next up previous [pdf]

Next: A fair warning Up: Two solvers Previous: The slow steepest-descent method

The fast L-BFGS method

The L-BFGS method is a member of the quasi-Newton family: it updates at each iteration an approximation of the inverse Hessian $ \bold{Q}$ . This technique is very cost effective: given the most recent history of gradient and model vectors (usually around 5) kept in memory, the quasi-Newton search direction $ \bold{Qdu}$ (inverse Hessian times the gradient) is computed directly with simple vector multiplications. Therefore, the L-BFGS solver can be used for large non-linear problems (Nocedal, 1980).

Contrary to steepest descent where the step length is estimated with a Newton-search technique, the step length in L-BFGS is computed such that sufficient decrease of the error and of the local curvature is attained (so called ``Wolfe conditions''). The appendix shows the L-BFGS solver in more details. The L-BFGS code can be downloaded at http://users.eecs.northwestern.edu/~nocedal/lbfgs.html The pseudo-code below shows both the steepest-descent and L-BFGS algorithms.

U = 0.        # or other initializations
Remove the mean from U(omega).

Iteration {
     dU = 0
     Compute dU
     Remove mean from dU
     du = FT(dU)
     if (steepest descent) 
       {  
       Compute alfa with Newton iterations
       u  = u + alfa*du
       }
     else if (L-BFGS)
       {
       Compute alfa with More and Thuente method
       u  = u + alfa*Qdu # Q = inverse approximate Hessian
       }
          }



2012-10-29