Abstract:

The population diversity of conventional genetic algorithm can be easily destroyed, which further leads to premature convergence. To solve this problem, based on adaptive genetic algorithm (AGA) proposed by Srinivas, a modified adaptive genetic algorithm (MAGA) is presented by introducing a parameter measuring the population diversity. In this way, the probabilities of crossover and mutation are adjusted automatically according to both population diversity and the trends of fitness values. Since MAGA and back-propagation (BP) algorithm are good at searching global and local optimum respectively, an optimized BP neural network based on MAGA (MAGA+BP) is then presented for traffic classification. The Internet traffic dataset provided by university of Cambridge is introduced for experimental validation. Results show that: MAGA shows better performance on maintaining population diversity, overcomes the premature convergence of AGA and improves the fitness value of resulting optimum by 10.17%; MAGA+BP shows a better performance on Internet traffic classification.

Key words: adaptive genetic algorithm, population diversity, back-propagation neural network, traffic classification