Tokenized Game and Environment Models › Key Parameters

The key parameters for the count-based toy dynamics model are:

• Tokenizer: the function that maps each observation to its discrete code. Whether the key sprite is preserved determines whether the model can reason about the pickup rule.
• Codebook size $K$: the number of distinct code symbols. A small $K$ can encourage aliasing; a large $K$ may leave some codes poorly sampled. Measure actual code usage and task-relevant distinctions rather than inferring either outcome from $K$ alone.
• Tokens per frame $L$: the number of discrete codes emitted per observation. Larger $L$ can preserve more spatial detail, depending on the tokenizer, but it always lengthens the dynamics sequence and uses more context-window capacity.
• Training data support: the coverage of the training trajectories. A missing pickup transition leaves the count model without an entry; a neural model may produce a prediction, but its confidence and accuracy on that branch require separate assessment.
• Temporal and action interface: frame stacking determines the observation history, while action repeat, no-op controls, and control latency determine how controls align with transitions. A mismatch in either can make the learned dynamics wrong.