Super Mario as a String: Platformer Level Generation Via LSTMs

Researchers applied Long Short-Term Memory (LSTM) recurrent neural networks to the procedural generation of Super Mario Bros. levels, arguing that LSTMs, by mitigating the vanishing gradient problem, can capture longer-range structural dependencies than the Markov-chain and graph-based generators that have struggled with global coherence. The approach uses a three-layer architecture with 512 blocks per layer, trained with dropout on tile-level sequences. A central methodological contribution is the careful definition of 13 tile types, including distinctions between ?-Blocks that contain coins and those that contain power-ups, and the separation of pipe segments, alongside a comparison of eight candidate tile orderings such as bottom-to-top versus snaking traversal.

The study's findings center on how the encoding choices affect model performance. Among the configurations tested, including snaking ordering, player path information, and column depth markers, the inclusion of player path information produced the largest improvement. The best-performing configuration, combining snaking, paths, and depth, achieved a negative log-likelihood of 0.0177 and generated levels that a simulated agent could complete 97 percent of the time. This compares favorably with a previously reported best of 66 percent for machine-learning-based level generators.

Beyond completability, the generated levels broadly matched human-authored levels across most evaluated metrics, indicating that the model captured the stylistic and structural conventions of the original game design. A notable exception is that generated levels' path lengths were consistently longer than those of the original levels, a discrepancy the work identifies as a limitation of the current approach.

The scope of the work is confined to Super Mario Bros. level data, with the research situated within the broader literature on procedural content generation, including the 2010 Mario AI Championship and prior Markov-chain and Monte Carlo Tree Search-based generators. Its contribution lies in demonstrating that representation choices, particularly the encoding of player path information, can be as consequential as model architecture in generating playable, human-like platformer levels.