Tuning strategies

In some sense, hyperparameter tuning is the art and science of guessing and checking at scale. Using sophisticated algorithms and strategies, you can actually train whole fleets of models to test entire ranges of hyperparameters in a huge variety of configurations. Your tuning strategy will then help you find the best models in the end, eventually identifying critical hyperparameters to use at larger scales. I’ve seen hyperparameter tuning help customers get boosts in accuracy of anywhere from less than 1 all the way up to over 15 percentage points. If that’s a direct translation into business returns, you can see why it’s an attractive proposition.

There are many strategies and technical solutions for hyperparameter tuning. These are all similar in that you, as the end user, will need to pick hyperparameters and ranges for these that you’d like to test. Many hyperparameter tuning solutions will provide defaults for you as a starting...