This innovation uses a deep learning algorithm to learn a curvature parameterization to apply to the line-of-sight unit vector provided by a missile guidance navigation system. In one implementation the bias is computed using a deep neural network with a recurrent layer which allows the network to predict future target behavior from the history of target behavior during the engagement. The network is optimized using meta reinforcement learning (meta-RL) over an ensemble of engagement scenarios with varying target behavior. Other applications could include spacecraft guidance navigation and control. This method is compatible with both passive sensors and single loop integrated guidance and control. Potential hardware implementations of the innovation are software running on a CPU or tensor processing unit (TPU), and neuromorphic implementations. Background: Proportional navigation (PN) and related missile guidance laws are not optimal for the case of a maneuvering target guidance laws for maneuvering targets include the augmented PN (APN) guidance law, where the commanded acceleration from the PN guidance law is augmented with an additional term that is a function of the estimated target acceleration. However, although APN works well with targets maneuvering with a constant acceleration, its performance can actually be worse than standard PN for weaving targets and other more complex target maneuvers. A missile system company implementing the innovation would obtain a significant competitive advantage. Applications:
Advantages: