Genetic algorithm based on multipopulation competitive coevolution

Coevolutionary algorithms assess individuals by their performance in relation to others. The assessing offers the possibility of subduing premature convergence which is a long-standing problem of standard genetic algorithms (SGAs). This paper presents a novel genetic algorithm based on multipopulation competitive coevolution (GAMCC) with inter-population assessment. GAMCC comprises three simultaneously coevolving populations: the learner population, the evaluator population and the fame hall. Learners are assessed by their competitive performance relative to evaluators. Learners and evaluators take turns learning and evaluating, reciprocally driving one another to increase levels of performance. The fame hall saves the elites selected from the learner population. The competitive exclusion principle in ecological theory is applied in the fame hall to maintain the chromosome diversity. Different mutation probabilities are employed to balance the tradeoff between exploration and exploitation. Experimental results show that GAMCC is more likely to avoid the occurrence of premature convergence, and maintains the chromosome diversity more effectively, outperforming the competing genetic algorithms.