Investigation of neural networks for the scheduling and allocation problem in high-level synthesis

Abstract

In recent years neural network have been shown to be quite effective in solving difficult combinatorial optimization problems. In this work a Hopfield neural network is used to schedule operations in a dataflow graph. This is an important step in behavioral synthesis systems. These operations must be assigned to a limited number of control steps, functional units, and busses. Also, there is an objective to minimize the lengths of data paths. Current methods which do this type of scheduling typically rely on heuristic algorithms. The neural network devised to solve this problem is one of the most complex to date. A special mechanism, "flag" neurons, was developed to enable the neural network to encode a bussing constraint. The neural network has been tested with problems from literature and problems randomly generated. The results have been consistently superior to those produced by a heuristic algorithm called ALAP.