World Models are a recent model-based reinforcement learning framework that has been shown to achieve state-of-the-art results in a variety of popular reinforcement learning environments. However, training them requires extreme amounts of computing availability, often only found in very high profile private-sector companies.

In this research work, we attempt to make this framework accessible to a larger fraction of the reinforcement learning research community by investigating how scaled-down variants of World Models perform.

As a consequence, we discover competitive high-performance low-budget World Models that are trainable on limited computational resources and are individually competitive. However, when compared to the state-of-the-art on standardized evaluation metrics, our scaled-down World Models drop in performance. In this regard, we point towards modifications in our training algorithm that improve the evaluation score while maintaining reduced complexity.