sonalities in Stealth-Based Games
Modeling personalities with believable and consistent behaviour for NPCs in games using stealth gameplay style
Master’s thesis in Game Design & Technology
Marek ˇ Cernák
Orestis Lianoudakis
Department of Computer Science and Engineering C
HALMERSU
NIVERSITY OFT
ECHNOLOGYU
NIVERSITY OFG
OTHENBURGThe Tower: Design of Distinct AI Personalities in Stealth-Based Games
Modeling personalities with believable and consistent behaviour for NPCs in games using stealth gameplay style
Marek Černák Orestis Lianoudakis
Department of Computer Science and Engineering Chalmers University of Technology
University of Gothenburg
Gothenburg, Sweden 2020
using stealth gameplay style Marek Černák
Orestis Lianoudakis
© Marek Černák, Orestis Lianoudakis, 2020.
Supervisor: Staffan Björk, Department of Computer Science and Engineering Examiner: Olof Torgersson, Department of Computer Science and Engineering
Master’s Thesis 2020
Department of Computer Science and Engineering
Chalmers University of Technology and University of Gothenburg SE-412 96 Gothenburg
Telephone +46 31 772 1000
Typeset in L
ATEX
Gothenburg, Sweden 2020
using stealth gameplay style Marek Černák
Orestis Lianoudakis
Department of Computer Science and Engineering
Chalmers University of Technology and University of Gothenburg
Abstract
Computer games using stealth situations for defining their gameplay often rely on modeling an immersive, yet exploitable non-player character behaviour. Tradition- ally, the aim of the design of such a behaviour is to provide players with memorable behaviour patterns of the in-game agents in order to facilitate the means of devising strategies used by players to finish the game levels. Although certain behaviours for the agents are usually designed to make them reactive to environment events, it is not common for commercial game titles to include the reactions based on distinct personality traits of the agents. As a part of the thesis, we research game industry methods for creation of AI agents in games. Furthermore, we explore a research in the area of modelling autonomous agents with distinct personalities. We apply these concepts to design and implement a playable game prototype, including mod- els of AI agents with distinct personalities and perform several play-tests. From the outcome of the process, we formulate guidelines for designing distinct personalities with believable and consistent behaviour for NPCs in games using stealth gameplay style.
Keywords: stealth gameplay, games research, AI, NPC, autonomous agents, person-
ality modeling, thesis.
We would like the to thank our supervisor Staffan Björk for the advice and support during the development and writing of the thesis. We would also like to thank the people who tested the game and for the feedback they provided. We are also thankful to the University of Gothenburg for the opportunity of working on this project. Finally, we would like to thank Theodoris Michelis for evaluating the thesis’
text.
Marek Černák, Orestis Lianoudakis, Gothenburg, May 2020
List of Figures xiii
List of Tables xviii
List of Listings xix
Glossary xx
1 Introduction 1
1.1 Problem Description . . . . 1
1.2 Research Question . . . . 2
1.3 Aim . . . . 3
1.4 Solution Constraints . . . . 3
2 Background 5 2.1 Research Area . . . . 5
2.1.1 AI Personalities In Stealth Games . . . . 6
2.2 Related Work . . . . 6
2.2.1 Related Research . . . . 7
2.2.2 Application in Games . . . . 7
3 Theory 11 3.1 Perception . . . 13
3.1.1 Knowledge Communication . . . 13
3.1.2 Sensory Systems . . . 13
3.2 Knowledge . . . 17
3.3 Personality . . . 18
3.3.1 Personality Models . . . 19
3.3.2 Emotions . . . 20
3.3.3 Mood . . . 21
3.3.4 Motivation . . . 21
3.4 Behaviour . . . 22
3.4.1 Decision Making . . . 23
3.4.2 Decision Making Based on Personalities . . . 25
3.5 Design Frameworks . . . 29
3.5.1 Game Design Patterns . . . 29
3.5.2 MDA . . . 31
3.6 Stealth Gameplay Style . . . 32
4 Methodology 35 4.1 Iterative Process . . . 35
4.1.1 Managerial Practices . . . 36
4.1.2 Sprint . . . 36
4.1.3 Iteration Planning . . . 36
4.1.4 Daily Meeting . . . 36
4.2 Solution Requirements . . . 37
4.3 Computational Personality Model Prototype . . . 38
4.4 Personality Model Application . . . 39
4.4.1 Stealth Design Guidelines . . . 39
4.4.2 Stealth Design Methods . . . 40
4.5 Testing the game prototype . . . 41
4.5.1 Self-testing . . . 41
4.5.2 User Testing . . . 42
4.5.3 Data gathering . . . 43
4.5.4 Evaluation . . . 44
4.6 Wicked Problems . . . 45
4.6.1 A definition . . . 45
4.6.2 Ties to our question . . . 47
5 Planning 49 5.1 Planning Report . . . 49
5.2 Framework Prototype Development . . . 49
5.3 Thesis Finalisation . . . 51
6 Execution and Process 53 6.1 First Sprint . . . 53
6.1.1 Game AI . . . 53
6.1.2 Stealth Level Prototype . . . 55
6.2 Second Sprint . . . 60
6.2.1 Game AI . . . 60
6.2.2 Stealth Level Prototype . . . 63
6.3 Third Sprint . . . 65
6.3.1 Computational Personality Model . . . 65
6.3.2 Stealth Level Prototype . . . 69
6.3.3 Personality Models . . . 71
6.4 Fourth Sprint . . . 72
6.4.1 Game AI . . . 73
6.4.2 Stealth Level Prototype . . . 75
6.4.3 First Testing Session . . . 78
6.5 Fifth Sprint . . . 83
6.5.1 Game AI . . . 84
6.5.2 Stealth Level Prototype . . . 85
6.5.3 Personality Models . . . 89
6.5.4 Second Testing Session . . . 90
7 Results 93
7.1 Design Guidelines . . . 93
7.1.1 Consider which Computation Personality Model to work with 94 7.1.2 Consider order of design of personality and behaviour . . . 95
7.1.3 Consider using intents/motivations for NPCs . . . 97
7.1.4 Motivations and personality traits should conform each other . 98 7.1.5 Reactions of the same NPC given the same environment stim- uli should not conflict each other . . . 98
7.1.6 NPCs should be quite determined in finishing actions started . 99 7.1.7 Consider level design choices to fit the personality . . . 100
7.1.8 Consider in which order the personalities should be shown to the player . . . 101
7.1.9 Consider personalities differentiation via audiovisual represen- tation . . . 101
7.2 The Tower: A Playable Game Prototype . . . 102
7.2.1 Game System Architecture . . . 102
7.2.2 Game Level Design . . . 106
8 Discussion 113 8.1 Results . . . 113
8.1.1 Design Guidelines . . . 113
8.1.2 Game Prototype . . . 114
8.2 Methodology, Execution and Process . . . 116
8.3 Validity . . . 117
8.4 Future Work . . . 118
8.5 Ethical considerations . . . 119
9 Conclusion 121
References 123
A Computational Personality Model Supplement Material I
B Description of Designed Personalities III
B.1 Personalities . . . III B.1.1 Honorable Personality . . . III B.1.2 Pedantic Personality . . . IV B.1.3 Lazy Personality . . . . V B.1.4 Fearful Personality . . . VII B.2 Personalities Alternatives . . . VIII
B.2.1 Honorable Personality Alternatives . . . VIII B.2.2 Pedantic Personality Alternatives . . . IX B.2.3 Lazy Personality Alternatives . . . IX B.2.4 Fearful Personality Alternatives . . . IX
C First Gameplay User-Test Session Materials XI
D Second Gameplay User-Test Session Materials XVII
2.1 Screenshot from Sims3 [48]. At the bottom, the status UI shows the needs of the agent. These determine the motivation of the AI . . . . 8 2.2 In Batman: Arkham City [69] the player can at anytime view the
condition of their enemies. The guard in the figure is nervous with a high pulse indicating that they are aware that something is afoot.
They have either witnessed a passed-out fellow guard or they have seen the player character at one point and are now investigating . . . 9 2.3 An orc ally betraying the players. Taken from Shadow of War [57]
game. . . . . 9 3.1 Basics of an AI sensory system. Taken from [44] . . . 14 3.2 Sight cones. Image taken from [55] . . . 14 3.3 Field of view shapes used in Tom Clancy’s Splinter Cell: Blacklist.
Cone shape used for detection directly in front of the NPC for a quick detection, coffin-box shape used for more believable detection in larger distances from NPC. Image taken from [82]. . . 15 3.4 Examples of covering mechanisms used in games with stealth game-
play style. . . 16 3.5 Example of hiding mechanism applied to light brightness in Thief [27]. 16 3.6 Sound intensity attenuation. Image taken from [55]. . . 17 3.7 Decision making graph showing how internal and external data influ-
ence the decisions and how they can be retroactively modified based on the decided actions. Image taken from [55]. . . 18 3.8 Five-Factor personality model. The image shows what kind of per-
sonality falls under each trait (factor) depending on its value (high scorers or low scorers) within specific trait. Image taken from [20]. . . 19 3.9 The Sixteen Personality Factor Questionnaire personality traits (scales)
along with personality features described for low and high range for specific scale. Image taken from [18]. . . 20 3.10 Possible combinations of PAD along with description of the emotional
state they express. Image taken from [32]. . . 21 3.11 16 desires (named as striving in the picture) with personality features
based on the value of specific desire. Image taken from [20]. . . . 22 3.12 A simple state machine showing states and transition between them
defined by conditionals. Image taken from [55]. . . . 23
3.13 A class diagram showing simple solution for reusable state machines.
Image taken from [33]. . . 24 3.14 Example of Behaviour Tree. Question mark symbol represents Selec-
tor Composite node. Left-to-right arrow symbol represents Sequence Composite node. Diamond shaped node represents Until Fail Deco- rator node. Image taken from [55] . . . 25 3.15 Overview of Deliberative Module. Image taken from [72]. . . 26 3.16 iATTAC’s architecture. Image taken from [21]. . . 28 4.1 Minimap game state indicators in Deus Ex: Mankind Divided. Pic-
ture taken from [28]. . . 40 4.2 Blind spots in stealth gameplay style design as described and illus-
trated by Smith [73]. . . 41 4.3 Types of testers recommended for prototyping stages. Picture taken
from [31]. . . . 42 5.1 Gantt chart for Planning Report phase of the Master’s Thesis project.
Coloured box signifies that the task was worked on during that week. 49 5.2 Gantt chart for Framework Prototype Development stage of the Mas-
ter’s Thesis project. Coloured boxes specify which sprint takes part during that week. All of the tasks located below the Sprint’s name belong to that sprint. . . 50 5.3 Gantt chart for Thesis Finalisation stage of the Master’s Thesis project.
Coloured boxes specify which task takes part during that week. All of the sub-tasks located below the task’s name belong to that task. . 51 6.1 High-level diagram of the game AI showing the flow of data and
execution between the AI modules within one update process. Full lines represent the data flow, whereas dotted lines show the execution flow. . . 54 6.2 Sight perception system design. Green volume displays the main
vision of the NPC, magenta volumes display peripheral vision of the NPC. . . 55 6.3 Unity’s Navigation System. The blue overlay is a walkable area, the
red cube is an object with the NavMeshAgent component and the small cube has the NavMeshObstacle, as signified by the grey, non- walkable area surrounding it. . . . 56 6.4 A side by side comparison of the whiteboard and digital prototype
of the level. The green numbers, on the whiteboard prototype, pair key locations with doors they unlock. Black numbers indicate level elevation, with 0 being ground floor. . . . 57 6.5 Patrol design for Enemies in the first area of the level. Red numbers
are the patrol points and the lines that connect them show the path the guard should take. . . 58 6.6 Screenshot from the first scene. The player Character is laying down
on the floor drenched in a pool of spilled wine. . . 59
6.7 Guard model in idle position. The green light indicates their view range and radius. . . 59 6.8 High-level class diagram of the game AI system showing how each
parts of the system are associated. Some parts of the system can be thought of as modules (encapsulating larger parts of system), but for the sake of clarity, we introduce them here as classes. They are visualized with grey colour highlight. . . 60 6.9 A class diagram of the CPM part of the game AI system. . . 62 6.10 Personality model setup in Editor of the Unity engine. Setup includes
five personality traits characteristics and 16 motivation desires. . . 63 6.11 Two Guards patrolling in the first area of the level. Visible in the fig-
ure are also a key that can be picked up, doors that can be interacted with as well as hide-spots (blue planes) on which the player can crouch. 64 6.12 Structure of emotional event. Type specifies what change was noticed
in AI agent’s surrounding environment and the array specifies what type of emotions such a change triggers and at what strength. . . 66 6.13 Personality model setup in editor of the engine. Setup includes five
personality traits characteristics and 16 motivation desires. . . 68 6.14 Action trigger rules setup. On the left, we can see the logical oper-
ations setup, choosing specific logical operator, or operand, which is referred to by it’s specific index number. On the right, list of possible Knowledge Base rules is shown. The rule in this example translates to: (“playerSuspicion” == TRUE) OR (“playerHiding” == TRUE) OR (“noiseHeard” == TRUE). . . 69 6.15 Camera areas layout for the long hall area in the first level prototype.
Each area is assigned to specific camera position, and entering new area causes change to that camera position. Leaving the area, without entering new one, causes current camera position to follow the player Character. . . 70 6.16 Emotional reactions of the Guards displayed in a textual representation. 73 6.17 Sight perception system design. Green volume displays the main
vision of the NPC, magenta volumes display peripheral vision of the NPC. . . 74 6.18 The doors, similar to all objects of interest, get highlighted when the
cursor is over then to show that they can be interacted with. . . 75 6.19 The second tutorial level. Player starts from the top. There is a key
on the bottom left and a door that is unlocked with that key on the bottom right. Two Guards patrol the area, a lazy (Guard in corridor) and an honorable (Guard between key and door positions). . . 76 6.20 Guards looking directly into the shadows. They cannot see the player
hiding inside. The player is underneath the wooden balcony right in front of the Guard. . . 77 6.21 The UI at the top right of the screen shows how many items of a type
the player has at the moment. . . 77
6.22 The intro screen that appears at the start of the demo. . . 78
6.23 Results for two questions of online questionnaire. . . . 83
6.24 Example of hearing area of NPC to react only to noise within the predefined area of influence. Blue rectangle displays NPC’s position, red rectangles show interactible objects that can cause noise. . . 84 6.25 The Great Hall level. The starting area overlooks the final room with
the most Guards. . . 86 6.26 The Great Hall level overview. The player starts on the bottom right,
making their way to the left and then go through all the smaller rooms surrounding the large one in the middle. . . 87 6.27 Two agents Guard an object of interest (wooden shield). The left is
a pedantic type while the right is a lazy type. The player needs to exploit their weaknesses to make them move from their position. . . . 87 6.28 The chandelier dropping makes the Guards around it to react. . . 88 6.29 The lantern changes colours depending on the state of the agent.
Green colour indicates that they are unaware of the player’s presence.
Yellow indicates that they are investigating. Red indicates that they are alert. . . 89 7.1 Target motivations of the honorable personality setup. Only Honor,
Idealism and Physical Activity values are considered. . . 96 7.2 Large space design utilized by placing alternative honorable Guards
with quicker initial chase sequence. . . 97 7.3 Coherent reactions of the fearful personality being distracted twice in
a short amount of time. Orange light suggests the personality starts investigation, red light shows switch to shout and run away fearful behaviour. . . 99 7.4 The game system’s architecture overview. . . 102 7.5 When using a checkpoint, the player’s Character crouches in the mid-
dle of the hiding spot and a invisible wall appears from where the player cannot be spotted. Protecting them from the Guard’s view. . . 104 7.6 When interacting with an object of interest, a set of items might be
required for it to progress to the next message . . . 104 7.7 High-level class diagram of the game AI system used in the final
prototype. Some parts of the system can be thought of as modules (encapsulating larger parts of system), but for the sake of clarity, we introduce them here as classes. They are visualized with grey colour highlight. . . . 105 7.8 The player model. Taken from Unity’s asset store. . . 106 7.9 The first moments of the first level. The Tutorial is giving instructions
to the player to move towards the red mark on the floor. . . 107 7.10 In the second level, the player learns how the shadow works. When
being inside the UI shows that you are hidden from view . . . 108 7.11 When interacting with this bookcase a hidden door and staircase are
revealed leading to a previous visited room while providing the player
with a new key to open a previously locked door. . . . 109
7.12 The cauldron gives a stone key once its requirements have been met . 110
7.13 The player interacts with two interactive shields. Each with its own puzzles, to light up both of the gargoyle’s eyes. After doing so, inter- acting with the gargoyle rewards the player with a silver key. . . 111 A.1 Table showing relationships between motivation types and personality
traits expressed numerically. Table taken from [72]. . . . I A.2 Table showing relationships between emotion types and personality
traits, and also emotion types and motivation expressed in PAD space, expressed numerically. Table taken from [72]. . . . II B.1 Personality traits setup for honorable personality. . . IV B.2 Target motivation setup for honorable personality. . . IV B.3 Personality traits setup for pedantic personality. . . . V B.4 Target motivation setup for pedantic personality. . . . . V B.5 Personality traits setup for lazy personality. . . . VI B.6 Target motivation setup for lazy personality. . . VII B.7 Personality traits setup for fearful personality. . . VIII B.8 Target motivation setup for fearful personality. . . VIII C.1 First Part (Personalities) of the Interview questions for the first game-
play session. . . XI C.2 Second Part (level design) of the Interview questions for the first
gameplay session. . . XII C.3 Online questionnaire for the first gameplay session (part one). . . XIII C.4 Online questionnaire for the first gameplay session (part two). . . XIV C.5 Online questionnaire for the first gameplay session (part three). . . . XV C.6 Online questionnaire for the first gameplay session (part four). . . XVI D.1 First Part (Personalities) of the Interview questions for the second
gameplay session. . . XVII D.2 Second Part (level design) of the Interview questions for the second
gameplay session. . . XVIII
D.3 Online questionnaire for the second gameplay session (part one). . . . XIX
D.4 Online questionnaire for the second gameplay session (part two). . . . XX
D.5 Online questionnaire for the second gameplay session (part three). . . XXI
D.6 Online questionnaire for the second gameplay session (part four). . . XXII
3.1 Eight definitions of AI thought process and behaviour divided based on the success of the AI measured either by a rational or "human-like"
performance. Table taken from [70]. . . 11 A.1 Table showing relationships between mood PAD dimensions and per-
sonality traits expressed numerically. Table taken from [32] . . . . II
3.1 Sample code for reusable state machine. Sample taken from [33] . . . 24 6.1 Patrolling Motivation Action definition using Behaviour Trees by
NPBehave [58]. . . 67
AI Artificial Intelligence. xiii, xviii, 1–3, 5–8, 11–15, 17, 18, 20, 22, 23, 26, 28, 32, 34, 35, 41, 53, 54, 56, 63–66, 68, 75, 84, 92, 94, 98, 99, 103–106, 117, 119–121 CPM Computational Personality Model. xv, 25, 41, 53, 60–62, 72, 93–95, 98, 100,
103, 112, 114, 115, 117, 118, 121, 122 FOV Field of View. 14, 15
FPS Frames Per Second. 85
MDA Mechanics, Dynamics, Aesthetics. 31–33
NPC Non-Playable Character. xiii, xv, xvi, III, VII, IX, 1–3, 5, 6, 8, 9, 12–17, 22–24, 29–31, 33–35, 37–41, 47, 56, 60, 61, 64–66, 71–74, 80–86, 88, 89, 91–96, 98–102, 112–122
PAD Pleasure, Arousal, Dominance. 21, 26 POI Person of Interest. 1
UI User Interface. xiii, xv, xvi, 8, 77, 108
1
Introduction
As the computational power of consumer-affordable gaming consoles, personal com- puters and other devices used for gaming grows, the computational time for per- forming specific tasks of game engines rises, too [55]. Hand in hand with this trend, expectations of the gaming community from game designers and developers to pro- vide more immersive experience strengthens. One of the areas that gains such an attention is capturing the behaviour of AI agents interacting with the players within the games. Although several smart and established techniques to simulate the AI behaviour have been developed [55, 47], exploring the possibilities gained by the ad- ditional computational power could open door to further improvements in providing more believable experience for players. Games focusing on, or making substantial use of stealth in their gameplay design such as Metal Gear Solid V: The Phantom Pain [40], Hitman [37], Batman: Arkham City [69], Deus Ex: Mankind Divided [28]
or Assasin’s Creed [78] all rely on modeling levels that try to capture the immersive stealth situations, including AI NPC patterns that make for believable gameplay experience. The design usually focuses on observable and exploitable patterns in AI behaviour, such that the player can learn and devise a successful strategy to get through the level.
Moving one step forward, exploring the idea of extending the AI design in stealth situations by modeling personality of the AI agents could add an additional layer of immersion for players that would be expected to combine the NPC stealth patterns with their reactions dependent on their individual personality traits. It could also widen the possibilities of game and level designers on how to approach the design of stealth situations in games.
1.1 Problem Description
Application of AI in games usually relies on modelling the unified, shared behaviour
of the NPCs. For example in Batman: Arkham Asylum [69], the panic rate esca-
lates amongst the AI NPC, depending on the environment changes around them
(for instance discovering knocked out NPC, briefly seeing the player character or
reacting to noises). In Hitman (IO Interactive, 2016), AI NPCs can belong into
one of three predefined categories: a mob AI, POI (Person of Interest) AI and a
bodyguard AI. As explained above, NPCs react to the same environmental impulses
with the same response, or as in the later case, with the same response within the same NPC category. While the approaches were admittedly proven reliable and gen- erally welcomed by the player community, such approaches do not consider distinct individual personalities each NPC might have.
We believe that by applying distinct personality traits to the game AI agents by using personality models we can:
• Provide more believable or immersive experience for players.
• Support gameplay designers with a wider range of stealth gameplay design possibilities.
The topic of modeling the personality behaviour is an active area of research [20] and there have been several proposals on how to approach the problem [29, 21, 45, 72].
The proposed solutions, however, usually target the general solution for modeling the AI behaviour. Finding the suitable models and refining them to suit the needs of the stealth gameplay style might bring new possibilities for designing the game AI within the gameplay style.
1.2 Research Question
In the thesis, we focus on exploration of possibilities of a more varied behaviour of NPCs in stealth gameplay design. Our idea is to explore and choose the parts of the existing personality models that could be relevant for stealth based NPC creation.
We approach the problem with usability in mind, exploring the ways to support conveying the information about the NPC personality in an understandable and immersive fashion.
Trying to reach the thesis goal, our focus will mainly revolve around answering the following research question:
“What guidelines should be considered when designing distinct personal- ities for NPCs with consistent and believable behaviour in stealth-based gameplay?”
Trying to answer this question generates a number of challenges:
• Conveying the information about the NPC’s personality and making it ex- ploitable by the players to be able to form a strategy.
• Combining the personality model with environmental and situational context whilst keeping the final result of any interactions generated by this combination believable and immersive.
• Making sure the AI decisions are consistent (deterministic) and that the inde-
cisiveness is not present or minimal.
• Making a playable prototype to test and present our research findings.
• Making the prototype as modular and extendable as possible, i.e. making kind of framework that should be functional with or without the personality model.
Such a framework should also allow to replace personality model of choice with different ones with as lowest effort as possible.
• Finding suitable methods or metrics for evaluation of the results and their comparison to other established approaches.
1.3 Aim
The aim of this work is a creation of set of guidelines for designing the distinct personalities of NPCs in games using stealth gameplay style. More specifically, we plan to analyse the current state of research within the area of modeling AI personality and explore the options in application of these within the specific area of stealth gameplay style games. To connect the research with the applications of AI techniques used in game industry, we also aim to analyze the common approaches in designing the AI of NPCs in the games using the stealth gameplay style used in the game industry. The output of the analysis should serve as a background knowledge for designing and implementing a playable prototype with research findings applied to test the possibilities of personality features application and formulate the results.
1.4 Solution Constraints
The focus of the thesis is to formulate a set of guidelines for designing the person- ality models for NPCs used in the games with stealth gameplay style. As a part of research, general parts of AI design in games are presented, but only to the ex- tent needed for understanding of how game AI is generally designed and created.
Similarly, for building a prototype to test the possibilities of personality features application, we aim at developing only parts necessary for basic AI functionality expected in games using stealth gameplay style.
Furthermore, the implementation of the prototype shall be provided for testing purposes only, again, focusing mainly on testing the possibilities of using distinct personalities for individual NPCs. Therefore, we strive to develop only parts neces- sary for basic AI functionality expected in games using stealth gameplay style, or reuse the existing tools to implement supporting AI techniques such as movement and sensory systems.
Additionally, we plan to design only a small section of stealth situation (part of a level) to be able to satisfy the needs of user testing to produce measurable results.
It is also not our intention to develop a fully functional tool for public usage, but
rather provide a concept for the functionality of a prototype developed for the needs
of reaching our goals.
2
Background
This chapter will contextualize solutions explored in reports and existing research on commercial products that relate to the problem description and provide inquiry support for the upcoming theory chapter.
2.1 Research Area
In the field of AI in games, there are numerous cases we can draw examples from in order to see their benefits in how a world can feel immersive and alive for the player to experience. AI agents are programmed to have a set of states and reactions they provide to the player depending on situation and status. Some games have it more complex than others, introducing a behaviour to their agents that is appro- priate for the world the game is being set in. Such examples one can find in games like Assassin’s Creed [78], Shadow of War [56], Thief [27] and Deus-Ex: Mankind Divided [28] among others. Because of the nature of the style, agents need to ex- emplify an immersive, yet predictable behaviour[55]. The player can then exploit the NPCs in order to maneuver around them, sneak past them or use other tools in their disposal [85]. The solutions provided by the developers are focused on how to achieve a level of immersion without making it unexploitable for the player, thus ruining the experience.
A common practice for AI agents to have in a stealth game are alert states [85].
Agents cycle through these states depending on whether they see the player, hear a noise, notice movement etc. The number of states and their level of alertness may differ from game to game, but in general, the AI agents usually make use of three base states [47, 85]:
• Relaxed
• Suspicious
• Alert
In these solutions the agents work in unison. Depending on the category that they
belong in (referring to NPC categories differentiation examples provided in Sec-
tion 1.1) or their state of alertness they would behave somewhat similarly. This is
key for their predictable behaviour [85] [55] that the player can exploit when dealing with these NPCs. However, we believe that this behaviour can be expanded upon, without breaking their needed predictability.
2.1.1 AI Personalities In Stealth Games
Introducing individuality, through different and distinct personalities to the agents can bring another layer of complexity to the NPCs. On top of their usual behaviour, the application of personalities can provide a plethora of different reactions an agent can resort to when dealing with different scenarios. For instance, a cowardly guard can be startled easier than others and call for backup regardless if they reach their alert state. An arrogant persona can dismiss the possibility of calling in assistance and try to resolve the threat of the intruder (the player in this supposed stealth game scenario) by themselves.
Previous research on how to implement personalities to AI agents talks about the grouping of different factors that affect the final outcome of an agent’s reac- tion [72, 21]. The models suggested in the papers goes in detail on how one can approach the application of unique personalities on agents and to have them re- act differently depending on their personality, motivation, emotion, mood, rituals or memories. Similarly, such research could be conducted in more specific areas of game development, such as specific gameplay style. Therefore, we see a possibility of exploration of different personality features used for NPCs in stealth gameplay style.
Regarding the predictability of the agents, a communication between the game and the player needs to be maintained in order for the player to understand what type of personality the agent portrays. This can be achieved for example by visual rep- resentation of the agent, similar to how a change of alertness is communicated by change of posture. Another example is a verbal approach, with soft dialogue ex- changes amongst the agents that can give a clue on the type of personality each agent has. It is critical for the personalities to be consistent and not randomized in each agent so that the player can expect same personality features recurring in NPCs when facing them repeatedly.
Understanding what type of personality each NPC has is important for the player to know how to approach the situation, thus giving an additional level of complexity that we believe can enhance the level of immersion in the game.
2.2 Related Work
Regarding related work that is in correlation with the thesis’ goals and research questions, there are a plethora of available sources both from the scientific and the commercialised community that have contributed on the subject in different ways.
Related studies have given a framework on how to approach the individuality for
each agent. The commercial products are, for now, mostly relying on established
methods that strive for agent behaviour unison rather that distinction. Additionally, it is not uncommon for the scientific community to focus its research on how these commercial products have achieved their immersive and yet exploitable behaviour in their games.
A combination of the two approaches is important for our research. By looking into the relevant aspects of the individuality in agents and how they can be introduced into the established framework of the commercialized products, we can provide a more extensive answer to the subject at hand.
2.2.1 Related Research
For more realistic social interactions amongst agents in games the iATTAC [21]
was introduced. The authors introduce the importance of agent memory, personal agenda, emotions and personality. They describe that because of the system’s gen- eral design it is implementable in other games other than its original intended usage.
In the article by Shvo at el. [72] the authors attempt to model the implementation of personalities in AI agents by combining the components of Five-Factor personality model [49], Reiss’ 16 basic desires model [66], list of possible emotions as presented by Ortony at al. [62] and Mehrabian’s model of mood [50] [50]. By their model, they try to address root cause of the emotions experienced by the agent.
Additionally, Evans, R. [29] attempts to represent personality traits in synthetic agents, extending upon the model used in Sims 3 [48] described in Section 2.2.2. He describes that the computational model of personality should satisfy a set of goals in order to achieve a level of believably that is to be considered enough for player immersion.
2.2.2 Application in Games
Some of the more standardized methods mention in the beginning of this section 2.1 have been used in commercial products such as, Batman: Arkham City [69], Hit- man [37], the Sims 3 [48], Facade [64] and Shadow of War [57]. Additional games are mentioned throughout the thesis when more specific techniques are applied.
In Sims 3 [48], developers used a wide pool of personality traits and assigned a
combination of them to each agent [16]. In turn, agents interacting with objects or
other agents they could affect their emotional state and satisfy their motives (when
these actions were suited for a trait, then they were considered satisfying for the
corresponding motive). Additionally, affects taking place in the agent’s emotional
state were expressed in a simple declarative language. Figure 2.1, shows a screenshot
of the game where the motivation status bars are visible.
Figure 2.1: Screenshot from Sims3 [48]. At the bottom, the status UI shows the needs of the agent. These determine the motivation of the AI
Facade [64] is a game where the player walks into the home of a couple in which they can discuss by typing to the NPC and the drama unfolds depending on what they player is typing. The developers of the game introduced a dynamic dialogue system where the agents were operating autonomously and moved based on their personal goals. An external manager would, depending on the situation, add or remove behaviours to the agents. Meaning that depending on what has been mentioned before and what the player had last typed in, the agents would dynamically adjust their behaviour and available responses based on what the external manager was directing them with.
In Hitman [37], the developers grouped together the AI agents depending on their status in the world in two main categories:
• Civilians
• Guards
Their status determines their basic reaction to whatever the player was doing. For
instance, if the player draws a weapon in the open public and starts firing, the
civilians would react by moving away from him. Guards, on the other hand, would
begin to fire back and start a combat sequence with the player. These two differences
show how an agent can react differently to a similar situation in order for the game
world to align with how the player perceives it. Since the Hitman games take place
in a world similar to the real one, it would not be an immersive decision to make
Figure 2.2: In Batman: Arkham City [69] the player can at anytime view the condition of their enemies. The guard in the figure is nervous with a high pulse indicating that they are aware that something is afoot. They have either witnessed a passed-out fellow guard or they have seen the player character at one point and are now investigating
both of these groups to react the same way.
In Batman: Arkham City [69], the player can activate a specific mode that scans different points of interest and NPCs. Scanning reveals their nervousness in a car- diograph, Figure 2.2. The nervousness, in this case, is another way of showing their aforementioned alertness state. This can help with the feeling of the situation esca- lating and the situation of the guard’s evolving with the players activity in the area.
The memory of the guards increases the level of believability in the game.
Figure 2.3: An orc ally betraying the players. Taken from Shadow of War [57]
game.
In Shadow of War [57], the enemies follow, similar to the previous game in series,
Shadow of Mordor [56], a system called the nemesis system [43]. This system is too
complex to cover fully in this thesis, however, in this paragraph we mention some
of the key points the system provides. In this system, the enemies can switch sides
and join the player’s army. They can betray the player as shown in Figure 2.3,
ambush them, seek revenge for themselves or a fallen companion from a previous
encounter. They gain ranks among their own military system, they have unique
abilities and weaknesses that tie into their rank and personality. The intriguing
part of this system is that every enemy can become important, meaning they can be
assigned personality, name and rank. If a random enemy kills the player’s character,
the game shows their name, they get added into the military hierarchy and can now
be tracked down for the player to seek revenge against, or recruit them in their own
army, or even disgrace them, lowering their status amongst their fellow orcs and
making them more susceptible to interrogation and fear-tactics.
3
Theory
In this chapter we discuss the knowledge important to understand the representation of an intelligent agent in games using stealth gameplay style and also important in order to proceed with the design of a technical framework for modelling the individual personality of the AI in games.
Before we start with the overview of specific parts that AI is usually modeled by in games using stealth gameplay style, let us focus on the definition of the AI and intelligent agent itself.
In the book by Russel and Norvig [70], authors divided eight definitions of AI in two dimensions: definitions describing the thought process and reasoning and definitions concerned with behaviour.
The division with definitions can be seen in the Table 3.1. In the same table, another grouping can be seen, that of separating the measured success of AI into
Thinking Humanly Thinking Rationally
"The exciting new effort to make com- puters think ...machines with minds, in the full and literal sense." [81]
“The study of mental faculties through the use of computational models." [22]
"The automation of] activities that we associate with human thinking, activi- ties such as decision-making, problem solving, learning ..." [6]
"The study of the computation that make it possible to perceive, reason, and act." [84]
Acting Humanly Acting Rationally
“The art of creating machines that per- form functions that require intelligence when performed by people." [41]
"Computational intelligence is the study of the design of intelligent agents." [63]
“The study of how to make computers do things at which, at the moment, peo- ple are better." [67]
“AI ...is concerned with intelligent be- haviour in artifacts." [59]
Table 3.1: Eight definitions of AI thought process and behaviour divided based on the success of the AI measured either by a rational or "human-like" performance.
Table taken from [70].
"human-like" performance and into rational (ideal) performance.
In order for a computer to think in a more humane way, it must be first established how a human would think in a specific situation. AI can get closer to human thinking by producing expected outputs for provided inputs (matching cognitive models).
To better increase the believability and a human nature of how the AI behaves, it requires the computer to be able to process the external impulses, communicate, to represent knowledge (a memory), reason, learn and adapt. Being able to give out the impression of rational thinking requires the agent to proceed with logic, following the rules of argumentation that would always lead to a correct conclusion.
Rational acting emphasize on the agent acting in alignment with reaching the best outcomes. [70]
An agent can be reasonably called as intelligent when behaving successfully in a certain environment, given certain criteria [70]. More specifically, an intelligent agent is able to perceive signals from its surrounding environment and make decisions and subsequent actions that would specify the best possible outcome given the state of the environment.
In stealth gameplay, however, the agent is expected to act intelligently, but also to be fair the to players given their knowledge about the in-game world state. Walsh [82]
mentions four characteristics that the agent model needs to display in order to be successful:
• Fairness - designing the AI model such that player does not feel overwhelmed and discouraged by the NPC’s abilities.
• Consistency - building a kind of predictability in the AI behaviour, enabling a player to construct strategies.
• Good feedback - the player must be able to comprehend why the NPCs behave in the way they do in all situations.
• Intelligence - perception of an intelligent agent, however, not necessarily a smart NPC behaviour is required, but always acceptably sensible.
In the remainder of this chapter, first, we focus on describing the intelligent agent
perception and knowledge representation approaches used in the games with stealth
gameplay style by the industry. Next, we continue with description of the personal-
ity models developed in research to define the distinct personality, followed by the
research application of the models in games trying to simulate distinct personality
behaviour. Then we proceed with overview of the techniques for behaviour simula-
tion, covering AI decision making and actions execution. The chapter continues with
the description of game design frameworks that helps us analyse common features of
NPC design in games using stealth gameplay style and validate the design decisions
used in order to formulate the framework for modelling the individual personality
of the AI. Finally, the chapter ends with the description of stealth gameplay style.
3.1 Perception
One of the important properties of the game agents is to be able to interact with the game world. For a game agent to be able to react to environment changes and also communicate with other agents active in the game world, communication channels that update the NPCs with the relevant information need to be built.
Furthermore, in case of designing a believable AI behaviour, agent are supposed to perceive the world information in means explainable when applied to real-world scenarios. Specifically, in-game NPCs need to be able to use various senses to perceive the information about the current state of the game world.
3.1.1 Knowledge Communication
Communication if the world state information between the game world and agent, or between agents between themselves, is a mean of AI NPC knowledge retrieval that is necessary for simulating autonomous behaviour. There are two generally used techniques for knowledge transfer between the game world and its agents [55]:
• Polling
• Events management
In the first mentioned approach, the AI agent knows what information it is interested in and polls for it accordingly. In the later approach, instead of actively asking about the information needed, agents are instead notified by other systems/agents when a change of the state happens. Distribution of the events about the game world state is usually handled by an event manager [55] which serves as a centralized event system that delivers new knowledge to all of its subscribers. We will not go further into technicalities of the events management, more detailed description about the event management functionality and potential implementation can be found here [55].
3.1.2 Sensory Systems
While in Section 3.1.1 we talked about the means of transferring the knowledge amongst the game world entities, it is also important to describe how these infor- mation are first gathered. The usual approach to model the perception abilities of the agents is via sensors that simulate the human senses.
The diagram shown in Figure 3.1 created by Leonard, T. in [44] describes the basic
functions of a sensory systems. Following the diagram, AI senses are depicted as
expressed through its Awareness. Awareness in this example functions as a set
of possible states representing knowledge about the current state of an object of
interest. Collecting the knowledge about the object of interest through the AI
Senses is decoupled from the AI’s decision making that is influenced by the data
gathered. The Awareness state of an AI agent is then stored in the Sense Links that
can associate the AI agent to another Game Object in the game. Sense Links serve
Visibility Game Object
1 1
Sense Link
1 n
Awareness
AI Senses AI
Player
1 n
1 1
Configuration View Cones
Alertness
1
n
Figure 3.1: Basics of an AI sensory system. Taken from [44]
Figure 3.2: Sight cones. Image taken from [55]
as a mean of storing and sharing relevant sensory details such as time, location.
In the remainder of this section we talk about several common senses typically used in game AI agents models and their representation. The theory provided in the remainder of the section, if not stated otherwise, is based on the information gathered from [55].
3.1.2.1 Sight
Sight is one of the most used senses for AI perception modelling. It also requires to consider several important features to make it appear behaving believable by the players. It is usually the sense of the most importance, as players can usually notice the bad sight design easily.
Field of View
Field of view (FOV) represents the area in which the AI NPC is able to see an object
of interest. The most common technique of representing the FOV is a cone, such as
the ones shown in Figure 3.2. The cone restricts the field of view to a certain degree
horizontally, vertically or in both directions.
Box
Cone
Peripheral NPC
Figure 3.3: Field of view shapes used in Tom Clancy’s Splinter Cell: Blacklist.
Cone shape used for detection directly in front of the NPC for a quick detection, coffin-box shape used for more believable detection in larger distances from NPC.
Image taken from [82].
Even though sight cones are an easy to use technique, it has a drawback of modeling the correct behaviour at larger distances [82]. In Tom Clancy’s Splinter Cell: Black- list [79], the developers used a coffin-like shape, which can be seen in Figure 3.3, to model the phenomena. In this solution, first, as the distance from the NPC grows, the field of view widens, but only up to a certain point in depth. After that point, the field of view gradually narrows down.
Other influences
While FOV is very important sight feature to consider when designing stealth AI, there are other considerations worth mentioning. Distance limitation, for example, is sometimes considered to restrict the NPC visibility, although usually in stealth gameplay the distance is not restricted. This is mainly because of the stealth game- play areas are usually smaller, hence making the distance restriction would not seem plausible. Stealth gameplay design usually tries to provide means for player to tackle the NPC sight, such as cover mechanisms as shown in Figure 3.4.
Another frequently used feature to model the sight perception is light brightness.
Such a feature can be seen for example in Thief [27] with example usage in Figure 3.5.
Brightness is commonly used to provide additional hiding mechanism for players.
3.1.2.2 Hearing
In comparison with sight, sound distribution can travel through non-transparent physical obstacles. Volume of the sound is gradually lowered down as the sound progresses through the space. The sound wave progression depends on the nature of the sound (e.g. it’s frequency) and material it is passing through.
In the video games the previously described phenomena is usually simplified to split
(a) Hiding behind the walls Deus Ex: Mankind Divided [28].
(b) Hiding in bushes in Rise of the Tomb Raider [24].
Figure 3.4: Examples of covering mechanisms used in games with stealth gameplay style.
Figure 3.5: Example of hiding mechanism applied to light brightness in Thief [27].
the game objects materials to those that transmit sound and those that do not.
Materials that transmit the sound act like an air. Sound then progresses through the transmissive material and attenuates depending on its distance from the source.
Such sound interpretation in games can be seen in Figure 3.6.
In Tom Clancy’s Splinter Cell: Blacklist, developers instead chose to calculate the distance to the sound through the NPC navigable area [82], calculated as
Distance = dist(source, A
1) + dist(A
1, A
2) + ... + dist(A
n−1, destination) (3.1) where A
1..A
nare the “choke points”, points that describe the connection between areas explorable by NPCs such as doors or windows.
3.1.2.3 Other senses
Other notable senses such as touch or smell are currently not extensively used in
stealth gameplay style (with the exception of collision detection simulating touch)
and we will leave them mentioned here without any further description, as there
exploration is out of the scope of this thesis.
Figure 3.6: Sound intensity attenuation. Image taken from [55].
3.2 Knowledge
Knowledge represents every aspect of the game that AI agent can perceive using the senses, store it and use later in decision making process. Knowledge can be represented as an interface between the agent’s perception and decision making systems.
Knowledge could be also described as a tool for storing short or long term memory.
Therefore, a kind of data storage is usually used to remember the current world state, called Knowledge Base. Knowledge Base is a database for knowledge representation, that can be updated by perception data, polled for the current state by decision making systems, and serves as an interface for sharing knowledge with other agents.
Knowledge Base is also updated continuously, to simulate a memory functionality, forgetting certain events, or make combinations of different world state changes to derive more complex knowledge facts.
Knowledge in game can be represented in many forms, and how it is represented often reflects the game AI needs for a specific game, as it can have impact on development time and game run-time efficiency [15]. Knowledge can be represented as simple current event spotted by the perception system, overriding the current event using priorities system of events, such as knowledge representation described by Miles [54]
in the article describing NPC awareness in the Mark of the Ninja [39]. For shooters, for example, the knowledge can be represented in the means of potential shooting targets, such as the one described by Welsh [83] in the article presenting target tracks perception system originally written for Crysis 2 [25].
More general approach is described in the article by Carlisle [15], providing design
requirements and implementation details. In this approach, a property-based system
is used, storing named data in a variant format allowing for retrieval of the data by
Internal knowledge
Decision maker
Internal changes
External changes External knowledge
Action request
Figure 3.7: Decision making graph showing how internal and external data influ- ence the decisions and how they can be retroactively modified based on the decided actions. Image taken from [55].
property name, in any suitable format.
In general, knowledge in games can be separated into Internal and External knowl- edge, as shown in the diagram in Figure 3.7. Internal knowledge consists of the factors important for the given agent, for example:
How long has it been since the agent spotted a player ?
Which can be used by agent’s internal systems to indicate whether to return to routine behaviour, or choose different actions instead.
External knowledge, on the other hand, is a knowledge that is shareable within the enemies, such as:
Do I have knowledge of player whereabouts?
Which can be used to alarm other agents to act accordingly.
Note that the example used for internal knowledge representation, could be, in certain cases used as an external knowledge, too. The separation between those is usually game specific.
3.3 Personality
Personality analysis is a much focused subject in the field of psychology. In the attempts made by various researchers to understand such an abstract matter, there have been a number of models made that attempt to contextualize and make an understanding of this complex subject.
Based on these findings, researchers have explored further ways to apply them to
model AI behaviour. In this section, we provide overview of some of the established
personality models, including models of personality traits, moods, emotions and
motivations.
Factors Low scorers High scorers Extroversion Loner
Quiet Passive Reserved
Joiner Talkative Active Affectionate Agreeableness Suspicious
Critical Ruthless Irritable
Trusting Lenient Soft-hearted Good-natured Conscientiousness Negligent
Lazy Disorganised Late
Conscientious Hard-working Organised Punctual
Neuroticism Calm
Even-tempered Comfortable Unemotional
Worried Temperamental Self-conscious Emotional Openness to
experience Down-to-Earth Uncreative Conventional Uncurious
Imaginative Creative Original Curious
Figure 3.8: Five-Factor personality model. The image shows what kind of per- sonality falls under each trait (factor) depending on its value (high scorers or low scorers) within specific trait. Image taken from [20].
3.3.1 Personality Models
Personality models aim to offer means to describe a unique individual personality via unified set of properties. Evans in his article [29] argues that a requirement on any such personality computational model is it’s components atomicity and reusability.
Having personality model consisting of reusable atomic units provide for fast and easy creation of variety of new personality types.
One of the widely used personality models is the Five-Factor personality model [49].
In this model, the personality is classified into five personality traits that can be seen in Figure 3.8 as factors.
Neuroticism is a tendency for quick arousal upon stimulation and slow relaxation from aforementioned arousal. With emphasis in negative emotional arousal, neu- roticism is defined by its emotional instability, negativity and maladjustment.
Extraversion is the level of enthusiasm, talkativeness, assertiveness and gregarious- ness that people tend to project onto others whilst being part of social interaction.
An extraverted person finds more reward on time spent in social activities rather than the contrary.
Conscientiousness implies a desire for a person to perform on a task with care and diligence. People characterized as conscientious tend to aim for achievement, are dependable and rather thorough in their approach and planning when undertaking any task.
Agreeableness is the characteristics of compassion, kindness, sympathy, cooperation
and consideration. A key factor for concepts like cooperation and social harmony.
Descriptors of Low Range Primary Scales Descriptors of High Range
Reserved,Impersonal,Distant Warmth (A) Warm-hearted,Caring,Attentive To Others Concrete,Lower Mental Capacity Reasoning (B) Abstract,Bright,Fast-Learner
Reactive,Affected By Feelings Emotional Stability (C) Emotionally Stable,Adaptive,Mature Deferential,Cooperative,Avoids Conflict Dominance (E) Dominant,Forceful,Assertive Serious,Restrained,Careful Liveliness (F) Enthusiastic,Animated,Spontaneous Expedient,Nonconforming Rule-Consciousness (G) Rule-Conscious,Dutiful
Shy,Timid,Threat-Sensitive Social Boldness (H) Socially Bold,Venturesome,Thick-Skinned Tough,Objective,Unsentimental Sensitivity (I) Sensitive,Aesthetic,Tender-Minded Trusting,Unsuspecting,Accepting Vigilance (L) Vigilant,Suspicious,Skeptical,Wary Practical,Grounded,Down-To-Earth Abstractedness (M) Abstracted,Imaginative,Idea-Oriented Forthright,Genuine,Artless Privateness (N) Private,Discreet,Non-Disclosing Self-Assured,Unworried,Complacent Apprehension (O) Apprehensive,Self-Doubting,Worried Traditional,Attached To Familiar Openness to Change (Q1) Open To Change,Experimenting Group-Orientated,Affiliative Self-Reliance (Q2) Self-Reliant,Solitary,Individualistic Tolerates Disorder,Unexacting,Flexible Perfectionism (Q3) Perfectionistic,Organized,Self-Disciplined
Relaxed,Placid,Patient Tension (Q4) Tense,High Energy,Driven
Global Scales
Introverted,Socially Inhibited Extraversion Extraverted,Socially Participating Low Anxiety,Unperturbable Anxiety Neuroticism High Anxiety,Perturbable Receptive,Open-Minded,Intuitive Tough-Mindedness Tough-Minded,Resolute,Unempathic Accommodating,Agreeable,Selfless Independence Independent,Persuasive,Willful
Figure 3.9: The Sixteen Personality Factor Questionnaire personality traits (scales) along with personality features described for low and high range for specific scale. Image taken from [18].
Openess describes the ability of one to be able to accept new experiences. This factor, when scored high, characterizes people that prefer unfamiliar routines, new experiences and a wider range of interests.
By reworking the Five-Factor personality model, The Sixteen Personality Factor Questionnaire [18] classifies personality types in more detail by introducing 16 pri- mary personality traits. The authors of the questionnaire also constructed second- order traits, or global factors, that would describe the personality more generally, grouping together the 16 primary traits. These global factors are closely related to Five-Factor personality model, each of the factor having similarity with factors of the Five-Factor model. Primary and global traits can be both seen listed in Figure 3.9.
3.3.2 Emotions
Bates in his article [5] stated that emotions are key for believable AI agent behaviour.
The idea is further argued by Evans [29], who claims that emotions are tied to rational decision making processes.
Basic emotions are, according to Ortony et al. [62], separated into three basic classes.
• pleased and displeased
• approving and disapproving
• liking and disliking
All of these classes refer to reactions to different events, agents and objects accord-
Table 1: Mood octants of the PAD space.
+P+A+D Exuberant -P-A-D Bored +P+A-D Dependent -P-A+D Disdainful +P-A+D Relaxed -P+A-D Anxious +P-A-D Docile -P+A+D Hostile
