Super Mario as a String: Platformer Level Generation Via LSTMs (1603.00930v2)

Abstract: The procedural generation of video game levels has existed for at least 30 years, but only recently have machine learning approaches been used to generate levels without specifying the rules for generation. A number of these have looked at platformer levels as a sequence of characters and performed generation using Markov chains. In this paper we examine the use of Long Short-Term Memory recurrent neural networks (LSTMs) for the purpose of generating levels trained from a corpus of Super Mario Brothers levels. We analyze a number of different data representations and how the generated levels fit into the space of human authored Super Mario Brothers levels.

• The paper introduces LSTM networks to generate Super Mario levels by representing level tiles as sequences, achieving high playability rates.
• It leverages LSTM memory and gating mechanisms to overcome global coherence challenges common in traditional procedural generation methods.
• The approach integrates player path data with level geometry, paving the way for enhanced level design analytics and adaptive difficulty features.

Procedural Generation of Platformer Levels Using LSTM Networks

The presented paper by Summerville and Mateas explores the application of Long Short-Term Memory (LSTM) recurrent neural networks in procedural content generation, specifically in the creation of levels for the classic game Super Mario Brothers. This research contributes to the field of procedural content generation (PCG) by exploiting LSTM networks' sequence learning capabilities to produce complex and coherent game levels without explicitly predefined generation rules.

Key Aspects of the Research

1. Background and Motivation: The procedural generation of video game levels has a rich history, with traditional methods often relying on predefined rules and constraints. More recent approaches have ventured into machine learning, including Markov chains and other statistical methods, to generate content based on learned patterns from existing game levels. However, these approaches often struggle with global coherence, a challenge addressed in this research using LSTM networks.
2. RNN and LSTM Framework: Recurrent Neural Networks (RNNs) and particularly LSTMs are a natural fit for sequential data due to their architecture, which can accommodate dependencies over long sequences. This capability is essential for level generation, where local and global coherence is required. The integration of mechanisms like input gates, forget gates, and memory cells in LSTMs helps address the vanishing gradient problem, allowing the networks to learn from both immediate and distant events in a sequence.
3. Data Representation: A crucial innovation in this research is the conceptualization of Super Mario Brothers levels as sequences of tile types, essentially treating level geometry as a series of characters in a string. The authors explored several representations, considering the ordering of tile placements, player paths, and the depth information of columns within levels. The integration of player path data, which captures potential routes through the level, proved critical, significantly enhancing the playability of generated levels.
4. Experimental Validation: The authors evaluated their method on datasets comprising levels from both the original Super Mario Brothers and its Japanese sequel, Super Mario Brothers 2. They demonstrated significant improvements in expressed metrics such as playability, linearity, leniency, and path coverage. Specifically, the inclusion of path information resulted in levels with a 97% playability rate, a noteworthy achievement surpassing other machine learning-based generators and even some human-authored systems.

Implications and Future Directions

The use of LSTMs for level generation represents a shift towards more data-driven approaches in PCG. The notable success in generating high-playability levels implies that neural network-based methods can understand both local and global level structures effectively, offering a balance between novelty and playability.

For practical applications, this work highlights the possibility of leveraging such AI systems for both the generation of new game content and the analysis of existing levels for playability and other characteristics. The integration of path data opens avenues for further developments in level analytics and automatic difficulty adjustment based on play data. Additionally, the system's ability to produce human-like paths could inform adaptive AI agents for in-game tasks.

Future research could explore extending these techniques to genres with less linear mapping between time and space, such as Metroidvania style games, where recursive exploration is a key feature. Moreover, leveraging advanced neural architectures, like attention mechanisms, could refine the technique's applicability to these non-linear contexts.

In summary, Summerville and Mateas provide compelling evidence for the capabilities of LSTM networks in game level generation, laying the groundwork for further advancements in procedurally generated content through machine learning.