| dc.description.abstract | This paper examines how recent advances in sequence modeling translate for machine learning assisted procedural level generation. We explore the use of Transformer based models like DistilGPT-2 to generate platformer levels, specifically for the game Super Mario Bros., and explore how we can use reinforcement learning to push the model towards a task like generating levels that are actually beatable. We found that large language models (LLMs) can be used without any major modifications from the original NLP focused models to instead generate levels for the aforementioned game.
However, the main focus of the research is connected to how advancement in the area of NLP by the use of reinforcement learning (RL) algorithms, specifically PPO, translates to the arena of procedural level generation in cases where the levels can be treated as token sequences.
We did however not find any combinations of hyperparameters that allowed the PPO to reach higher better results than our baseline model trained for next token prediction.
Despite its success in the area of NLP, we failed to find a combination of hyperparameters that improved upon the level generation by applying an reward for the whole level. However there are methods that we did not try yet, like finding specific parts of the level to reward and penalize. | |
