An EA Framework to Avoid Premature Convergence in Stochastic Search

Premature convergence to suboptimal solutions is one of the prime concerns of using evolutionary algorithms (EA) in high complexity real world optimization problems. As the evolutionary search progresses, it is important to avoid reaching a state where the genetic operators can no longer produce superior offspring before reaching an asymptotic optimum. This is likely to occur when the search space reaches a homogeneous or near-homogeneous configuration converging to a local optimal solution. Maintaining a certain degree of population diversity is widely believed to help curb this problem. This paper presents a brief analysis of premature convergence and diversity in the context of evolutionary search process. An informed operator-based technique [6] to introduce constructive diversity has also been described. The proposed technique presented here uses informed genetic operations to reach promising, but un-explored or under-explored areas of the search space, while discouraging local convergence. Elitism is used at a different level aiming at convergence. The proposed technique's improved performance in terms solution precision and convergence characteristics is observed on a number of benchmark test functions with a genetic algorithm (GA) implementation.