Understanding coarse-to-fine search
Coarse-to-Fine Search (CFS) is a combination of grid and random search hyperparameter tuning methods (see Chapter 3, Exploring Exhaustive Search). Unlike grid and random search, which are categorized in the uninformed search group of methods, CFS utilizes knowledge from previous iterations to have a (hopefully) better search space in the future. In other words, CFS is a combination of sequential and parallel hyperparameter tuning methods. It is indeed a very simple method since it is basically a combination of two other simple methods: grid and random search.
CFS can be effectively utilized as a hyperparameter tuning method when you are working with a medium-sized model, for example, a shallow neural network (other types of models can also work) and a moderate amount of training data.
The main idea of CFS is just to start with a coarse random search from the whole hyperparameter space, then gradually refine the search in more detail, either...
