A combination of local and global optimization strategies seems very
promising since they combine the fast convergence of the gradient method
with the robustness of the global algorithms regarding the
initial guess. One could think of a scenario where a global optimizer
tries to find an initial guess for a local optimizer. The global optimizer
thereby performs a small number of iterations. Some of the best
individuals of this global optimization are then used as initial guess
for the local optimizer. The local optimization is terminated as
soon as no important reduction of the target function is
achieved. Several iterations between global and local algorithms are
possible where the global optimizer could either restart with the
individuals that were used as initial guess or the result from the
local optimizer could be used as a new starting point for a global
refining step.
2003-03-27
