Abstract: A vehicle management system comprising a computer system and an agent. The agent comprises a machine learning model and a rule system. The machine learning model system is trained to receive observations for the vehicle system and select a behavior for the vehicle system in response to receiving the observations. The rule system is configured to select a set of actions to execute the behavior for the vehicle system in response to a selection of the behavior by the machine learning model system.

Abstract: A vehicle behavior system comprises a computer system, an observation processor, and neural networks. The observation processor and the neural networks are located in the computer system. The observation processor is configured to receive observations for a vehicle system. The observations are for a current time. The observation processor is configured to extract features from the observations. The neural networks are configured to receive the features extracted from the observations and estimate a behavior for the vehicle system for time steps in response to receiving features extracted from the observations processed by the observation processor. Each of the neural networks is trained to estimate the behavior for the vehicle system for a different time step in the time steps.

Abstract: A local agent of a multi-agent reinforcement learning (MARL) system is disclosed, the local agent comprising a MARL network comprising at least one local hidden layer responsive to a plurality of local observations. A transmitter is configured to transmit an output of the local hidden layer to at least one remote agent, and a receiver is configured to receive an output of a remote hidden layer from the at least one remote agent. A combiner module is configured to combine the local hidden layer output with the remote hidden layer output to generate a combined hidden layer output, wherein the MARL network is configured to process the combined hidden layer output to generate at least one action value for the local agent.

Abstract: Described is a system for controlling multiple autonomous platforms. A training process is performed to produce a trained learning agent in a simulation environment. In each episode, each controlled platform is assigned to one target platform that produces an observation. A learning agent processes the observation using a deep learning network and produces an action corresponding to each controlled platform until an action has been produced for each controlled platform. A reward value is obtained corresponding to the episode. The trained learning agent is executed to control each autonomous platform, where the trained agent receives one or more observations from one or more platform sensors and produces an action based on the one or more observations. The action is then used to control one or more platform actuators.

Abstract: An example includes a method for training an agent to control an aircraft. The method includes: selecting, by the agent, first actions for the aircraft to perform within a first environment respectively during first time intervals based on first states of the first environment during the first time intervals, updating the agent based on first rewards that correspond respectively to the first states, selecting, by the agent, second actions for the aircraft to perform within a second environment respectively during second time intervals based on second states of the second environment during the second time intervals, and updating the agent based on second rewards that correspond respectively to the second states. At least one first rule of the first environment is different from at least one rule of the second environment.

Abstract: An example includes a method for training an agent to control an aircraft. The method includes: selecting, by the agent, first actions for the aircraft to perform within a first environment respectively during first time intervals based on first states of the first environment during the first time intervals, updating the agent based on first rewards that correspond respectively to the first states, selecting, by the agent, second actions for the aircraft to perform within a second environment respectively during second time intervals based on second states of the second environment during the second time intervals, and updating the agent based on second rewards that correspond respectively to the second states. At least one first rule of the first environment is different from at least one rule of the second environment.

Abstract: A method includes receiving, by machine-learning logic, observations indicative of a states associated with a first and second group of vehicles arranged within an engagement zone during a first interval of an engagement between the first and the second group of vehicles. The machine-learning logic determines actions based on the observations that, when taken simultaneously by the first group of vehicles during the first interval, are predicted by the machine-learning logic to result in removal of one or more vehicles of the second group of vehicles from the engagement zone during the engagement. The machine-learning logic is trained using a reinforcement learning technique and on simulated engagements between the first and second group of vehicles to determine sequences of actions that are predicted to result in one or more vehicles of the second group being removed from the engagement zone. The machine-learning logic communicates the plurality of actions to the first group of vehicles.

Abstract: A computer-implemented method for predicting behavior of aircraft is provided. The method comprises inputting a current state of a number of aircraft into a number of hidden layers of a neural network, wherein the neural network is fully connected. An action applied to the aircraft is input into the hidden layers concurrently with the current state. The hidden layers, according to the current state and current action, determine a residual output that comprises an incremental difference in the state of the aircraft resulting from the current action. A skip connection feeds forward the current state of the aircraft, and the residual output is added to the current state to determine a next state of the aircraft.

Abstract: 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.

Abstract: Training a compressive encoder is provided. The method comprises calculating a difference between a current state of an aircraft and a previous state. The current state comprises a reduced state wherein the value of a specified parameter is missing. The difference is input into compressive layers of a neural network comprising an encoder. The compressive layers learn, according to the difference, a value for the missing parameter. The current state and learned value are concurrently fed into hidden layers of a fully connected neural network comprising a decoder. An action applied to the aircraft is input into the hidden layers concurrently with the current state and learned value. 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.

Abstract: Training adversarial aircraft controllers is provided. The method comprises inputting current observed states of a number of aircraft into a world model encoder, wherein each current state represents a state of a different aircraft, and wherein each current state comprises a missing parameter value. A number of adversarial control actions for the aircraft are input into the world model encoder concurrently with the current observed state, wherein the adversarial control actions are generated by competing neural network controllers. The world model encoder generates a learned observation from the current observed states and adversarial control actions, wherein the learned observation represents the missing parameter value from the current observed states. The learned observation and current observed states are input into the competing neural network controllers, wherein each current observed state is fed into a respective controller.

Abstract: An example method for training a machine learning algorithm (MLA) to control a first aircraft in an environment that comprises the first aircraft and a second aircraft can involve: determining a first-aircraft action for the first aircraft to take within the environment; sending the first-aircraft action to a simulated environment; generating and sending to both the simulated environment and the MLA, randomly-sampled values for each of a set of parameters of the second aircraft different from predetermined fixed values for the set of parameters; receiving an observation of the simulated environment and a reward signal at the MLA, the observation including information about the simulated environment after the first aircraft has taken the first-aircraft action and the second aircraft has taken a second-aircraft action based on the randomly-sampled values; and updating the MLA based on the observation of the simulated environment, the reward signal, and the randomly-sampled values.

Abstract: A method includes obtaining multiple sets of trajectory data, each descriptive of trajectories of two or more objects (e.g., first and second objects). The method also includes generating transformed trajectory data based on the trajectory data. Each set of transformed trajectory data is descriptive of the trajectories of the two or more objects in a normalized reference frame in which a movement path of the first object is constrained. The method further includes generating feature data, performing a clustering operation based on the feature data to generate a set of trajectory clusters, and generating training data based on the set of trajectory clusters. The method further includes using the training data to train a machine learning classifier to classify particular trajectory patterns.

Abstract: A system is provided. The system includes a first platform including a first platform level agent configured to direct one or more actions of the first platform based on at least one of a selected target or a selected goal. The system also includes a computer system in communication with the first platform level agent. The computer system programmed to a) execute a supervisor level agent configured to select at least one of a target or a goal for one or more platforms including the first platform, b) receive targeting information including one or more targets, c) receive platform information for the one or more platforms, d) select, by the supervisor level agent, a target of the one or more targets based on the target information and the platform information, and e) transmit, to the first platform level agent, the selected target.

Abstract: A method includes receiving, by machine-learning logic, observations indicative of a states associated with a first and second group of vehicles arranged within an engagement zone during a first interval of an engagement between the first and the second group of vehicles. The machine-learning logic determines actions based on the observations that, when taken simultaneously by the first group of vehicles during the first interval, are predicted by the machine-learning logic to result in removal of one or more vehicles of the second group of vehicles from the engagement zone during the engagement. The machine-learning logic is trained using a reinforcement learning technique and on simulated engagements between the first and second group of vehicles to determine sequences of actions that are predicted to result in one or more vehicles of the second group being removed from the engagement zone. The machine-learning logic communicates the plurality of actions to the first group of vehicles.

Abstract: Apparatus and methods for training a machine learning algorithm (MLA) to control a first aircraft in an environment that comprises the first aircraft and a second aircraft are described. Training of the MLA can include: the MLA determining a first-aircraft action for the first aircraft to take within the environment; sending the first-aircraft action from the MLA; after sending the first-aircraft action, receiving an observation of the environment and a reward signal at the MLA, the observation including information about the environment after the first aircraft has taken the first-aircraft action and the second aircraft has taken a second-aircraft action, the reward signal indicating a score of performance of the first-aircraft action based on dynamic and kinematic properties of the second aircraft; and updating the MLA based on the observation of the environment and the reward signal.

Abstract: A method includes obtaining multiple sets of trajectory data, each descriptive of trajectories of two or more objects (e.g., first and second objects). The method also includes generating transformed trajectory data based on the trajectory data. Each set of transformed trajectory data is descriptive of the trajectories of the two or more objects in a normalized reference frame in which a movement path of the first object is constrained. The method further includes generating feature data, performing a clustering operation based on the feature data to generate a set of trajectory clusters, and generating training data based on the set of trajectory clusters. The method further includes using the training data to train a machine learning classifier to classify particular trajectory patterns.

Abstract: Described is a system for autonomous behavior generation. The system includes both a high-level controller and a low-level controller. The high-level controller receives observations from an environment and, using a neural net, selects a high-level behavior based on the observations from the environment. The low-level controller generates an output command for a scripted action based on the selected one high-level behavior. After generating the output command, the system can implement an action, such as causing a device to perform the scripted action.

Abstract: Apparatus and methods for training a machine learning algorithm (MLA) to control a first aircraft in an environment that comprises the first aircraft and a second aircraft are described. Training of the MLA can include: the MLA determining a first-aircraft action for the first aircraft to take within the environment; sending the first-aircraft action from the MLA; after sending the first-aircraft action, receiving an observation of the environment and a reward signal at the MLA, the observation including information about the environment after the first aircraft has taken the first-aircraft action and the second aircraft has taken a second-aircraft action, the reward signal indicating a score of performance of the first-aircraft action based on dynamic and kinematic properties of the second aircraft; and updating the MLA based on the observation of the environment and the reward signal.

Abstract: Described is a learning system for multi-agent applications. In operation, the system initializes a plurality of learning agents. The learning agents include both tactical agents and strategic agents. The strategic agents take an observation from an environment and select one or more of the tactical agents to produce an action that is used to control a platform's actuators or simulated movements in the environment to complete a task. Alternatively, the tactical agents produce the action corresponding to a learned low-level behavior to control the platform's actuators or simulated movements in the environment to complete the task.