| Summary: | 博士 === 國立中正大學 === 電機工程研究所 === 93 === In recent years, the advanced communications technologies supply more and more network bandwidth. However, Internet users increase rapidly result in the network bandwidth is exhausted. When too many users are present in the Internet cause performance degrades. This situation is called congestion. Low throughput, high packet loss rate and high transmit delay result from congestion. The problem of insufficient bandwidth can be improved by way of enhancing the congestion control mechanism so that it can work more efficiently. Traditional methods for congestion control always monitor the queue length, on which the source rate depends. This paper is meant to explore the proposed congestion control with cooperative reinforcement learning (RL) scheme that differ from control method of AIMD, to adapt to the variant network environment. The RL scheme, mainly implemented by artificial neural networks (ANNs), consists of two subsystems: the expectation-return predictor is a long-term policy evaluator and the other is a short-term action selector, which is composed of an action-value evaluator and a stochastic action selector. In this research, we divide the study of proposed congestion control into three applications. The first application applies a RL scheme to congestion control in multimedia networks. The proposed RL scheme receives reinforcement signals generated by an immediate reward evaluator and takes the best action in the sense of state value evaluation to control source rates in consideration of system performance. The second application is the study for an adaptive multi-agent RL scheme on solving congestion control problems on dynamic high-speed networks. After receiving cooperative reinforcement signals generated by a cooperative fuzzy reward evaluator, the proposed cooperative multi-agent congestion control can learn to take correct actions adaptively under time-varying environments. The last one is the study for a cooperative congestion control for multimedia networks based on learning approach. In order to make the best of link utilization, a cooperative fuzzy reward evaluator provides cooperative reinforcement signals based on game theory are included. Simulation results have shown that these proposed approaches can increase system utilization and decrease packet losses simultaneously.
|
|---|
