Vehicle control systems including neural network control policies with non-linear h-infinity robustness

A method of generating a neural network control policy, the method comprising:

• obtaining a performance objective parameter associated with a control system of a device;
• obtaining a policy optimizing algorithm corresponding to the performance objective parameter;
• defining a system state space, a control action space, and a disturbance space, each associated with the control system;
• generating a first candidate neural network control policy based on the system state space, a control action space, and a disturbance space, using a first value for a tunable robustness parameter;
• generating a second candidate neural network control policy based on the system state space, a control action space, and a disturbance space, using a second value for the tunable robustness parameter, wherein the first value is different than the second value;
• comparing a robustness output of the first candidate neural network control policy with a robustness output of the second candidate neural network control policy;
• selecting the first candidate neural network control policy or the second candidate neural network control policy according to comparing the robustness output of each candidate neural network control policy, wherein the selected one of the first candidate neural network policy or the second candidate neural network control policy has a non-linear H-infinity robustness guarantee;
• and
• automatically controlling the device according to the selected one of the first candidate neural network policy or the second candidate neural network control policy.

A controller configured to implement a neural network control policy for a system, the controller comprising:

• memory configured to store computer-executable instructions and a neural network control policy;
• and
• at least one processor configured to:
• receive an input;
• provide a projection from the neural network control policy onto a robust action set, wherein the neural network control policy has a non-linear H-infinity robustness guarantee;
• and
• automatically control a component associated with the system according to the neural network control policy and the robust action set, wherein the system includes at least one of a vehicle control system, an aerial device control system, and a power grid control system.