Neural On-line Learning In Missile Guidance

In this work we investigate the use of neural networks in providing control signals to solve the target intercept problem. The approach taken here is based on an architecture that contains an adaptive critic network which evaluates previous control actions and produces a complementary control to counteract target acceleration. In previous work we used a linear optimal control law to produce the primary missile command accelerations. In this work we replace the optimal control law with a neural network approximation of the optimal control law and modify network weights on-line to react to target acceleration. Results of this study are encouraging, however, they show that proper network training is a key issue. Design issues involving the use of adaptive critic networks are currently being investigated.