Global optimization properties of parallel cooperative search algorithms: A simulation study

Abstract

Cooperative search is a parallelization strategy where parallelism is obtained by concurrently executing several search programs for the same optimization problem instance. The programs cooperate by exchanging information on previously explored regions of the solution space. When the sharing of information overlaps among several programs, changes in the search behavior of one program can propagate over time to several other programs; this is a process called diffusion in physical systems. The optimization properties of diffusion dynamics in cooperative algorithms have not been formally established. However, it is generally believed that when the selection of shared information is biased by the cost (objective) function, diffusion dynamics help to improve the search of cooperating programs. In this study, we simulate this aspect of cooperative algorithms using cellular automata (CAs) (these are artificial dynamical systems often used to simulate the dynamics of complex systems). Our results show that the sharing of information based on the cost function does not affect the diffusion dynamics and therefore does not seem to help the optimization strategy of cooperating programs. However, this study increases our understanding of the role played by diffusion processes in cooperative algorithms. We suggest new approaches that can help to subordinate diffusion dynamics to the optimization goals of the search programs.