Active Neural Networks

The aim of this project is the adoption of a neural network computing paradigm on the Active Networks.
Currently, a prototype of an elementary neural unit has been designed, implemented and submitted to the ABONE Server Authority.
Neurald is a daemon capable of simulating the behavior of an elementary  unit of a ìmultilayerì neural network; it adopts the ìback propagationî algorithm as learning rule. This kind of neural network is constituted by input, hidden and output nodes. The first ones (the inputnodes)receive and forward the input patterns, the last ones (the output nodes) produce the network results, whilst the hidden units, interposed between input and output nodes, contribute to the learning process of the whole network.
A Neurald process can simulate each of above nodes depending on the initialization options given at startup.
An external process, the Front-End,has in charge the management and the synchronization of the whole network: it coordinates the actions of the distributed neural computing engine. In particular it distributes the input patterns to the front layer of the network, computes the global error and stops the learning phase when this error is less than a given threshold.

In our project the underlying idea is the adoption of a distributed neural network for approaching tasks which are specific of the computer network environment.

Two different possible applications are currently under investigation.
 

1. The first one is concerned with the application of a neural approach to the communication routing, as for instance the  detection of the shortest path towards all the network destinations or the setup of a minimum cost multicast tree.
2.  The second application is devoted to the classification of conforming packets in a congestion control scenario. This is, for instance, the case of a policing agent in a resource reservation environment.

Experiments and simulations are currently carried out over ABONE, the Active Network backbone.
 

Project Team

Antonio Chella
Giuseppe Di Fatta
Donatella Guarino
Giuseppe Lo Re
Giuseppe Favaro'

• Jari Hakkinen, Martin Lagerholm, Carsten Peterson, Bo Soderberg, "A Potts Neuron Approach to Communication Routing", Neural Computation, 10, 1587-1599 (1998) © 1998 Massachusetts Institute of Technology.

 
• Jari Hakkinen, Martin Lagerholm, Carsten Peterson, Bo Soderberg, "Local Routing Algorithms Based on Potts Neural Networks", LU TP 98-5 February, 28, 1998

 
• ...more to be added