Saeed Karimi

In solving large scale problems, the quasi-Newton method is known as the most efficient method in solving optimization problems. We introduce a hybrid optimization approach in- volving Gauss-Newton (GN) and limited-memory BFGS (L-BFGS) with projection and Wolfe line search for solving the nonnegative tensor least squares (NN-TLS) problems. The method adaptively combines GN and L-BFGS directions with a mixing parameter based on the size of residual, such that the method globally converges with a faster local convergence property. Convergence is proved under mild conditions in theory. Numerical experiments demonstrate the efficiency of the new method.