Neural Networks for Switching

T. X Brown

Neural Networks in Telecommunications, eds. B. Yuhas, et al. Kluwer,
Boston, 1994. pp. 11--36.

Abstract:
This chapter presents a framework for designing neural network
solutions using extensions of the winner-take-all circuit. Unlike the
Hopfield energy function approach that requires the researcher to first
define the constraints of the problem and then go through an imprecise
and obscuring energy function to define the weights, these networks
have properties that can be directly defined and controlled. This
direct approach allows efficient implementations that are scalable to
large sizes and as long as the external inputs are within defined
limits, the network will always satisfy the constraints embodied in the
winner-take-all circuits.

The winner-take-all circuit serves as an efficient means for
communicating the many constraints of the problems between the neurons
in our solutions.  The neural network design, based on simple
electrical components, has a direct implementation in hardware.  Thus,
the neural components can be concentrated where they can excel: as
highly interconnected massively parallel feedback elements.

While applicable to a variety of optimization problems, a review of the
literature reveals that more than 10 results on neural network switch
control (all that could be readily found) can be placed within this
framework with the immediate benefits of reduced connectivity by a
factor of at least one order of magnitude, the elimination of invalid
computation outputs, and for the crossbar packet switch, an improved
controller that should surpass the performance of any known high-speed
controller.

The solutions described show that neural networks can be applied to
practical problems.  We emphasize that although each instance of these
problems requires a different neural network, the solution is well
defined, even for large-size problems where the parallelism of the
neural network is most advantageous.