numpy-random

Overview

NumPy's random module shifted from the legacy RandomState to the modern Generator API. This new approach provides better statistical properties, faster algorithms, and a robust system for parallel random number generation using SeedSequence.

When to Use

• Stochastic simulations requiring high-quality random bits.
• Shuffling datasets for machine learning training.
• Generating independent random streams for parallel computing workers.
• Creating reproducible experiments across different runs.

Decision Tree

1. Starting a new project?
   • Use np.random.default_rng(). Do not use np.random.seed().
2. Need independent streams for multiple CPUs?
   • Use SeedSequence.spawn() to create children.
3. Shuffling in-place?
   • Use rng.shuffle(arr). For a copy, use rng.permuted(arr).

Workflows

1. Parallel Random Stream Generation

   • Initialize a SeedSequence with a high-quality entropy source.
   • Use the .spawn(n) method to create independent seed sequences for workers.
   • Instantiate a new Generator for each worker using its specific child sequence.

2. Reproducible Simulation Setup

   • Obtain a 128-bit seed (e.g., using secrets.randbits(128)).
   • Initialize the generator: rng = np.random.default_rng(seed).
   • Log the seed to allow exact reproduction of the stochastic results in future runs.

3. In-Place Array Shuffling

   • Create a Generator instance.
   • Pass an existing array to rng.shuffle(arr) to modify it in-place.
   • Specify the axis parameter if only certain dimensions (e.g., rows) should be rearranged.

Non-Obvious Insights

• Legacy Discouragement: RandomState is essentially in maintenance mode; Generator is faster and has better statistical distribution qualities.
• Small Seed Limitation: Seeding with small integers (0-100) limits the reachable state space; SeedSequence ensures high-entropy starting states.
• Bitstream Instability: Even with the same seed, the bitstream is not guaranteed to be identical across different NumPy versions due to algorithmic improvements.

Evidence

• "In general, users will create a Generator instance with default_rng and call the various methods on it to obtain samples." Source
• "SeedSequence mixes sources of entropy in a reproducible way to set the initial state for independent and very probably non-overlapping BitGenerators." Source

Scripts

• scripts/numpy-random_tool.py: Implements parallel seed spawning and reproducible RNG.
• scripts/numpy-random_tool.js: Basic random sampling logic.

Dependencies

• numpy (Python)

References

• references/README.md