Add common optimization algorithms #3
Description
To implement various classification and regression algorithms, we need optimization algorithms as solvers. Different solvers come with different benefits, usually based on data-characteristics (sparse solvers, small/large number of features/examples, etc.). I suggest we implement those solvers as library functions that can then be used by other algorithms for model training.
One example would be a L-BFGS algorithm that can be used in a Linear Regression algorithm to solve the optimization problem. Other algorithms could include direct solvers (for small data, possibly could use some library) and (stochastic) gradient descent (sgd) to solve the (weighted) least squares problem.
A description of Spark's implementations can be found here.
I suggest to open separate issues for each algorithm and track/discuss general progress here.