Optimization by pulsed recursive neural networks

This paper proposes a new recursive neural network, called pulsed neural network, for solving optimization problems. The motion equations of the neurons are directly derived from the constraints and the cost criteria. By alternating constraint satisfaction steps and pulsation steps to escape from local minima, the neural dynamics regularly provides valid solutions minimizing the cost criteria. Applying the previous neural network has the following advantages: it converges towards feasible solutions in finite time; the quality of the solutions and the satisfaction of the constraints are insensitive to a fine tuning of some parameters; it can propose a feasible solution in a bounded time; the network proposes various feasible solutions whose quality statistically increases with the number of iterations. We use a complex resource allocation problem to demonstrate the performances of this method and compare the performances to a simulated annealing algorithm.