Neural network controller for a high-speed packet switch

A neural network input access scheme in a high-speed packet switch for broadband ISDN is presented. In this switch each input port maintains a separate queue for each of the outputs, thus there are n ² queues in an (n × n) switch. Using synchronous operation at most one packet per input and output port will be transferred in any slot in such a way as to maximize the throughput of the switch. The high transmission rates in broadband ISDN, with slot durations of the order of microseconds, demand that the choice of the packets be performed in real time. A conventional solution of this optimization problem cannot meet timing constraints. We propose a recurrent neural network maximizing the throughput of the switch, and determine the corresponding energy function, its optimized parameters, and the connection matrix. The energy function has linear cost terms with excellent convergence properties. The neural network has null programming complexity that avoids readjusting the parameters before presenting new inputs. From the hardware implementation point of view, because the neural network has O(n ²) neurons and O(n ³) connections, the sparse connection matrix will help in implementation. Finally, comparing simulation results with analytically derived upper and lower bounds we show close to optimal throughput.