This PDF file contains the front matter associated with SPIE Proceedings Volume XXXXX, including the Title Page, Copyright information, Table of Contents, and Committee Page.

Today’s AI, namely deep learning and neural networks, are fundamentally changing how software can be created. Deep neural networks are becoming more prevalent in processing the tsunami of data collected by sensors to extract important information. In this talk, we will overview Department of Defense’s sensors and processing Science and Technology portfolio and highlight a few examples of the exciting work.

Existing artificial intelligence (AI) agents are most successful on narrow, well-defined tasks, where training data are plentiful, well-labeled, and match the deployment scenarios. It is also possible to train an AI agent to do multiple tasks (such as identifying IEDs from image data as well as recognizing faces), given sufficient time to craft the training regime and network architecture. However, data and time are often in short supply -- multi-domain operations involve rapidly shifting and adaptive compositions of capabilities, against adversaries that will likely be adapting on the fly. We suggest that to be robust in such scenarios, AIs need to be capable of learning in the field with opportunistically-available data, limited human oversight, and limited or no access to ground truth. These challenges also apply to reinforcement learning agents, with the additional challenge that for such agents, bad decisions cascade and pose a further difficulty in learning new tasks on the fly. We present an overview of the challenges in enabling AI systems for multi-domain operations, current algorithmic approaches for developing lifelong learning agents, and potential techniques for evaluating them.

This paper presents an effective pipeline for generating cross-domain plan options for Multi-Domain Operations through the coordinated use of legacy planning systems. We accomplish this by combining the intuitive interface of our Distributed InteRactivE C2 Tool (DIRECT) with the efficient distributed AI reasoning of our Multi-Domain Adaptive Request Service (MARS). Existing military planning processes are typically stove-piped and use a variety of manual and domain-specific legacy planning tools. This makes it challenging to generate a plan that draws on resources from different domains for different tasks, resulting in underutilized resources and unfulfilled tasks. MARS uses a marketplace approach to address this problem by conducting auctions that enable legacy planners to participate in a cross-domain planning process. Moreover, MARS provides a framework for developing new automated domain planners in cases with limited existing automated planning support. The process works for either pre-execution or in-execution planning. DIRECT provides an interface to the operator to navigate and compare these cross-domain plan options in the context of the original plan to provide transparency and build trust with the operator. Our experiments demonstrate the efficacy of the approach using a combination of our Air Planner and surrogate legacy planners for space and ground fires, as compared to stove-piped systems. For example, we are able to increase the percentage of requests fulfilled for sequencing strike and battle damage assessment effects from 38% using only single domain air assets to 54% when using the auction to draw in air, land, and space assets.

The pace of innovation in Artificial Intelligence (AI) is completely unprecedented, enabling AI to augment humans and increase productivity and efficiency for critical tasks and operations across the Department of Defense. AI is envisioned to support Command and Control in Joint All-Domain Operations by significantly enhancing situational awareness from heterogenous platforms, systems, and sensors deployed across multiple operational domains, and enabling more rapid and improved decision-making. To achieve these benefits, new AI architectures and capabilities are rapidly evolving and being developed that are transforming the AI landscape – with core functions and technology layers of the AI Stack being distributed between the enterprise, the edge, and embedded on-platform. This concept paper will analyze and compare centralized vs. distributed AI architectures in support of all-domain operations and explore key attributes and capabilities to directly impact the resiliency and adaptability of the AI, and its ability to provide insights and decision-support at a speed and scale of relevance to and to converge effects across all warfighting domains to overwhelm the adversary and present them with multiple dilemmas. By overcoming the traditional dependency of Centralized AI architectures on human supervision to aggregated and engineer data for algorithmic processing, Distributed AI can drastically accelerate AI processing and integrate AI capabilities and insights from the enterprise to the edge of the battlefield that will maximize mission effectiveness, reduce risk, and save lives.

Absent the use of artificial intelligence (AI), each stage of the targeting process requires human input, which consumes time – a precious resource during high-intensity conflict. To counter that problem, the application of AI and related technologies could rapidly automate certain steps within the targeting process. The challenge, however, becomes how to responsibly speed up the process using AI, while also maintaining appropriate levels of human oversight based on level of risk acceptable to the commander. We address this challenge with the introduction of a fuzzy logic controller that is designed to automatically adapt and optimize the level of human-machine interaction based on the current targeting conditions. The logic controller uses the confidence of the AI algorithm at each stage of the targeting process, as well as the commander’s risk tolerance at the time, to automatically determine what stages of the targeting process can be trusted for AI to accelerate and which ones require explicit human verification. The final stage of verifying and authorizing a fire mission is reserved for humans only to make.

We propose a novel approach to explainable AI (XAI) based on the concept of "instruction" from neural networks. In this case study, we demonstrate in a human subject experiment how a superhuman neural network might instruct human trainees as an alternative to traditional approaches to XAI. Specifically, an AI examines human actions and calculates variations on the human strategy that lead to better performance. Experiments with a JHU/APL-developed AI player for the cooperative card game Hanabi suggest this technique makes unique contributions to explainability while improving human performance. One area of focus for Instructive AI is in the significant discrepancies that can arise between a human's actual strategy and the strategy they profess to use. This inaccurate self-assessment presents a barrier for XAI, since explanations of an AI's strategy may not be properly understood or implemented by human recipients. As an alternative, we propose a method of translating insights from an AI into corrections on human decision making. With neural networks, this allows a direct calculation of the changes in network weights needed to improve the human strategy to better emulate an AI that outperforms humans on certain metrics. Subjected to constraints (e.g. sparsity) these weight changes can be interpreted as recommended changes to human strategy (e.g. "value A more, and value B less"). Instruction from AI such as these functions both to help humans perform better at tasks, but also to better understand, anticipate, and correct the actions of an AI.

Computer vision, enabled by artificial intelligence and deep learning, has a nearly limitless number of possible applications, military and civilian. Object detection methods are a particularly notable type of computer vision, with broad usefulness in a variety of systems, such as autonomous vehicles, robotics, and security. Development of effective object detection methods faces many challenges; one such challenge of significance is a lack of good labeled data for the target domain, as hand labeled data is time-consuming and expensive to produce. Synthetic data generation seeks to solve this problem by programmatically generating both training data and labels simultaneously, allowing for the creation of arbitrarily large training datasets. However, synthetic data has several drawbacks { generating realistic imagery is challenging and computationally expensive, and models trained with synthetic data frequently suffer in accuracy when applied to real test data.

In our research, we use model explainability techniques to connect model predictions back to the model training data, in order to identify the most important features that need to be represented accurately in synthetic training data. Influence functions score model training samples based on how influential each sample was to a particular prediction, by approximating the effect of retraining the model after leaving each individual training sample out of the training set. In this work, we seek to extend influence functions to identify the most valuable features in real and synthetic training data for use in improving our synthetic data generation tools.

Explainable Artificial Intelligence (XAI) is the capability of explaining the reasoning behind the choices made by the machine learning (ML) algorithm which can help understand and maintain the transparency of the decision-making capability of the ML algorithm. Humans make thousands of decisions every day in their lives. Every decision an individual makes, they can explain the reasons behind why they made the choices that they made. Nonetheless, it is not the same in the case of ML and AI systems. Furthermore, XAI was not wideley researched until suddenly the topic was brought forward and has been one of the most relevant topics in AI for trustworthy and transparent outcomes. XAI tries to provide maximum transparency to a ML algorithm by answering questions about how models effectively came up with the output. ML models with XAI will have the ability to explain the rationale behind the results, understand the weaknesses and strengths the learning models, and be able to see how the models will behave in the future. In this paper, we investigate XAI for algorithmic trustworthiness and transparency. We evaluate XAI using some example use cases and by using SHAP (SHapley Additive exPlanations) library and visualizing the effect of features individually and cumulatively in the prediction process.

Robust defenses to the new threats require early determination of the adversary’s plan of attack. At the same time, automated path-planning systems often behave predictably and produce paths with recognizable characteristic features. The increasing adoption of autonomous systems changes the defense landscape, thwarting traditional defenses with rapid decision speeds, but opening up new weaknesses. In this work, we investigate several possibilities to exploit artifacts of path planning algorithms that might be used as tells, giving a defending commander early warning and advantage in mounting a defense. With the application of an integrated air-defense system in mind, we examine a use case of several high-value targets which are obstructed by known defending threats. We use incoming time-series track data to predict the most probable targets and future trajectories of an enemy platform. One such approach is to attempt to directly learn the mapping from threat track to target. However, such an approach is likely brittle, requiring large volumes of data and substantial retraining for each target laydown. By contrast, we attempt to exploit predictable features from the path-planning algorithm itself, by first classifying the path-planning algorithm being used, and including that knowledge in our target prediction algorithm. We demonstrate that it is possible to differentiate classes of path planning algorithms with high accuracy based on track data alone. Utilizing the underlying model, we can then predict likely track updates and likely targets. We discuss strengths and limitations of this approach with respect to the aim of adding a robust tool to the air-defense use case.

In the context of multi-domain operations (MDO), artificial intelligence (AI) systems support human operators by processing large volumes of electro-optical/infrared (EOIR) sensor data. In this paper, we demonstrate how confidential computing technology, incorporating a hardware-based root of trust, can provide systemic identity verification through mutual attestation, secure the integrity of AI models, and preserve the confidentiality of processed sensor data. Using the example of aircraft detection and classification, we describe how confidential computing can defend against adversarial machine learning (AML) attacks by providing intrinsic security at the tactical edge and within the distributed applications environment that characterizes MDO.

The use of deep learning in multi-domain operations to analyze satellite imagery is becoming particularly important. As deep learning models are computationally expensive to train and require vast amounts of data, there is an increasing trend towards the outsourcing of model training to the cloud, relying on pre-trained models and use of third party datasets. This poses serious security challenges and exposes users to adversarial attacks that aim to disrupt the training pipeline and insert Trojan behavior (backdoors) into the AI system. In this work, we demonstrate a method based on Generative Adversarial Networks (GANs) to automatically detect Trojans in deep learning computer vision models with a high detection accuracy (89%). We pick a land usage classification problem on satellite imagery for this demonstration. These results can easily be extended to other computer visons problems such as object detection. This technique is agnostic to the internal architecture of the deep learning network in question. We make no hard assumptions about the nature of the Trojan - size or pattern of the trigger, the targeted classes and the method of trigger injection.

In today’s battlefield, military operations are conducted in all domains including air, land, maritime, space, and cyberspace in addition to the electromagnetic spectrum, and the information environment. The data, network connections, sensors, computer systems and other information technology infrastructures utilized must be protected as adversaries are competing for dominance across all domains. Cyberspace attacks can cause significant disruption to a nation’s power grid and other critical infrastructure components resulting in collateral damage and even loss of life. With the continued shortage in Cybersecurity professionals, the government, universities and companies look to Artificial Intelligence Machine Learning (AI/ML) techniques to fill the gap. AI applications have been successfully applied to Department of Defense (DoD) to detect and mitigate cyber-attacks. Moreover, cybercriminals, state actors and other adversaries have used AI to plan and launch cyber-attacks.

The Army Research Laboratory is working to employ AI services using a novel framework that proactively detects and responds to attacks on a remote sensor network. The components of the framework include AI/ML for detecting malware and intrusion detection and prevention (IDP), cloud AI tools for sensor and system monitoring, log monitoring for detection of user behavior anomalies and sensor resiliency characteristics, and a visual mapping application to display sensor and target locations supporting geographically and social distancing of observers and parties of interest. Notifications are part of the framework to ensure system administrators are made aware of possible threats in near real-time.

This paper provides a detailed description of components of this framework and the initial use-case for a small distributed remote sensing network application.

This paper determines optimal strategies for transmitting messages in a mobile ad-hoc network (MANET) in a communications-limited and lossy communications environment. When an agent generates or receives a message it must decide to which neighbors, and how many times, that message is to be passed. The opposing goals are to (1) propagate all messages throughout the MANET quickly and (2) to minimize the total number of messages sent. We compare two optimized decision strategies for the agents, reinforcement learning (RL) and game theory (GT) methods. For the RL framework, each node in the MANET acts as a reinforcement learning agent who must learn optimal decisions for when to send messages to whom. For the GT framework, we create a game tree where the nodes encompass message knowledge and connectivity information, and the decision branches represent sending messages to neighbors. We solve the game using a Monte-Carlo Tree Search (MCTS) variation to determine the probability that a message is sent to a neighbor. Performance is assessed in terms of the total number of messages sent, and the length of time for a given percentage of messages to reach a given percentage of nodes. Experiments with MANETs of varying size and connectivity are considered, and the RL and GT performance and training speed are compared. The decision strategies are domain agnostic and may be applied to ground, air, surface, sub-surface, or satellite networks.

The multi-agent Value of Information (MAVoI) problem is concerned with the capacity of autonomous agents to evaluate the estimated value of available information from and for other agents. Such a paradigm would allow agents to limit information sharing under network constraints, environmental and adversarial, and also to decide when to prioritize inter-agent communications over further information-gathering. In our work, we contribute to the conceptual development of an MAVoI framework by expanding upon a previous paradigm [1] that connected the notion of the Value of Information (VoI), as it originates from decision theory, to the body of work on the Quality of Information (QoI/IQ). Within this paradigm, QoI characterizes the intrinsic attributes of information objects that may be used to judge them, while VoI judges their context- and user-specific utility and fitness-for-use. For teams of multiple, autonomous agents with distributed intelligence, we propose an additional category: the Affinity of Information (AoI). Since classical VoI quantifies the gain in value that comes from more information regarding the expected value of available actions, we propose AoI as those metrics and features that characterize the change in an agent’s state model. We then conceptually illustrate our proposed MAVoI taxonomy in the context of a distributed Simultaneous Localization and Mapping (SLAM) task, in which the mission objective is the localization of multiple objects of interest.

As the autonomy of intelligent systems continues to increase, the ability of humans to maintain control over machine behavior, work effectively in concert with them, and trust them, becomes paramount. Ideally, a machine’s plan of action would be accessible to and understandable by human team members, and machine behavior would be modifiable in real time, in the field, to accommodate unanticipated situations. The ability of machines to adapt to new situations quickly and reliably based on both human input and autonomous learning has the potential to enhance numerous human-machine teaming scenarios. Our research focuses on the question, “Can robots become competent and adaptive teammates by emulating human skill acquisition strategies?” In this paper we describe the Robotic Skill Acquisition (RSA) cognitive architecture and show preliminary results of teaming experiments involving a human wearing an augmented reality headset and a quadruped robot performing tasks related to reconnaissance. The goal is to combine instruction and discovery by integrating declarative symbolic AI and reflexive neural network learning to produce robust, explainable and trusted robot behavior, adjustable autonomy, and adaptive human-robot teaming. Humans and robots start with a playbook of modifiable hierarchical task descriptions that encode explicit task knowledge. Neural network based feedback error learning enables human-directed behavior shaping, and reinforcement learning enables discovery of novel subtask control strategies. It is anticipated that modifications to and transitions between symbolic and subsymbolic processing will enable highly adaptive behavior in support of enhanced situational awareness and operational effectiveness of human-robot teams.

We present multi-agent navigation methods for autonomous robots maneuvering in complex environments. Our approach considers the agent’s dynamics to compute smooth paths that avoid collisions with other agents and static and dynamic obstacles in a decentralized fashion. We assume each agent performs local sensing and has knowledge of its neighboring obstacles in its sensing region. For each agent, its local obstacle knowledge is used to construct collision avoidance constraints and is incorporated into an Model Predictive Control (MPC). The MPC considers the agent’s dynamics and the collision avoidance constraints to compute a smooth, locally optimal path for the agent. Additional task objectives, such as improved coverage, can be considered in the optimization problem for task-specific navigation. Real-world scenarios, such as military applications, can suffer from compromised sensing capabilities, resulting in noisy sensing data. Thus, we express the linear collision avoidance constraints as chance constraints and solve a probabilistic optimal control problem to improve safe operation. The chance constraints are reformulated as deterministic constraints assuming a Gaussian Mixture distribution. The computed control input safely navigates the agent toward its goal using the noisy sensing data. We compared the performance of our proposed methods with other state-of-the-art decentralized methods in dense scenarios with multiple obstacles. Our method produces smoother paths with improved collision avoidance performances even under noisy sensing. We observe that our method scales to tens of agents owing to its decentralized computation.

In this paper we explore the application of recent advances in AI/algorithmic game theory for autonomous tactical maneuvering in aerial combat mission scenarios. Tactical maneuvering requires autonomous aircrafts to learn defensive/offensive tactics while reasoning about an intelligent adversary, making it a challenging decision making problem. We consider a one-vs-one aerial dogfighting scenario and formulate it as a two-person zero-sum perfect information game. To solve this game online we apply simultaneous move Monte Carlo Tree Search (MCTS) since both aircrafts simultaneously take maneuvering decisions to gain tactical advantage. Compared to other techniques, MCTS enables efficient search over long horizons and uses self-play to select best maneuver in the current state while accounting for the opponent aircraft tactics. We explore different algorithmic choices in MCTS and demonstrate the framework numerically in a simulated 2D tactical maneuvering application.

We develop a hierarchical approach for controlling a team of aircraft in multi-agent adversarial environments. Each individual aircraft is equipped with a high-level agent that is solely responsible for target assignment decisions, and a low-level agent that generates actions based only on the selected target. We use distributed deep reinforcement learning to train the high-level agents, and neuroevolution to train the low-level agents. This approach leverages centralized training for decentralized execution to enable individual autonomy when communication is limited. Simulation results confirm the superiority of our proposed approach as compared to non-hierarchical multi-agent reinforcement learning methods.

Secure data communication is crucial in contested environments such as battlefields. In such environments, there is always risk of data breach through unauthorized interceptions. This may lead to unauthorized access to tactical information and infiltration into the systems. In this work, we propose a detailed training setup in the federated learning framework for object classification where the raw data will be maintained locally at the edge devices and will not be shared with a central server or with each other. The server sends a global model to edge devices, which is then trained locally at the edge, and the updated parameters are sent back to the central server, where they are aggregated, which takes place iteratively. This setup ensures robustness against malicious cyberattacks as well as reduce communication overhead. Furthermore, to tackle the irregularity in object classification task with a single data modality in such contested environment, a deep learning model incorporating multiple modalities is used as the global model in our proposed federated learning setup. This model can serve as a possible solution in object identification with multi-modal data. We conduct a comprehensive analysis on the importance of multi-modal approach compared to individual modalities within our proposed federate learning setup. We also provide a resource profiling based on memory requirements, training time, and energy usage on two resource constrained devices to demonstrate the feasibility of the proposed approach.

We develop a network synthesis scenario, which is built around a concrete perimeter surveillance application, yet we believe captures a number of the challenges and requirements that are common to other tactical communication and computational network applications. The proposed scenario addresses the problem of binary population identification within a perimeter: our goal is to synthesize a sensing and computing network that classifies people moving within a given perimeter into one of two categories (e.g., friend or foe). We discuss several open challenges that we organize across the following clusters: sensor placement, communication network provisioning and optimization, computational task placement, dynamic re-synthesis and resilience under adversarial settings. We also briefly discuss approaches that attempt to address such challenges.

To increase efficiency in all-domain military operations, such as achieving desired effects within severely compressed decision cycles, a multi-prong approach to technology is required. Traditional approaches provide a single algorithmic solution that keeps users, and acquisition efforts, “locked-in” to a result that may not be ideal for a particular mission, or possibly hinders technological advancements. Combining algorithms through ‘plug and play’, users can select the optimal algorithm(s) for their mission, and the acquisition community can easily improve existing or introduce new algorithms, thereby increasing performance while reducing cost. Our solution provides a set of heterogeneous independent optimization algorithms (IOAs) developed separately by three defense contractors, coordinated by a central Meta- Optimizer (MO) that is connected to a simulation and testing (S&T) environment.

The Internet of Battlefield Things (IoBT) will advance the operational effectiveness of infantry units. However, this requires autonomous assets such as sensors, drones, combat equipment, and uncrewed vehicles to collaborate, securely share information, and be resilient to adversary attacks in contested multi-domain operations. CAPD addresses this problem by providing a context-aware, policy-driven framework supporting data and knowledge exchange among autonomous entities in a battlespace. We propose an IoBT ontology that facilitates controlled information sharing to enable semantic interoperability between systems. Its key contributions include providing a knowledge graph with a shared semantic schema, integration with background knowledge, efficient mechanisms for enforcing data consistency and drawing inferences, and supporting attribute-based access control. The sensors in the IoBT provide data that create populated knowledge graphs based on the ontology.

This paper describes using CAPD to detect and mitigate adversary actions. CAPD enables situational awareness using reasoning over the sensed data and SPARQL queries. For example, adversaries can cause sensor failure or hijacking and disrupt the tactical networks to degrade video surveillance. In such instances, CAPD uses an ontology-based reasoner to see how alternative approaches can still support the mission. Depending on bandwidth availability, the reasoner initiates the creation of a reduced frame rate grayscale video by active transcoding or transmits only still images. This ability to reason over the mission sensed environment, and attack context permits the autonomous IoBT system to exhibit resilience in contested conditions.

In modern industrial settings, the quality of maintenance efforts directly influence equipment’s operational uptime and efficiency. Condition monitoring is a common process employed for predicting the health of a technical asset, whereby a predictive maintenance strategy can be adopted to minimize machine downtime and potential losses. Throughout the field, machine learning (ML) methods have become noteworthy for predicting failures before they occur, thereby preventing significant financial costs and providing a safer workplace environment. These benefits from predictive maintenance techniques, are particularly useful in the context of military equipment. Such equipment is often significantly expensive, and untimely machine failure could result in significant human endangerment. In this paper, a prognostic model (PROGNOS) is proposed to predict military equipment’s remaining useful life (RUL) based on their monitoring signals. The main considerations of PROGNOS are expectation maximization tuned Kalman Filter (EM-KF) for signal filtering, a recently introduced feature extraction algorithm (PCA-mRMR-VIF), and predictive LSTM model with an adaptive sliding window. The viability and performance of the proposed model were tested on a highly complex competition dataset: the NASA aircraft gas turbine engine degradation dataset, wherein readings from multiple sensor channels were recorded for degrading machines. According to testing results, we can confidently say that the proposed PROGNOS model was viable and robust overall, proving its general usefulness on all military equipment that emit signals.

Building on growth in the commercial IoT space, the Internet of Battlefield Things (IoBT) paradigm envisions mixed application of commercial and military IoT technologies to support Multi-Domain Operations via next-generation Artificial Intelligence and Machine Learning (AI/ML) applications. Operational environments for IoBT are expected to feature significant data volume and variety to support mission needs at varying echelons, while simultaneously requiring coordinated asset management and AI/ML usage across distributed coalitions. These conditions motivate the design and implementation of next-generation infrastructures intended to facilitate: (1) Data exchange between systems according to broader mission needs; (2) Coordinated AI/ML processing across systems; (3) Intelligent prioritization of cross-infrastructure processing and dissemination. Towards achieving these aims, this work looks into the design and prototyping of an AI/ML pipeline, aimed at: (1) Ingest and processing of video feeds; (2) Prioritization of video feed object detections based on Value of Information (VoI) assessment; (3) Support for information exchange with external systems. This submission will cover implementation details of the AI/ML pipeline, a reference usage scenario for IoT data processing and information exchange, and planned future work.

Modeling Command and Control (C2) problems is notoriously complex across operating environments, systems, and missions. We seek to systematically decompose the complexities of C2, and present the foundational problems in a form that is accessible to the larger research community. Our first objective is to “flatten” core C2 problems by formulating and developing a surrogate environment that can foster new insights and evaluation. Using this framework, the second objective is to host a public challenge to acquire a diversity of solutions from outside the DoD community.

End-users of Command and Control (C2) systems can struggle to effectively determine which Course of Action (CoA) should be chosen out of many complex options. A novel top-K recommender system has been developed to provide C2 end-users with comprehensive CoA recommendations through four critical measures named for explanation purposes only: Optimality, Diversity, Feedback, and Preference. The Optimality measure introduces a novel “Pareto-mesh” technique for quantitative CoA metrics, which offers robust and computationally efficient comparisons of Pareto-front percent coverages among ranking options. The Diversity measure also introduces a novel linchpin-based qualitative/quantitative technique that considers qualitative data’s Hamming distances, quantitative data’s Euclidean distances, and optimizes diversity of top-K CoA’s based a chosen top linchpin CoA using Maximin optimization. The Feedback measure is based on implicit user choices, peripheral simulation results, and real-world scenario results, with benchmarks of state-of-the-art collaborative filtering methods SVD, BiVAE, and NCF for this domain. The Preference measure is based on explicit user preference input of quantitative data, and ranks CoA’s based on a simple weighted sum against quantitative CoA metrics. Representative data related to C2 end-users, request attributes, and CoA’s is randomly generated to perform evaluations of each recommender system measure. Hyper-parameters used to calculate each measure are also globally tuned by means of a Genetic Algorithm multi-objective optimization technique, relying on a per-measure, inversion-minimizing optimization function. The designed recommender system’s hyper-parameter training time is on the order of hours and the system inference time is on the order of milliseconds.

Traditionally, learning from human demonstrations via direct behavior cloning can lead to high-performance policies given that the algorithm has access to large amounts of high-quality data covering the most likely scenarios to be encountered when the agent is operating. However, in real-world scenarios, expert data is limited and it is desired to train an agent that learns a behavior policy general enough to handle situations that were not demonstrated by the human expert. Another alternative is to learn these policies with no supervision via deep reinforcement learning, however, these algorithms require a large amount of computing time to perform well on complex tasks with high-dimensional state and action spaces, such as those found in StarCraft II. Automatic curriculum learning is a recent mechanism comprised of techniques designed to speed up deep reinforcement learning by adjusting the difficulty of the current task to be solved according to the agent's current capabilities. Designing a proper curriculum, however, can be challenging for sufficiently complex tasks, and thus we leverage human demonstrations as a way to guide agent exploration during training. In this work, we aim to train deep reinforcement learning agents that can command multiple heterogeneous actors where starting positions and overall difficulty of the task are controlled by an automatically-generated curriculum from a single human demonstration. Our results show that an agent trained via automated curriculum learning can outperform state-of- the-art deep reinforcement learning baselines and match the performance of the human expert in a simulated command and control task in StarCraft II modeled over a real military scenario.

A barrier to developing novel AI for complex reasoning is the lack of appropriate wargaming platforms for training and evaluating AIs in a multiplayer setting combining collaborative and adversarial reasoning under uncertainty with game theory and deception. An appropriate platform has several key requirements including flexible scenario design and exploration, extensibility across all five elements of Multi-Domain Operations (MDO), and capability for human-human and human-AI collaborative reasoning and data collection, to aid development of AI reasoning and the warrior-machinelike interface. Here, we describe the ARL Battlespace testbed which fulfills the above requirements for AI development, training and evaluation. ARL Battlespace is offered as an open source software platform (https://github.com/USArmyResearchLab/ARL_Battlespace). We present several example scenarios implemented in ARL Battlespace that illustrate different kinds of complex reasoning for AI development. We focus on ‘gap’ scenarios that simulate bridgehead and crossing tactics, and we highlight how they address key platform requirements including coordinated MDO actions, game theory and deception. We describe the process of reward shaping for these scenarios that will incentivize an agent to perform command and control (C2) tasks informed by human commanders’ courses of action, as well as the key challenges that arise. The intuition presented will enable AI researchers to develop agents that will provide optimal policies for complex scenarios.

The information that decision-makers use for command and control has uncertainty. Previous work has described different types of uncertainties, and the methods for using information to evaluate and rank alternative courses of action vary based on the type of uncertainty that occurs. Thus, when developing ways to generate automated decisions to support Soldier tactical planning, including multi-criteria decision making (MCDM) with different types (“modalities”) of uncertain information, no single method or algorithm will be optimal for all situations. Metareasoning is reasoning about reasoning, which is a type of self-adaptation, and it has been closely studied in AI and in logic due to its relevance to autonomous decision making; it is also of interest in cognitive science under the rubric of executive reasoning. A software agent or autonomous system can use metareasoning to monitor and control the procedures that it uses to process sensor information, evaluate potential courses of action, and plan its actions. This concept paper presents a metareasoning framework that can enhance artificial reasoning about uncertain information in the context of generating and ranking alternative courses of action. In this framework, the decision support agents will use rules to select the MCDM algorithms that are most appropriate for the type of information and the uncertainty modalities that are present. The rules may be curated by experts or generated from machine learning algorithms. We expect that using metareasoning will improve the ability to make complicated decisions with uncertain information.

The DEVCOM C5ISR (Command, Control, Computers, Communications, Cyber, Intelligence, Surveillance and Reconnaissance) Center describes an ongoing focus to provide decision makers with mission command capabilities and platforms that aid the management of resources, seamlessly integrate across the six Warfighter functions, and complement the skills and experience of the Warfighter. The challenge is to provide “the right information to the right person at the right time”, regardless of the complexities of the operation. As an extension of our research in artificial reasoning, particularly for adaptive decision making, we have designed the Enhanced Tactical Inferencing (ETI) framework. The goal of ETI can be thought of “as a recommender to a recommender” where both explicit and implicit data is used to reason over information and generate decision point and course of actions recommendations. This paper will discuss the ETI framework and a use-case to explore potential recommendations for the current implementation and future enhancements.

We explore the application of single image super-resolution technique to satellite image and its effect on object detection performance. This technique uses a deep convolutional neural network to learn transformations between different zoom levels of image pyramids, also referred to as Resolution Set (Rset). The network can learn the transformations from the 2:1 RSet at a Ground Sample Distance (GSD) of 60cm to the full resolution image at a GSD of 30cm by minimizing the differences between ground-truth full resolution and the derived 2x zoom. After training, the learned transformation is applied to the 1:1 full resolution image transforming the pixels to 2x resolution. The learned transformations has intelligence built in and can infer higher resolution images. We find super-resolution images significantly improve object detection accuracy, improve manual feature extraction accuracy, and also benefit imagery analysis workflows and derived products which use satellite images.

The field of Reinforcement Learning continues to show promise in solving old problems in new and innovative ways. Thanks to the algorithms’ ability to learn without an explicit set of labeled training data, the action, environment, reward approach has lured many researchers into framing old problems in this manner. This work focuses on utilizing a multi-agent reinforcement learning approach to develop an optimization algorithm which can be used for training gradient descent-based machine learning models. The work focuses on the agents’ abilities to collaboratively navigate a multi-dimension error space and locate an optimal solution for the model in training, and compares and contrasts the collaborative Reinforcement Learning approach with popular optimizers used today such as the Root Mean Square Propagation (RMSProp) [1] algorithm and the Adaptive Moment Estimation (Adam) algorithm [2]. The work will also discuss how the agents’ performance varies as the model in training gains complexity. Furthermore, the work will examine rare conditions under which the agents failed to find an optimal solution. Finally, the work will discuss similarities between machine learning model training and Multi-Domain Operations decision making. Additionally, the work will discuss how lessons learned during the agent training can be applied to developing decision agents for Multi Domain Operations.

In real world applications, machine learning classifiers are only as good as their performance on real world data. In practice, this data comes without truth labels — that is why you need a classifier. In this paper we consider the scenario where we have a trained classifier, and a set of unlabeled test data that we wish to run through it. If the unlabeled data is not within the distribution of the classifier’s training data, we should not expect the classifier to work well on the test set. We explore the use of the Henze-Penrose divergence, a measure of separation between two multivariate distributions, as a way to predict performance of a classifier on a dataset and to detect distributional shift, and we find that by computing Henze-Penrose scores between the training and test sets first in the input space, and then in the feature space of the classifier, we can get an indication that the test data is out of distribution and that classification accuracy will be unreliable.

The current airborne ISR (Intelligence, Surveillance, and Reconnaissance) collection planning and management is designed for the employment of a limited number of large ISR platforms with respect to a set of collection requirements created through an assembly line-like process. This process is often very cumbersome and mentally exhaustive for intelligence analysts when dealing with a large volume of dynamic all-source intelligence. Consequently, more human time and resources are spent on processing and digesting raw data, instead of strategizing and designing the most effective ISR collection plans. These problems could be amplified significantly in potential future conflicts with a near-peer adversary because the complexity, scale, and intensity of an airborne ISR operation could be several orders higher.

This paper discusses our new research concepts and preliminary results in applying artificial intelligence (AI) and machine learning (ML) as proof-of-concepts for facilitating future ISR collection planning and management. We explored a dynamic knowledge graph to represent historical and current all-source intelligence data as well as the corresponding contextual relationships and temporal states. Such an accumulative knowledge base allows us to apply machine learning and other analytical methods to infer and visualize multi-layered intelligence on an adversary’s capabilities and operations. These analytical results could assist human ISR analysts in understanding adversary’s operational tactics, techniques, and procedures. We further look into AI-enabled virtual agents in assisting mission planners to best manage a group of ISR assets for fulfillment of collection requirements. Applying deep reinforcement learning to Intellection, a new cooperative multi-player ISR collection board game, we train the virtual agents into AI players who can follow the game rules and estimate the desirable flight paths for potentially maximizing successful collections over prioritized targets. Together, these explorations may potentially create decision advantages and relieve possible manpower bottlenecks in future fastpaced ISR collection planning and management.

Only limited studies and superficial evaluations are available on agents' behaviors and roles within a Multi-Agent System (MAS). We simulate a MAS using Reinforcement Learning (RL) in a pursuit-evasion (a.k.a. predator- prey pursuit) game, which shares task goals with target acquisition, and we create different adversarial scenarios by replacing RL-trained pursuers' policies with two distinct (non-RL) analytical strategies. Using heatmaps of agents' positions (state-space variable) over time, we are able to categorize an RL-trained evader's behaviors. The novelty of our approach entails the creation of an influential feature set that reveals underlying data regularities, which allow us to classify an agent's behavior. This classification may aid in catching the (enemy) targets by enabling us to identify and predict their behaviors, and when extended to pursuers, this approach towards identifying teammates' behavior may allow agents to coordinate more effectively.

In recent years, Wireless Rechargeable Sensor Networks (WRSNs) have adopted wireless energy transfer technology, which has emerged as a promising solution to the limited energy issues in traditional Wireless Sensor Networks (WSNs). Since the sensor’s lifetime is decided by the charging scheme of the Mobile Charger (MC), designing an effective charging algorithm is challenging. Although many efforts have been devoted to optimizing charging schemes in WRSNs, the studies still face several critical issues. Firstly, most previous studies assumed that the MC’s battery capacity is sufficient or unlimited, resulting in the MC can move and charges all sensors in a charging cycle. That may cause a long waiting time for energy-hurry sensors and significant exhaustion of the MC’s energy. Secondly, existing works often optimize the MC’s charging path, whereas the charging time has not been thoroughly considered. This work aims to solve the limitations above by optimizing both the charging path and charging time simultaneously under the MC ’s limited-energy constraint. Our objective is to minimize the energy depletion of sensor nodes. To this end, we leverage the advantage of the bi-level optimization approach and propose a charging algorithm with two levels: the charging path optimization at the upper level and the charging time optimization at the lower level. The proposed charging scheme combines Genetic Algorithm (GA) and Differential Evolutionary (DE) to identify the optimal charging path and time. We conducted extensive experiments to demonstrate the effectiveness of our charging scheme in comparison to the related studies.

Future conflicts involving Joint All-Domain Operations (JADO) will require an unprecedented scale of standoff targeting against hardened, mobile, and dynamic targets that will only be possible through effective use of sensing, command and control (C2), and targeting assets across multi-domain (space/air/ground/sea) kill chains. Technologies such as Artificial Intelligence and Autonomy are critical to enabling effective, scalable, and timely kill chains, but only if applied appropriately to address kill chain limitations, informed by quantitative system of systems modeling and simulation at the data level.

To address these challenges, BAE Systems is developing a Virtual System of Systems Testbed to enable detailed quantitative mission level analysis of kill chain performance in a variety of multi-domain scenarios, combining high fidelity platform and sensor simulation with detailed modeling of information flows through the kill chain. The Testbed uses Bohemia Interactive Simulation’s Virtual Battlespace 4 (VBS4) to model virtual physical entities, and uses the VBS4 Simulation SDK to pass dynamic scenario data messages for real-time processing by advanced algorithm plug-ins including AI-based information dissemination and Autonomous mission management. In this paper, we will describe the challenges posed by JADO and the critical importance of system of system analysis to inform the application of advanced algorithmic technologies. We discuss the rationale and architecture of the Virtual System of Systems testbed, the technologies incorporated, and the methodologies for system of systems analysis. We will present results assessing system of systems contribution of these technologies to improving mission-level metrics such as latency, scalability, and robustness of kill chain response.

The electromagnetic environment is becoming increasing cluttered and complex, therefore the need to perform operations such as modulation recognition on wide bandwidth signals in real time is becoming crucial. Deep learning solutions have shown state-of-the-art performance for modulation recognition, however they are computationally expensive. In this paper we address this demand on resources by applying Quantized Neural Networks (QNNs) to the problem of modulation recognition. We demonstrate that QNN implementations of state-of-the-art Neural Networks (NNs) with different bit widths achieve varying levels of performance, on both our own 10 class dataset (up to 80% classification accuracy) and the open source 24 class DeepSig dataset (up to 75% classification accuracy at 10dB signal-to-noise ratio (SNR)). We then explore the accuracy-hardware trade-off of scaling NN size and bits when using QNNs. It was found that scaling the NN size had similar accuracy-hardware trade-offs as scaling the NN bits. However, 1-bit NNs had an unfavorable trade-off, therefore suggestions are made on how to improve the 1-bit NN architecture and training procedure. To demonstrate the potential benefits of low-precision quantized NNs and encourage further work we propose novel optimizations for 1bit and 2bit QNNs in FPGAs and ASICs that eliminate the need for weights memory and simplify the hardware required to implement a QNN. This provides further savings in hardware, allowing QNNs to provide favorable accuracy-hardware trade-offs, and supports the implementation of significantly larger NNs with a throughput of 100s of millions of inferences per second.

Advancements in computer science, especially artificial intelligence (AI) and machine learning(ML) have brought about a scientific revolution in a plethora of military and commercial applications. One of such area has been data science, where the sheer astronomical amount of available data has spurred sub-fields of research involving its storage, analysis, and use. One such area of focus in recent years has been the fusion of data coming from multiple modalities, called multi-modal data fusion, and their use and analysis for practical and employable applications. Because of the differences within the data types, ranging from infrared/radio-frequency to audio/visual, it is extremely difficult, if not flat-out impossible, to analyze them via one singular method. The need to fuse multiple data types and sources properly and adequately for analysis, therefore arises an extra degree of freedom for data-science. This paper provides a survey for multi-modal data fusion. We provide an in-depth review of multi-modal data fusion themes, and describe the methods for designing and developing such data fusion techniques. We include an overview for the different methods and levels of data fusion techniques. An overview of security of data-fusion techniques, is also provided which highlights the present gaps within the field that need to be addressed.

We present a method for applying deep reinforcement learning to maritime platform defense, showing how to successfully train agents to schedule countermeasures for defending a fleet of ships against stochastic raids in a simulated environment. Our Schedule Evaluation Simulation (SEvSim) environment was developed using extensive input from subject matter experts and contains realistic threat characteristics, weapon efficacies, and constraints among weapons. Our approach includes novelty in both the representation of the system state and the neural network architecture: threats are represented as vectors containing information on the projected effect of different scheduling actions on their viability and fed to network input “slots” in randomized locations. Agents are trained using Proximal Policy Optimization, a state-of-the-art method for model-free learning. We evaluate the performance of our approach, finding that it learns scheduling strategies that both reliably neutralize threats and conserve inventory. We subsequently discuss the remaining challenges involved in bringing neural-network-based control to realization in this application space. Among these challenges are the needs to integrate humans into the loop, provide safety assurances, and enable continual learning.

Multi-domain operations drastically increase the scale and speed required to generate, evaluate, and disseminate command and control (C2) directives. In this work we evaluate the effectiveness of using reinforcement learning (RL) within an Army C2 system to design an artificial intelligence (AI) agent that accelerates the commander and staff’s decision making process. Leveraging RL’s superior ability to explore and exploit produces novel strategies that widen a commander’s decision space without increasing cognitive burden. Integrating RL into an efficient course of action war-gaming simulator and training hundreds of thousands of simulated battles using the DoD supercomputing resources generated an AI that produces acceptable strategic actions during a simulated operation. Moreover, this approach played an unexpected but significant role in strengthening the underlying wargame simulation engine by discovering and exploiting weaknesses in its design. This highlights a future role for the use of RL to test and improve DoD systems during their development.

Autonomous systems, including self-driving cars and unmanned underwater/surface/ground/air/space vehicles, have caught the imagination of the press, the public, and personnel of the U.S. Department of Defense (DoD). Moreover, these systems increasingly harness the power of Artificial Intelligence and Machine Learning (AI/ML), more specifically Deep Learning (DL), for their operation. The stunning performance of deep learning algorithms in comparison to extant methods, including pattern matching, computational linguistics, statistical inferencing, and legacy machine-learning, has taken the world by storm. However, adoption of such systems in safety-critical applications has been the subject of intense debate and scrutiny, because of the propensity of these algorithms for erroneous operation, such as misdirection with seemingly innocuous signage, and their vulnerability to data poisoning and data sparsity attacks, leading to incorrect operation which can be both embarrassing and damaging. This has naturally led the DoD community to ask, “How do we harness this technology being unleashed upon the world, and yet keep our nation and warfighters safe?” Before we answer this question, however, it is important to note that trust is integral to DoD systems, including autonomous systems, and ensuring reliable system operations is paramount. Therefore, we need strategies to equip the DoD to design, build, deploy, and sustain autonomous systems that are trustworthy, secure, reliable, and dependable. In this paper, we investigate issues leading to poor performance and lack of robustness of autonomous systems based on machine learning, and discuss the state of the art for their mitigation.

Convolution Neural Network (CNN) models have demonstrated remarkable success for many applications including computer vision in recent years. However, the increase in accuracy achieved so far is at the cost of memory and complex computations. The CNN models are growing deeper and wider making it difficult to be fit in a single device that has limited resources. Moreover, the inference time of such models is large enough to be unfit to apply for the real-time mission critical applications like the Internet of Battlefield Things (IoBT). In IoBT, where the unmanned aerial vehicles (UAVs) flying in the battlefield zone and capturing images, require accurate learning and immediate inference. It becomes problematic if the learning model does not fit in a single UAV when it has limited resources. Considering the aforementioned issues, in this paper, we study a formal approach to improve the inference time and memory utilization in resource constrained IoBT. We consider that multiple UAVs are involved in the inference process in which we apply spatially parallel convolution and pooling operations for all the convolution layers and pooling layers as well as model parallelism for fully connected (FC) layers. Finally, we present the numerical results for varying number of participating UAVs, input data/image sizes, and communication speeds.

This paper explores the application of adversarial machine learning (AML) in RF communications, and more specifically the impact of intelligently crafted AML perturbations on the accuracy of deep neural network (DNN) based technology (protocol) and modulation-scheme classifiers. For protocol classification, we consider multiple heterogeneous wireless technologies that operate over shared spectrum, exemplified by the coexistence of Wi-Fi, LTE LAA (Licensed Assisted Access), and 5G NR-Unlicensed (5G NR-U) devices in the unlicensed 5 GHz bands. Time-interleaving-based spectrum sharing is assumed. Given a window of received I/Q samples, a legitimate DNN-based classifier (called the defender’s classifier ) is often used to identify the underlying protocol/technology. Similarly, DNN classifiers are often used to discern the underlying modulation scheme. For both types of classifiers, we study an attack model in which an adversarial device eavesdrops on ongoing transmissions and uses its own attacker’s classifier to generate low-power AML perturbations that significantly degrade the accuracy of the defender’s classifier. We consider several DNN architectures for protocol and modulation classification (based on recurrent and convolutional neural networks) that normally exhibit high classification accuracy under random noise (i.e., AWGN). By applying AML-generated perturbations, we show how the accuracy of these classifiers degrades significantly, even when the signal-to-perturbation ratio (SPR) is high. Several attack vectors are formulated, depending on how much knowledge the attacker has of the defender’s classifier. On the one extreme, we study a “white-box” attack, whereby the attacker has complete knowledge of the defender’s classifier and its training dataset. We gradually relax this assuming, ultimately considering an almost “black-box” attack. Mitigation techniques based on AML training are presented and are shown to help in countering AML attacks.

Multi-domain operations (MDO) are characterized by simultaneous and sequential operations; rapid and continuous integration; and surprise. Machine learning (ML) for MDO is no different. Translated into ML, MDO requires highly assured yet rapid data and model fusion. Assurance demands robustness, reliability, and explain-ability, while speed demands computational efficiency and sample efficiency. Combinatorial interaction testing offers explainable and rigorous techniques to ML for fusing data and models with runtime guarantees. But such methods are underexplored in the literature. Combinatorial coverage has been applied to neuron- and layer-levels of neural networks, but only recently to ML in general. There are also ongoing debates of efficacy in the literature, but these debates are scoped to explainable deep learning. This work presents a framework for using combinatorial coverage for multi-domain operations. We discuss how coverage metrics can incorporate multi-modal meta-data and mission context into fusion processes, how coverage is oriented towards identifying gaps in and between sets of data, and how coverage can identify cases where performance is expected to be difficult. We conclude that combinatorial coverage should be considered a core capability for supporting ML in MDO.

NextG networks are intended to provide the flexibility of sharing the spectrum with incumbent users and support various spectrum monitoring tasks such as anomaly detection, fault diagnostics, user equipment identification, and authentication. For that purpose, a network of wireless sensors is needed to monitor the spectrum for signal transmissions of interest over a large deployment area. Each sensor receives signals under a specific channel condition depending on its location and trains an individual model of a deep neural network accordingly to classify signals. To improve the accuracy, individual sensors may exchange sensing data or sensor results with each other or with a fusion center (such as in cooperative spectrum sensing). In this paper, distributed federated learning over a multi-hop wireless network is considered to collectively train a deep neural network for signal identification. In distributed federated learning, each sensor broadcasts its trained model to its neighbors, collects the deep neural network models from its neighbors, and aggregates them to initialize its own model for the next round of training. Without exchanging any spectrum data, this process is repeated over time such that a common deep neural network is built across the network while preserving the privacy associated with signals collected at different locations. Signal classification accuracy and convergence time are evaluated for different network topologies (including line, star, ring, grid, and random networks) and packet loss events. In addition, the reduction of communication overhead and energy consumption is considered with random participation of sensors in model updates. The results show the feasibility of extending cooperative spectrum sensing over a general multi-hop wireless network through federated learning and indicate the robustness of federated learning to wireless network effects, thereby sustaining high accuracy with low communication overhead and energy consumption.

Future Multi-Domain Operations (MDO) will require the coordination of hundreds—even thousands—of devices and component services. This will demand the capability to rapidly discover the distributed devices/services and combine them into different workflow configurations, thereby creating the applications necessary to support changing mission needs. Motivated by neuromorphic processing models, in previous work it was shown that this can be achieved by using hyperdimensional symbolic semantic vector representations of the services/devices and workflows. Using a process of vector exchange the required services are dynamically discovered and inter-connected to achieve the required tasks. In network edge environments, the capability to perform these tasks with minimum energy consumption is critical. This paper describes how emerging spiking neural network (SNN) neuromorphic processing devices can be used to perform the required hyperdimensional vector computation (HDC) with significant energy savings compared to what can be achieved using traditional CMOS implementations.

The commercial industry has been working to develop ever-larger and more capable machine learning (ML) models (such as recent models from OpenAI, Microsoft and Google with more than ten billion parameters) for everything from general language processing to computer vision that have larger and larger computational, memory and other resource requirements. These powerful models generally need hardware accelerators to accommodate their workloads. GPU systems have been a popular choice among users and deep learning model designers either for their ability to run the inherently parallel deep learning workloads efficiently or for their large memory resource that would fit large deep learning models. We profile and analyze different deep learning workloads using various GPUs and configurations, emphasizing how deep learning architectures have diverse compute requirements. We analyze three popular deep learning workloads: Ultralytics’s You-Only-Look-Once model (Yolov5) on COCO dataset, Bidirectional Encoder Representations from Transformers (BERT), and Deep Learning Recommendation Model (DLRM) on the Criteo Kaggle Display Advertising Challenge Dataset. This work aims to shed light on the performance bottlenecks when using GPU systems as accelerators for training recent deep learning models.

With the outstanding accomplishments achieved in recent years, Reinforcement Learning (RL) has become an area where researchers have flocked to in order to find innovative ideas and solutions to challenge and conquer some of the most difficult tasks. While the bulk of the research has been focused on the learning algorithms such as SARSA, Q-Learning, and Genetic, not much attention has been paid to tools used to help these algorithms (e.g. the Experience Replay Buffer). This paper goes over what is believed to be the most accurate Taxonomy of the AI field and briefly covers the Q-Learning algorithm, as it is the base algorithm for this study. Most importantly, it proposes a new Experience Replay Buffer technique, the Round Robin Prioritized Experience Replay Buffer (RRPERB), which aims to help RL agents learn quicker and generalize better to rarely seen states by not completely depriving them of experiences which are ranked as less priority.

This paper presents a distributed multi-class Gaussian process (MCGP) algorithm for ground vehicle classification using acoustic data. In this algorithm, the harmonic structure analysis is used to extract features for GP classifier training. The predictions from local classifiers are then aggregated into a high-level prediction to achieve the decision-level fusion, following the idea of divide-and-conquer. Simulations based on the acoustic-seismic classification identification data set (ACIDS) confirm that the proposed algorithm provides competitive performance in terms of classification error and negative log-likelihood (NLL), as compared to an MCGP based on the data-level fusion where only one global MCGP is trained using data from all the sensors.

Many studies show that using synthetic data or mixed synthetic and real data might improve machine learning (ML) performance but it is difficult to draw generalizable conclusions. A contribution to this problem is the fact that the synthetic data from most vendors are improperly filtered and contain aliased (wrapped-around) high frequency information which they should not possess. Most vendors use spatial-domain, low-pass FIR filters to generate synthetic images at various ranges. Unfortunately, these FIR filters aim for interpolation with a desired frequency domain cutoff and spatial spacing (general non-integer scale factor and/or decimation). Hence, instead of a sharp cutoff at the desired low-band, they produce aliased data. This erroneous information in synthetic imagery could actually mislead an ML algorithm. In addition, most synthetic images do not account for a camera’s MTF (Modulation Transfer Function). A Fourier-based filtering can easily incorporate any MTF in the frequency domain based on Rayleigh resolution theory, properties of a camera lens and digital image properties. The spectral properties of images which are acquired with real sensors are studied and compared them with synthetic images from several vendors. We have also developed a metric that exhibits that the camera system’s MTF shows the same spectral property for real images at different ranges. The metric can help us to determine if a synthetic image generation engine violates this property and, hence, produces erroneous information.

Object detection on imagery captured onboard aerial platforms involves different challenges than in ground-to-ground object detection. For example, images captured from UAVs with varying altitude and view angles present challenges for machine learning that are due to variations in appearance and scene attributes. Thus, it is essential to closely examine the critical variables that impact object detection from UAV platforms, such as the significant variations in pose, range to objects, background clutter, lighting, weather conditions, and velocity/acceleration of the UAV. To that end, in this work, we introduce a UAV-based image dataset, called the Archangel dataset, collected with a UAV that includes pose and range information in the form of metadata. Additionally, we use the Archangel dataset to conduct comprehensive studies of how the critical attributes of UAV-based images affect machine learning models for object detection. The extensive analysis on the Archangel dataset aims to advance optimal training and testing of machine learning models in general as well as the more specific case of UAV-based object detection using deep neural networks.

With the ever-increasing adoption of interconnected technologies and rapid digitization observed in modern-day life, many online networks and applications face constant threats to the security and integrity of their operations or services. For example, fraudsters and malicious entities are continuously evolving their techniques and approaches to bypass current measures in place to prevent financial fraud, vandalism in online knowledge bases and social networks like Wikipedia, and malicious cyber-attacks. As such, many of the supervised models proposed to detect these malicious actions face degradations in detection performance and are rendered obsolete over time. Furthermore, fraudulent or anomalous data representing these attacks are often scarce or very difficult to access, which further restricts the performance of supervised models. Generative adversarial networks (GANs) are a relatively new class of generative models that rely on unsupervised learning. Moreover, they have proven to effectively replicate the distributions of real data provided to them. These models can generate synthetic data with a degree of quality such that their resemblance to real data is almost indistinguishable, as demonstrated in image and video applications – like with the rise of DeepFakes. Based on the success of GANs in applications involving image-based data, this study examines the performance of several different GAN architectures as an oversampling technique to address the data imbalance issue in credit card fraud data. A comparative analysis is presented in this paper of different types of GANs used to fabricate training data for a classification model, and their impact on the performance of said classifier. Furthermore, we demonstrate that it is possible to achieve greater detection performance using GANs as an oversampling approach in imbalanced data problems.

Amidst the extensive global integration of computer systems and augmented connectivity, there have been numerous difficulties within ensuring confidentiality, integrity and availability across all systems. Malware is an ever-present and persistent challenge for security systems of all sorts. Numerous malware detection methods have been proposed, with traditional approaches no longer providing the necessary protection against evolving attack methodologies and strategies. In recent years, machine learning for malware detection has been investigated with great success. In addition, the analysis of application operation code, or opcode, due to its unavoidable nature, can reveal necessary information about software intention. Visualization of opcode data allows for simple data augmentation and texture analysis. The proposed approach utilizes a simple visual attention module to perform a binary classification task on program data, focusing on visualized application opcode data. The proposed model is tested with an ARM-based Internet of Things (IoT) application opcode dataset. In addition, a comparative analysis, using numerous metrics, is conducted on the proposed model’s performance along with several other algorithms. The results indicate that the proposed method outperformed all other tested techniques in accuracy, recall, precision, and F-score.

Artificial feedforward neural networks (ANN) have been traditionally trained by backpropagation algorithms involving gradient descent algorithms. This is in order to optimize the network’s weights and parameters in the training phase to minimize the out of sample error in the output during testing. However, gradient descent (GD) has been proven to be slow and computationally inefficient in comparison with studies implementing the extended Kalman filter (EKF) and unscented Kalman filter (UKF) as optimizers in ANNs. In this paper, a new method of training ANNs is proposed utilizing the sliding innovation filter (SIF). The SIF by Gadsden et al. has demonstrated to be a more robust predictor-corrector than the Kalman filters, especially in ill-conditioned situations or the presence of modelling uncertainties. In this paper, we propose implementing the SIF as an optimizer for training ANNs. The ANN proposed is trained with the SIF to predict the Mackey-Glass Chaotic series, and results demonstrate that the proposed method results in improved computation time compared to current estimation strategies for training ANNs while achieving results comparable to a UKF-trained neural network.

Predictive maintenance refers to the ability to predict when machinery or systems need to be maintained. Making an accurate prediction is quite challenging given the costs for both over-estimating (unnecessary maintenance and reduction in availability of assets) and under-estimating (untimely breakdowns and possible loss of equipment or lives). To address these challenges researchers were able to develop new approaches for analyzing oil samples taken extracting samples from oil-wetted machinery that may provide information critical to developing predictive capabilities. We consider the problem from both supervised (though data limited) and unsupervised approaches and provide a first look into a data driven approach for identification of condition indicators. Through this work we identify a collection of candidate features that can form the basis of condition indicators for both a high level discrimination of failure vs. normal operation as well as a set for potential failure mode identification. Finally, we present an anomaly detection framework for detecting failures which can be a viable solution for an onboard analysis tool in deployed systems.

Cross-domain face matching between the thermal and visible spectrum is a desired technology to recognize the visible face image with probe images captured in the thermal spectrum for the night-time surveillance and security applications. However, the large modality gap between faces captured in different spectrum makes thermal-to-visible face recognition quite a challenging problem. Classical Independent Components Analysis (ICA) and its variants have been successfully used for feature learning, and the projection matrix can be used for forming the basis of images. In this paper, we present a deep joint independent component analysis network (DJICAN), a deep architecture based on multilayer independent component analysis, which learns the mutual mappings between visible and thermal face images. Joint ICA assumes the sources are the same for a visible image and its corresponding thermal image. The goal of DJICAN is to obtain the basis matrices of visible and thermal face images, which represent the individual imaging systems for the two domains. First, a forward multilayer ICA is performed. Then, we use a novel backpropagation algorithm based on a reconstruction loss function to optimize the ICA basis and the sources. Extensive experiments are performed on the ARL polarimetric thermal facial datasets that contain face images that have been taken at three different ranges and with different face expressions. The results show that the proposed method performs better than the state-of-the-art methods in terms of both synthetic image quality and thermal-to-visible face recognition accuracy.

Artificial intelligence (AI), or more specifically deep learning approaches to AI, have led to astonishing results in recent times, which makes them a prime candidate for guiding agent actions in military domains. However, it is often difficult to train multiple agents with deep learning approaches when a task is sufficiently complex, or the state space is huge, as is often the case in military domains. One possible way to alleviate the difficulties associated with military tasks is to leverage military doctrine to assist in the guidance of multi-agent systems. Military doctrine is a guide to action rather than hard rules for the execution of military campaigns, operations, exercises, and engagements. Doctrine, written by experts in their respective domains, is used to make sure that each task associated with an engagement, for example, is executed according to military standards. Such standards ensure coordination between different tasks, resulting in a greater likelihood of Mission success. In addition, the efficacy of combining doctrine with deep learning must be tested to determine any realized benefit for AI driven military engagements under adversarial conditions. Further, the inherent complexities associated with military engagements demand coordination between heterogeneous resources and teams of Soldiers which are often geospatially separated. In this work, we establish a baseline of doctrine-based maneuvers for a military engagement scenario with a multi-agent system (MAS) embedded in the StarCraft Multi-Agent Challenge (SMAC) simulation environment, now a standard test environment for Multi-Agent Reinforcement Learning (MARL). We introduce a hybrid training approach that combines MARL with doctrine (MARDOC) to test whether doctrine-informed MARL policies produce more realistic behaviors and/or improved performance in a simple military engagement task. We compare this hybrid approach to both doctrine-only (i.e., supervised learning to mimic doctrine) and MARL-only approaches to evaluate the efficacy of the proposed MARDOC approach. Our experiments show that MARDOC approaches produce desired behavior and improved performance over supervised approaches or MARL alone. In summary, the experimental results suggest that MARDOC approaches provide a sufficient advantage over MARL alone due to agent doctrinal guidance of MARL exploration to overcome the complexities in military domains.

Machine learning algorithms are being widely used in different fields such as image recognition, speech recognition, traffic and weather prediction, recommendations, spam filtering, self-driving cars, stock market prediction, medical diagnosis, and more. The ability of machines to feed in years of data and predict the outcome has helped humans in unimaginable ways. Machine learning has automated half of human work requiring very little human intervention and saving time and energy. However, sometimes individuals end up paying the price and falling victim to the unfair and biased outcome of machine learning algorithms. Machines learn through data what they are provided with, but the data that machines learn from does not come free from human biases. Human biases based on race, sex, ethnicity, skin color, and other sensitive attributes are reflected in the dataset which, when fed to the machine, results in a similar biased prediction. The years of data represent the bias that has been present in society, and the machine learning model simply mimics the pattern. There has been constant research and experiments being done on how to prevent these biases from reflecting on the prediction. In this paper, we will investigate if there is any bias present in the benchmark Statlog “Australian Credit Approval” dataset and take necessary measures to mitigate the bias present in the data. The paper shows how the AIF360 tool can identify and mitigate bias in the data and eventually in the learning algorithms.

A problem of interest for nuclear nonproliferation is monitoring activities at nuclear facilities, where proliferation events may only take place a few times and often under variable conditions. Machine learning has revolutionized data analytics by enabling the use of measurable signatures to generate predictive models of facility operations. However, traditional methods for training these models require large, reliable data sets with labeled observations, a challenge for nonproliferation. Highly variable conditions further complicate this as events from training data may have occurred in conditions quite different from the event of interest. Our hypothesis is that when events occur in a highly variable environment, careful training data selection for each test event could outperform the standard approach of using all available training data. We developed a method to optimize training data selection for the given test event and applied it to predicting the power level of the High Flux Isotope Reactor (HFIR) at Oak Ridge National Laboratory. In this study, the reactor startup exhibits variability between occurrences due to natural variability in environmental conditions and operational procedures. Using a combination of analysis techniques, a similitude assessment was performed on data collected from HFIR to isolate clusters that were optimal for training a predictive model. Concepts such as dynamic time warping and Jaccard similarity were used in conjunction with clustering analysis. In order to validate this approach, the model was trained on every combination of unique training events and the predictive performance was compared to the performance using a subset of the training data selected by isolated clusters found through the similitude assessment.

The US Army has existing challenges associated with command and control and the execution of the operations processes, data processing, information management, and knowledge management. Continued limitations in the capability to combine and reason across explicit and tacit knowledge, due to the increased flow of data from multiple domains, is one shortfall associated with these challenges. We have created a framework to ingest various modalities of data enabling reasoning particularly for decision making tasks. The Enhanced Tactical Inferencing (ETI) framework is designed to have components that send and receive data from different information systems and sources to a reasoning module. The reasoning module is composed of sub-modules with different reasoning model profiles and functions. These sub-modules can work independently or interdependently. The output is a series of recommendations for decision making. One key model is for Uncertainty of Information (UoI). This model incorporates a series of rules and algorithms to associate uncertainty across the multiple data sources. The intent of the ETI framework is to provide recommendations to humans, intelligent systems, and combinations of both. This paper will present the details of the ETI framework, focusing on the UoI model, as well as potential refinements and applications.

As defense organizations integrate artificial intelligence (AI) into evermore critical operations, especially those near the tactical edge with real-time decision making, the necessity of a standardized, robust framework for deployment and management of AI systems is increasing. In this paper, we propose a Prediction Management Framework (PMF) that aligns with the Department of Defense’s Ethical Principles for AI for ethical, governable, and interpretable deployments. We explore different requirements for the framework with inspiration drawn from various regulatory, safety, and communication standards. In support of these requirements, we offer recommendations and implementation guidance to provide comprehensive visibility into the system.

In this article, an Artificial Neural Network (ANN) has been used to model the relationship between the gender of human subject to its response to whole-body vibration (WBV). To train, validate and test the model, an experiment was conducted on 20 female and 20 male subjects at different vibration frequencies in the range of 20 to 45 Hz. The response was measured by taking the ratio between the subject head’s horizontal acceleration to the platform’s vertical acceleration. The subjects’ body mass index, mass, height, gender, and age, and the excitation frequency were used as inputs to the ANN. The ANN model showed a good performance of 98.645% matching regression, and RMSE and MAE of 0.0128 and 0.0483, respectively.

The state-of-charge (SOC) of Li-ion batteries is an important parameter for regulating and ensuring the safety of batteries, especially for Electric Vehicle applications, where accurate SOC estimation is important for remaining driving range prediction. The SOC is conventionally obtained through different indirect measurement methods that are either impractical or inaccurate. Data-driven approaches for SOC estimation are becoming more popular as an alternative to the conventional estimation methods due to their accuracy and low complexity. In this work, we apply 4 machine learning algorithms: Multiple Linear Regression (MLR), Multilayer Perceptron (MLP), Support Vector Regression (SVR), and Random Forest (RF) to predict the SOC using voltage, current, and temperature measurements from mixed driving cycles datasets, with 50000 instances. Out of the 4 models, the Random forest model performed the best with an MAE of only 0.82%.