Predictive modeling of aircraft dynamics
Training an encoder is provided. The method comprises inputting a current state of a number of aircraft into a recurrent layer of a neural network, wherein the current state comprises a reduced state in which a value of a specified parameter is missing. An action applied to the aircraft is input into the recurrent layer concurrently with the current state. The recurrent layer learns a value for the parameter missing from current state, and the output of the recurrent layer is input into a number of fully connected hidden layers. The hidden layers, according to the current state, learned value, and current action, determine a residual output that comprises an incremental difference in the state of the aircraft resulting from the current action.
1 . A computer-implemented method for training an encoder, the method comprising: using a number of processors to perform the steps of:
inputting a current state of a number of aircraft into a recurrent layer of a neural network, wherein the current state comprises
trajectory data and heading angle data, and wherein the current state comprises a reduced state in which a value of a specified parameter is missing;
inputting an action applied to the aircraft into the recurrent layer concurrently with the current state;
learning, by the recurrent layer, a value for the parameter missing from current state;
inputting an output of the recurrent layer into a number of fully connected hidden layers;
determining, by the hidden layers according to the current state, learned value, and current action, a residual output that comprises an incremental difference from one timestep to the next timestep in the state of the aircraft resulting from the current action, wherein
the current state is an accumulation of residual outputs from at least one hundred timesteps;
inputting the residual output into a loss function;
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 recurrent layer and hidden layers; and
updating connection weights in the recurrent layer and hidden layers according to the computed gradients.
2 . 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.
3 . 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.
4 . The method of claim 1 , wherein the neural network models the behavior of multiple aircraft.
5 . The method of claim 4 , wherein the action is one-hot encoded for input into the recurrent layer.
6 . The method of claim 1 , further comprising applying a modulo operation to the residual output, wherein the modulo operation ensures the angle changes within a specified range.
7 . A system for training an encoder, the system comprising: a storage device configured to store program instructions; and one or more processors operably connected to the storage device and configured to execute the program instructions to cause the system to:
input a current state of a number of aircraft into a recurrent layer of a neural network, wherein the current state comprises
trajectory data and heading angle data, and wherein the current state comprises a reduced state in which a value of a specified parameter is missing;
input an action applied to the aircraft into the recurrent layer concurrently with the current state;
learn, by the recurrent layer, a value for the parameter missing from current state;
input an output of the recurrent layer into a number of fully connected hidden layers;
determine, by the hidden layers according to the current state, learned value, and current action, a residual output that comprises an incremental difference from one timestep to the next timestep in the state of the aircraft resulting from the current action, wherein
the current state is an accumulation of residual outputs from at least one hundred timesteps;
input the residual output into a loss function; 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 recurrent layer and hidden layers; and
update connection weights in the recurrent layer and hidden layers according to the computed gradients.
8 . The system of claim 7 , 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.
9 . The system of claim 7 , wherein the action is selected randomly from a set of possible actions according to a number of control policies for a controller.
10 . The system of claim 7 , wherein the current state comprises at least one of:
trajectory data; or
a heading angle.
11 . A computer program product for training an encoder, the computer program product comprising: a computer-readable storage medium having program instructions embodied thereon to perform the steps of:
inputting a current state of a number of aircraft into a recurrent layer of a neural network, wherein the current state comprises
trajectory data and heading angle data, and wherein the current state comprises a reduced state in which a value of a specified parameter is missing;
inputting an action applied to the aircraft into the recurrent layer concurrently with the current state;
learning, by the recurrent layer, a value for the parameter missing from current state;
inputting an output of the recurrent layer into a number of fully connected hidden layers;
determining, by the hidden layers according to the current state, learned value, and current action, a residual output that comprises an incremental difference from one timestep to the next timestep in the state of the aircraft resulting from the current action, wherein
the current state is an accumulation of residual outputs from at least one hundred timesteps;
inputting the residual output into a loss function;
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 recurrent layer and hidden layers; and
updating connection weights in the recurrent layer and hidden layers according to the computed gradients.
12 . The computer program product of claim 11 , further comprising instructions for:
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.
13 . The computer program product of claim 11 , wherein the action is selected randomly from a set of possible actions according to a number of control policies for a controller.
14 . The computer program product of claim 11 , wherein in the case of heading angle, further comprising applying a modulo operation to the residual output, wherein the modulo operation ensures the angle changes within a specified range.