#random grid

Compare search methods: "In Search Of..." It seems clear that nonlinear programming methods have great potential to find better tuning parameter values. However, there are some nontrivial considerations:

How can your estimate the fitness value efficiently? If you have a ton of data, a single holdout would not be a bad choice for evaluating whatever metric you have chosen to measure performance (e.g. RMSE, accuracy, etc). If not, resampling is most likely the answer and that might not be very efficient. Direct search methods like genetic algorithms (GA), Nelder-Mead and others can require a lot of function evaluations and that can take a while when combined with resampling. Other metrics (i.e. AIC, adjusted R 2 ) could be used but rely on specifying the number of model parameters and that can be unknown (or greater than the sample size).

If you are going to evaluate the same data set a large number of times, even with resampling, there is great potential for overfitting.