Model Evaluation

Goal

How do you pick the best model?
How can you justify the added complexity of a model over simpler/existing solutions?