Choose an optimiser and its hyperparameters based on the problem rather than habit, and know what each actually does. Use when training is unstable, slow to converge, or generalising poorly.
