On Sun, Dec 30, 2012 at 12:16 PM, Konstantin Berlin <kber...@gmail.com>wrote:
> > ... > > There would be no burden on the user's side: the visitor pattern has been > > implemented for RealVectors in version 3.1. Besides, we could provide all > > the relevant visitors (addition, scaling, …) > > There is an additional burden to the user, since if you require > implementation of the full RealVector or RealMatrix interface, which > includes a large set of functions not need for the optimizer, since the > user has no idea which function you will use inside the optimizer and which > he can leave blank. For example, if a user needs to create their own > implementation of a vector multiplication, because they have a special case > which can be handled faster, or because they use a GPU, they are still > burdened with implementing unnecessary support for the dozens of others > functions which will never be used. For a GPU example, like Ted has > pointed out, they can implement a GPU version for basic operations (add, > multi, etc), but to guarantee general support for any Java function using > the visitor pattern the user would also need to implement a Java version of > the visitor pattern. > With a good abstract class to inherit from, this isn't much of a problem in practice. You should need to implement very little and you should be able to over-ride what you will without much danger. It also helps to have standardized tests that can pretty exhaustively test new implementations for correctness. To a surprising extent, this allows new implementations to be well tested nearly on the day they are written.
