Neural networks in real time control of telecommunication networks

A neural network based controller allocating capacity in a telecommunication network was proposed in order to overcome `real time´ response constraints. Two basic architectures were evaluated, a feedforward network-heuristic and a feedforward network-recurrent network, and compared against the benchmark-a linear programming optimizer, which was also used as a teacher to label the data samples for the feedforward network. It was found that the test systems were able to provide control with 99% and 95%, respectively, of the quality of the `optimal´ linear programming solution (measured in terms of expected call blocking), in a fraction of the time required by the linear programming optimizer