An alternative approach was proposed by Bergstra and Bengio (http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf) in 2012. They demonstrated that a random search across a large hyperparameter space is more effective than a manual approach, as we did for Multinomial Naive Bayes, and often as effective—or more so—than GridSearch.
How do we use it here?
Here, we will build on top of the results such as that of Bergstra and Bengio. We will break down our parameter search into the following two steps:
- Using RandomizedSearch, go through a wide parameter combination space in a limited number of iterations
- Use the results from step 1 to run GridSearch in a slightly narrow space
We can repeat ...