A neural net arbitrator for large crossbar packet-switches

Guaranteeing efficiency and safety of a Hopfield network for crossbar switching

“…Simulations demonstrated that the network E 2 achieved 98% optimal switching [4]. Note that representation E 2 is multilinear, so that network E 2 always converges to a vertex of a hypercube.…”

“…Many trials of crossbar switching by Hopfield networks have been made, and have shown good performance, achieving optimal switching of nearly 100% [1,4,1113]. …”

“…Hence, the neural representation E 1 is based on too strong an assumption and is not practical. Thus, the next representation, which consists only of the first two terms of E 1 and does not require a priori knowledge of the number of packets at maximum throughput, is sometimes employed [4].…”

“…In telecommunications, for example, these include packet switching and circuit routing in interconnection networks, and channel assignment in cellular radio systems [15]. In particular, simulations have shown Hopfield networks for packet crossbar switchings achieve very good performance, obtaining optimal solutions nearly 100% [1,4,1113]. Thus, with this good performance, neural systems for highspeed computation by parallel processing are promising.…”

Using recurrent neural networks to learn the structure of interconnection networks

“…In that paper, Hillis makes use of simulated evolution to construct sorting networks. It should be also mentioned that neural networks have been previously used for interconnection network routing: for example, see (Brown, 1989;Brown & Liu, 1990;Funabiki et al, 1991;Funabiki et al, 1993;Goudreau & Giles, 1992;Hakim & Meadows, 1990;Lee & Chang, 1993;Marrakchi & Troudet, 1989;Melsa et al, 1990a;Melsa ct al., 1990b;Thomopoulos et al, 1991;Troudet & Waiters, 1991). However, none of these methods learned the structure of the interconnection networks; the structure of the interconnection network was always directly mapped into the neural network.…”