Design of differential game controllers using adaptive critic neural networks

An adaptive critic structure including three neural networks was developed to solve the two point boundary value problem of differential games. Two control neural networks were used to optimize the controllers on two sides of the differential games, and a co-state neural network was used to approximate the co-state variables in Hamiltonian function. The output of co-state network was used to correct the output of the control networks, and the two convergent control networks can be used as feedback controllers on two sides of the differential games system respectively. The solution of differential games based on neural networks was compared with the one based on Chebyshev technique. The simulation results of the pursing-escaping differential games show that the neural network controllers are consistent with the optimal solution and present good robustness with respect to the initial conditions and measuring noises.