All Model Types, Training & Tuning

1. We reference our blog post about random search because we believe that to be the more common definition of random search within model tuning literature. The Wikipedia entry more closely hews towards a randomized trust region method, which can be expanded to a simulated annealing strategy to deal with nonconvex functions that are cheap to evaluate or to conduct MCMC sampling from a multimodal function. Return

2. For simplicity in generating this plot, we ignore the fact that during each 1D slice, some number of desired function values may already be present from previous slices. Correctly storing and managing those values would reduce the cost of this method. Return

3. Some literature uses Halton sequences starting at the origin, but this example skipped that first point. The (2, 3, 5) Halton sequence refers to the prime numbers used to generate the sequence in each of the dimensions. Return

4. https://www.github.com/sigopt/evalset is a Python implementation of real-valued functions that SigOpt uses to conduct regression testing as we try to deploy new optimization strategies. Some of these functions were used for testing in our ICML AutoML submission. The functions provided in the eval set are meant to be minimized, so these experiments instead maximize the inverse of the functions. Return