Gradient descent

An optimization method that uses calculus to find which way is downhill on a loss such as the negative log-likelihood, then adjusts parameters step by step to lower it. It is how models with too many parameters for a grid search are trained.