The interacting-particle algorithm with dynamic heating and cooling

We consider an interacting-particle algorithm which is population-based like genetic algorithms and also has a temperature parameter analogous to simulated annealing. The temperature parameter of the interacting-particle algorithm has to cool down to zero in order to achieve convergence towards global optima. The way this temperature parameter is tuned affects the performance of the search process and we implement a meta-control methodology that adapts the temperature to the observed state of the samplings. The main idea is to solve an optimal control problem where the heating/cooling rate of the temperature parameter is the control variable. The criterion of the optimal control problem consists of user defined performance measures for the probability density function of the particles' locations including expected objective function value of the particles and the spread of the particles' locations. Our numerical results indicate that with this control methodology the temperature fluctuates (both heating and cooling) during the progress of the algorithm to meet our performance measures. In addition our numerical comparison of the meta-control methodology with classical cooling schedules demonstrate the benefits in employing the meta-control methodology.