3 ms·

Is unguided random sampling more likely to find a solution than in-order enumeration?

• PartiallyTyped 3y ago
  In order enumeration of parameters / candidates? If so, then I'd argue that yes random search is more likely. When we do hyper parameter tuning in ML, random search beats grid search in efficiency. The higher the dimensionality of the problem, the worst grid search is going to be because you will spend a lot of time at the boundaries. If you are doing Grid Search, then you might as well use a space filling curve, find a promising block, and increase the resolution. This is how GeoHash [1] works more or less. On the other hand, if a single solution exists, then it is improbable that you will find it via random search. There are more complex solutions to exact constraint solving, but perhaps they don't scale particularly well. [1] https://en.wikipedia.org/wiki/Geohash https://en.wikipedia.org/wiki/Geohash