Glossary · Computing
Random seed
Also called seed, seeds, random seed
The number that starts a program's random number generator. Two training runs that differ only in their seed show how much of a result is luck.
Training a policy uses randomness everywhere: the network’s starting weights, the actions it explores, where each episode starts, when the shoves come. A computer’s “random” numbers come from a generator started from a number, the seed. Same seed, same code, same machine: the same run. A different seed: a different run of the same experiment.
That makes seeds the cheapest way to see how much of a result is the method and how much is luck. If two seeds land within a few percent of each other, the number means something; if one solves and the other doesn’t, it doesn’t yet.
In this project
Reinforcement learning varies a lot between seeds. In Muscles instead of motors every steps-to-solve number has two seeds: the servo body 12.9 M and 12.7 M (close), the stage 1 muscle body 37.8 M and 26.6 M, MyoLeg26 33.1 M and 49.6 M. Comparisons are made seed against seed, and given as ranges.