Discrete World Models via Regularization

Researchers introduce Discrete World Models via Regularization (DWMR), a new method for learning Boolean representations of environments without requiring reconstruction or contrastive learning. The approach uses specialized regularizers to maximize entropy and independence while enforcing locality constraints, showing superior performance on benchmarks with combinatorial structure.

• →DWMR eliminates the need for decoder-based reconstruction or contrastive learning in world model training.
• →The method uses novel regularizers that maximize entropy and independence of representation bits through variance, correlation, and coskewness penalties.
• →A locality prior is enforced to handle sparse action changes in the environment.
• →Experiments demonstrate more accurate representations and transitions compared to reconstruction-based alternatives.
• →The approach can be combined with auxiliary reconstruction decoders for additional performance gains.