Computational intelligence in rate-based control for data networking

The work presented in this thesis developed a negative acknowledgment based transport layer protocol called the versatile transport protocol (VTP). In addition to this a fuzzy neural network based fuzzy controller for the VTP was evaluated. The controller was developed by extending the Brain Construction Kit neural network simulation software to support fuzzy neural networks. The VTP controller was then used together with a client/server bulk transfer software called Phat (developed at the University of Manitoba), which acted as the emulation environment for VTP.; To implement a generic multi-layered feed-forward fuzzy neural network, the generalized delta rule for back-propagation was extended to include both logic-based as well as arithmetic neural descriptions. To implement training of the network it was also necessary to develop a fuzzy back-propagation algorithm to guarantee the correctness of the connection strength parameters for logic-based neurons. In doing this a different view of the learning rate was required, one that incorporates both a network independent as well as a network sensitive part.; This thesis represents one of the first attempts to use a fuzzy neural network to control or broker bandwidth within a rate-based transport protocol running over IP. As opposed to models based on simulation, a fuzzy neural controller was demonstrated on top of the Phat protocol which was used to emulate a rate-based transport layer protocol.; In addition to contributions made to fuzzy neural network control theory, this work illustrates the viability of fuzzy control as applied to telecommunication applications such as bandwidth brokering within the confines of an IP network. In addition, the study illustrates the role that fuzzy control can play in improving end-to-end performance by adaptation to host and/or network bottlenecks.