Piergiulio Maryni

A neural strategy for optimal multiplexing of circuit- and packet-switched traffic

Hybrid fixed-length frames that are used to carry two basic traffic types, a circuit-switched isochronous and a packet-switched asynchronous one, are considered in a time-division multiplexing (TDM) structure. The available bandwidth, in terms of slots/frame, is dynamically allocated between the two traffic types, by means of a randomied decision rule, in order to minimize a cost function that takes into account circuit blocking probability and packet loss rate (for a finite packet buffer), or average packet delay (for an infinite buffer). The dynamics of the circuit-switched traffic can be represented by a discrete-time birth-death process, and the optimization problem is posed by considering the evolution of the corresponding Markov chain. The optimal decision rules over a finite future time horizon are approximated by a set of neural networks, whose parameters are determined by a method that exploits the finite state Markov chain model. Some preliminary numerical examples showing the convergence behavior of the parameters are reported.<<ETX>>

A two-level stochastic approximation for admission control and bandwidth allocation

In an access node to a multiservice network [e.g., a base station in an integrated services cellular wireless network or the optical line terminal (OLT) in a broad-band passive optical network (PON)], the output link bandwidth is adaptively assigned to different users and dynamically shared between isochronous (guaranteed bandwidth) and asynchronous traffic types. The bandwidth allocation is effected by an admission controller, whose goal is to minimize the refusal rate of connection requests as well as the loss probability of cells queued in a finite buffer. Optimal admission control strategies are approximated by means of backpropagation feedforward neural networks, acting on the embedded Markov chain of the connection dynamics; the neural networks operate in conjunction with a higher level bandwidth allocation controller which performs a stochastic optimization algorithm. The case of unknown, slowly varying input rates is explicitly considered. Numerical results are presented that evaluate the approximation and the ability to adapt to parameter variations.

Developing a distance learning system using Java applets

An overview of an experimental environment that is able to support distance learning sessions is presented. In particular the system is capable of carrying out distance learning sessions between two workstations using Netscape graphic user interface and a number of Java applets for handling the control of the system.

Load Estimation and Control in Best-Effort Network Domains

A mechanism for the estimation of the available bandwidth between two end-points of a best-effort network is presented. The estimation is obtained by a simple statistical analysis of the effect that the network has on a synchronous train of packets. The possibility of exploiting self-similar characteristics of the delay jitters is also discussed, and a possible use of the estimates for management actions is suggested.

An adaptive neural network admission controller for dynamic bandwidth allocation

In an access node to a hybrid-switching network (e.g., a base station handling the downlink in a cellular wireless network), the output link bandwidth is dynamically shared between isochronous (guaranteed bandwidth) and asynchronous traffic types. The bandwidth allocation is effected by an admission controller, whose goal is to minimize the refusal rate of connection requests as well as the loss probability of packets queued in a finite buffer. Optimal admission control strategies are approximated by means of backpropagation feedforward neural networks, acting on the embedded Markov chain of the connection dynamics. The case of unknown, slowly varying, input rates is explicitly considered. Numerical results are presented, comparing the approximation with the optimal solution obtained by dynamic programming.

Real-Time Estimation of the Link Capacity in Multimedia Networks

In order to guarantee quality of service, we must be able to calculate the link capacity, i.e., the total number of calls of different types that can be admitted on a single link at any given time. In this paper we use the schedulable region to represent the link capacity, and, we present a new approach for computing the schedulable region. Our methodology relies on real—time quality of service measurements to dynamically compute the size and shape of the schedulable region. We do not make any assumptions on traffic source models or scheduler operations.

Estimating the available bandwidth for real-time traffic over best effort networks

A mechanism for the estimation of the available bandwidth between two end-points of a best-effort network is presented. The estimation is obtained by a simple statistical analisys of the effect which the network has on a synchronous train of packets. The possibility of exploiting self-similar characteristics of the delay jitters is also discussed.

A call admission control strategy for multiservice wireless cellular packet networks

A simple connection control system for multiservice cellular wireless networks is presented. Mobile stations are classified depending on the traffic they generate (e.g., voice, data). Within each class, two subclasses are also identified: stations which have originated inside the cell and stations which come from adjacent cells. The connection control mechanism is carried out by considering a number of priorities among the various classes and their subclasses. It works on two levels: static and dynamic. The static level looks at “packet-level” quality of service (QoS), such as cell loss and delay, while the dynamic level takes care of connection dynamics and allows the load of the system to be driven with respect to the various subclasses. Results that illustrate the performance of this control mechanism are presented.

A call-level access control strategy for integrated services wireless packet networks

A simple call admission control system for multiservice wireless networks is presented. By means of it, connection requests coming from different “classes” of traffic (i.e., voice, data) are regulated in order to meet a set of Quality of Service (QoS) constraints. The admission mechanisn works on two levels: a static and a dynamic one. The static level looks at “packet-level” QoS, such as cell loss and delay, while the dynamic level takes care of call dynamics and allows to drive the load of the system with respect to the various traffic classes. Whenever a connection recpiest for the n-th traffic class appears, both levels of control intervene and the call is accepted if both of them do not reject it. The decision at the static level is deterministic, and it is based on the concept of Schedulable Region; the decision at the dynamic level is probabilistic and it is based on the concept of Input Rate Region, which is explained in the paper. Results that show the performance of this control mechanism are presented.

Neural approximations of optimal allocation policies for hybrid multiplexing

A multiplexing structure is considered, where TDM frames are used to carry both isochronous, circuit-switched, and asynchronous, packet-switched traffic. Control functions are sought, whose task is that of deciding the allocation of the frame capacity between the two traffic types. The problem is defined in the context of Markov decision processes, and multilayer feedforward neural networks are used to approximate the optimal control laws. A backpropagation algorithm is described, which exploits the finiteness of the systems state. The procedure is conceived in the framework of repetitive (receding horizon) control schemes. Numerical results are presented, as well as comparisons with the control laws obtained by applying a dynamic programming algorithm.