Neural networks for exact solution of constrained optimal control problems

Abstract

A novel kind of neural network for solving constrained optimal control problems is proposed in this paper. The major difference from other related neural networks is that the local inequality constraints on state and control variables are dealt with by means of the Kuhn-Tucker multiplier neurons which operate in a one-sided saturated mode so that the additional slack variables for converting the local inequality constraints into the equality ones can be avoided and an exact optimal solution to constrained optimal control problems can be found without requiring a sufficiently large value of penalty parameter. It can be shown that under suitable conditions the state trajectory of the neural network converges to an equilibrium point which corresponds to a local optimal solution to the original problem. The simulation results are given which illustrate the feasibility and performance of the proposed neural network in solving constrained optimal control problems.