A Method to Accelerate Convergence and Avoid Repeated Search for Dynamic Optimization Problem

Most of the optimization problems are dynamic in real world. When dealing with the dynamic optimization problems, the evolutionary algorithms always suffer from low accuracy and diversity loss. One of the main reasons of low accuracy is that the population cannot convergent to the optima in limit computational cost. And one of the main reasons of diversity loss is that some areas are searched repeatedly while leave the others unsearched deal to the unbalanced attraction from local optima. To cope with the deficiency, two strategies are proposed in this paper. One is called Searching Gbest, which searches for a better solution along each dimension of the best one in the population to accelerate the convergence, and the other is predicting convergence, which deletes the population if it has the trend of converge to the searched area to avoid the repeatedly search. The proposed methods are tested on PSO with multiple populations. The experiments on the Moving Peaks Benchmark show that the methods can improve optima tracking ability, avoid repeatedly search and save the computing resources effectively.