The Benefits of Basing Optimization for Hyperparameter Optimization

In the real world, people have less good priors on what good hyperparameters are going to be and often many more hyperparameters that they could be changing. So we implemented Bayesian optimization with Gaussian priors and also TPE for the modeling and put in reasonable defaults. And then we worked with customers to even figure out what are better, reasonable defaults for real-world situations.