Abstract. This paper presents a new evolution control method to reduce the number of computationally expensive simulations for evolution strategies with fitness function models. A feedforward neural network is used as a fitness model and constructed with the help of some previously evaluated solutions in the search space. On-line learning is implemented during searching process. In the evolution strategy with the proposed method the number of controlled individuals is changed during optimization and the choice of parents for the next generation is always made out of controlled individuals. The results of the evolution strategy implementation with the selective evolution control method for three standard test functions in comparison with other known evolutionary strategies are presented.

Keywords: evolution strategy, neural network, metamodel, evolution control

ACM Classification Keywords: I.2 Artificial Intelligence: I.2.6 Learning: Connectionism and neural nets

Link:

SELECTIVE EVOLUTION CONTROL METHOD FOR EVOLUTION STRATEGIES WITH NEURAL NETWORK METAMODELS

Pavel Afonin

http://foibg.com/ijitk/ijitk-vol05/ijitk05-2-p08.pdf