List Of Optimizations

Implementation Optimizations

• Used tables of tables and external functions for efficient tensor calculations instead of space-intensive object-oriented formats.

• Some models can use solvers and these solvers tend to cache the result of repetitive matrix calculations. This allows the solvers to “shortcut” certain matrix calculations by retrieving the precomputed matrix from the cache.

Mathematical Optimizations

• Used sufficient statistics for statistical models to avoid computationally expensive batch training.

• If available, the more computationally efficient formula is used instead of the original formula.

Roblox-Specific Optimizations

• Almost all model parameters are stored as tables of tables. This reduces the number of data needed to store inside Roblox’s DataStores compared to dictionaries.

• All models have task.wait() per iteration to avoid eating up all computational resources.