---
title: 'Games Mapper: Mapping Game Structures'
url: https://www.emergentmind.com/topics/games-mapper
type: topic
---

# Games Mapper: Mapping Game Structures

Games Mapper denotes a family of methods that represent, analyze, generate, extract, edit, or visualize structure in games through explicit mapping formalisms. In the narrowest published sense, it is the name of a Steam market-analysis method based on the Mapper algorithm from topological data analysis, extended with automated cluster labelling to produce interpretable visualizations of genre evolution over time [2606.14376]. In a broader technical sense suggested by adjacent work, the term also covers machine-usable extraction of playable-world structure from execution traces, graph-based summaries of combat flow, location-based remapping of physical space into game space, designer-centric procedural level generation, semantics-preserving scene graph merging, and AI-assisted editing of environment material maps rather than gameplay topology [1707.03908].

## 1. Games Mapper as a topological market-analysis method

In the paper that explicitly bears the name, Games Mapper is a Steam market-analysis workflow built on the Mapper algorithm from topological data analysis, with an automated cluster-labelling component tailored to video game tags [2606.14376]. Its problem setting is strategic rather than geometric: studios, publishers, and analysts want to understand how a market is structured, how subgenres evolve, and where viable entry points or competitive barriers exist. The paper argues that static genre labels and year-by-year clustering are insufficient because they do not explicitly represent continuity, branching, persistence, or transition over time.

The data objects are Steam games represented by their capital tag priorities. For the Simulation case study, the dataset consists of games released from 2015 to 2025, with at least 100 reviews, and carrying the tag Simulation with priority at least \(0.6\) [2606.14376]. Tag priority is defined as a score from 0 to 1 proportional to the number of players assigning that tag to the game. The analysis excludes purely aesthetic tags such as Anime and Pixel Graphics because the clustering is intended to reflect gameplay features rather than visual style.

The Mapper construction follows database construction, filtering, cover construction, local clustering, graph building, and interpretation. The filter is release year, effectively
\[
f : X \to \mathbb{R}, \qquad f(x)=\text{release year of game }x.
\]
The cover is temporal and overlapping: each interval contains two consecutive years, and consecutive intervals overlap by one year, producing layers
\[
(2015,2016), (2016,2017), \dots, (2024,2025), (2025).
\]
Within each interval, the paper applies the clustering method from Grelier and Kaufmann (2024), then constructs a graph whose vertices are local clusters and whose edges connect clusters in consecutive layers that share games. Operationally, for clusters \(C_i^{(r)}\) and \(C_j^{(r+1)}\),
\[
w\!\left(C_i^{(r)}, C_j^{(r+1)}\right)=\left| C_i^{(r)} \cap C_j^{(r+1)} \right|,
\]
and an edge is drawn when this quantity is nonzero [2606.14376].

A distinctive contribution is the naming machinery. For a tag \(t\) and cluster \(C\), the method computes \(h(t,C)\), the Cohen’s \(h\) comparing the proportion of tag \(t\) inside the cluster to its proportion in the whole interval. The naming score for a clustering \(\mathcal C = \{C_1,\dots,C_k\}\) is
\[
n(\mathcal{C}) := \frac{1}{|S|} \sum_{1\leq i \leq k} |C_i| \max_t h(t,C_i).
\]
The paper uses the elbow method on this score to choose the number of clusters per layer. In the final Simulation study, it chooses 6 clusters for layers from (2015–2016) through (2022–2023), and 7 clusters for the last three layers. Cluster labels are then expanded beyond a single dominant tag: each cluster is labelled with all tags \(t\) satisfying
\[
h(t,C) \ge 0.8\, h(t_{\max},C),
\]
sorted by \(h(t,C)\), with at most the top three tags displayed [2606.14376].

This formulation makes Games Mapper a temporal topological representation of market structure rather than a conventional clustering plot. A plausible implication is that its core contribution is not only segmentation, but the explicit encoding of continuity between adjacent market states.

## 2. Structure, interpretation, and findings in the Steam Simulation case study

The Simulation case study illustrates what Games Mapper means by a mapped market topology. The resulting graph segments Simulation into several coherent subgenres, with Management and Sports identified as especially persistent regions, an Open World to Exploration trajectory interpreted as a broad evolving lineage, and shorter trajectories such as Horror, Story Rich, and Relaxing appearing with different degrees of persistence [2606.14376].

Management appears in every interval and grows from 80 games in (2015–2016) to 331 in (2024–2025). Representative titles include Cities: Skyline I and II, RimWorld, Crusader Kings III, Manor Lords, and Schedule I. Sports appears in every interval except (2022–2023), grows from 63 to 93 games over the same span, and is interpreted as a relatively stable classic rather than a strongly trending subgenre. Open World and Exploration coexist as separate clusters in several intervals, but strong inter-layer connections suggest one broad lineage, with the path growing from 64 games in (2015–2016) to 221 in (2024–2025). Horror shows a stronger recent trajectory from 2023 onward, growing from 106 games in (2023–2024) to 121 in (2024–2025) [2606.14376].

The interpretive value lies in the topology of connections rather than in labels alone. Strong high-weight vertical paths indicate continuity of a subgenre over time. Bridges reveal proximity between nominally distinct subgenres. Weakly connected or isolated nodes suggest unstable or temporary niches. The paper repeatedly emphasizes that standard independent yearly clustering would show local groups but not these cross-year structural relationships, and that dimensionality-reduction snapshots would not naturally encode persistent trajectories, transitions, or overlap-induced continuity.

Several limitations are explicit. The 100-review threshold is arbitrary and biases against newer releases. Steam tags are partly user-generated and therefore reflect community perception as well as game design. The pipeline inherits stochasticity from both K-means and elbow-based model selection. The graph is sensitive to typical Mapper design choices such as filter, interval cover, overlap, and cluster count. The authors state that there is currently no automatic method they use for setting these hyperparameters in this domain, except for the number of clusters per layer [2606.14376].

A common misconception is that Games Mapper refers only to spatial level maps. In the cited paper, it instead denotes a topological analysis of market and genre structure. This suggests that the phrase has become polysemous across game research: sometimes it denotes economic or taxonomic mapping, sometimes playable-world extraction, and sometimes spatial or visual level representations.

## 3. Automatic extraction of playable-world maps and route structure

A second major sense of game mapping is automatic extraction of machine-usable world structure from execution traces. Mappy is a canonical example: it produces a good approximation of a linked map of rooms from a Nintendo Entertainment System game program and a sequence of button inputs exploring its world [1707.03908]. The goal is not merely a stitched screenshot, but a richer representation containing rooms, scrolling geometry, temporal tile changes, object placements, transition links, and candidate room merges.

Mappy’s inputs are the game program, access to an emulator exposing graphics state and savestates, and a sequence of button inputs exploring the world. The system observes PPU state frame by frame, recovers visible background tiles, estimates scrolling, and accumulates room-aligned tile observations. It stores a temporal history rather than overwriting changed tiles, because tiles may change due to scrolling artifacts, animations, or gameplay events such as breakable blocks and collapsing bridges. The paper describes the room map representation as a dictionary whose keys are spatial coordinates and observation times, and whose values are tile keys.

Transition detection is central because it segments a trace into rooms and links them. For smooth transitions, Mappy uses loss of control plus significant scrolling, with speculative execution from savestates to estimate whether the player had control. For teleport-style transitions, it uses a heuristic based on sudden drastic changes in screen appearance. Once transitions are detected, it inserts directed room links. It then offers similarity-based candidate merges among rooms that may correspond to the same location, while leaving the final merge decision to a human analyst [1707.03908].

The outputs include visual room maps, tile grids, temporal tile histories, object placements in room coordinates, a graph-like room-linkage structure, and clusters of rooms that may be aliases of one another. The paper positions this as a way to generate the VGLC’s tile-based and graph-based formats. No formal benchmark metrics are given; evaluation is qualitative across titles such as Super Mario Bros., The Legend of Zelda, Metroid, and Mega Man 2.

This line of work clarifies an important distinction inside Games Mapper. Here, mapping means recovering the structure of a playable world from execution. It is neither market topology nor texture editing, and it emphasizes rooms, links, and temporal state.

## 4. Spatial remapping, movement graphs, and gameplay analytics

Other strands of the literature map game-relevant structure onto physical or behavioral spaces rather than extracting room graphs from ROM execution. PacMap remaps PacMan into the player’s actual urban surroundings by treating the surrounding road network as game action corridors, with intersections as graph nodes, cookies placed along road segments, and ghosts moving as virtual agents over the geographic network [1501.02659]. The playable area is an imaginary circle centered on the user with a radius of 200 meters, populated from OpenStreetMap via a client-server architecture. The red ghost’s pursuit is computed locally on the client using Dijkstra’s algorithm on a weighted road graph whose edge costs are physical distances, avoiding repeated use of commercial routing services. On an average game space with 420 nodes, shortest-path computation takes 95 msec on a Samsung Galaxy S4.

A different mapping problem appears in battle-map analytics. “Enhancing Battle Maps through Flow Graphs” introduces an intermediate aggregation step that converts representative troop trajectories into weighted directed acyclic graphs before rendering [1906.04435]. The purpose is retrospective summarization of team movement in games such as World of Tanks. Representative trajectories of the form
\[
o=l_1 \rightarrow l_2 \rightarrow l_3 \rightarrow \ldots \rightarrow l_k
\]
are turned into directed edges
\[
(l_i, l_{i+1})
\]
with weights counting how many representative trajectories use the transition. The resulting graph is rendered with cubic Hermite splines and \(C_1\) continuity, reducing occlusion and making troop strength, splits, and merges more legible.

Mixed-reality analytics extend mapping into embodied and multimodal telemetry. GAMR is a Unity plugin for HoloLens 2 that records and reconstructs gameplay sessions, provides 2D heatmaps using X and Z coordinates from 3D vectors, and tracks hands, camera, input, audio, and customizable objects [2408.01573]. It supports replay, filtering, annotation, and multi-session comparison. In an experimental study, participants found GAMR more useful for the Action-Adventure prototype than for the FPS prototype, with overall usefulness medians of 7 versus 6, respectively, and the audio tracker was rated more useful in the FPS game. The tool’s heatmaps, replay, and annotation system are therefore part of a Games Mapper tradition in which player behavior is remapped into spatially situated analytic views.

These systems share a graph-centered and spatially explicit view of games, but they differ in what is being mapped: urban road networks into gameplay graphs, player trajectories into flow summaries, or mixed-reality sessions into replayable spatial telemetry.

## 5. Generative and editing-oriented map representations

A further sense of Games Mapper concerns the generation or modification of game environments themselves. “Procedural Generation of 3D Maps with Snappable Meshes” presents a designer-controlled, gridless 3D map generator that assembles authored meshes by snapping compatible connectors together [2108.00056]. A piece carries one or more connectors, each with a heading, color, and pin count. Legal attachment depends on unused compatible connectors, optional color-matrix and pin-tolerance rules, and optional overlap checks. The generator is stochastic, constructive, greedy, and has no backtracking. High-level map shape is controlled by selection methods such as arena, corridor, star, and branch. Playability is checked after generation using navigation points and reachability metrics including
\[
\overline{c} = \frac{c_t}{c_{all}}
\]
for average relative connectivity and \(A_r^\text{max}\) for the relative area of the largest fully connected region. Reported mean generation times range from 8.6 ms to 42.4 ms in the benchmark scenes.

AAA environment editing raises a different representational issue. “In the Blink of an Eye: Instant Game Map Editing using a Generative-AI Smart Brush” does not change topological gameplay layout or geometric navigation graphs; it edits terrain and material appearance data in highly detailed 3D environments [2503.19793]. In the World of Tanks setting, each map consists of a \(10 \times 10\) grid of chunks, each chunk has 8 material layers, and each layer has a grayscale tile mask controlling how one RGB material texture contributes to the final rendered result:
\[
M_{final}\left(x,y\right) = \sum_{i=1}^{N} T_i \left(x,y\right) \cdot M_i\left(x,y\right).
\]
The Smart Brush regenerates masked regions in this material-mask representation, conditioned on neighboring content, global albedo map, height map, and object masks. BrushGAN is the strongest model in the paper, producing the sharpest and most detailed outputs and taking around 2 seconds per chunk on an A100 GPU, while BrushCLDM takes about 6 seconds per chunk. This distinction matters because “map editing” in this context means context-aware texture and material inpainting, not layout remapping.

Collaborative editing introduces yet another mapping formalism. LevelMerge models a game level as a labeled directed acyclic graph,
\[
G = (V, E_d, E_i, \lambda, A),
\]
where \(E_d\) are direct-dependency edges, \(E_i\) are indirect-dependency edges, \(\lambda\) gives node types, and \(A(v)\) gives node attributes [1603.00713]. Merge is then treated as a semantics-preserving three-way merge over these graphs rather than over text files. In a user study, LevelMerge received an average rating of \(4.11\) versus \(1.85\) for Git merge, and Git merge corrupted \(23.6\%\) of scenes. Here, the map is the dependency structure of the level itself: objects, assets, hierarchy, and references.

Together, these works show that generative and editing-oriented Games Mapper systems are often defined by their underlying representation: connectorized meshes, material-layer masks, or labeled dependency graphs.

## 6. Topological data analysis, illumination, and future directions

Topological and quality-diversity methods generalize the mapping idea beyond market analysis. Ball Mapper is introduced as a Mapper-inspired descriptor that replaces the lens-and-cluster pipeline of classical Mapper with a cover built directly from metric balls centered at selected landmarks [1901.07410]. Given a metric dataset \(X\), radius \(\varepsilon\), and centers \(C\subset X\) such that
\[
X \subset \bigcup_{c\in C} B(c,\varepsilon),
\]
Ball Mapper constructs a nerve-like graph in which vertices correspond to balls and edges indicate shared coverage of at least one data point. This metric-only construction motivates later Mapper-type methods in game-adjacent settings.

A game-specific use appears in “Mapper-type algorithms for complex data and relations,” where Equivariant Ball Mapper is applied to the UCI Tic-Tac-Toe endgame dataset [2109.00831]. Terminal board configurations are encoded in \(\mathbb{R}^9\), the dihedral symmetry group acts by isometries, and with \(l_1\) distance and \(\epsilon = 2.5\), the resulting graph perfectly separates winning states, losing states, and ties. This is not a level-map tool, but it demonstrates how Mapper-type constructions can visualize combinatorial game-state spaces with symmetry.

Quality-diversity mapping appears in dungeon and FPS generation. “Illuminating the Space of Dungeon Maps, Locked-door Missions and Enemy Placement Through MAP-Elites” stores elites in a \(5\times5\) archive indexed by leniency and exploration coefficient, where leniency is the fraction of safe rooms and exploration coefficient averages normalized coverage between start-goal and key-lock reference-room pairs [2202.09301]. “Procedural Generation of First Person Shooter Maps using Map-Elites” treats FPS map design as an archive of elites organized by descriptor pairs such as area–maxSymmetry and pace–averageEccentricity [2605.30570]. Its quality objective is the entropy of duel-match balance:
\[
entropy = -\sum_{i=1}^{n} \left(\frac{k_i}{k_{tot}}\right)\log_2\left(\frac{k_i}{k_{tot}}\right).
\]
The paper reports that the new Point-Line and Spatial-Layout representations generate maps with higher diversity and quality than All-Black and Grid-Graph. In this usage, a Games Mapper is literally a map of map-space.

A final emerging direction is visual transition inference. “Adaptive Visual Navigation Assistant in 3D RPGs” formalizes Spatial Transition Points (STPs) and Main STPs (MSTPs) from a single frame, using Faster R-CNN for candidate detection and a local-global MSTP selector with optional retrieval-augmented late fusion [2508.18539]. On a custom dataset from five Action RPG titles, the MSTP selector reaches 76.62% \(\pm\) 2.96% mean accuracy across seeds, and in a real-time pilot study achieves an approximate success rate of 82.5% over 63 navigational decisions. This suggests a future Games Mapper stack in which connector detection, temporal tracking, region modeling, and graph maintenance are coupled into richer client-side auto-mapping systems.

Across these lines of work, Games Mapper is not one algorithm but a convergent research program. It includes temporal topological market graphs, emulator-based room extraction, route and behavior analytics, procedural content illumination, semantics-preserving scene graphs, AAA material-mask editing, and visual transition discovery. The unifying principle is that game-relevant structure becomes explicit, queryable, and operational once it is mapped into a formal representation.

Source: https://www.emergentmind.com/topics/games-mapper