Source-linked AI summary
Deep Learning for Procedural Content Generation
Jialin Liu, Sam Snodgrass, Ahmed Khalifa, Sebastian Risi, Georgios N. Yannakakis, Julian Togelius
TL;DR
Procedural game-content generation has longstanding methods, while deep learning introduces newer generative capabilities that must satisfy game-specific functionality constraints and often face limited training data. This article surveys direct and indirect deep-learning approaches, their combinations with other PCG methods, and methods that remain underused. It concludes that the field is expanding across many content types, while small-data training, functional constraints, underexplored content, and cross-game generalization remain important boundaries and directions.
Problem
Deep learning methods developed for content production are not always directly applicable to games because game content must satisfy functional constraints and training data is often limited.
Method
The article surveys deep learning methods used to generate, analyze, evaluate, search, repair, and transfer game content, while also discussing underused methods and combinations with traditional PCG techniques.
Results
The survey finds that deep learning has been applied alone or with other PCG methods across levels, text, character models, textures, music, and sound, using architectures matched to content characteristics.
Takeaways & Limitations
The review identifies underexplored content such as events, goals, and character features, and highlights parallel training of agents and generators, cross-game generalization, and real-time personalization as future directions.
Takeaways & Limitations
Most deep-learning PCG approaches are constrained by access to training data, especially when a game is still in development and only a few hand-designed examples exist.
Abstract
from arXiv · showhide
Procedural content generation in video games has a long history. Existing procedural content generation methods, such as search-based, solver-based, rule-based and grammar-based methods have been applied to various content types such as levels, maps, character models, and textures. A research field centered on content generation in games has existed for more than a decade. More recently, deep learning has powered a remarkable range of inventions in content production, which are applicable to games. While some cutting-edge deep learning methods are applied on their own, others are applied in combination with more traditional methods, or in an interactive setting. This article surveys the various deep learning methods that have been applied to generate game content directly or indirectly, discusses deep learning methods that could be used for content generation purposes but are rarely used today, and envisages some limitations and potential future directions of deep learning for procedural content generation.
1 Introduction
Deep learning is bringing new capabilities to procedural game-content generation, but game-specific functionality, coherence, and limited training data require careful adaptation. This survey organizes existing approaches and relevant deep learning methods across game-content types.
- Opportunities: Deep learning methods for generating audio, images, 3D objects, and network layouts could support the large quantities of high-definition media required by modern games.Potential applications include generating game characters from photorealistic face-generation methods.
- Background: Procedural content generation has a four-decade history in games, alongside a research community applying grammar expansion, constraint solving, and evolutionary computation.The dedicated research community has existed for roughly the last decade and a half.
- Opportunities: Deep learning has only recently been brought systematically to game content generation, including GANs, VAEs, and LSTMs.These methods create new opportunities but are not automatically applicable to game content.
- Challenges: Game content requires adaptations beyond generic generation because levels must be playable, sprites must convey intended character or emotion, and training data is often limited.Reliable models require sufficient data quantity and quality, which many games lack.
- Survey scope: The article surveys deep learning approaches for direct and indirect game-content generation, then reviews applicable deep learning methods and content-specific constraints.It covers content types, their affordances and constraints, and examples where available.
2 Scope of The Review
The review focuses on deep learning for broadly defined games and distinguishes its coverage from earlier PCG surveys. It systematically collects recent work to assess a rapidly expanding research area.
- Scope: The review defines deep learning as neural networks with at least two layers and some nonlinearity, and includes board, card, arcade, role-playing, shooter, and puzzle games.Its inclusive game definition covers any games a human would conceivably play.
- Scope: The article delineates its scope by comparing the review with existing books and surveys and describing its paper-selection methodology.These discussions appear in Sections 2.1–2.3.
- Novelty: Unlike prior PCG surveys, this review focuses on deep learning, covers more content types, includes indirect support for PCG, and incorporates newer work.The comparison specifically notes narrative text, graphical textures, content-quality prediction, and work published after the earlier PCGML survey.
- Motivation: The review responds to rapid growth in deep-learning PCG, including substantial publication activity after 2018 and the need for an updated state-of-the-art review.The authors identify latent variable evolution as an important trend beginning in 2018.
- Methodology: The authors searched Google Scholar and Web of Science through the end of August 2020 using game-design and procedural-content-generation queries.They systematically examined returned papers, including work in games and computational-intelligence venues.
3 Content Types
Game content spans levels, text, characters, textures, music, and sound, but each type imposes different functional, coherence, and aesthetic constraints. Levels and narrative text demand especially strong correctness, whereas textures and sound often permit softer constraints.
- Constraints: Game content differs from non-interactive media because generated content must satisfy functionality constraints, not merely appear coherent.A level with a missing key for its final door becomes unplayable rather than merely less visually convincing.
- Game levels: Levels are common PCG targets that require traversability, necessary items, and enemies that are not unbeatable.Platform, roguelike, shooter, and battle-map levels combine hard constraints with genre-specific quality requirements such as balance.
- Game levels: Music-game levels can be generated from existing music, with player actions timed to level features and the musical track.These are treated as two-dimensional levels in the review.
- Narrative text: Narrative text has strong truthfulness constraints because statements about the game world must match what actually happens.Traditional game text generation has generally relied on simple substitution or grammar-based approaches.
- Character models: Deep learning has advanced face and character-model creation, but these methods have generally not transferred into games, where models may also need believable animation.Controllable character design remains a notable application case for machine-learning-based PCG.
- Textures and sound: Textures have relatively few functionality constraints, while soundtracks mainly require appropriateness to actions and emotional tone.Texture generation has long used procedural methods such as Perlin Noise, and procedural soundtracks are common in some games.
4 Training Methods and Neural Architectures of DLPCG
Deep-learning PCG systems use different training paradigms and neural architectures according to the content being generated. The surveyed work spans supervised, unsupervised, reinforcement, and evolutionary approaches, with substantial attention to 2D levels and maps.
- The reviewed systems are categorized by machine-learning method, target content, generality, and combinations of evolutionary computation with deep learning.
- Neural architectures: LSTMs mainly generate time-dependent sequences and language, while convolutional networks commonly handle image-like content; GANs use generator–discriminator training for realistic content.
- Supervised learning: Supervised models predict gameplay outcomes, classify structures, generate personalized levels, tune difficulty or emotion, and model NPC behavior or communication.
- Unsupervised learning: Unsupervised methods learn content representations or sequential relationships, then sample or decode new levels, segments, sprites, characters, cards, and text-like content.
- Reinforcement learning: Reinforcement-learning PCG frames generation as an MDP in which actions iteratively modify content to maximize expected future quality, but no standard formulation exists.
- Reinforcement learning: RL-based examples include deck construction, city planning, and PCGRL, whose Narrow, Turtle, and Wide transformations control where a generator modifies 2D levels.
5 Using Deep Learning to Evaluate Content and Content Generators
Deep learning can evaluate generated content and generators indirectly through statistical analysis, simulated playtesting, human-like agents, and player-experience models. These approaches complement content-based measures with learned semantic or subjective evaluations.
- The surveyed indirect uses include analyzing generated content, constructing human-like playing bots, modeling player experience, and evaluating the generator itself.
- Analyzing content: Statistical and similarity measures compare generated content with training distributions, but they mainly capture quantifiable properties.
- Analyzing content: Deep-learning latent representations can support semantic analysis and style transfer beyond measures based only on directly quantifiable features.
- Playtesting: Simulated playtesting provides quick insight into generated content and its generative space, including balance, team advantage, and desired outcomes.
- Player experience: Human trials and surveys capture subjective qualities such as perceived quality, but collecting those evaluations can be laborious and unsuitable for rapid iteration.
- Player experience: Neural models learn mappings from content and gameplay to player-experience measures such as difficulty, enjoyment, aesthetics, believability, and affect.
6 Discussion and Outlook
The discussion emphasizes more controllable, broader, and data-efficient DLPCG, while identifying underexplored content, game types, and multi-domain generation as major opportunities.
- Mixed-initiative DLPCG: Mixed-initiative PCG incorporates designer or player preferences to make generation more controllable, but DLPCG use in this setting remains limited.Level-generation tools can use sketches or interaction data to guide generated content.
- Style Transfer and Breeding: Style transfer and model blending have been applied to game maps, sprites, landscapes, characters, and levels from different games, but only a few studies target game content.Examples include sketch-conditioned map and sprite generation and blending models trained on separate games.
- Underexplored Content Types: Most reviewed work generates 2D tile- or pixel-based content, while text, narrative, music, rhythm, weapons, and other content types receive limited attention.Characters, cards, rules, events, and goals are also identified as underexplored content categories.
- Underexplored Content Types: Platformer and dungeon-like games are over-represented, with Super Mario Bros and Zelda commonly used to test GAN-based level generation.The survey identifies broader game-type coverage as an open need.
- Online Generation: Real-time generation could select level segments during gameplay according to player preferences and performance measured in earlier segments.For Super Mario Bros, different MarioGAN models can target jumps, coins, or difficulty through fitness functions.
- Learning from Small Data: Training-data access is a major PCGML limitation, especially for games still in development that lack sufficient hand-designed content.Bootstrapping can add generated content satisfying functionality constraints to the training set, while data needs vary with model, data, and training complexity.
- Orchestration for Game Generation: Cross-game training could support conditional generators for new games, while orchestration remains a future direction because few systems generate across multiple content domains.Orchestration combines outputs from domains such as visuals, audio, narrative, levels, rules, and gameplay.
7 Conclusions
The surveyed work reflects converging advances in deep learning and machine learning for PCG, producing a broad range of direct and indirect applications. The authors conclude that the field has expanded rapidly while retaining important opportunities in underexplored content, online generation, parallel agent-generator training, and cross-game generalization.
- Conclusions: The survey emerged from converging trends: deep learning for non-game generative tasks and increasing machine learning use in PCG.These trends built on advances that made machine learning effective for new problem classes.
- Conclusions: Interest in deep learning for PCG has exploded, motivating this survey’s effort to systematize existing work and identify fertile research problems.The authors frame the survey as timely given the pace of the field.
- Conclusions: Reviewed deep learning methods generate, analyze, play, evaluate, predict, search, recognize, and repair game content, alone or with other PCG methods.CNNs, LSTMs, VAEs, and GANs are associated with different content types and tasks.
- Conclusions: Future work should expand generation of events, goals, and skill-dependent character features, while pursuing online generation and generalization across games.The authors also identify parallel evolution or training of game-playing agents and content generators as a research direction.
Conflict of interest
The authors declare a financial interest in modl.ai, a company developing AI technologies for games.
- S. Snodgrass, S. Risi, G. N. Yannakakis, and J. Togelius declare a financial interest in modl.ai.
- modl.ai develops AI technologies for games.
- The disclosure identifies modl.ai as the authors’ financial-interest connection.