Neural network with node gates and its application to nonlinear system control

Neural networks with node gates are proposed to solve complicated or large problems with `divide and conquer' approach. Each hidden node of the network has a node gate on its output channel which controls the flow of the output from the node. By opening and closing depending on situations, the node gates form a sub-network dynamically which gives the solution suited for the current situation. When the situation changes, the gate openings are also changed accordingly, and a different sub-network will emerge to give a new solution. In the paper, a mechanism that controls the gate opening is proposed as well as the learning method of the network weights and the parameters contained in the gates. The network is applied to nonlinear system control problem where a number of different situations occur and demand for different control strategies. The results show that the proposed network can deal with the change of the situations appropriately.