Procedural content generation (PCG) algorithms pervade the world of game development. They are useful tools that allow creators to unshoulder a degree of authorial burden of non-trivial virtual worlds. Such a goal necessitates a welldefined m odel o f w hat t he a lgorithm i s p roducing. I n this work, a PCG system called TileTerror is presented that uses a model of horror constructed from a ludological analysis of well-known horror games to embed horror story features in a procedurally generated 2D tilemap. The horror features found in the ludology are separated into a distinct hierarchy: low-level features and high-level features. These features are used in conjunction with a procedurally generated tilemap to create an annotated map denoting candidate locations for strong horror moments. This system is evaluated using a series of pathfinding agents which score the annotated tilemap. TileTerror represents a first step into exploring what story elements can be embedded in procedurally generated constructs like 2D tilemaps.
From fully-generated simulations with Dwarf Fortress
PCG techniques are necessary for generating non-trivial virtual worlds. We say this to focus on how each game treats design differently; for example, a platformer is concerned with a different set of design goals than a shooter game. In turn, the virtual worlds generated by PCG algorithms are only as good as the model, and how well the model maps to the original design constraints and goals. In a similar vein, our work attempts to build cohesive worlds by modeling the horror genre for the procedural generation of annotated 2D tilemaps.
The horror genre has received academic attention from
film and literary studies to interactive experiences
Thus, this paper contributes a PCG system that generates 2D tilemaps and evaluates their opportunities for horror. First, this paper describes a domain decomposition of horror aspects in video games, which features level design models that promote specific horror types a player experiences. Next, this paper describes the system architecture of TileTerror, which consists of two major parts: a preprocessing algorithm and a generator. Finally, five pathfinding agents analyze the tilemap for evaluation. Each pathfinding agent maps to a possible player type, inspired by the persona evaluations by Holmga˚rd et al. (2019). Our system offers a unique bottom-up approach to this daunting task and an initial evaluation system for our generated artifacts.
From the drama manager of Left 4 Dead
Another inspiration for this work is Tanagra
Yet another inspiration for TileTerror is Subcutanean
Games are the confluence of multiple creative fields, as
This section discusses how TileTerror treats horror. First, we define horror from a games studies perspective in terms of two overlapping horror subgenres: slasher and psychological. Subgenres were chosen over the overall horror genre to avoid cluttering the high-level features with non-related genre tags when evaluating for subgenres. Then, an analysis of well-known horror games is provided which showcases how each game uses our identified set of low- and high-level features contained in the model.
Since the horror genre is the focus of TileTerror, this work
defines the term genre as a loose definition, as described by
Slasher horror features a character or cast of characters
that face a threat that stalks them, eliminating characters
throughout the story. The name has cinematic roots, with
Psycho
In contrast, psychological horror focuses on a character’s
mental, emotional, and psychological states to frighten and
unsettle the player. Psychological horror often uses
mystery, uncertainty, and unreliability surrounding the
characters, plot, and setting to heighten tension and paranoia. A
cinematic exemplar of this genre is The Thing
Below are game examples that showcase both slasher and
psychological aspects and the importance of environmental
features, such as ambient creep, low visibility, jumpscare,
and isolation—the four features used in our model.
Amnesia: The Dark Descent Amnesia: The Dark
Descent
Alien: Isolation Alien: Isolation places the player in the
shoes of Amanda Ripley, who becomes trapped on a
damaged space station, the Sevastopol. It is revealed to the
player that Sevastopol has become the hunting ground for
the Xenomorph, the antagonist from the Alien
Misao In the 2D world of Misao
Transcending the Z-Dimension Alien Isolation’s Sevastopol station, Amnesia’s Brennenburg Castle, and Misao’s supernatural school succeed in environmental storytelling—in that they provide enough set decoration to establish a horror atmosphere. Even though Sevastopol and Brennenburg Castle benefit from 3D environments and lighting systems, 2D horror manages to create a horror atmosphere by using appropriate tilesets and similar horror techniques, like jumpscares and isolation.
This section provides a walkthrough of our technical implementation, as shown in Figure 1. We begin by describing the various components involved in the generation process. First, there is a degree of preprocessing involved with the selection of a tileset that exhibits the low-level features identified in our ludological analysis; identifying the intrinsic local constraints of the chosen tileset; and designing any partial patterns of tiles we wish to be present in the output. This work uses a graveyard-themed tileset1. Second, both the original tileset with its intrinsic constraints and the added constraints of any partials are given as input to WFC, which generates the remaining unsolved sections of the output. Third, a feature detector analyzes the solved tilemap, creating an annotated map of low- and high-level features combined, which are visualized in Figure 3. Lastly, we provide a description of our evaluation process, which involves a series of pathfinding agents that solve the fully annotated tilemap and score it based on its potential for horror.
Critical components of WFC are hand-authored neighbor pairings that are made with design domain knowledge. In our work that uses the graveyard tileset, for example, it makes reasonable sense that tombstone tiles can appear next to each other. Thus, WFC uses these neighbor constraints 1https://angrysnail.itch.io/pixel-art-graveyard-tileset and the tileset to generate graveyard tilemaps. The simple stress test for this work was generating a tilemap with a row of tombstones and one mausoleum; however, TileTerror wasn’t always able to generate this particular construction. To both help WFC generate appropriate levels and to allow the designer to have more agency in the generation process, our system includes support for additional preprocessing in the way of designerly partial patterns, or more simply, partials.
To overcome multi-tile design constraints, TileTerror introduces partially solved tilemaps. Figure 2 shows an example of a tilemap partial; they are hand-authored designs, like the set of tiles that collectively illustrate a unified mausoleum. To achieve this effect, TileTerror takes the tilemap partials and places the partials within the map, ensuring no overlap. TileTerror generates a 2D matrix, representing a 2D tilemap, setting each cell to false, stating that the cell hasn’t been used yet for the partial generation. The algorithm randomly chooses a location, checking if the tilemap partial can fit. If the partial is too big, or the space already has another partial, the algorithm will move to a new location, trying to find space for the partial. If there is no space, the solver will move on to another partial until there are no more. If there is space, the algorithm will resolve the undecided sub-matrix to the partial. After generating this partially solved tilemap, TileTerror inputs the results into WFC.
WaveFunctionCollapse then takes in both the partially solved map and the local neighbor constraints. Using the partially solved map, WFC generates a new 2D matrix, where each cell contains an array of true booleans, effectively creating a 3D tensor. Each boolean represents a tile from the tileset.
WFC generates a choice heuristic by using an entropy calculation for each cell. The value represents how stable a cell is, and the lower the entropy value is, the closer a cell is to stability. Stability in WFC refers to how close a boolean array is to containing only one true value. WFC uses the entropy value when choosing matrix cells, prioritizing cells with the lowest entropy score.
Upon generating the entropy value, WFC generates a set of neighbors for each tile. This set contains the four cardinal directions for each tile. Each tile must have a neighbor for all four directions, and if a tile does not have a neighbor for one of the four, it will be incompatible with whatever neighbors it borders, at least with the WFC variant TileTerror uses.
WFC chooses a random cell and chooses a random boolean index to keep true, switching all other booleans to false. WFC updates the adjacent neighbors based on the neighbor constraints, validating each neighbor and ensuring it is still compatible with the newly chosen tile. If there are any incompatible tiles, WFC sets their boolean flags to false within the boolean array, propagating the removed tiles to the adjacent tile’s neighbors, beginning the validation process once more. This removal propagation stops when all matrix cells are compatible with their adjacent neighbors; or a cell reaches an incompatibility point, where every boolean flag within a cell’s array is false. If this happens, then the generation process has failed, and WFC will return nothing.
If there are no incompatibilities, WFC chooses another cell, using the entropy calculation as a guide. This choicepropagation pattern continues until every cell’s array has only one true value or an incompatibility occurs. After generation, feature annotations are added to the map, searching for high-level features and appending them as they are found.
This paper identifies four features an environment can use to generate horror: ambient creep, low-visibility, jumpscare potential, and isolation. As a caveat, these are not the only features a horror environment can contain, but the four we believe are the basis for other features. All features we describe here are added to an annotation matrix which directly maps to the generated tilemap, cell by cell.
Low-Level Features For this paper, low-level features are intrinsic qualities, like a tombstone on the tile. Tiles can also have multiple low-level features, such as a tree being creepy and blocking the line of sight, i.e. low-visibility. The first low-level feature is ambient creep which describes the tile’s atmospheric horror quality. This work assigns ambient creep to tombstones, dead trees, and bone tiles. Figure 2 gives an example of ambient creep with the first four tiles on the left: bones and tombstones. The next low-level feature is lowvisibility, which describes how a tile obscures a player’s line of sight. Given that the horror genre makes heavy use of obscuring sight, low-visibility becomes the more important of the two low-level features for high-level feature detection. Figure 2 contains an example of low-visibility with the far right tile, a dead tree. Finally, the low-level feature set includes a traversability feature. It should be noted that traversability is not a horror-specific domain feature, but a feature needed for evaluation.
High-Level Features While low-level features focus on single tiles, high-level features focus on the relationship between tiles. To find these relationships, detectors are used on each tile within the tilemap, inspecting both the tile and its neighbors. If the detector discovers a high-level feature, then the detector appends the feature to the tile’s feature set. This work uses the following high-level features: jumpscare and isolation. For this work, the combination of low-visibility and traversability determines the jumpscare potential and isolation potential of a tile. Figure 3 shows the high-level features through heatmaps of a generated tilemap.
To determine a tile’s jumpscare value, a tile is checked for its low-visibility. If there is no low-visibility property, no changes occur. If the tile has a low-visibility feature, traversability becomes the next checked feature. If the tile is also traversable, then it is marked as a location for a potential jumpscare. If the tile is not traversable, the tile’s immediate neighbors are checked, seeing if any of them are traversable. If those neighbors are traversable, the detector marks the neighbor as a potential jumpscare location.
The next high-level feature is isolation, which can be elicited in two ways: one, if a tile is surrounded by a number of low-visibility tiles, it is marked as isolated; and two, if a tile is not neighboring any tiles with features other than traversability, it is marked as isolated. This duality comes from the idea that isolation is not only linked to claustrophobia, but also to a general sense of vastness. To detect isolation, the detector moves through the matrix, checking if a tile is traversable. If the tile is not traversable, the detector moves on to the next tile. If the tile is traversable, the immediate neighbors are gathered, evaluating how many of them have a low-visibility feature. If the number of lowvisibility neighbor tiles reaches a threshold, the detector annotates the chosen tile with an isolation feature. This work sets the threshold to 3. If a tile is next to another isolation tile, then it will be marked with a temporary lonely tag. This tag is used for increasing the tile’s isolation score.
Our evaluation focuses on the expressive range of TileTerror
through pathfinding agents that solve the map, scoring for
both slasher and psychological horror potential. Five agents
start at the same point: the origin, or the upper left
corner of the map. Each agent navigates the map, reaching the
goal of the furthest corner from the start point: the bottom
right corner for these experiments. The pathfinding
algorithm TileTerror uses is the A* algorithm
The first agent employed is a speed-running pathfinder, which takes the shortest path to the goal. This agent takes the Euclidean distance function as its score. The second agent, the completionist, takes the longest route by scoring the reciprocal of the Euclidean distance. These two agents showcase what parts of the level might need a designer’s attention. The shortest path helps show developers where speedrunners might go, and the longest path gives the developers a chance to place rewards for those who complete the entire level or add more opportunities for horror.
The three other agents solve both for the goal and prioritizing either slasher or psychological horror features. The first of the horror-focused agents is a horror-avoidance agent or the scaredy-cat agent. The scaredy-cat agent’s scoring function sums up all the feature scores across ambient creep, low-visibility, jumpscare, and isolation, using the sum as the score. If a tile has more horror features, its score is higher, and the agent tries to avoid the tile. If there is no better option, then the agent must traverse over the scary tile. The distance from the goal is also a part of the score, making the agent find the shortest path if there is no nearby horror. This score is given as distance times the total horror score.
The fourth agent is a slasher horror-prioritized agent, which searches for all horror, but it prizes jumpscares over all else. Unlike the scaredy-cat agent, the horror-prioritized agents use the reciprocal of the total horror score, prioritizing horror potential rather than avoiding it. By not removing the total horror score (THS), TileTerror co-opts all of the horror values together, summing them up for each tile. But, each feature is weighted differently. The total horror score for the slasher agent is given as: T HS = (1 AC) + (4 J S)+(2 LV )+(0:1 I). For this formula, we denote ambient creep with the variable AC, jumpscare with the variable J S, low-visibility with the variable LV , and isolation with the variable I. Finally, The THS is multiplied by the distance from the goal for the total score.
The final agent is a psychological horror-prioritized agent that seeks out isolation and ambient creep, devaluing jumpscares. Below is a formula for this agent’s valuation of total horror, altering the weights to reflect those features we think are salient to psychological horror: T HS = (2 AC) + (0:1 J S) + (1 LV ) + (4 I). Its final score is given by the distance times the reciprocal of the above total horror score.
TileTerror reconstructs each agent’s path and evaluates each tile for its slasher and psychological score. The first is the slasher score, which only takes into account the two low-level features and jumpscare. The psychological score takes into account the two low-level features and isolation. These scores are the basis of TileTerror’s expressive range, further elaborated in the results section.
We ran TileTerror for 10,000 iterations, producing 9,613 maps, which the pathfinding agents then evaluate. Figure 4 contains a summary of our results. Because the starting and the ending locations may not have a reachable path and WFC can fail, there is a probability of failure within TileTerror—which in our experimentation was approximately 0.04. In case of failure, TileTerror can produce a new result since we envision this tool for offline generation.
The completionist path results seem strange at first, with psychological being the greater of the two. Yet, TileTerror’s output often has large swatches of empty tiles that increase the psychological horror, which can be seen as a more diffused experience. And slasher horror requires tiles with objects on them, like gravestones or buildings or fences, to appear within a level. Thus, an agent with more moves taken will, on average, experience more psychological horror than a slasher because TileTerror seems to generate more open levels. But, these parameters are mutable, letting the developer change how often a tile appears within a level. Thus, the completionist path scoring gives developers a warning sign if the generation parameters need tuning for either psychological or slasher horror.
Surprisingly, the shortest path agent was only a few points away from the scaredy-cat agent, but the scaredy-cat agent took twice as long to complete the maps. This increase in moves could indicate that when the scaredy-cat agent sensed horror, they ran away, increasing steps in the map but not increasing their overall score. And thus, it seems there is often a less terrifying path within the level that developers can discover. With this path, developers can alter the level’s flow by adding more scares in that section or blocking the path entirely.
Between slasher and psychological, it was unexpected to see a slight bump for the slasher score. However, when comparing the moves taken, the same argument from the longest path can be used here: The slasher agent roams the map longer, hitting more psychological horror, while the psychological agent spends less time on the map overall. Figure 4 illustrates each agent’s slasher and psychological scores normalized by distance taken.
As shown by the normalized scores, each of the agents does what is expected. What was interesting seems to be the tradeoff between moves taken and the overall slasher score. It seems the longer an agent is in the level, the more psychological horror they gain. This result corroborates the shortest path agent having a lower overall psychological score while having a higher slasher score and the longest path agent having a higher psychological score than the slasher.
We view this work as a stepping stone to more robust
bottom-up generation and feature detection algorithms. The
first step is to create a denser set of features for our model,
looking towards game studies literature such as Perron’s
study of videoludic horror games (2009a) and integrating
temporal features such as tension into generation and
evaluation. By introducing temporal features, we hope to eke out
a sense of flow
Given that horror often employs adversaries that characters encounter, encoding features like enemy placement would enhance the tilemap and be another reasonable next step. To this end, we envision including a set of adversarial pathfinders in our collection of navigational agents. By doing so, the evaluation agents can attempt to move through the map given a starting and ending location, while the adversaries give chase or lie in wait in certain areas. By discovering where the adversaries and evaluation pathfinders meet, a possible “confrontation” module can be used for the generation, nudging the map towards generating a more deliberate setting for the confrontation.
This work describes a bottom-up architecture for evaluating horror story devices in procedurally generated tilemaps. First, TileTerror examines two subgenres of horror—slasher and psychological—to acquire a baseline understanding of low- and high-level features for tilemap evaluation. TileTerror defines these low-level features as ambient creep, traversability, and low-visibility—horror qualities discovered from a preliminary examination of the horror genre. Combining the low-level features, TileTerror creates a set of high-level features for evaluation.
These high-level features map directly to commonly used horror devices: jumpscares and isolation. Our technical implementation uses WFC to generate a tilemap, the template for the low-level matrix, and detectors run through the feature matrix, finding high-level features. Finally, TileTerror uses a system of evaluation that scores the resulting tilemaps on each map’s horror potential through a set of navigational agents.