An improved artificial fish-swarm algorithm and its application in feed-forward neural networks

Artificial fish-swarm algorithm (AFSA) is a novel method to search global optimum, which is typical application of behaviorism in artificial intelligence. In order to improve the algorithm´s stability and the ability to search the global optimum, we propose an improved AFSA (IAFSA). When the artificial fish swarm´s optimum value is not variant after defined generations, we add leaping behavior and change the artificial fish parameter randomly. By the way, we can increase the probability to obtain the global optimum. A new feed-forward neural networks optimization module based on IAFSA is presented. The comparative result between BP algorithm, AFSA and IAFSA demonstrates that the IAFSA has better global stability and avoids premature convergence effectively.