Global Optimization
Both deterministic algorithms like Newton's method and stochastic
ones like simulated annealing can converge very slowly to a global extremum
or become trapped at a local one with high probability. By restarting
the algorithm properly, one can be assured that
the probability the goal has not been attained by the nth iteration
converges to 0 geometrically fast. The Perron-Frobenius theory of non-negative
matrices or generalizations are used to prove this. In
Restarting Search Algorithms with Applications to Simulated Annealing
the authors study the case of simulated
annealing. In both the deterministic and stochastic cases, parallelization leads to more rapid convergence.
In Estimating the convergence rate
of a restarted search process. real-time estimation of crucial parameters to
determine the number of
processors to add in order to achieve high probability of attaining the goal by a given time is studied.