The well-known difficulty with a wide array of static optimization problems is that their search spaces become too large to search for an optimal solution within reasonable time and computational resource bounds once the size of these static optimization problems increases to real-life dimensions. Algorithms implemented on conventional computing systems lack the computational power to search for an optimal solution of a large set of real-life size static optimization problems in a time efficient manner since they are not able to utilize the inherent parallelism that may exist. Artificial Neural Networks offer key advantages over conventional computing systems as optimization problem solvers: they can take advantage of the inherent parallelism of the optimization problem and offer very fast computation cycles for a hardware realization of the neural network algorithm. These advantages make Artificial Neural Networks a very attractive choice for addressing the static optimization problems of real-life complexity. At the present time, the scaleability properties of current Artificial Neural Network algorithms employed for static optimization are far from the desired state, which establishes the motivation for the proposed research project.
The goal of this project is to address and solve the scaling problem which Artificial Neural Network algorithms currently experience for static optimization problems. Towards that goal, computational promise of the Simultaneous Recurrent Network, a trainable and recurrent Artificial Neural Network algorithm, for static optimization problems will be explored and assessed. Existing neural optimizer algorithms are either trainable feedforward architectures or recurrent architectures with preprogrammed weight structures. This proposed research project will use a neural algorithm which is both a recurrent architecture and trainable will be employed to address the difficulties related to the scaling problem. It is hoped that recurrency and learning will form a very potent combination to address the challenging problem of scaling.
