Neuron design and stability analysis of neural network for TSP

The factors that result in unstability and local optimization of Hopfield neural network employed by combinational optimization problem are investigated and a progressive algorithm is proposed. The presented paper proves that, for Hopfield neural network with complete negative connection, the system moving trajectory can be well controlled by modifying the activation function such that the computational energy function of the network can rapidly converge to its local minimum along the current maximum gradient direction. When the local optimal solution is obtained, a simulated annealing (SA) like method is imposed to produce enough disturbance to enforce the solution to move from local optimization to global optimization. The proposed algorithm consumes less time than that of general SA and gets better experimental results than other known methods