Hierarchical Evolutionary Strategy forComplex Fitness Landscapes

摘要

Evolutionary Strategies (ES) are effective forms of Evolutionary Algorithms that enable solving optimization problems. Effective use of ES algorithm has been made in numerous fields. Major optimization problems of today possess a very complex fitness landscape with numerous modalities. The optimization in these complex landscapes is much more difficult as it is possible only to explore a relatively small section of the entire landscape. Also the fitness function behaves in a very sensitive manner with a large amount of change for small changes in the parameter values. We hence propose a hierarchical ES to optimally explore the fitness landscape and return the optima. The inner or the slave ES is controlled by a controlling algorithm or the master. The master has a number of slave ES, each trying to find a solution at some different part of the complex high dimensional fitness landscape. Each ES tries to find the optimal point in its local surroundings. Hence the variable step size is initially kept low. As the iterations of the master increase, we keep reducing the number of ESs and increase the step size to give it a global nature. This is the local to global nature search performed by the algorithm. Since the fitness landscape is complex, the master mutates the locations of the ESs and adds new ESs (deleting the non-optimal ones) as iterations or generations proceed. The novelty of the suggested approach lies in the tradeoff between the search for global optima and conver-gence to local optima that can be controlled between the two hierarchies. Experimental analysis shows that the proposed algorithm gives a decent per-formance in simple optimization problems, but a better performance as we increase the complexity, when compared with the conventional Genetic Algo-rithm, Particle Swarm Optimization and conventional ES.