Parallel Genetic Algorithms, Premature Convergence and the nCUBE

Genetic Algorithms (GAs), rst proposed by John Holland in the early seventies, are growing in stature as tools in the elds of machine learning and function optimization. GAs model evolution of life. To solve a particular task, a genetic algorithm creates and maintains a population of organisms, probabilistically modifying the population , while seeking a near-optimal solution to the task at hand. While GAs have been extensively used in designing semiconductor layouts, in factory control and in the solution of optimization problems, they suuer from premature convergence. Under conditions of premature convergence, the entire population looks almost alike and hence cannot produce oospring which are radically diierent. When this happens, the GA, will never explore other regions of the search space (since it has no way of generating oospring reeecting search in those regions) and hence it might miss out on the global solution of the problem. In this paper we explore the possibility of alleviating (not overcoming!) this problem by maintaining independent subpopulations, each performing a search in the solution space (hopefully in diierent regions). We also allow migration of individuals from one population to another as a mode of communication of current search status of the subpopulations. These migrated individuals introduce diversity and variation in otherwise prematurely converged populations, thereby keeping the search process alive. Intuitively, such an approach enhances the chances of the distributed GA nding the global solution and we show through simulations that this is indeed true.