Predictive modeling of aircraft dynamics

1 . A computer-implemented method for training a compressive encoder overcoming missing a value for a specified parameter of a current state of each aircraft in a number of aircraft and directing a next action of a host aircraft, the method comprising using a number of processors to perform the steps of:

sensing, through a sensor system of the host aircraft, parameters of each aircraft in the number of aircraft flying in a sensor cone of the host aircraft indicating the current state, respectively, for each of the number of aircraft in the sensor cone, wherein the current state comprises a velocity, respectively, of each of the number of aircraft;

calculating a difference between the current state, respectively, of each of the number of aircraft in the sensor cone and a previous state, respectively, of each of the number of aircraft in the sensor cone, wherein the current state comprises a reduced state in which the value of the specified parameter is missing;

inputting the difference into compressive layers comprising an encoder of a neural network, wherein the compressive layers, each comprise respectively, a rectified linear unit activation;

generating, by the compressive layers according to the difference, a learned value for the specified parameter missing from the current state;

inputting the current state into four fully connected hidden layers, each applying a second rectified linear unit activation, in communication with a neuroevolution controller and the compressive layers, comprising a decoder of the neural network;

inputting the learned value into a controller and into the hidden layers concurrently with the current state;

inputting, as one-hot encoding from control policies converted to a vector of values, an action applied to the aircraft into the hidden layers concurrently with the current state and learned value;

determining, by the hidden layers according to the current state, learned value, and current action, a residual output that comprises an incremental difference in the current state of the aircraft resulting from the current action;

calculating a ground truth residual consisting of the difference between the current state and a previous state of at least one of the aircraft;

computing, using the difference between the residual output and a ground truth residual, a loss function;

modifying, using a gradient descent optimization, at least some neural network weights;

generating, by the controller using the current state and the learned value, the next action;

executing, by the host aircraft, the next action.

2 . The method of claim 1 , further comprising:

the number of aircraft comprising adversary aircraft;

the next action of the host aircraft comprising a guidance for a weapon system against the adversary aircraft;

inputting the residual output into a loss function;

applying, for any numerical conditioning challenges encountered during training, a loss function only to a residual output;

inputting a number of ground truth residual values into the loss function concurrently with the residual output;

inputting an output from the loss function into a gradient descent optimizer that computes gradients for the compressive layers and hidden layers; and

updating connection weights in the compressive layers and hidden layers according to computed gradients.

3 . The method of claim 1 , further comprising:

inputting the current state into a neural network controller;

inputting the learned value into the controller concurrently with the current state; and

determining, by the controller, a next action to be applied to the aircraft according to the current state and learned value.

4 . The method of claim 1 , wherein the action is selected randomly from a set of possible actions according to a number of control policies for a controller.

5 . The method of claim 1 , wherein the neural network models a behavior, respectively, of each of the number of aircraft in the sensor cone.

6 . The method of claim 1 , wherein the current state comprises trajectory data.

7 . The method of claim 1 , wherein the current state comprises a heading angle.

8 . The method of claim 7 , further comprising applying a modulo operation to the residual output, wherein the modulo operation ensures the angle changes within a specified range.

9 . The method of claim 1 , wherein the action is one-hot encoded for input into the hidden layers.

10 . An environment comprising:

a host aircraft and

a system configured to train a compressive encoder to direct next action of a the host aircraft, wherein the system comprises:

a sensor system that comprises a sensor cone of the host aircraft configured to gather parameters that indicate a current state of each aircraft in a number of aircraft in flight in the sensor cone, wherein the current state comprises a velocity, respectively, of each of the number of aircraft;

a storage device configured to store program instructions; and

one or more processors that comprise a controller configured to determine the next action of the host aircraft and operably connected to the storage device and configured to execute the program instructions to cause the system to:

calculate a difference between the current state, respectively, of each of the number of aircraft in the sensor cone and a previous state, respectively, of each of the number of aircraft in the sensor cone, wherein the current state comprises a reduced state in which a value of a specified parameter is missing;

input the difference into compressive layers comprising an encoder of a neural network, wherein the compressive layers, each comprise respectively, a rectified linear unit activation

generate by the compressive layers according to the difference, a learned value for the parameter missing from current state;

input the current state into four fully connected hidden layers, each applying a second rectified linear unit activation, comprising a decoder of a neural network, in communication with a neuroevolution controller and the compressive layers;

input the learned value into the controller and into the hidden layers concurrently with the current state;

input, as one-hot encoding from control policies converted to a vector of values, an action applied to the aircraft into the hidden layers concurrently with the current state and learned value;

determine, by the hidden layers according to the current state, learned value, and current action, a residual output that comprises an incremental difference in the current state of the aircraft resulting from the current action;

calculate a ground truth residual consisting of the difference between the current state and a previous state of at least one of the aircraft;

compute, a loss function based on the difference between the residual output and a ground truth residual;

modify, based upon a gradient descent optimization, at least some neural network weights;

generate, by the controller based upon the current state and the learned value, the next action; and

execute, by the host aircraft, the next action.

11 . The environment of claim 10 , wherein the processors further execute instructions to:

input the residual output into a loss function;

apply, for any numerical conditioning challenges encountered during training, a loss function only to a residual output;

input a number of ground truth residual values into the loss function concurrently with the residual output;

input an output from the loss function into a gradient descent optimizer that computes gradients for the compressive layers and hidden layers; and

update connection weights in the compressive layers and hidden layers according to computed gradients.

12 . The environment of claim 10 , wherein the processors further execute instructions to:

input the current state into a neural network controller;

input the learned value into the controller concurrently with the current state; and

determine, by the controller, a next action to be applied to the aircraft according to the current state and learned value.

13 . The environment of claim 10 , wherein;

the action is selected randomly from a set of possible actions according to a number of control policies for a controller;

the number of aircraft comprises adversary aircraft; and

the next action of the host aircraft comprises a guidance for a weapon system against the adversary aircraft.

14 . The environment of claim 10 , wherein the current state comprises at least one of:

trajectory data; or

a heading angle.

15 . A computer program product configured to train a compressive encoder wherein the computer program product comprise a controller configured to direct a next action of a host aircraft and a computer-readable storage medium that comprises program instructions configured to:

receive, through a sensor system of the host aircraft, parameters of a number of aircraft in a sensor cone of the host aircraft that indicate a current state, respectively, for each of the number of aircraft in the sensor cone, wherein the current state comprises a velocity, respectively, of each of the number of aircraft;

calculate a difference between the current state, respectively, of each of the number of aircraft in the sensor cone and a previous state, respectively, of each of the number of aircraft in the sensor cone, wherein the current state comprises a reduced state in which a value of the specified parameter is missing;

input the difference into compressive layers comprising an encoder of a neural network, wherein the compressive layers, each comprise respectively, a rectified linear unit activation;

generate, by the compressive layers according to the difference, a learned value for the specified parameter missing from the current state;

input the current state into four fully connected hidden layers, each applying a second rectified linear unit activation, comprising a decoder of the neural network in communication with a neuroevolution controller and the compressive layers;

input the learned value into a controller and into the hidden layers concurrently with the current state;

input, as one-hot encoding from control policies converted to a vector of values, an action applied to the aircraft into the hidden layers concurrently with the current state and learned value;

determine, based upon the hidden layers according to the current state, the learned value, and current action, a residual output that comprises an incremental difference in the current state of the aircraft resulting from the current action;

calculate a ground truth residual that consists of the difference between the current state and a previous state of at least one of the aircraft;

compute, a loss function based on the difference between the residual output and a ground truth residual;

modify, based upon a gradient descent optimization, at least some neural network weights;

generate, by the controller using the current state and the learned value, a the next action; and

execute the next action by the host aircraft.

16 . The computer program product of claim 15 , further comprising instructions for:

inputting the residual output into a loss function;

applying, for any numerical conditioning challenges encountered during training, a loss function only to a residual output;

inputting a number of ground truth residual values into the loss function concurrently with the residual output;

inputting an output from the loss function into a gradient descent optimizer that computes gradients for the compressive layers and hidden layers; and

updating connection weights in the compressive layers and hidden layers according to computed gradients.

17 . The computer program product of claim 15 , further comprising:

inputting the current state into a neural network controller;

inputting the learned value into the controller concurrently with the current state; and

determining, by the controller, a next action to be applied to the aircraft according to the current state and learned value.

18 . The computer program product of claim 15 , wherein the action is selected randomly from a set of possible actions according to a number of control policies for a controller.

19 . The computer program product of claim 15 , wherein;

the current state further comprises at least one of:

trajectory data; or

a heading angle;

the number of aircraft comprises adversary aircraft; and

the next action of the host aircraft comprises a guidance for a weapon system against the adversary aircraft.

20 . The computer program product of claim 19 , further comprising a modulo operation applied to the residual output for the heading angle, wherein the modulo operation ensures the angle changes within a specified range.