Random Seeds

Data Skeptic

• Episode: Random Seeds
• Website: Data Skeptic
• Feed URL: https://dataskeptic.com/api/blog/rss
• Published: 2017-02-07 01:00

b'## Random Seeds\n\nI had a listener write in and ask a follow up question related to a discussion I had with Daniel Whitenack on a recent episode. The listener was just starting to get into machine learning, and asked an interesting question people in a similar situation might benefit from hearing the answer to.\n\nI asked Daniel about reproducibility on models that have inherently non-deterministic characteristics. If you\'re new to the field, you\'re going to encounter a lot of jargon around this point. The words stochastic, random, non-deterministic, and ergodic might all be used to describe things. Each of these ideas is distinct and important in it\'s own right. Yet, they all more or less mean the same thing when discussing the behavior of algorithms that generate machine learning models.\n\nI\'m going to adopt the phrase non-deterministic as my own personal word choice, and I\'ll explain precisely what I mean by it first.\n\nA deterministic system is one which behaves in an entirely predictable way. If you know it\'s current state and any changes being done to it, then the next step is knowable. At no point does a deterministic system flip a coin to pick between left and right. The choices are locked in forever from the moment of the system\'s creation.\n\nA deterministic system is a bit like chess, given the players moves. The pawn never rolls a dice and moves that many spaces ahead. If the move is declared, the next board configuration is unarguable.\n\nNon-deterministic systems, on the other hand, have some decision step made by chance. Whether this decision is made by an agent interacting with the system, or is a metaphorical coin tossed by the system itself, a prediction of the future cannot be perfectly done. One may have a perfect understanding of the statistical nature of the system, but this just quantifies the uncertainty. A backgammon player can strategize about the game, but their strategy must account for the fact that a dice roll is going to introduce uncertainty about their fut