users also conveyed ”these changes had made

the game feel more active and homey” and ”feels like really ﬁlling in the space”.

These comments convey that, although there were some issues, most users that commented on the

page, enjoyed the modiﬁcations. However, since the questionnaire did not contain a minimum number

of responses for statistical analysis, no conclusive result could be deducted.

6.2 User Tests

User tests were made in order to perceive the success of these modiﬁcations. The metrics aimed to

improve from this social system are:

• Player immersion

• NPC Believability

All user tests were performed in IST, Instituto Superior T

ecnico, at the Taguspark campus. All of the

participants were college students and young adults. A total of thirty one students participated in the

tests. After ﬁnishing the user test, each participant received a 5C gift card on the Steam platform

, as a

reward for participating in this experiment. The user’s data is anonymous and participants had to sign a

consent form to participate in the tests.

6.2.1 Instruments

Users had to answer a prepared questionnaire and perform some tasks in-game in a controlled scenario.

In order to have a more controlled playing experience, a save data was prepared so that all participants

would start from the same exact condition. The save had the main character lead a basic party with a

level one clan. It contained 20 troops so that the participants were not attacked by bandits while testing.

The party was left at a location close to the settlements needed to visit. The entire procedure took

around twenty to thirty minutes per participant.

The questionnaire used in these tests was created and answered digitally using Google Forms, which

provides a simple, fast and effective platform to create forms or questionnaires.

The questions used in the questionnaire were crafted speciﬁcally for these tests. Although already

established questions used in metrics like enjoyment, ﬂow, presence and social believability [16] were

studied in the making of this questionnaire, these were not found to be accurate matches to the testing

situation. Therefore some questions were adapted from existing certiﬁed questionnaires, while others

were created from scratch in order to measure the determined metrics with as much accuracy as possi-

ble.

A – Demographic data - consists of collecting the age group, gender, general games playtime as

hours per week and the user’s afﬁnity for social aspects of games. Playtime can give an insight into

some problems occurring when analyzing the data. A population not used to playing games could skew

data since they would not be as accustomed to the testing environment as desirable. Users with very

low afﬁnity for the social aspect of games could also skew the data since they would be uninterested in

the testing environment present in these user tests. One of the questions to test this problem, was if the

user usually ﬁnds characters they dislike, in order to test their usual emotional attachment. The other

question targeted to this problem was if the user usually skips dialog, to test if they are uninterested in

the dialog or get bored easily in this type of tasks.

B – Player Immersion - is one of the main metrics used to test the success of CiF-Bannerlord.

Questions in this metric were related to the social environment encountered in the test and how well the

characters developed ﬁt into the game or activities performed. Two questions were related to the social

environment, which was the tavern inside a settlement. One question asked if they found the tavern to

Steam, https://store.steampowered.com/

be immersive in the sense of enjoying and relating to the tavern environment to a real one. Another

question was if the users found the tavern interesting in the sense of having several occurrences and

interactions. The other two questions were related to the characters. One of them asked users if they

would like to meet new characters, to test if they were interesting enough. The last question asked in

this metric was if they think the personalities in NPCs ﬁt in the game, testing if they are unusual and

break the player’s immersion.

C – NPC Believability - is the other main metric used to test the success of CiF-Bannerlord. Ques-

tions in this metric were used to test the user’s perception of the characters, how they behaved and

communicated. Users were asked if it was easy to manipulate conversations, in order to determine if

the characters offer meaningful conversations and are believable in their interactions. Another question

asked was if the users identiﬁed with the characters, proving if CiF-Bannerlord NPCs are believable

enough to make users relate to their actions or reactions. Users were also asked if they felt like the

characters had emotions. Testing if their personality was believable enough to convey emotions.

Four closed questions were used in the questionnaire to test the user’s perception of the NPCs

personalities and the user’s attachment to the characters in-game. The ﬁrst two questions were, if they

found someone nice/friendly and if they found someone annoying. These two traits should be relatively

easy to track in CiF-Bannerlord, based on the interactions between the player and character and their

frequency. The last two questions asked in this metric were if the users felt like they made a friend,

or if they felt like they made a character dislike them. These questions were aimed to test the user’s

attachment to the connection and relations created when performing the study.

D – Interaction Enjoyment - Another metric used in this study was enjoyment related to the inter-

actions performed. The main objective was to improve the users’ experience, through better Social AI.

For this purpose interactions with NPCs should be enjoyable. There was one question in this metric that

asked if users would like to perform more interactions with characters, like the ones they did in the test.

Figure 6.2: Questions present in the questionnaire and their respective metrics

The table 6.2 illustrates all the questions asked in the questionnaire and maps them to their respective

explanation, labels and study metric. The labels help to identify the questions in the next graphs, tables

and respective explanations.

Due to the structure of the tests, it is necessary to separate data into two versions. Version A

corresponds to the task with the CiF-Bannerlord modiﬁcations. Version B corresponds to the task in

the base game, without the CiF-Bannerlord modiﬁcations. Every question, except proﬁling ones, is

present in each part of the questionnaire related to the tasks. Therefore there are two versions, A and

B, for each question.

6.2.2 Procedure

The procedure has ﬁve steps, three of each are answering the questionnaire and two tasks playing

the game. Participants alternate between ﬁlling in the questionnaire and playing. The ﬁrst step, from

the questionnaire, is related to the proﬁling questions. Afterwards the participant starts playing the

game and performing a task. At the end of each task, the participant answers a set of questions in the

questionnaire related to the task.

Figure 6.3: Graph detailing the procedure

One of the tasks is performed in the base game without the CiF modiﬁcations/New Lords mod (Ver-

sion B) and the other one is a similar task but with the modiﬁcations of CiF-Bannerlord (Version A).

Participants were asked to travel to a nearby tavern. Upon arriving at the tavern they are tasked to

observe the characters and their behaviour. Afterwards, players need to speak to at least three different

characters inside the tavern. After ﬁnishing the conversations, participants are asked to repeat these

sub-tasks, in another tavern, close by to the one they are in. This tavern has a different culture than the

previous one. These simple objectives conclude the ﬁrst task, players are asked to close the game and

answer the questionnaire’s part of the ﬁrst task.

The other task consists of closing the game and booting it with the mod New Lords enabled. The

second task is very similar to the ﬁrst task, besides requiring CiF-Bannerlord modiﬁcations, participants

are also asked to converse at least three times with the same character, instead of just different ones.

This difference is due to the lack of conversation choices within the base game, it was not possible to

ask the subtask of conversing at least three times with the same character in the ﬁrst task. Usually, in

the user tests, participants would speak with several characters, most of them would, in fact, converse

with more than ten different characters in the second task.

Due to this procedure having repeated measures, two versions of the procedure were created with

the tasks being performed with different orders.

6.2.3 Issues

This procedure for the user tests results in a relatively quick test, averaging about twenty to thirty minutes

per participant. The procedure was aimed to test the NPC Believability and player immersion, however,

it was not possible to include testing for one of CiF-Bannerlord’s main features, rumors. This feature is

most notable and impactful to users playing over a relatively large amount of time. The gossip created

around the player’s or other characters’ actions, notions of battles and wars are all elements connected

to longer playtime which cannot be achieved through the short gameplay in these user tests.

One of the main issues with these user tests was that each test could develop different results

since the participants were not given instructions on speciﬁc characters to talk to. This problem arises

with the nature of the New Lords mod itself, since the load of the tavern (or any social environment)

is procedurally generated, where it is impossible to foresee where each character/personality is going

to be. Forcing participants to converse with speciﬁc characters would also hinder their immersion by

removing their choice to interact with characters of their choosing.

6.2.4 Data Analysis

In order to simplify the analysis process, a cronbach’s alpha test was performed, grouping questions

with their respective metric. Only scale-type questions were considered for this test.

Table 6.1: Metrics and their respective cronbach’s alpha

A B

Immersion 0.675 0.903

Believability 0.434 0.902

Considering 0.7 as the minimum value for accepted to aggregate questions, we can conclude based

on the results displayed in the table above, that these questions should not be grouped as a single vari-

able. Although questions in version B both represent excellent cronbach’s alpha values, these cannot be

grouped since they will be directly analyzed with the version A questions. In version A Immersion ques-

tion group the value is close to the minimum, however, it was not possible to raise it further. Removing

any question from this group would lower the value. With this information in mind, the questions will be

evaluated individually.

A shapiro-wilk normality test was applied to determine if the data was parametric. The results showed

that the data was not parametric, therefore the non-parametric test selected to perform statistical anal-

ysis was the Wilcoxon signed-rank test. The chosen test for closed questions was the chi-square for

association test.

6.2.4.1 Demographic Data

There were a total of thirty one participants, were 97% where male and 3% were female. Since the

participants were college students, the entire population of this study is contained in the 18 to 25 age

group. In this set there were no participants with prior knowledge or experience in the Bannerlord base

game, meaning this test was their ﬁrst impression of the game.

Figure 6.4: Playtime in hours of gaming per week intervals and their count

The graph above represents the participants’ dedicated playtime hours per week. We can observe a

balanced set of participants, where the majority is somewhat comfortable with games.

Figure 6.5: Usual emotional attachment to characters in-game

The graph above represents how participants usually view characters and how attached they can

become to them. By usually ﬁnding characters the user dislikes we can conclude they are usually

minimally attached emotionally to games. A population that is normally apathetic to this type of situations

could skew data, since they would be mostly uninterested in the testing environment. Fortunately, as we

can observe in the graph this is not the case for this study.

Figure 6.6: Likelihood to skip dialog

The graph above represents the frequency in which participants skip dialog. Similar to the previous

question, a population that usually skips dialog could skew data. Luckily, this set does not seem to have

a large proportion of possibly disinterested participants.

6.2.4.2 Player Immersion

In order to perceive the difference between the versions A and B questions regarding Player Immersion,

a boxplot was created using median values. The questions represented in the graph were Likert scale

questions with values between 1 and 7, where 1 meant the participants disagreed completely with the

statement while 7 meant the participants agreed completely.

Figure 6.7: Boxplot illustrating immersion question comparison based on version

The graph demonstrates that version A questions have consistently higher median values than ver-

sion B. It is positive to have higher median and quartiles values since for example, in the Immersive

Tavern, the higher the value the more participants agree that the tavern was immersive.

We can also see version A questions have their quartiles close to the median value, demonstrating

a concise set. The ends of the box are close as well, except in the question Meet Character version A

where it demonstrates a more distributed set of results than the others.

Table 6.2: Wilcoxon signed-rank test with their respective Z-values, p-values and r-values

Interest Tavern Fitting Personalities Meet Character Immersive Tavern

z-value 3.833 2.960 2.388 4.145

p-value 0.000 0.003 0.017 0.000

r-value 0.688 0.531 0.429 0.744

The p-value represented in the table indicates whether or not the results have statistical signiﬁcance.

Values below 5% or 0.05 are considered positive, which prove signiﬁcance. As we can observe in

the table, all of the questions present an p-value below 0.05, meaning all of them are statistically

signiﬁcant. The effect size, r-value, was calculated in order to test the impact of this signiﬁcance.

r =

√

(6.1)

This formula presents r, the effect size, where z is the z-value obtained from the Wilcoxon signed-

rank test and N is the number of samples, that is the number of participants. N, in this case, is 31.

The interval to consider the size of r-value is 0.1 to 0.3 for a small effect, 0.3 to 0.5 for a moderate

effect and larger than 0.5 for a large effect. Considering this and the results presented in the table for

the r-value, we can conclude most questions have a large effect size. The question Meet Character

has a medium, but close to strong effect.

Based on these tests, we can conclude CiF-Bannerlord was successful in improving the Player

Immersion.

6.2.4.3 NPC Believability

The NPC Believability has two sections in data analysis, the ﬁrst regarding Likert scale questions and

the second regarding closed questions.

The question Manipulate Conversation in particular had its results inverted into Neg-Manipulate Con-

versation. Since the question was formulated in a negative sense where the higher the value, the more

the participants agreed that it was easy to manipulate the characters. Although it is positive to have an

environment where it is possible to manipulate characters and conversations, it is not ideal to be easy

to do it. Characters very easily manipulated break their believability as emotional and thinking entities.

Since the ease for manipulation is considered a negative value, and in order to maintain the consistency

where it is positive having higher values, the question was inverted.

Figure 6.8: Boxplot illustrating believability question comparison based on version

The graph above demonstrates that participants found it was much easier to manipulate conver-

sations in CiF-Bannerlord than the base game since the value for version A is signiﬁcantly lower than

version B, which is a negative value for the study. We can observe that in both versions the data is

dispersed.

Regarding the questions Have Emotion and Identify with Character, we can conclude the mod has

made a positive impact. Both boxplots represent much higher values for version A than version B. In

the graph we can also observe a concise set of data regarding version A questions. Version B presents

a more dispersed set, but nevertheless signiﬁcantly lower values.

Table 6.3: Wilcoxon signed-rank test with their respective Z-values, p-values and r-values

Have Emotion Identify with Character Neg-Manipulate Conversation

z-value 4.730 3.745 -4.065

p-value 0.000 0.000 0.000

r-value 0.850 0.672 -0.730

In the table above we can see that all p-values are 0 and therefore below 5%, which means all

questions are statistically signiﬁcant. Given that the questions have signiﬁcance, the next step is

ﬁguring the effect size of it. Using the formula 6.1, the r-values were calculated. All questions contain a

large effect size. As we distinguished previously, Neg-Manipulate Conversation presents lower values

in version A. The z-value of this question is negative because it was calculated based on the negative

ranks instead of the positive as other questions. Based on the signiﬁcance and large effect size, we can

conclude this question did indeed players feel conversations were easier to manipulate. This occurrence

might be due to the fact that participants did not ﬁnd any possibility to manipulate without the mod, since

the content in conversations is much more limited.

In the graph 6.8 we can observe that the Have Emotion question version B has an extremely dis-

perse and low set of values. Comparing it to version A in this question we can see one of the largest

differences accounting for all box plots in this study. Pairing this information with the largest effect size,

we can conclude CiF-Bannerlord seems to have the most impact in improving the emotional perception

of players.

The purpose of the four closed questions is also to test NPC Believability, but more speciﬁcally, to

account for the participant’s perception of the characters’ personality set and of the emotional attachment

developed with characters.

Bar graphs were developed to demonstrate the difference between versions. Blue bars mark version

A, while red mars version B. The questions Nice and Annoying are related to the perception of the

personality and below are presented the respective graphs.

Figure 6.9: Count of the Nice question

Figure 6.10: Count of the Annoying question

As we can observe in both questions, version A questions have a substantially higher percentage of

ﬁnding friendly/nice and annoying characters than version B. Because of this disparity in the values

analysed in the two graphs above, we can conclude the mod had a positive impact in the user’s

perception of the characters personality. In order to test the signiﬁcance of these values, a chi-square

for association test was made. Both Nice and Annoying questions have a p-value of 0.000. Since the

value is below 0.05 or 5% we can conclude these results are statistically signiﬁcant.

Figure 6.11: Count of the Friend question

Figure 6.12: Count of the Enemy question

Similar to previous closed questions, in the graphs above we can clearly see version A questions

have substantially higher percentage of participants that feel like they made a friend or an enemy in

CiF-Bannerlord version B. Considering this, we can conﬁrm the mod has a strong positive impact in

the user’s emotional attachment to characters in-game. The chi-square for association tests present

p-values of 0.000 in both Friend and Enemy questions. Since the value is below 0.05 we can conclude

these questions are statistically signiﬁcant.

Considering the scale and closed questions both present positive results, with signiﬁcant statisti-

cal meaning and respective large effect size, we can conclude CiF-Bannerlord was successful at im-

proving NPC Believability when compared to the original game. Excluding the question Manipulate

Conversation, since it has contrary values and they might be due to the fact as conversations became

complex and with enough content to manipulate. Nevertheless, it seems conversations might be too

easily manipulated which can be considered a ﬂaw in CiF-Bannerlord.

6.2.4.4 Interaction Enjoyment

Although it was not one of the main metrics, in the questionnaire there was a question measuring the

enjoyment participants had with the interactions in the New Lords mod.

Figure 6.13: Boxplot with enjoyment difference regarding version

Once more it is displayed in the graph, version A with substantially higher values than version B.

This translates to a positive impact in the user enjoyment. Excluding some outliers in version A, we

can see it has a concentrated set of data, contrary to version B with a more dispersed set. In order to

test the signiﬁcance of these results, the Wilcoxon signed-rank test showed a z-value of 3.962 and a

p-value of 0.000. We can conclude that this question is statistically signiﬁcant. The effect size was

calculated using the formula 6.1, which returned a r-value of 0.712. Considering these values we can

conclude this data has a large effect size and it is statistically signiﬁcant. Although we can obverse

that CiF-Bannerlord has provided more enjoyable interactions with NPCs, there was only one question

related to measuring enjoyment therefore, conclusions in this metric are limited.

6.2.5 Participant Feedback

At the end of the procedure, some of the participants decided to add some informal feedback regarding

the experiment.

A – Relation progress indication - Some participants conveyed they wanted some sort of progress

bar or indication of the relation between the player and other characters. The base game already has

this mechanic, with a number indicating the relation between the player and notables or lords. However

other characters do not, like the townsmen, CiF-Bannerlord allows these relationships to exist with every

other character. The progress indication translates to some lack of immersion regarding the emotional

state of the characters. Although a solid and desirable gameplay feature, this would not reﬂect realistic

personalities, since people do not have a progress bar and would instead rely on conveying characters

as game tools/pieces.

B – Impactful Relations - Some participants mentioned they wanted relationships to have more

impact on the gameplay aspect. They felt the relationship they were forming was useless since they

would not be gaining or achieving anything with it. Once more demonstrates a great gameplay feature,

that would possibly convey characters as tools to gain bonuses. Participants gave examples such as

being possible to recruit units at cheaper (or steeper) prices, gain promotions or punishments such as

more or fewer quests and receive items.

6.3 Results Conclusion

There were two different methods of evaluating the success of the CiF-Bannerlord modiﬁcations, an

online mod and user tests. The mod was available to anyone willing to install and play. Mod users were

asked to answer a questionnaire that, unfortunately, did not reach an appropriate number of responses

to have statistical signiﬁcance. However, the online feedback provided by these users was positive.

The second method was user tests that accrued a total of thirty one participants. The purpose of

these tests was to test the impact on player Immersion and NPC Believability.

Regarding the player Immersion metric, there were four questions present in the questionnaire that

tested it. Considering the all questions present statistical signiﬁcance with large effect size and consider-

ing the results presented in the ﬁgure 6.7, we concluded CiF-Bannerlord was successful in improving

player Immersion.

Regarding the NPC Believability metric, there were two sets of questions. There were three scale-

type questions that tested this metric. The 6.8 illustrates two of the questions as having a positive

impact and the question Neg-Manipulate Conversation with negative impact. These questions presented

statistical signiﬁcance with large effect sizes. We deduced the negative result as the fact users could

not manipulate simple conversations, and the CiF-Bannerlord complexity made this feature possible.

However, based on these three questions, we concluded an overall positive impact on NPC Believability.

The closed questions present statistical signiﬁcance and demonstrate a clear improvement based on

the results analyzed. Considering all the questions in this metric, we concluded the CiF-Bannerlord was

successful in improving NPC Believability.

With only one question to test the Interaction Enjoyment, with statistical signiﬁcance and respective

large effect size, we concluded this metric to have improved Interaction Enjoyment. However, this

metric presents limited conclusions due to the low number of questions.

Based on all of the results studied in section 6.2, we can conclude that CiF-Bannerlord was success-

ful in creating a set of NPCs that provide satisfactory social interactions.

Conclusion

Contents

7.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

CiF-Bannerlord was developed with the objective to create a set of believable Non-Playable Char-

acters in the game Mount and Blade II: Bannerlord in order to improve the immersion and provide

satisfactory social interactions for the player.

CiF-Bannerlord has brought changes to the base game through:

• Characters - The NPCs gained unique personality sets that allowed them to stand out from each

other. They gained statuses, memory and feelings that allowed them to have volitions and de-

sires. Characters have cultural differences, based on the base game’s cultures, and gain speciﬁc

attributes to be distinct from one another. These features make NPCs believable to players.

• Conversations - Players can interact with NPCs, with more engaging conversations. Besides

all the options contained in the base game, CiF-Bannerlord offers new options for dialog, being

possible to perform new interactions with all characters.

• Environments - After entering a social environment like the tavern, players can observe the NPCs

interact with each other. CiF-Bannerlord transforms these environments to become more lively

places with NPCs performing actions independent from the player.

• Reactive World - CiF-Bannerlord helps to make the world a reactive place since characters talk

about recent events. Interactions with or between NPCs have consequences and can generate

gossip. Players enter a world that is independent of them, where characters react to news.

Considering the Online feedback and the results from User tests we can conclude CiF-Bannerlord

modiﬁcations were successful in improving users Immersion and NPC Believability.

CiF-Bannerlord is a CiF model inspired in the previous iterations, CiF-CK and CiF-Ex, and it was

developed speciﬁcally into the Mount and Blade II: Bannerlord. However, this model can be applied to

other games in an attempt to improve the player’s Immersion and NPC Believability.

7.1 Future Work

Although CiF-Bannerlord is considered a successful implementation of a Social Artiﬁcial Intelligence

architecture, it is an implementation with room to grow and become more robust.

A – Rules and Interactions - The CiF-Bannerlord engine was built to be extensible and efﬁcient. It

is easy to create and add any kind of simple rules, interactions or dialog within the code of the project.

It is also straightforward to add more complex rules or interactions, depending only on the complexity

of the rule itself. Adding more content has the problem of authoring every step, and it also requires to

be added within the code of the project. A signiﬁcant improvement on this system would be to create a

system that reads and stores these rules and interactions in conﬁguration ﬁles. This could provide users

or modders with a way to upgrade the model without the need to modify the base code.

B – Dialog system - Similar to the previous point of rules and interactions, the CiF-Bannerlord en-

gine was built to be extensible and efﬁcient regarding dialog. The same easiness to add dialog lines

or dialog rules can be seen in the system. However, even more important than the previous point, a

signiﬁcant improvement would be to create a system that stores the dialog in conﬁguration/dialog ﬁles.

Dialog was written in English in CiF-Bannerlord, and the only possible method of changing it is through

the code ﬁles. With dialog contained in ﬁles, it would give access to users and modders to change, add

or replace dialog, as well as offering the possibility of different languages support.

C – Rumors/Lies - A scrapped feature from CiF-Bannerlord was Lies. These were supposed to be

rumors that did not happen, NPCs would invent these to generate confusion, to lower or raise the stature

of other characters or just to try to seem interesting. This feature could introduce a layer of immersion

where the player would have to ﬁgure out what rumors are true or not. Maybe it would be possible for

the player to generate rumors themselves, in order to manipulate situations as they see ﬁt.

D – More testing More tests would prove useful to better understand the impact of CiF-Bannerlord.

Utilizing users that are experienced with the base game could tell us more accurate results. These users

could try the modiﬁcations for longer periods of time, having a greater understanding of how repetitive

or interesting the changes are. With experienced users, we could also test the impact of rumors or the

cultural differences implemented, that could not be tested using other users.

Bibliography

[1] Arakji, R.Y., Lang, K.R.: Digital consumer networks and producer-consumer collaboration: inno-

vation and product development in the video game industry. Journal of Management Information

Systems 24(2), 195–219 (2007)

[2] Bogdanovych, A., Trescak, T., Simoff, S.: What makes virtual agents believable? Connection Sci-

ence 28(1), 83–108 (2016)

[3] Guimaraes, M., Santos, P., Jhala, A.: Cif-ck: An architecture for social npcs in commercial games.

In: 2017 IEEE Conference on Computational Intelligence and Games (CIG). pp. 126–133. IEEE

(2017)

[4] Guimar

aes, M., Santos, P., Jhala, A.: Prom week meets skyrim. In: AAMAS. pp. 1790–1792 (2017)

[5] Jennett, C., Cox, A.L., Cairns, P., Dhoparee, S., Epps, A., Tijs, T., Walton, A.: Measuring and

deﬁning the experience of immersion in games. International journal of human-computer studies

66(9), 641–661 (2008)

[6] Marchand, A., Hennig-Thurau, T.: Value creation in the video game industry: Industry economics,

consumer beneﬁts, and research opportunities. Journal of interactive marketing 27(3), 141–157

(2013)

[7] Mascarenhas, S., Guimar

aes, M., Prada, R., Dias, J., Santos, P.A., Star, K., Hirsh, B., Spice, E.,

Kommeren, R.: A virtual agent toolkit for serious games developers. In: 2018 IEEE Conference on

Computational Intelligence and Games (CIG). pp. 1–7. IEEE (2018)

[8] Mateas, M., Stern, A.: A behavior language for story-based believable agents. IEEE Intelligent

Systems 17(4), 39–47 (2002)

[9] Mateas, M., Stern, A.: Fac¸ade: An experiment in building a fully-realized interactive drama. In:

Game developers conference. vol. 2, pp. 4–8 (2003)

[10] McCoy, J., Treanor, M., Samuel, B., Reed, A., Mateas, M., Wardrip-Fruin, N.: Prom week: Designing

past the game/story dilemma. In: FDG (2013)

[11] McCoy, J., Treanor, M., Samuel, B., Reed, A.A., Mateas, M., Wardrip-Fruin, N.: Social story worlds

with comme il faut. IEEE Transactions on Computational Intelligence and AI in Games 6(2), 97–112

(2014). https://doi.org/10.1109/TCIAIG.2014.2304692

[12] McCoy, J., Treanor, M., Samuel, B., Mateas, M., Wardrip-Fruin, N.: Prom week: social physics as

gameplay. In: Proceedings of the 6th International Conference on Foundations of Digital Games.

pp. 319–321 (2011)

[13] McCoy, J., Treanor, M., Samuel, B., Reed, A.A., Wardrip-Fruin, N., Mateas, M.: Prom week. In:

Proceedings of the International Conference on the Foundations of Digital Games. pp. 235–237

(2012)

[14] McCoy, J., Mateas, M., Wardrip-Fruin, N.: Comme il faut: A system for simulating social games

between autonomous characters (2009)

[15] Morais, L., Dias, J., Santos, P.A.: From caveman to gentleman: a cif-based social interaction model

applied to conan exiles. In: Proceedings of the 14th International Conference on the Foundations

of Digital Games. pp. 1–11 (2019)

[16] Nadolski, R., UOB, L.Y., Humphreys, S., Bazzanella, B., Mascarenhas, S.: Ms8–ﬁrst pilot validation

instruments. Integration 5, 09–11 (2016)

[17] Samuel, B., Reed, A.A., Maddaloni, P., Mateas, M., Wardrip-Fruin, N.: The ensemble engine:

Next-generation social physics. In: Proceedings of the Tenth International Conference on the Foun-

dations of Digital Games (FDG 2015). pp. 22–25 (2015)

[18] Scherer, K.R.: Appraisal theory (1999)

[19] Treanor, M., McCoy, J., Sullivan, A.: A framework for playable social dialogue. In: Twelfth Artiﬁcial

Intelligence and Interactive Digital Entertainment Conference (2016)

[20] Zhong, Z.J.: The effects of collective mmorpg (massively multiplayer online role-playing games)

play on gamers’ online and ofﬂine social capital. Computers in human behavior 27(6), 2352–2363

(2011)

Appendix

In this appendix, there is the material used for the user tests, which was performed in IST, Instituto

Superior T

ecnico, at the Taguspark campus. The questionnaire, consent form and procedure were

developed in Portuguese. The ﬁrst six pages are from the questionnaire, which was answered digitally.

The printing of the questionnaire pages raised some issues like not being able to see the full scale on

some questions. After the questionnaire, there is the consent form which is one page long. The last four

pages are from the procedure, where the ﬁrst two pages correspond to the Order BA (the ﬁrst task is

without the mod), and the last two pages correspond to the Order AB (the ﬁrst task is with the mod).

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

1/7

Informação Pessoal

Mark only one oval.

< 18 anos

18-25 anos

26-35 anos

36-50 anos

51-65 anos

> 65 anos

Mark only one oval.

Masculino

Feminino

Other

New Lords Mod

Obrigado pela sua participação nesta experiencia. Este questionário foi feito especificamente

para uma estudo sobre o impacto da Inteligência Artificial nos jogos.

Todas as respostas são anónimas e toda e qualquer data recebida será usada estritamente e

apenas neste projeto.

O questionário não deve demorar mais de 10 minutos para responder. Ao continuar está a

consentir que as suas respostas sejam recolhidas e analisadas.

Versão 1

*Required

Quantos anos têm? *

Qual é o seu género? *

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

2/7

Mark only one oval.

< 1 hora

1 - 5 horas

6 - 10 horas

11 - 20 horas

21 - 30 horas

> 30 horas

Mark only one oval per row.

Tarefa 1

As perguntas seguintes pertencem à tarefa 1. Quando estiver feita clique para continuar.

Bannerlord

Normalmente, dedica quantas horas por semana a jogar? *

Por favor escolha a resposta que se adecua mais relativamente às frases abaixo em relação

à sua experiência com jogos em geral. *

Discordo

totalmente

2 -

Discordo

3 -

Discordo

levemente

4 - Não

concordo

nem

discordo

5 -

Concordo

levemente

6 -

Concordo

7 -

Concordo

totalmente

Sempre que

posso, evito

diálogo

Encontro

sempre

personagens

que

desgosto

Sempre que

posso, evito

diálogo

Encontro

sempre

personagens

que

desgosto

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

3/7

Mark only one oval per row.

Por favor escolha a opção que se adecua melhor com a sua experiência em relação a ultima

tarefa que fez. *

Discordo

totalmente

2 -

Discordo

3 -

Discordo

levemente

4 - Não

concordo

nem

discordo

5 -

Concordo

levemente

6 -

Concordo

A taberna é um local

interessante e cheio de

vida

As personalidades das

personagens parecem

enquadrar bem no jogo

Estava ansioso para

conhecer novas

personagens

Senti que as

personagens tinham

emoções

Senti me imerso na

taberna

Senti que era fácil

manipular as

conversas/personagens

Gostaria de interagir

deste modo com mais

personagens

Consigo me identificar

com personagens

presentes

A taberna é um local

interessante e cheio de

vida

As personalidades das

personagens parecem

enquadrar bem no jogo

Estava ansioso para

conhecer novas

personagens

Senti que as

personagens tinham

emoções

Senti me imerso na

taberna

Senti que era fácil

manipular as

conversas/personagens

Gostaria de interagir

deste modo com mais

personagens

Consigo me identificar

com personagens

presentes

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

4/7

Mark only one oval.

Other:

Sim

Não

Mark only one oval.

Other:

Sim

Não

Mark only one oval.

Other:

Sim

Não

Mark only one oval.

Other:

Sim

Não

Tarefa 2

As perguntas seguintes pertencem à tarefa 2. Quando estiver feita clique para continuar.

Conseguiu encontrar alguém simpático? *

Conseguir encontrar alguém irritante?

Sentiu que fez algum amigo?

Sentiu que fez alguém não gostar da sua personagem?

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

5/7

New Lords Bannerlord

10.

Mark only one oval per row.

Por favor escolha a opção que se adecua melhor com a sua experiência em relação a ultima

tarefa que fez. *

Discordo

totalmente

2 -

Discordo

3 -

Discordo

levemente

4 - Não

concordo

nem

discordo

5 -

Concordo

levemente

6 -

Concordo

A taberna é um local

interessante e cheio de

vida

As personalidades das

personagens parecem

enquadrar bem no jogo

Estava ansioso para

conhecer novas

personagens

Senti que as

personagens tinham

emoções

Senti me imerso na

taberna

Senti que era fácil

manipular as

conversas/personagens

Gostaria de interagir

deste modo com mais

personagens

Consigo me identificar

com personagens

presentes

A taberna é um local

interessante e cheio de

vida

As personalidades das

personagens parecem

enquadrar bem no jogo

Estava ansioso para

conhecer novas

personagens

Senti que as

personagens tinham

emoções

Senti me imerso na

taberna

Senti que era fácil

manipular as

conversas/personagens

Gostaria de interagir

deste modo com mais

personagens

Consigo me identificar

com personagens

presentes

10/28/21, 1:10 AM

New Lords Mod

https://docs.google.com/forms/d/1sy04CLm-3n7UW6UssoJeZwy12Jjg93D4GwuO0bQAIbs/edit

6/7

11.

Mark only one oval.

Other:

Sim

Não

12.

Mark only one oval.

Other:

Sim

Não

13.

Mark only one oval.

Other:

Sim

Não

14.

Mark only one oval.

Other:

Sim

Não

Obrigado por participar.

Acabou tudo no teste. Muito obrigado pela sua contribuição.

Conseguiu encontrar alguém simpático? *

Conseguir encontrar alguém irritante?

Sentiu que fez algum amigo?

Sentiu que fez alguém não gostar da sua personagem?

New Lords Teste de utilizador

Investigador responsável:

Filipe Silveira

flipslv98@gmail.com

Caro/a participante,

Este teste de utilizadores pertence a um estudo que pretende analisar o impacto da

Inteligência Artificial nos jogos. Neste teste vai testar a Inteligência Artificial

adicionada pelo mod New Lords ao jogo Mount & Blade II: Bannerlord.

Nesta sessão terá de completar duas tarefa semelhantes e completar um questionário

em paralelo. Uma das tarefas é no jogo base e a outra tarefa é com o mod New Lords.

Cada tarefa terá uma duração de 5 a 10 min, o questionário terá também uma duração

de 5 a 10 min e a experiencia terá uma duração total entre 20 a 40 mins.

A qualquer momento poderá colocar qualquer dúvida que tenha. A sua participação

é voluntária e poderá desistir a qualquer momento sem qualquer penalização ou

consequência.

Para participar nesta experiência, pedimos-lhe que leia o consentimento informado e

caso concorde em participar de acordo com os termos abaixo, pedimos-lhe que

assine o formulário no local indicado.

Obrigado pela sua colaboração!

1 - Li e compreendi o significado deste estudo. Tive a oportunidade de colocar

questões, caso necessário, e recolher as respetivas respostas.

2 - Compreendo que a participação neste estudo é voluntária e que posso desistir a

qualquer momento, sem apresentar qualquer explicação. Caso tal aconteça, não serei

alvo de qualquer penalização e os dados relativos à minha experiência serão

removidos e destruídos.

3 - Autorizo o processamento dos dados no âmbito deste projeto para fins de análise,

investigação e disseminação de resultados em publicações científicas ou

conferências na área do projeto, pelos investigadores deste projeto.

4 - Compreendi que os dados recolhidos neste estudo serão utilizados como

mencionado anteriormente.

5- De acordo com o descrito acima, autorizo a minha participação neste estudo e

aceito as suas condições.

O/A participante Data

__________________________________ ___________________________

New Lords Teste de utilizador

Investigador responsável:

Filipe Silveira

flipslv98@gmail.com

1. Abra e faça o questionario até a parte da primeira tarefa.

https://docs.google.com/forms/d/e/1FAIpQLSe-

tsZm673uaY64altfWRrO2fSdlFB3NpSnP77ySJEFK0LMNg/viewform?usp=sf_link

2. Primeira tarefa (5-10 min)

2.1. Inicie o jogo e selecione “Continue Campaign”

2.2. Desloque-se até a cidade de Amprela e entre na taverna (Enter tavern district)

2.2.1. Observe as personagens e os seus comportamentos (1 or 2 min)

2.2.2. Fale com 3 personagens diferentes

2.3. Saia da taverna (Fique a pressionar Tab -> Leave tavern -> Leave)

2.4. Desloque-se até a cidade de Makeb (à direita de Amprela) e entre na taverna

2.4.1. Observe as personagens e os seus comportamentos (1 or 2 min)

2.4.2. Fale com 3 personagens diferentes

2.5. Feche o jogo sem gravar (Esc -> Exit to main menu -> Exit Game)

3. Acabou a tarefa, prossiga para o questionário até a parte da segunda tarefa.

4. Segunda Tarefa (10 min)

New Lords Teste de utilizador

Investigador responsável:

Filipe Silveira

flipslv98@gmail.com

4.1. Inicie o jogo com o mod New Lords ativado e selecione “Continue Campaign”

(antes de carregar no play precisa de ativar o mod na tab de mods)

4.2. Desloque-se até a cidade de Amprela e entre na taverna

4.2.1. Observe as personagens e os seus comportamentos (1 or 2 min)

4.2.2. Faça 3 interações de “Converse” com a mesma personagem (à sua

escolha)

4.2.3. Fale com 3 personagens diferentes

4.3. Desloque-se até a cidade de Makeb (à direita de Amprela) e entre na taverna

4.3.1. Observe as personagens e os seus comportamentos (1 or 2 min)

4.3.2. Faça 3 interações de “Converse” com a mesma personagem (à sua

escolha)

4.3.3. Fale com 3 personagens diferentes

5. Acabou a tarefa, prossiga para o questionário.

6. Acabou as tarefas. Se quiser pode continuar a conversar com as personagens

(maximo 10 minutos).

New Lords Teste de utilizador

Investigador responsável:

Filipe Silveira

flipslv98@gmail.com

1. Abra e faça o questionario até a parte da primeira tarefa.

https://docs.google.com/forms/d/e/1FAIpQLSdmjlVr0dXtzv_w6FAq8jM9qtWbkxLSE7

4lD2NhfFZZIA4pQg/viewform?usp=sf_link

2. Primeira Tarefa (10 min)

2.1. Inicie o jogo com o mod New Lords ativado e selecione “Continue Campaign”

quando entrar o menu principal

2.2. Desloque-se até a cidade de Amprela e entre na taverna

2.2.1. Observe as personagens e os seus comportamentos (1 or 2 min)

2.2.2. Faça 3 interações de “Converse” com a mesma personagem (à sua

escolha)

2.2.3. Fale com 3 personagens diferentes

2.3. Saia da taverna (Fique a pressionar no Tab -> Leave Tavern -> Leave)

New Lords Teste de utilizador

Investigador responsável:

Filipe Silveira

flipslv98@gmail.com

2.4. Desloque-se até a cidade de Makeb (à direita de Amprela) e entre na taverna

2.4.1. Observe as personagens e os seus comportamentos (1 or 2 min)

2.4.2. Faça 3 interações de “Converse” com a mesma personagem (à sua

escolha)

2.4.3. Fale com 3 personagens diferentes à sua escolha)

2.5. Feche o jogo sem gravar

3. Acabou a tarefa, prossiga para o questionário até a parte da segunda tarefa.

4. Segunda tarefa (5-10 min)

4.1. Inicie o jogo com o mod New Lords desativado e selecione “Continue

Campaign” quando entrar o menu principal

4.2. Desloque-se até a cidade de Amprela e entre na taverna

4.2.1. Observe as personagens e os seus comportamentos (1 or 2 min)

4.2.2. Fale com 3 personagens diferentes

4.3. Saia da taverna (Fique a pressionar no Tab -> Leave Tavern -> Leave)

4.4. Desloque-se até a cidade de Makeb (à direita de Amprela) e entre na taverna

4.4.1. Observe as personagens e os seus comportamentos (1 or 2 min)

4.4.2. Fale com 3 personagens diferentes

5. Acabou a tarefa, prossiga para o questionário.

6. Acabou as tarefas. Se quiser pode continuar a conversar com as personagens

(maximo 10 minutos).

107