1

2
Game Mechanic Alignment Theory 59
60
3 61
4
Michael Cerny Green Ahmed Khalifa Philip Bontrager 62
5
mike.green@nyu.edu ahmed@akhalifa.com pbontrager@gmail.com 63
6
New York University | OriGen.AI Game Innovation Lab | Modl.ai New York University 64
7 New York City, New York, USA New York City, New York, USA New York City, New York, USA 65
8 66
9 Rodrigo Canaan Julian Togelius 67
10 rmc602@nyu.edu julian@togelius.com 68
11 New York University New York University 69
12 New York City, New York, USA New York City, New York, USA 70
arXiv:2102.10247v2 [cs.AI] 10 Aug 2021

13 71
14 ABSTRACT game’s rules. Without this guidance, the player may become frus- 72
15 We present a new concept called Game Mechanic Alignment theory trated, unable to figure out how to play or feel a sense of progression. 73
16 as a way to organize game mechanics through the lens of systemic At worst, a tutorial is unhelpful and confusing. But if done correctly, 74
17 rewards and agential motivations. By disentangling player and a tutorial excites and encourages the player to keep playing. 75
18 systemic influences, mechanics may be better identified for use To be effective teachers, tutorials need to contain the crucial bits 76
19 in an automated tutorial generation system, which could tailor of information needed to play, such as what the controls are, how 77
20 tutorials for a particular playstyle or player. Within, we apply this to win, and how to lose. This information is typically defined by 78
21 theory to several well-known games to demonstrate how designers the game’s mechanics, i.e. the events within the game triggered by 79
22 can benefit from it, we describe a methodology for how to estimate game elements that impact the game state [47]. “Critical mechan- 80
23 “mechanic alignment”, and we apply this methodology on multiple ics” are the mechanics that need to be triggered in order to win a 81
24 games in the GVGAI framework. We discuss how effectively this level [23]. Therefore, it makes logical sense that critical mechanics 82
25 estimation captures agential motivations and systemic rewards be explained within a tutorial. Previous work has proposed several 83
26 and how our theory could be used as an alternative way to find solutions to automatically find critical mechanics (coined “critical 84
27 mechanics for tutorial generation. mechanic discovery methods”) using uninformed [21, 22] and in- 85
28 formed [23] tree search methods and a graph of game mechanic 86
29 CCS CONCEPTS relationships. 87
30 However, player enjoyment cannot and should not be limited to a 88
• Applied computing → Computer games; • Human-centered
31 binary choice of “winning” and “losing”. Players engage with games 89
computing → Information visualization; • Mathematics of com-
32 with their own biases and motivations, which impact the enjoyment 90
puting → Nonparametric statistics.
33 they receive from play. For example, in the game Minecraft (Mojang 91
34 2007), a voxel-based open-world sandbox game, the designers may 92
35
KEYWORDS have intended for players to travel to the “End,” a dangerous zone 93
36 tutorial, player behavior, video game, mechanic, game mechanic, of monsters and treacherous terrain, to defeat the Ender Dragon, 94
37 playstyle which could act as the game’s final boss. However, many players 95
38
ACM Reference Format:
may choose to never travel to the End, instead selecting to build 96
39 Michael Cerny Green, Ahmed Khalifa, Philip Bontrager, Rodrigo Canaan, large castles and design elegant structures that are aesthetically 97
40 and Julian Togelius. 2021. Game Mechanic Alignment Theory. In FDG ’21: pleasing. In an extreme case, some players employ TNT in an explo- 98
41 Foundations of Digital Games, August 03–06, 2021, Montreal, Canada. ACM, sive attempt to blow up as much of the rendered world as possible 99
42 New York, NY, USA, 11 pages. https://doi.org/10.1145/1122445.1122456 without crashing their server 1 . Admittedly, defeating the Ender 100
43 Dragon does not explicitly end the game either, as the player can 101
44 keep on playing in their world with the ability to slay the Ender 102
1 INTRODUCTION
45 Dragon repeatedly. 103
46
A player’s first experience with a video game is often with its tuto- Elias, Gutschera and Garfield [18] distinguish between “systemic” 104
47
rial. Tutorials provide a way for a game designer to communicate and “agential” properties of game characteristics of a game. Sys- 105
48
with the player, to train them, and to help them understand the temic properties derive from the game’s rules, and agential prop- 106
49 erties depend on the players. In this work, we consider that the 107
50
Permission to make digital or hard copies of all or part of this work for personal or choice of a player to trigger a mechanic can be made through a 108
classroom use is granted without fee provided that copies are not made or distributed
51 for profit or commercial advantage and that copies bear this notice and the full citation combination of systemic and agential factors: a mechanic provides 109
52 on the first page. Copyrights for components of this work owned by others than ACM “systemic rewards” if it results in the player obtaining an external 110
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, (to the player) reward signal created by the designer to encour-
53 111
to post on servers or to redistribute to lists, requires prior specific permission and/or a
54 fee. Request permissions from permissions@acm.org. age certain behavior (e.g. points) or if it directly contributes to 112
55 FDG ’21, August 03–06, 2021, Montreal, Canada progression or winning. However, mechanics can also be pursued 113
56
© 2021 Association for Computing Machinery.
1 Example video of TNT explosion gameplay: https://www.youtube.com/watch?v=o9V2FOyD- 114
ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00
57 https://doi.org/10.1145/1122445.1122456 GM 115
58 1 116
 FDG ’21, August 03–06, 2021, Montreal, Canada Green, et al.

117 due to “agential motivations” that are internal to the player. These level faster, but it is not necessary to reach to the desired end state. 175
118 motivations may be aligned with or run against systemic rewards Another example is Pacman (Namco 1980), where eating pellets is 176
119 By categorizing mechanics within a framework that provides primary and eating ghosts is secondary. 177
120 spaces for both systemic and agential influence, we can better under- Indeed core mechanics have been defined by others [1, 31, 46], 178
121 stand game mechanics and how to teach them to players. Automated but lack a systemic perspective. For example, Salen and Zimmerman 179
122 tutorial generation [22] is a relatively unexplored artificial intelli- define them as “the essential play activity players perform again 180
123 gence application that could greatly benefit from such a framework. and again in a game (...) however, in many games, the core mechanic 181
124 Critical mechanic discovery [23], i.e. the process to automatically is a compound activity composed of a suite of actions” [46]. Jarvi- 182
125 find which mechanics to teach inside a tutorial, provides a useful nen indirectly touches upon reward systems with “the possible or 183
126 family of methods to feed an automated tutorial generator. How- preferred or encouraged means with which the player can interact 184
127 ever, previous tutorial generation methods are limited and have with game elements as she is trying to influence the game state at 185
128 drawbacks, chief among those being a reliance a complex game hand towards the attainment of a goal” [31], but still does not tell 186
129 graph of mechanical relationships similar to the graphs created by us whose goals and if the core mechanics are actually essential to at- 187
130 Machinations framework2 [16]. These game graphs require innate taining it. Although Sicart’s definitions provide precision with this 188
131 knowledge of the game-in-question and how different mechanics useful formalism, they admittedly lack the ability to classify game 189
132 relate to one another. Such analysis can be difficult without detailed mechanics outside of those oriented around environmentally de- 190
133 insight about the game elements themselves, making them difficult fined goal/reward structures. Our proposed framework attempts to 191
134 for developers to quickly deploy them. These methods also do not cover some of this undefined space, providing a method to analyze 192
135 leave room for differences between players, who may play the same mechanics through the lens of player and environment. 193
136 level differently from one another depending on their goals. 194
137 This paper presents a Game Mechanic Alignment theory, a frame- 2.2 Tutorials and Tutorial Generation 195
138 work in which mechanics can be analyzed in terms of agential Developers have experimented with multiple tutorial formats [51]. 196
139 motivations and systemic rewards. This work includes examples For simple games meant to be picked up and played quickly, me- 197
140 of Game Mechanic Alignment applied to some well-known games, chanics tend to be intuitive: “Press space to shoot”, “Press up to 198
141 a methodology that can estimate mechanic alignment given play- jump”, and so on. As a result, these games usually lacked a formal 199
142 trace data, and an application of this methodology on several video tutorial. As game complexity increased and home consoles started 200
143 games in the General Video Game Artificial Intelligence (GVGAI) to explode in popularity, formal tutorials became more common. 201
144 framework [41]. We conclude how the theory can be applied as Tutorials have since evolved to incorporate different design and 202
145 input to automated tutorial generation systems as an alternative presentation styles depending on the taste and conviction of game 203
146 method to critical mechanic discovery. designers and their perception of players [21]. Tutorials can adapt 204
147 205
to different learning capabilities of the users who use them. Sheri
148 2 GAME MECHANICS AND TUTORIALS Graner Ray [44] discusses different knowledge acquisition styles. 206
149 207
Sicart defines a “game mechanic” as an event within the game that Explorative Acquisition follows a “learning by doing” philosophy,
150 208
is fired by a game element that impacts the game’s state [47]. Me- while Modeling Acquisition, is about “reading before doing.” Green
151 209
chanics allow players to interact with and impact the game’s state. et al. [21] proposes three different presentation styles within the
152 210
Tutorials are meant to help players understand game mechanics same vein: Text, Demonstrations, and Well-Designed Experiences.
153 211
and, ultimately, learn how to play with them. Several projects have addressed challenges in automated tu-
154 212
torial generation, such as heuristic generation for Blackjack and
155 213
2.1 Game Mechanics Poker [12–14] or quest/achievement generation in Minecraft [2].
156 214
In addition to his definition of game mechanics, Sicart defines “core,” Mechanic Miner [11] evolves mechanics for 2D puzzle-platform
157
“primary,” and “secondary” mechanics, which are also relevant to games, using Reflection3 to find a new game mechanic then generate 215
158 216
this work: levels that utilize it. The Gemini system [48] takes game mechanics
159 217
as input and performs static reasoning to find higher-level mean-
160 • Core mechanics are (repetitively) used by agents to achieve 218
ings about the game. Mappy [40] can transform a series of button
161 a systemically rewarded end-state. 219
presses into a graph of room associations, transforming movement
162 • Primary mechanics are the subset of core mechanics that 220
mechanics into level information for any Nintendo Entertainment
163 can be directly applied by agents to solve challenges that 221
System game.
164 lead to a desired end-state. 222
The AtDelfi system [22] attempts to solve the challenge of au-
165 • Secondary mechanics are the subset of core mechanics that 223
tomatically generating tutorials for video games using two dif-
166 make it easier for the agent to reach the desired end-state 224
ferent formats: text-based instructions and curated GIF demon-
167 but are not essential like primary mechanics. 225
strations. This has been later expanded upon to include small
168 For example, in Super Mario Bros (Nintendo, 1985) World 1-1, Jump- sub-levels [24, 33] and entire levels in Mario [25] and 2d arcade 226
169 ing and Running are both core mechanics. Jumping is a primary games [10]. To develop these tutorials, each system requires an 227
170 mechanic as you can’t finish the level without it. Running is a input set of game mechanics, referred to as the critical mechanics: 228
171 secondary mechanic as it makes jumping easier and finishing the the set of mechanics that are necessary to trigger in order to win 229
172 230
173 2 https://machinations.io/ 3 https://code.google.com/archive/p/reflections/ 231
174 2 232
 Game Mechanic Alignment Theory FDG ’21, August 03–06, 2021, Montreal, Canada

233 the level. In addition to presenting a method to automatically find 4 GENERAL VIDEO GAME ARTIFICIAL 291
234 critical mechanics, AtDelfi also includes mechanics that give the INTELLIGENCE FRAMEWORK (GVG-AI) 292
235 player points or cause a loss. Talin [4] is a Unity-based tutorial 293
GVG-AI is a research framework for general video game play-
236 generation system which dynamically presents information to a 294
ing [41, 42], aimed at exploring the problem of creating artificial
237 player based on their skill-level. Novice players will be presented 295
players capable of playing a variety of games. Organizers host an
238 with more information, whereas experienced players will be spared 296
annual competition where AI agents are scored on their perfor-
239 unneeded tooltips. Talin differentiates itself from AtDelfi in that it 297
mance in unseen games. Each agent is given 40 milliseconds to
240 does not attempt to automatically discover which mechanics are 298
submit an action provided with a forward model for the current
241 critical, but instead which of the manually-selected mechanics need 299
game. The framework’s environment has grown over years [41],
242 to be displayed for the user’s consumption. 300
including new competition tracks such as level generation [36], rule
243 301
generation [34], learning agents [53], and two-player agents [19].
244 302
To date, the framework contains a diverse set of games numbering
245 3 PLAYER BEHAVIOR AND PLAYER over 100, including familiar titles such as Pacman (Namco 1980)
303
246 MODELING and Sokoban (Imabayashi 1981), and brand new games such as Wait
304
247 305
By choosing to expose or omit certain game mechanics, a tutorial For Breakfast.
248 306
may cater to some player’s motivations over others. In this work,
249 307
250
we suggest that automated tutorial generative systems should incor- 5 GAME MECHANIC ALIGNMENT THEORY 308
porate not only different presentation formats and learning styles, This paper presents our theory of “Game Mechanic Alignment”.
251 309
but also playstyles . Within this framework, game mechanics can be categorized in
252 310
253 terms of the system (the game) and the agent engaging with it. 311
254 Usually, games contain designer-defined reward systems which 312
255
3.1 Player Behavior are consistent regardless of who is playing. The impact these reward 313
256 Ribbens et al. wrote that studying player behavior should not oc- systems have on play and the form they take highly depends on 314
257 cur in isolation of the game environment [45]. A player cannot the conscious decisions of the designers. In a general sense, these 315
258 play without an environment to interact within, and gameplay is reward systems can be interpreted as systemic penalties versus 316
259 intimately intertwined with the particular player. systemic rewards, which usually guide the player toward winning 317
260 Gameplay environments are created and shaped by the game and away from losing. This is a separate concept from the en- 318
261 designer. Conscious design decisions can influence player behavior vironmental reward explicitly defined in reinforcement learning 319
262 and therefore the player experience [6]. But player behavior is not environments [49]. In games that do not have explicit winning con- 320
263 just influenced by the environment. Players carry their own biases ditions, such as Minecraft (Mojang, 2008) or The Sims (Maxis, 2000), 321
264 into the game that influence their in-game behavior, stemming this condition can be substituted with another win-like condition, 322
265 from individual motivations [37], culture [7], and even age [50]. By such as defeating the Ender Dragon in Minecraft, getting a job in 323
266 analyzing behaviors, one can categorize players into a taxonomy The Sims, having 𝑋 amount of happy customers in RollerCoaster 324
267 of playstyles [5, 56], each category not being mutually exclusive Tycoon (Chris Sawyer, 1999), etc. 325
268 of the rest. Artificial intelligence systems may assist with player In addition to built-in systemic rewards, player-specific incen- 326
269 behavioral analysis [30]. Our work presents a novel method to ana- tives also influence how a player engages with the game. For exam- 327
270 lyze player behavior by focusing on game mechanic usage during ple, speed runners and casual players have very different goals and 328
271 play. motivations. A speed runner may bypass, skip, or glitch their way 329
272 to the end of the game. A casual player may take their time and 330
273 explore the environment. They may even test out losing to better 331
274 3.2 Player Modeling understand how certain failure mechanics work. Thus, a player’s 332
275 Player modeling is the study of computational models of players in goals may be independent of the game’s systemic rewards. A game 333
276 games, including their incentives and behavior [55]. It is often used may reward a player when they move toward the right side of the 334
277 to study and even mimic the styles of players. This can be done using screen, but a player may want to first collect every powerup before 335
278 methods such as supervised learning using real playtraces [39, 52] moving on to the next checkpoint. In an extreme case, a player 336
279 or utility-function formulation [26–29]. Player modeling is relevant interested in exploring the losing mechanics of a game may find 337
280 to this work as AI gameplaying agents are used in place of humans themselves moving counter to the systemic rewards. If agential 338
281 for the sake of rapidly studying the efficacy of this method. incentives and systemic rewards are in agreement for a specific 339
282 Observations of human play data has been used to bias tree mechanic, we consider this mechanic to be in alignment. 340
283 search agents to play card games [15]. Outside of learning from We can think of mechanics as existing in a 2D space according 341
284 human data, several projects have demonstrated that human behav- to how they correlate to systemic reward systems and agential 342
285 ior can be mimicked by limiting computational resources [38, 57]. incentives. The extremes of the “Systemic Rewards” axis would 343
286 Khalifa et al. [35] identify another method that can be used to ma- coincide with critical and fatal mechanics as defined by previous 344
287 nipulate tree search to act more like humans. In this work, we utilize work in automated game mechanic discovery [23]: 345
288 player modeling by using gameplaying agents in lieu of human • Critical Mechanics: The set of game mechanics which must 346
289 players to demonstrate the method’s efficacy. be triggered to win, or the equivalent to winning. 347
290 3 348
 FDG ’21, August 03–06, 2021, Montreal, Canada Green, et al.

349 misaligned. The player may be motivated to take actions that cause 407
350 environmental penalties, or else refuse to take actions that would 408
351 otherwise give them rewards. It is important to note that being “in 409
352 alignment” is not a judgement of player ability, nor is it a statement 410
353 on what is “correct” behavior in the game. It is simply a way to 411
354 categorize mechanics relative to the systemic rewards of the game 412
355 and how often the player is motivated to trigger them during play. 413
356 In the following sections, we propose how this theory can be 414
357 applied in practical situations. Section 6 describes how a game 415
358 designer might apply this theory to some video game examples. 416
359 Section 7 explains a methodology which can be used to estimate 417
360 mechanic alignment using playtrace data. 418
361 419
362 420
363 6 VIDEO GAME EXAMPLES 421
364 Figure 1: Game Mechanic Alignment Theory To demonstrate the usefulness of this framework, in this section we 422
365 present examples using several well-known games: Super Mario Bros 423
366 (Nintendo 1985), Minecraft (Mojang 2009), and Bioshock (Bioware 424
367 • Fatal Mechanics: The set of game mechanics which result in 2007). For each game, we highlight aspects relevant to the discussion 425
368 losing, or the equivalent harshest environmental penalty. of player incentives and systemic rewards and provide mechanic 426
369 At the extremes of the “Agential Incentives” axis lie mechanics alignment charts as examples of how player incentives associated 427
370 that the player feels interanally incentivized to pursue or to avoid: with certain actions could differ for two hypothetical player pro- 428
371
• Incentive Mechanics: The set of mechanics that a specific files. Since these examples are simply for illustration purposes, the 429
372
player is incentivized to trigger over the course of the level values on the x and y axis are meant to be taken qualitatively. In 430
373
to accomplish that player’s subgoals. each example, we simplify the number and description of the game 431
374
• Avoidance Mechanics: the opposite of incentive mechanics. mechanics: each has many more mechanics and they may be de- 432
375
The set of mechanics that a specific player avoids trigger- scribed subjectively different depending on the individual doing 433
376
ing over the course of the level to accomplish that player’s analysis. In sections 7 and 8 we propose a method for estimating 434
377
subgoals. these values given playtraces of a game or level by many different 435
378 players. 436
379
The intersection of agential incentives and systemic reward axes 437
380
creates an origin point, aka the “neutral” zone, where the player’s be- 438
381
havior is neither influenced by systemic reward/penalty or agential 6.1 Super Mario Bros 439
motivations/aversions or only marginally influenced. The quadrant
382 While much has been said about how the Mario series teaches the 440
that a mechanic inhabits will embody the relationship that a me-
383 player through careful level design [20], a lot of the discussion 441
chanic has in terms of the environment and the player. Figure 1
384 around the series focuses on critical and fatal mechanics such as 442
shows the player incentives and systemic rewards space plotted as
385 moving to the right, collecting power ups and jumping over hazards. 443
2D axes where x-axis represents systemic rewards and the y-axis
386 These mechanics directly help the player progress towards winning 444
represents agential incentives. These two axes divide the space into
387 states or avoid losing states. However, the game also acknowledges 445
four quadrants. They are, in counter-clockwise order:
388 and incorporates in its design various agential motivations that 446
389 • Quadrant 1 (Green): Both the environmental rewards and players are likely to exhibit. 447
390 the player’s motivations encourage the player to trigger this For example, the scoring system featured in the early games of 448
391 mechanic. the franchise rewards the players both for actions that directly help 449
392 • Quadrant 2 (Yellow): The environment punishes the player the player progress and avoid losing, such as killing enemies and 450
393 but the player wants to trigger this mechanic anyways. picking up power-ups, and also for actions that display mastery 451
394 • Quadrant 3 (Red): Both the systemic rewards and the agen- over the game which have high appeal to advanced players, such as 452
395 tial motivations discourage the player from triggering this grabbing the flagpole at a higher spot and finishing the level faster. 453
396 mechanic. The collectible coins fulfill multiple roles. At first, they provide no 454
397 • Quadrant 4 (Blue): The environment rewards the player but immediate benefit other than serving as a token of player-motivated 455
398 the player wants to avoid triggering this mechanic regard- achievements such as mastering a tricky jump or uncovering secrets. 456
399 less. But once a certain number of coins are collected, the player is given 457
400 When game mechanics are within Quadrant 1 or Quadrant 3 an extra life, which helps avoid a game-over state. Finally, coins 458
401 (green and red zones), they can be considered to be “in alignment”. In and other collectibles in Mario and similar games can be used to 459
402 other words, both the environment and the player are in agreement lead the player guide the player toward secrets or suggest alternate 460
403 in regards to rewards and incentives. Perfect alignment would be at paths that may require a power-up, which Khalifa et al. [32] call 461
404 the 𝑦 = 𝑥 line. However, if mechanics lie within Quadrant 2 and/or guidance and foreshadowing, respectively. Following these cues can 462
405 Quadrant 4 (yellow and blue zones), they are player-environment lead to both systemic rewards and fulfilling player goals. 463
406 4 464
 Game Mechanic Alignment Theory FDG ’21, August 03–06, 2021, Montreal, Canada

465 523
466 524
467 525
468 526
469 527
470 528
471 529
472 530
473 531
474 532
475 533
476 534
477 535
478 536
479 537
480 538
481 539
482 540
483 541
484 542
485 543
486 544
487 Figure 3: An example of alignment axes for two different 545
488 players in Minecraft 546
489
Figure 2: An example of alignment axes for two different 547
490
players in Super Mario Bros 548
491 549
492 550
493 551
494 6.2 Minecraft 552
495 While there is no explicit in-game tutorial for these reward sys- Minecraft presents an interesting case of study for our framework 553
496 tems, taken as a whole they show that designers can take steps to since much of the appeal of the game comes from fulfilling player- 554
497 align systemic rewards with the actions that satisfy the player’s incentivized goals. While triggering the credits by beating the Ender 555
498 motivations. Figure 2 suggests a possible Agential-Systemic Me- Dragon could be considered “winning” the game in a traditional 556
499 chanical Alignment chart for the mechanics of the game and the sense, players are free to ignore this goal (potentially indefinitely) 557
500 hypothetical player profiles of “explorer” and “speedrunner”. Both and focus on exploring, mining, crafting and building structures. 558
501 players are ultimately interested in beating the level, but the ex- Figure 3 illustrates the mechanical alignment for two hypotheti- 559
502 plorer player does this in a slower-paced and safer way. The explorer cal player profiles: the “builder”, who takes enjoyment from building 560
503 enjoys eating Mushrooms for extra safety, breaking Bricks out of structures, and the “adventurer”, who seeks challenging mob en- 561
504 curiosity for what’s inside, and collecting coins that are easily acces- counters, including the Ender Dragon and other opponents found in 562
505 sible. The speedrunner, as the name suggests, wants to beat the level the Nether and the End dimensions. Both players place equal value 563
506 with the lowest possible in-game time and will avoid collectibles on crafting equipment as it enables both in reaching their goals. 564
507 and bricks unless this results in a faster clear. The speedrunner will Neither player desires to be hit, but the adventurer is more comfort- 565
508 even occasionally take damage on purpose to shrink size in order to able with the possibility, and gets hit more often as consequence of 566
509 go through narrow paths, making its y-axis value higher than the the more frequent combat encounters. 567
510 explorer player. While neither player wants to purposefully waste For the x-axis values, we assigned them from the perspective of 568
511 time, the explorer will only choose to run when it is safe to do so, the designer who wants the player to reach the credit scene and 569
512 while the speedrunner will run most of the time, even at the risk finishing the game. As you can see, all mechanics provide progress 570
513 of losing a life (and thus restarting the level and the timer). That is toward reaching this goal are on the extreme right side such as 571
514 why the speed runner has a higher value for “Run” mechanic over “Visit End/Nether,” “Mine Ores,” and of course Defeating the Ender 572
515 explorer. Dragon. While all the mechanics that hinders this progress lies 573
516 For the x-axis values, we assigned relative order for these me- on the extreme left such as “Getting hit” and “Death.” Minecraft 574
517 chanics for what we think the game designer wants the player to do contains hundreds of block types, and several different types of 575
518 to finish the game. That’s why “Touch Flag” and “Move Right” have ores. To simplify this exercise, we have elected to merge all blocks 576
519 the highest x-axis value, as they provide progress toward finishing and ores into generic mechanics. However, in theory a designer 577
520 the game, while “Losing Life” and “Shrink one size” have the lowest could make each of these their own mechanic for more granular 578
521 values on the x-axis as they prevent progress. analysis. 579
522 5 580
 FDG ’21, August 03–06, 2021, Montreal, Canada Green, et al.

581 7 COMPUTATIONAL ESTIMATION OF 639

582 MECHANIC ALIGNMENT 640
583 641
In this section, we propose a method for automatically estimating
584 642
agential incentives and systemic rewards for the game mechanics
585 643
from a certain game and level. This method uses a sizable distri-
586 644
bution of playtraces from a diverse set of artifical agents/human
587 645
players where each agent/player plays the same level multiple times.
588 646
The large group of player data allows for the influence of player
589 647
incentives to be separated out from the environment specific moti-
590 648
vations and thus mechanic alignments to be calculated for many
591 649
types of players. Similar to any statistical technique, the bigger and
592 650
more diverse the input data, the more accurate are the estimations.
593 651
Without loss of generality, the method to calculate how criti-
594 652
cal/fatal a mechanic is is fairly straightforward and intuitive. We
595 653
simply want to know how often a mechanic occurs in a winning/losing
596 654
playtrace as opposed to how often it occurs in any playtrace in
597 655
general. For example, a critical mechanic would occur in every
598 656
winning playtrace while likely occurring infrequently in losing
599 657
playtraces (winning/losing doesn’t need to be literal, it could be a
600 658
certain condition that the player needs to trigger as discussed in
601 659
section 5). The frequency of a mechanic across playtraces can be
602 660
viewed as a probability density function (PDF) that represents how
603 661
likely a mechanic is to occur in a playtrace. It then follows that the
604 Figure 4: An example of alignment axes for two different 662
difference between the PDF of the mechanic and the PDF of the
605 players in Bioshock 663
mechanic given winning/losing would give a quantitative measure
606 664
of how critical a mechanic is. This can be applied directly to player
607 665
6.3 Bioshock motivations by simply conditioning the mechanic on playtraces of
608 666
the given player. We use the Wasserstein distance [54] to calculate
609 In Bioshock, the player is faced with an important choice regard- 667
the distance between the two distributions. The following equation
610 ing the fate of characters known as Little Sisters which incurs in 668
shows the general form of the distance calculation.
611 both moral (player incentivizes) and systemic implications. When 669
612 meeting one of these genetically-modified young girls, the player 𝐷𝑚,𝑐 = 𝑊1 (𝑃𝐷𝐹 (𝑚|𝑐), 𝑃𝐷𝐹 (𝑚)) (1) 670
613 can choose to either harvest or save them. Harvesting them yields where 𝑚 is the current mechanic, 𝑐 is the current condition (𝑤𝑖𝑛 in 671
614 more ADAM, a systemic reward that serves as in-game currency for case of systemic rewards and 𝑎𝑔𝑒𝑛𝑡 in case of agential motivations), 672
615 upgrades, but results in the child’s death. Saving them yields less and 𝑊1 is the first Wasserstein distance, 𝑃𝐷𝐹 (𝑚|𝑐) is the distri- 673
616 ADAM but can create other environmental benefits and can lead bution of the mechanic 𝑚 given the condition 𝑐 happening, and 674
617 to a “happier” ending. Thus the game pits a player’s motivations 𝑃𝐷𝐹 (𝑚) is the distribution of the mechanic 𝑚 in all the playtraces. 675
618 for taking a moral path and reaching the “happy” ending against The PDF can be calculated directly from the discrete data. The 676
619 environmental considerations. frequencies for each playtrace can be normalized to form a rough, 677
620 Figure 4 illustrates a possible alignment chart for two hypothet- discrete, approximation of the true PDF. The Wasserstein distance 678
621 ical player profiles: the “morally good” player and the “thriller” can then be computed directly on this discrete data to compute the 679
622 player. Both these players have the same final goal which is beating estimated distance. This approach has the benefit of not needing 680
623 the game, but the morally good player is trying to reach it with the to fit any distribution to the data and therefore does not require 681
624 least amount of killing, while the thriller only cares about power any assumptions about the data distribution. Calculating the result 682
625 and destruction. That is why they differs on the “Little Sisters”, the over a PDF also would allow for this approach to be used when the 683
626 thriller since it needs more power, it is highly motivated on killing PDF is known through other means. 684
627 the sisters as they provide too much power compared to saving The distance proposed in equation 1 is a scalar with value be- 685
628 them. On the other hand, the morally good player don’t care about tween 0 and 1, and doesn’t expression direction in the space. To 686
629 the power but feels bad about killing these little girls (even if they calculate the direction, we compare the mean of 𝑃𝐷𝐹 (𝑚). If the 687
630 are not real) and they might be motivated to reach the good ending conditional distribution has a higher average, it means it is more en- 688
631 of the game. couraged to happen, while if it has lower average, it is discouraged 689
632 Similarly to all the previous games, we sort the mechanics on from happening. 690
633 the x-axis such as mechanics near the right are pushing the player 691
634 to progress in the game such as “Shooting Enemies” and “Hacking 𝑆𝑚,𝑐 = 𝑆𝑖𝑔𝑛(𝜇𝑃𝐷𝐹 (𝑚 |𝑐) − 𝜇𝑃𝐷𝐹 (𝑚) ) (2) 692
635 Turrets”. While, mechanics near the left blocks that progress and To calculate the final rewards, Systemic reward (𝐸𝑚 ) and Agential 693
636 cause the player to replay certain areas such as “Death” and “Getting reward (𝐼𝑚 ), we combine both equations 2 and 1 to get the final 694
637 hit”. values. Equation 3 shows the final equation for both rewards where 695
638 6 696
 Game Mechanic Alignment Theory FDG ’21, August 03–06, 2021, Montreal, Canada

697 the difference is the agential incentive estimation is conditioned randomly, it is impossible for an agent to accurately predict if it will 755
698 on a certain agent(s) are playing, while the systemic reward is move and in what direction at any particular game tick. Slaying 756
699 condition on having a winning playtraces. monsters requires the player to risk themselves by being right next 757
700
𝐼𝑚 = 𝑆𝑚,𝑎𝑔𝑒𝑛𝑡 · 𝐷𝑚,𝑎𝑔𝑒𝑛𝑡 to one. 758
701 (3) Adrienctx (circle) far outperforms other agents when it comes 759
𝐸𝑚 = 𝑆𝑚,𝑤𝑖𝑛 · 𝐷𝑚,𝑤𝑖𝑛
702 to collecting the key and unlocking the door. In fact, Adrienctx 760
703 appears to be aligned along 𝑦 = 𝑥, and its mechanics are in perfect 761
8 EXPERIMENTAL EVALUATION
704 agential-systemic alignment. Greedy Search (diamond) seems more 762
705 We ran a diverse set of 26 agents on 3 game levels from the General inclined to chase after and slay monsters than the other agents, 763
706 Video Game Artificial Intelligence Framework (GVGAI). The set while the Do Nothing (cross) agent does nothing at all. The MCTS 764
707 of agents come with the public version of the framework. One (square) agent is the least inclined to slay monsters while being 765
708 of these game levels is 𝑍𝑒𝑙𝑑𝑎, a demake of The Legend of Zelda often slain by them. 766
709 (Nintendo 1986) dungeon system. The other is 𝑃𝑎𝑐𝑚𝑎𝑛, a port of We can use this graph to analyze the agent’s gameplay perfor- 767
710 Pacman (Namco 1980). The last one is 𝐵𝑢𝑡𝑡𝑒𝑟 𝑓 𝑙𝑖𝑒𝑠, which is a new mance like a human player: 768
711 game where the player tries to collect all the butterflies before all 769
the cocoons hatches. Figure 5 shows the levels used from each of • Adrienctx understands the game mechanics very well. They
712 770
these three games. All 26 agents play each game 100 times for a total could take more risks and kill more monsters to get a higher
713 771
of 2600 playtraces per game level. We recorded every mechanic score. It could be that they enjoy the maze side of the game
714 772
triggered by an agent, such as movement, collisions, and (in the case and would enjoy more challenging levels with complex wall
715 773
of Zelda) swinging a sword. Presented below are the results of 4 of patterns
716 774
these agents, which we believe best showcase a variety of playstyles • MCTS knows to collect the key but not what to do with
717 775
to demonstrate the efficacy of our method: Adrienctx, Monte Carlo it after. They also do not seem inclined to slay monsters,
718 776
Tree Search (MCTS), a Greedy Search agent, and Do Nothing, which although they seem to avoid them well enough. They might
719 777
always takes the same action (neutral). These 4 agents are sorted need an easier level.
720 778
based on their performance on these games [8]. Adrienctx uses • Greedy Search could benefit from learning to collect the key
721 779
Open Loop Expectimax Tree Search (OLETS) algorithm to play and unlock the door, but they seem to be heavily motivated
722 780
the games, it is also a previous winner of the planning track in to slay monsters. They might enjoy levels that contain large
723 781
the GVGAI competition [43]. MCTS algorithm comes with the quantities of monsters for a greater challenge.
724 782
GVGAI framework where it is a vanilla implementation of Upper • Do Nothing is entirely out of alignment with the systemic
725 783
Confidence Bounds of Trees (UCT) algorithm [9]. Greedy Search rewards. We could assume that they are purposefully dying
726 784
agent looks only for one step ahead and pick the action that will a lot. But most likely this is because the player may not
727 785
either make it win or increase the score. Finally, Do Nothing agent, understand the game rules. Perhaps they should be given
728 786
as its name implies, just stands still without executing any actions the opportunity to play a very simple series tutorial levels
729 787
until it dies or the game times out. that teaches them the basic mechanics.
730 788
731 789
732 8.1 Zelda 8.2 Butterflies 790
733 The goal of Zelda is to collect a key and unlock the door on the far Butterflies is considered a “deceptive game” [3], i.e. one where the 791
734 side of the level (as seen in figure 5a). Along the way the player reward structure is designed to lead a player away from globally 792
735 encounters monsters, who can destroy the player if they collide. The optimal decisions. The goal of butterflies is to clear the level of all 793
736 player can swing a sword in front of them, destroying any monster butterflies, which fly around randomly (figure 5b). More butterflies 794
737 it touches. Monsters move randomly around the level every couple can be spawned from cocoons, which crack open after a butterfly 795
738 of game ticks. Figure 6 displays the mechanical alignment of the touches it. However, the player will lose if no more cocoons are in 796
739 four agents. the level. Therefore, the optimal strategy is to let all but one cocoon 797
740 “Unlock Door” and “Collect Key” are the two most systemically open to spawn as many butterflies as possible and collect them all. 798
741 rewarding mechanics, which makes sense considering Zelda’s na- Figure 7 displays the mechanical alignment of the four agents. 799
742 ture as an adventure game. “Slay Monster” follows behind them as Due to the deceptive nature of the game, two of the mechanics 800
743 also positively rewarding. Interestingly enough, “Space Bar” (when (cocoons bursting and spawning butterflies) are heavily associated 801
744 the player presses the space bar) and “Swing Sword” are negatively with losing. Remember that if all the cocoons of the game pop, the 802
745 rewarding. The discrepancy between the two mechanic scores can player will lose. Most agents do not to engage with the risk-reward 803
746 be explained if we consider the frequency that these mechanics are of spawning more butterflies for a higher score, and therefore co- 804
747 triggered. The sword can only be swung every couple of seconds coons do not open nearly as often in winning playtraces. Collecting 805
748 due to the frame animation for the attack. Thus pressing space and a butterfly is slightly positively associated, as it happens slightly 806
749 swinging a sword are counted as separate mechanics, as pressing more in winning playtraces. 807
750 the space bar does not guarantee the sword will be swung. These Adrienctx is in near perfect alignment: while “catch butterflies” 808
751 mechanics are scored negatively in the environmental measure- is just under the x axis, it tries to catch all the butterflies as soon 809
752 ment. Agents who chase after monsters to slay them tend to also as possible which make it catch less butterflies overall compared 810
753 die from the stochastic nature of the game. Because monsters move to the MCTS agent. MCTS collects a lot of butterflies as it allow 811
754 7 812
 FDG ’21, August 03–06, 2021, Montreal, Canada Green, et al.

813 871
814 872
815 873
816 874
817 875
818 876
819 877
820 878
(a) Zelda (b) Butterflies (c) Pacman
821 879
822 880
823
Figure 5: GVGAI game levels used to test our estimation method. 881
824 882
825 883
826 884
827 885
828 886
829 887
830 888
831 889
832 890
833 891
834 892
835 893
836 894
837 895
838
Figure 6: The Agential-Systemic Mechanical Alignment Figure 8: The Agential-Systemic Mechanical Alignment 896
839
graph for 4 agents on GVGAI’s Zelda. graph for 4 agents on GVGAI’s Pacman. 897
840 898
841 899
842
Similar to Zelda, we could look at the graph and analyze these 900
843
agents like a human player: 901
844 • Adrienctx appears to not let many cocoons burst open, which 902
845 limits its ability to score high. They may be trying to play 903
846 it safe and not risk losing the game. They also might not 904
847 understand that by waiting a little longer before catching all 905
848 the butterflies so they have higher score. They might benefit 906
849 from a prompt reminding them that cocoons can be good in 907
850 the short term. 908
851 • MCTS lets lots of cocoons burst open for a greater challenge. 909
852 They may be more risk taking and not afraid of losing. They 910
853 may appreciate playing larger levels with lots of cocoons. 911
854 • Neither Greedy Search nor Do Nothing seem to understand 912
855 the mechanics of the game and could benefit greatly from 913
856 Figure 7: The Agential-Systemic Mechanical Alignment an easier level and a tutorial. 914
857 graph for 4 agents on GVGAI’s Butterflies. 915
858 8.3 Pacman 916
859 The Pacman level is made up of a 2-dimensional maze, which the 917
860 cocoons to burst into new butterflies (that is why that “Cocoon player traverses while collecting all the pellets and fruit in the level 918
861 spawns butterflies” and “Butterfly Destory Cocoon” are slightly (figure 5c). Four deadly ghosts chase the player in the maze and 919
862 positive on the y-axis). This behavior of MCTS is probably due to must be avoided. Some of the pellets are “power pellets” which 920
863 being less efficient compared to Adrienctx in collecting butterflies. grant the player a temporary invulnerability to the ghosts, allowing 921
864 Because it does not collect them as quickly, more butterflies causes the player to eat them and force them to respawn in the center of 922
865 more cocoons to pop, creating even more butterflies and allowing the maze. Figure 8 displays the mechanical alignment of the four 923
866 the agent to get a higher score. The rest of the agents (“Do Nothing” agents. 924
867 and “Greedy Search”) do not have any mechanics in alignment (all Eating fruit pellets, and power-pellets are the highest systemi- 925
868 of their cocoons would eventually pop and they would lose the cally rewarding mechanics of the game, which is in line with its 926
869 game). winning conditions. Eating ghosts are slightly positively rewarding, 927
870 8 928
 Game Mechanic Alignment Theory FDG ’21, August 03–06, 2021, Montreal, Canada

929 suggesting that this is not as aligned with winning and could be noisier data that the method can distinguish well between various 987
930 avoided to win. Getting eaten by a ghost results in the player losing, players. 988
931 so it makes sense that it is heavily penalizing. Another limitation of the estimation is that it relies on the corre- 989
932 Adrienctx seems to understand the basic premise of Pacman, lation between triggering a mechanic and winning the game (for 990
933 in that it must eat the pellets and fruit to win. However, all four the x axis) or between triggering a mechanic and belonging to a 991
934 agents seem nearly equivalent in getting eaten by ghosts, suggest- certain player profile (for the y axis). These correlations are not 992
935 ing that none of them can reliably escape losing relative to each necessarily causal/intentional and could be affected by spurious 993
936 other. This is supported by the fact that there are only 9 winning factors. One such factor is skill. Consider a hypothetical level where 994
937 playtraces out of the 2600. Adrienctx is relatively more capable of a cosmetic item with no functional value is hidden near the exit. 995
938 collecting power-pellets and proceeding to eat ghosts. DoNothing, The event of picking up this item would have a high correlation 996
939 on the other hand, is incapable of performing any of these actions. with winning, and therefore would show up far to the right in the 997
940 GreedySearch is only slightly better than DoNothing in this regard. x-axis, even though it does not improve a player’s chance to beat 998
941 MCTS, however, seems to be capable of engaging in these aspects the level. And if two player profiles A and B are equally internally 999
942 of the game, however not nearly as effectively as Adrienctx. incentivized to pick up the item, but profile A reaches the end of 1000
943 Looking at these agents as human players, we could deduce the the level more often, the event would show up higher on the y-axis 1001
944 following: for profile A, as profile A simply had more opportunities to trigger 1002
945 it due to its higher player skill. This can be seen in our examples, 1003
946 • Adrienctx’s incentives are in alignment with the systemic where all the automated agents except do-nothing are designed 1004
947 rewards of Pacman. They seem to lose nearly as often as with the goal to win the level but have different skill levels, causing 1005
948 other players though, so they are might playing levels that differences in the y-axis even with no discernible differences in 1006
949 match their skill level. motivation. 1007
950 • MCTS does not seem to be nearly as successful in collecting The Talin [4] system arguably touches upon this more directly. 1008
951 pellets or fruits. They might need less ghosts or maybe more In Talin, mechanics are dynamically taught to the player based on 1009
952 power pellets. their skill level. Mechanic mastery is represented using scalar val- 1010
953 • Greedy Search and Do Nothing is almost entirely out of ues initialized by the game designer. As the player plays and either 1011
954 alignment. It could be that they may not understand the triggers or fails to trigger that mechanic, its value rises or falls. A 1012
955 mechanics of the game. It could also be that they are in- similar system could be used to measure player skill/mastery of 1013
956 centivized to perform some other activity in the game, like particular mechanics to explain how much of a y-value difference 1014
957 leading the ghosts around in a specific pattern or seeing how between different players is a result of their skill versus incentives. 1015
958 long they can survive without power pellets. If we can be sure of the method’s accuracy, it may be more benefi- 1016
959 cial to designate another axis to account for player skill. However, 1017
960 we want to make it clear that being unable to differentiate between 1018
961 9 DISCUSSION incentive or skill when it comes to player behavior does not negate 1019
962 Game Mechanic Alignment allows us to visualise how different the utility of this technique. Two experiences could be offered to 1020
963 players engage with various game mechanics according to their the player to counter this: one being a tutorial giving an easier level 1021
964 internal incentives and systemic rewards. This has many potential to help “practice”, the other giving the player a more challenging 1022
965 applications, such as categorizing players based on the y-value of level with more complex uses of the mechanic to provide entertain- 1023
966 different events (e.g. through clustering), validating a designer’s ment. Regardless of it being due to skill or incentives, the game 1024
967 assumptions of what features of the game will be most enjoyed by system/designer has a deeper understanding of what the player 1025
968 players, building reward systems that are aligned (to the desired was doing during play than without the technique. 1026
969 extent) with player’s preferences and building tutorials that support Idiosyncrasies of a particular game or level could also contribute 1027
970 actions preferred by different types of players, even if they are not to spurious correlations. For example, if another hypothetical level 1028
971 all equally rewarded by the reward systems. contains a bifurcation where the player can choose between a red 1029
972 It is important, however, to distinguish between the theoretical path heading North and a blue path heading South, and these paths 1030
973 notion that agential incentives might be more or less aligned with feature opportunities to trigger different sets of events, then a naive 1031
974 systemic rewards and the quantitative methods used to estimate analysis of the prevalence of these events might be confounded 1032
975 this alignment. The first, and most obvious limitation of the quanti- by an aesthetic preference for the red or blue color (or, as is more 1033
976 tative estimation done in this paper is that we use artificial agents as likely for bots, by an arbitrary tie-breaker between the North and 1034
977 proxies for what a player’s behavior might look like. More accurate South directions). 1035
978 estimations could be done with either human play traces or using Level difficulty also may play a role here. In our examples, all 1036
979 artificial agents that attempt to emulate a specific persona. How- agents were compared on the same level. Therefore, we can be 1037
980 ever, these agents are suitable proxies for this experiment, as they assured that difficulty remains constant. However, many games 1038
981 are intentionally built to utilize different incentive systems during present the player with procedurally generated sets of levels, such as 1039
982 play (Adrienctx uses Hierarchical Open-Loop Optimistic Planning, Spelunky (Derek Yu, 2008). In games like these, a designer may not 1040
983 MCTS uses a systemic-reward biased tree, Greedy has no search be able to make the assumption that all levels are equally difficult. 1041
984 and takes the best immediate action, etc). By using agents which If we assume there exists a method for measuring level difficulty, 1042
985 were not emulating a specific persona, we have demonstrated with 1043
986 9 1044
 FDG ’21, August 03–06, 2021, Montreal, Canada Green, et al.

1045 then we could test our agents only against the levels with similar in similar manner to detecting highlights in soccer matches [17], 1103
1046 difficulty and the same set of game mechanics. or by using a method of automated mechanic discovery [11]. 1104
1047 It is possible, in principle, to attempt to correct for spurious 1105
1048 correlations. For example, if we know an event can only be triggered 10 CONCLUSION 1106
1049 at a certain point of the map, we could consider only play traces In this work, we present Game Mechanic Alignment theory, a frame- 1107
1050 that got within a certain distance of that point. Or we could, at each work which enables designers to organize game mechanics in terms 1108
1051 time-step of the play trace, simulate the game for a number of steps of the environment and a player engaging with it. We provide sev- 1109
1052 with the goal of triggering the event, and consider only play traces eral well-known games as examples for how a designer may apply 1110
1053 where it was possible to do so. We could also perform A/B tests this theory. To demonstrate its practicality, we propose a method- 1111
1054 where some features of the level (e.g. the events on each path) are ology to estimate reward values for Game Mechanic Alignment, 1112
1055 kept constant where others (e.g. the colors) are swapped. But these as well as an experimental evaluation using 3 games from the GV- 1113
1056 corrections would come at the cost of domain knowledge and/or GAI framework. We then point out shortcomings in regards to this 1114
1057 computational power, which is why they were not considered in methodology while discussing several ways they may be overcome. 1115
1058 our experiments. By taking both player and environment into account, tutorial 1116
1059 All of our examples in Sections 5 and 8 refer to games with generators can create highly personalized tutorials or experiences. 1117
1060 mechanics that have either immediate or delayed rewards/penalties. Rather than one tutorial for a single game or level, each player or 1118
1061 Let us consider the game Fallout 2 (Bethesda 1998), a first person playstyle could receive their unique tutorial. Players who are highly 1119
1062 shooter set in a post-apocalyptic Oregon, United States. In the skilled can be introduced to more complex mechanics whereas 1120
1063 game, the player may encounter a consumable substance called novice players can be given explanations of basic controls. Further- 1121
1064 “Jet,” a highly addictive meta-amphetamine which grants the player more, players with certain playstyles can be introduced to com- 1122
1065 temporary short-term bonuses to their combat abilities. After the pletely different styles, they might not have considered. Proposed 1123
1066 initial bonus period, however, the player will suffer heavy penalties future works are to validate this theory with real human players and 1124
1067 to these same skills. Additionally, the player may become addicted automatically design tutorials using this methodology to organize 1125
1068 to the substance, requiring them to keep dosing themselves every input, first for gameplaying agents using the same methodology 1126
1069 day or suffer additional penalties. In this example, consuming Jet from Section 8 and then for human players in a formal user study. 1127
1070 has opposing systemic reward/penalties, depending on the time 1128
1071 horizon: rewarding in the short-term, penalizing in the long-term. ACKNOWLEDGMENTS 1129
1072 To visualize this using Game Mechanic Alignment, we may need 1130
Michael Cerny Green would like to thank the OriGen.AI education
1073 to introduce a third axis to represent “time.” 1131
program for their financial support. Rodrigo Canaan gratefully
1074 One of our core motivations for this work was to find a new 1132
acknowledges the financial support from Honda Research Institute
1075 alternative method to critical mechanic discovery [23] for auto- 1133
Europe (HRI-EU).
1076 mated tutorial generation systems that considers the player and 1134
1077 the environment. While this method does not discover “critical me- 1135
REFERENCES
1078 chanics” as defined by Green et al, it does estimate the tendencies of 1136
[1] Ernest Adams and Andrew Rollings. 2007. Game design and development: Fun-
1079 the player to gravitate toward particular mechanic usage/triggers damentals of game design. New Jersey: Pearse Prentice Hall (2007). 1137
1080 (or lack there-of). The concept of critical mechanics also neglects [2] Ryan Alexander and Chris Martens. 2017. Deriving quests from open world 1138
mechanics. In Foundations of Digital Games. ACM, 12.
1081 to take different playstyles into account, as the definition allows [3] Damien Anderson, Matthew Stephenson, Julian Togelius, Christoph Salge, John
1139
1082 for only a single set of critical mechanics to exist for a given level Levine, and Jochen Renz. 2018. Deceptive games. In International Conference on 1140
1083 (only those needed to win). Consider the following use case: by the Applications of Evolutionary Computation. Springer, 376–391. 1141
[4] Batu Aytemiz, Isaac Karth, Jesse Harder, Adam Smith, and Jim Whitehead. 2018.
1084 building a variety of artificial agents with different utility func- Talin: A Framework for Dynamic Tutorials Based on the Skill Atoms Theory. In
1142
1085 tions (and therefore explicit agential incentives), a designer can Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital 1143
1086 quickly obtain a large collection of playtraces. Each agent would Entertainment, Vol. 14. 1144
[5] Richard A Bartle. 2004. Designing virtual worlds. New Riders.
1087 represent a potential “playstyle.” Then, any single human user’s [6] Kelly Bergstrom, Marcus Carter, Darryl Woodford, and Chris Paul. 2013. Con- 1145
1088 playtrace can be compared against the entirety of agent traces and structing the ideal EVE online player. Proceedings of DiGRA 2013: DeFragging 1146
Game Studies. (2013).
1089 the designer would know which archetype category(ies) best fit the [7] Mateusz Bialas, Shoshannah Tekofsky, and Pieter Spronck. 2014. Cultural influ-
1147
1090 player. This player could then be served a experience personalized ences on play style. In 2014 IEEE Conference on Computational Intelligence and 1148
1091 for that playstyle. This experience can be automatically built us- Games. IEEE, 1–7. 1149
[8] Philip Bontrager, Ahmed Khalifa, Andre Mendes, and Julian Togelius. 2016. Match-
1092 ing the information gleaned by Game Mechanic Alignment about ing games and algorithms for general video game playing. In Proceedings of the
1150
1093 that playstyle by creating or removing opportunities for triggering AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 1151
1094 certain sets of mechanics. Vol. 12. 1152
[9] Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I
1095 The mechanic-graph restriction of the previously mentioned Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samoth- 1153
1096 critical mechanic discovery methods [22, 23] is not a requirement rakis, and Simon Colton. 2012. A survey of monte carlo tree search methods. 1154
IEEE Transactions on Computational Intelligence and AI in games 4, 1 (2012), 1–43.
1097 for the Game Mechanic Alignment method. Therefore, users might [10] Megan Charity, Michael Cerny Green, Ahmed Khalifa, and Julian Togelius. 2020.
1155
1098 find it more generalized and easier to use. In our experimental Mech-Elites: Illuminating the Mechanic Space of GVG-AI. In International Con- 1156
1099 section, we used the output logs from agent gameplay in the GVGAI ference on the Foundations of Digital Games. 1–10. 1157
[11] Michael Cook, Simon Colton, Azalea Raad, and Jeremy Gow. 2013. Mechanic
1100 framework. However, mechanics could be automatically detected miner: Reflection-driven game mechanic discovery and level design. In European
1158
1101 Conference on the Applications of Evolutionary Computation. Springer, 284–293. 1159
1102 10 1160
 Game Mechanic Alignment Theory FDG ’21, August 03–06, 2021, Montreal, Canada

1161 [12] Fernando de Mesentier Silva, Aaron Isaksen, Julian Togelius, and Andy Nealen. dilemma. Journal of personality and social psychology 32, 5 (1975), 922. 1219
1162 2016. Generating heuristics for novice players. In Computational Intelligence and [38] Mark Nelson. 2016. Investigating vanilla MCTS scaling on the GVG-AI game 1220
Games. IEEE, 1–8. corpus. In Proceedings of the IEEE Conference on Computational Intelligence and
1163 [13] Fernando de Mesentier Silva, Julian Togelius, Frank Lantz, and Andy Nealen. Games. IEEE. 1221
1164 2018. Generating Beginner Heuristics for Simple Texas Hold’em. In Genetic and [39] Juan Ortega, Noor Shaker, Julian Togelius, and Georgios N Yannakakis. 2013. 1222
1165
Evolutionary Computation Conference. ACM. Imitating Human Playing Styles in Super Mario Bros. Entertainment Computing 1223
[14] Fernando de Mesentier Silva, Julian Togelius, Frank Lantz, and Andy Nealen. 2018. 4, 2 (2013), 93–104.
1166 Generating Novice Heuristics for Post-Flop Poker. In Computational Intelligence [40] Joseph Osborn, Adam Summerville, and Michael Mateas. 2017. Automatic map- 1224
1167 and Games. IEEE. ping of NES games with mappy. In Foundations of Digital Games. ACM, 78. 1225
[15] Sam Devlin, Anastasija Anspoka, Nick Sephton, Peter I Cowling, and Jeff Rollason. [41] Diego Perez-Liebana, Jialin Liu, Ahmed Khalifa, Raluca D Gaina, Julian Togelius,
1168 1226
2016. Combining Gameplay Data With Monte Carlo Tree Search To Emulate and Simon M Lucas. 2019. General video game ai: a multi-track framework for
1169 Human Play. In Proceedings of the AAAI Artificial Intelligence for Interactive Digital evaluating agents, games and content generation algorithms. Transactions on 1227
1170 Entertainment Conference. AAAI. Games (2019). 1228
[16] Joris Dormans. 2011. Simulating mechanics to study emergence in games. In [42] Diego Perez-Liebana, Spyridon Samothrakis, Julian Togelius, Tom Schaul, and
1171 Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Simon M Lucas. 2016. General video game ai: Competition, challenges and 1229
1172 Entertainment, Vol. 7. opportunities. In AAAI Conference on Artificial Intelligence. 1230
1173
[17] Ahmet Ekin, A Murat Tekalp, and Rajiv Mehrotra. 2003. Automatic soccer video [43] Diego Perez-Liebana, Spyridon Samothrakis, Julian Togelius, Tom Schaul, Si- 1231
analysis and summarization. IEEE Transactions on Image processing 12, 7 (2003). mon M Lucas, Adrien Couëtoux, Jerry Lee, Chong-U Lim, and Tommy Thompson.
1174 [18] George Skaff Elias, Richard Garfield, and K Robert Gutschera. 2012. Characteristics 2015. The 2014 general video game playing competition. IEEE Transactions on 1232
1175 of games. MIT Press. Computational Intelligence and AI in Games 8, 3 (2015), 229–243. 1233
[19] Raluca D Gaina, Diego Pérez-Liébana, and Simon M Lucas. 2016. General video [44] Sheri Graner Ray. 2010. Tutorials: learning to play. http://www.gamasutra.com/
1176 1234
game for 2 players: Framework and competition. In Computer Science and Elec- view/feature/134531/tutorials_learning_to_play.php?print=1.
1177 tronic Engineering. IEEE, 186–191. [45] Wannes Ribbens and Yorick Poels. 2009. Researching player experiences through 1235
1178 [20] Euro Gamer. 2015. Miyamoto on World 1-1: How Nintendo made Mario’s most the use of different qualitative methods.. In DiGRA Conference. 1236
iconic level. https://www.youtube.com/watch?v=zRGRJRUWafY. Last Accessed: [46] Katie Salen, Katie Salen Tekinbaş, and Eric Zimmerman. 2004. Rules of play:
1179 February 1, 2021. Game design fundamentals. MIT press. 1237
1180 [21] Michael Green, Ahmed Khalifa, Gabriella Barros, and Julian Togellius. 2017. " [47] Miguel Sicart. 2008. Defining game mechanics. Game Studies 8, 2 (2008), n. 1238
1181
Press Space to Fire": Automatic Video Game Tutorial Generation. In Proceedings of [48] Adam Summerville, Chris Martens, Sarah Harmon, Michael Mateas, Joseph Carter 1239
the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, Osborn, Noah Wardrip-Fruin, and Arnav Jhala. 2017. From Mechanics to Meaning.
1182 Vol. 13. AAAI. IEEE Transactions on Computational Intelligence and AI in Games. 1240
1183 [22] Michael Cerny Green, Ahmed Khalifa, Gabriella AB Barros, Tiago Machado, Andy [49] Richard S Sutton and Andrew G Barto. 2018. Reinforcement learning: An intro- 1241
Nealen, and Julian Togelius. 2018. AtDELFI: automatically designing legible, full duction. MIT press.
1184 1242
instructions for games. In Proceedings of the 13th International Conference on the [50] S. Tekofsky, P. Spronck, A. Plaat, J. van den Herik, and J. Broersen. 2013. Play
1185 Foundations of Digital Games. ACM, 1–10. style: Showing your age. In 2013 IEEE Conference on Computational Inteligence in 1243
1186 [23] Michael Cerny Green, Ahmed Khalifa, Gabriella AB Barros, Tiago Machado, and Games (CIG). IEEE, 1–8. https://doi.org/10.1109/CIG.2013.6633616 1244
Julian Togelius. 2020. Automatic Critical Mechanic Discovery Using Playtraces [51] Carl Therrien. 2011. " To Get Help, Please Press X" The Rise of the Assistance
1187 in Video Games. In International Conference on the Foundations of Digital Games. Paradigm in Video Game Design.. In DiGRA Conference. 1245
1188 ACM, 1–9. [52] Julian Togelius, Renzo De Nardi, and Simon M Lucas. 2007. Towards Automatic 1246
1189
[24] Michael Cerny Green, Ahmed Khalifa, Gabriella AB Barros, Andy Nealen, and Personalised Content Creation for Racing Games. In Proceedings of the IEEE 1247
Julian Togelius. 2018. Generating levels that teach mechanics. In Proceedings of Conference on Computational Intelligence and Games. IEEE.
1190 the 13th International Conference on the Foundations of Digital Games. 1–8. [53] Ruben Rodriguez Torrado, Philip Bontrager, Julian Togelius, Jialin Liu, and Diego 1248
1191 [25] Michael Cerny Green, Luvneesh Mugrai, Ahmed Khalifa, and Julian Togelius. Perez-Liebana. 2018. Deep Reinforcement Learning for General Video Game AI. 1249
2020. Mario level generation from mechanics using scene stitching. In 2020 IEEE In Computational Intelligence and Games. IEEE, 1–8.
1192 1250
Conference on Games (CoG). IEEE, 49–56. [54] Leonid Wasserstein. 1969. Markov processes with countable state space describing
1193 [26] Christoffer Holmgård, Michael Cerny Green, Antonios Liapis, and Julian Togelius. large systems of automata. IEEE Transactions on Image processing 5, 3 (1969). in 1251
1194 2018. Automated playtesting with procedural personas through MCTS with Russian. 1252
evolved heuristics. (2018). [55] Georgios N. Yannakakis, Pieter Spronck, Daniele Loiacono, and Elisabeth André.
1195 [27] Christoffer Holmgård, Antonios Liapis, Julian Togelius, and Georgios N. Yan- 2013. Player Modeling. In Artificial and Computational Intelligence in Games. 1253
1196 nakakis. 2014. Evolving Personas for Player Decision Modeling. In Proceedings of Dagstuhl Publishing, Saarbrücken/Wadern, 45–55. 1254
1197
the IEEE Conference on Computational Intelligence and Games. IEEE. [56] Nicholas Yee. 2002. Facets: 5 motivation factors for why people play MMORPG’s. 1255
[28] Christoffer Holmgård, Antonios Liapis, Julian Togelius, and Georgios N Yan- Terra Incognita 1 (2002), 1708–1723.
1198 nakakis. 2014. Personas versus clones for player decision modeling. In Proceedings [57] Alexander Zook, Brent Harrison, and Mark O Riedl. 2015. Monte-Carlo Tree 1256
1199 of the International Conference on Entertainment Computing. Springer, 159–166. Search for Simulation-based Strategy Analysis. In Proceedings of the Foundations 1257
[29] Christoffer Holmgård, Antonios Liapis, Julian Togelius, and Georgios N. Yan- of Digital Games Conference. ACM.
1200 1258
nakakis. 2015. Monte-Carlo Tree Search for Persona Based Player Modeling. In
1201 Proceedings of the AIIDE workshop on Player Modeling. 1259
1202 [30] Britton Horn, Amy K Hoover, Yetunde Folajimi, Jackie Barnes, Casper Harteveld, 1260
and Gillian Smith. 2017. AI-assisted analysis of player strategy across level
1203 progressions in a puzzle game. In Proceedings of the 12th International Conference 1261
1204 on the Foundations of Digital Games. ACM, 1–10. 1262
1205
[31] Aki Järvinen. 2008. Games without frontiers: Theories and methods for game studies 1263
and design. Tampere University Press.
1206 [32] Ahmed Khalifa, Fernando de Mesentier Silva, and Julian Togelius. 2019. Level 1264
1207 Design Patterns in 2D Games. In 2019 IEEE Conference on Games (CoG). IEEE, 1265
1–8.
1208 1266
[33] Ahmed Khalifa, Michael Cerny Green, Gabriella Barros, and Julian Togelius.
1209 2019. Intentional computational level design. In Proceedings of The Genetic and 1267
1210 Evolutionary Computation Conference. 796–803. 1268
[34] Ahmed Khalifa, Michael Cerny Green, Diego Perez-Liebana, and Julian Togelius.
1211 2017. General video game rule generation. In 2017 IEEE Conference on Computa- 1269
1212 tional Intelligence and Games (CIG). IEEE, 170–177. 1270
1213
[35] Ahmed Khalifa, Aaron Isaksen, Julian Togelius, and Andy Nealen. 2016. Modify- 1271
ing mcts for human-like general video game playing. In Proceedings of IJCAI.
1214 [36] Ahmed Khalifa, Diego Perez-Liebana, Simon M Lucas, and Julian Togelius. 2016. 1272
1215 General video game level generation. In Genetic and Evolutionary Computation 1273
Conference. ACM, 253–259.
1216 1274
[37] D Michael Kuhlman and Alfred F Marshello. 1975. Individual differences in
1217 game motivation as moderators of preprogrammed strategy effects in prisoner’s 1275
1218 11 1276