ATM TRAFFIC MANAGEMENT USING MODULAR NEURAL NETWORKS

Abstract

This thesis proposes methods based on predictions using the Hierarchical Mixtures of Experts (HME) to improve on Asynchronous Transfer Mode (ATM) flow control and Connection Admission Control (CAC). For flow control, many rate-based approaches suffer from beat-down and network congestion collapse problems. In [1], the 'intelligent congestion control technique' is shown to solve such problems. However, it only reacts to congestion after it occurs, which limits its effectiveness. In the proposed scheme, its performance is enhanced by integrating the HME to predict buffer queue length, so that reactions for congestion can be carried out before they occur. Through simulations, it is shown that the proposed scheme is approximately twice as effective in keeping the average queue lengths small, and that the advantages of having prediction capability increase as the link distances become longer. For CAO, HME-based schemes for two types of network situations are proposed. The first situation consists of a two-node network with solely single-bitrate applications. In the second situation, the network can have any arbitrary number of nodes and bottleneck links, and is also able to handle different traffic sources with varying traffic characteristics and cell loss ratio (CLR) requirements. In both cases, the HME is used to estimate the CLR resulting from the connection of new calls. For single-bit-rate CAC, four sets of simulations have been performed. An 'admit all' scheme has the largest throughput, but also the highest CLR. A 'simple' scheme, which rejects new calls if the buffer queue length exceeds a fixed threshold, reacts to congestion rapidly but lacks prediction capability. Even though high throughput is achieved, high OLR is inevitable because the scheme is "reactive" rather than "preventive". A Multilayer Perceptron (MLP)-based scheme, which replaces the HME with a MLP, learns the traffic patterns slowly and results in low throughput and high CLR. Finally, the proposed HME-based controller, with its swift adaptability, is able to learn fast and achieve very high throughput while keeping OLR low. For the multiservice situation, simulations are performed to compare the proposed scheme with bandwidth allocation methods using Peak Cell Rate allocation (PORA), Average Cell Rate allocation (AvCRA), as well as a leading equivalent bandwidth (EB) approach. The PORA scheme has zero cell loss but low link utilisation, while the A vCRA scheme has the largest throughput but also has the largest CLR. The EB scheme is found to be near optimum in terms of high throughput and low OLR. Even though the proposed HME-based scheme encounters high OLR during the learning stage, its link utilisation and CLR performance become comparable to those achievable using the EB approach after it has learnt. It is stressed that the HME scheme does not make any assumption about the traffic model used for the VBR sources. This is further illustrated by a simulation which uses different traffic models for the Variable Bit Rate (VBR) sources, in which the EB scheme cannot be applied. The HME scheme is expected to give very good performance if it is pre-trained prior to being used in real-world CAC applications.