Which "parameters" are you optimizing? Where did the corresponding model come from? When is a certain model even applicable? How do you implement the search procedure efficiently? Is the result actually meaningful? Can you expect future predictions to make sense or did you overfit to your training data?

To avoid any fluff, let's go with a concrete example: some of the best performing algorithms for segmentation are based on spectral clustering (http://en.wikipedia.org/wiki/Spectral_clustering). What does an eigenvector corresponding to the second-smallest eigenvalue of the normalized graph laplacian have to do with random walks? How do you compute it?