MULTISTEP AHEAD PREDICTION FOR REAL-TIME VBR VIDEO TRAFFIC USING DETERMINISTIC ECHO STATE NETWORK

摘要

Variable bit rate (VBR) video traffic, exhibiting high self-similarity and burstiness, has been a major traffic component in high speed network. However, its complex bit rate distribution makes VBR video traffic prediction, especially multistep ahead prediction, very difficult. Recently, deterministic echo state network with adjacent-feedback loop reservoir structure (ALR) was proved to have high prediction accuracy, good memory capacity, and simple structure. In the paper, we apply ALR to real-time VBR video traffic prediction. The proposed scheme makes use of loop reservoir with identity activation function to conduct sample learning in high dimension states. Experimental results show that the simplified ALR scheme can effectively capture dynamic characteristics of VBR video traffic with less training time. Its multistep prediction accuracy is improved by 2% on average, compared with the neural network based on multi-resolution learning.