Operational_Threads_Attachment.doc

23 JUL 2019

Eight Operation Threads

The Eight Operational Threads of Table 1 are defined in more detail in this section. Each figure represents a high-level description of Operational Thread, its process description, and potential application for ML. Please note that the applications are just a notional example list. Again, novel approaches for any one of the eight Operational Threads and strong technical content are encouraged. Furthermore, understand that the information below is not absolute, either operationally or technically. The figures are meant to merely suggest initial areas of applications, and does not force a specific approach.

Air Mobility Ops - Planning

Ground MOG Planning/Scheduling Problem: Air mobility planners represent in-theater airfield capacity and infrastructure using a measure called maximum-on-ground (MOG). MOG encompasses an airfield’s ability to both park and process airlift missions with the latter being the harder problem to solve. Processing aircraft includes material-handling equipment, refueling aircraft, and local air traffic control and is often referred to as Working versus Parking MOG. Collectively, working MOG is the number of aircraft that can be processed simultaneously for each aircraft’s total time at the airfield and are based on several factors that can change over time. This means working MOG is only an approximate measure because it aggregates several services into a single value. There is reasonable data available for parking MOG but not working MOG. In addition, a theater’s MOG is a dynamic snapshot of the aggregated MOGs of the individual air bases and not a simple summation. This theater view would also include the movement of cargo loaders and other resources from one base to another to alleviate resource contention and complicate the problem further.

Potential Solutions: Some solutions to this problem have employed linear programming optimization as an index to track locations of aircraft over time and model reduction techniques using data aggregation to prune the variables and constraints. The literature describes approaches that formulate a supervised machine learning model using mathematical optimization and linear programming that could be used in this problem, including clustering solutions and SVM models. SME techniques could be applied to generate an initial collection of Working MOG input datasets to help guide algorithm performance. If a RL approach is pursued, identify the operational systems and data sources required to develop an evaluation environment.

Possible Input Data: Available airfields, aircraft and delivery routes. Movement requirements and average on-ground cargo and parking space data. Approved mission take-off times, maximum payloads and aircraft utilization rates. Airfield data to include aircraft capacity over time and MOG efficiency factors.

Potential Evaluation Criteria:

Enhanced loading/unloading scheduling

(resources in use) / (available resources) per time

Airfield utilization efficiency

Air Mobility Ops - Execution

Flight Monitor Load Balancing

Problem: Flight monitors are responsible for monitoring, coordination, and management of currently executing air mobility missions. These missions may include: aeromedical evacuation, air mobility support, air refueling, and airlift. Their oversight persists until the mission has landed. Each flight monitor typically manages a dozen or more missions per shift. The number of monitored missions, mission complexity, and sudden, unanticipated spikes in workload can overwhelm the operator and exceed their capacity to make informed decisions. These could result in negative impacts on primary missions such as delays and even cancelations. Key to the successful management and coordination is the “communication” with the aircrew, advisory agencies, maintenance, and air traffic control. Effective flight monitors assist the aircrew, anticipate their needs, and offer alternatives and recommendations when situations change.

Potential Solutions: Of interest are ML applications supporting flight monitors’ tasks such as: modelling and predicting workload surges, improving operator collaboration, and balancing individual tasks. Predictive load monitoring can be achieved by monitoring mission management workload and surge cycles that can be used to train models that estimate completion-time and task metrics. Traditional approaches for predicting workload surges include weighted averaging and time series analyses. Potential AI/ML-based approaches could be explored under this effort such as: regression, random decision forests, support vector machines or any other applicable methods. Another candidate AI/ML application is focused task allocation based on operators’ situational skill levels, where the objective is to assign candidate flight monitors to manage missions that best meet their skills and experiences. Also, exploring the space of possible teaming structures and learning effective representations could maximize performance and minimize the risk and cost with respect to mission requirements. Matching mission tasks to specific flight monitors can possibly leverage the ideas introduced in DeepTriage , where the authors used an attention-based, deep bidirectional neural net to find best matches between a bug report and the appropriate developer that can fix the bug.

Possible Input Data: The same data for the MOG planning could be used as a starting point for this area. In addition, detailed air mobility mission data from legacy systems would be required, as well as the official airlift schedules for inter- and intra-theater missions. Individual flight planning data would be required such as diplomatic clearances and related coordination artifacts.

Potential Evaluation Criteria:

Mission categorization and task decomposition

Train models to monitor and predict spikes in workload

Train models to identify appropriate operator skill levels for a given mission

BMC2 - Planning

Operations Execution Management Problem: Battle Management (BM) is the coordination of activities within the operational environment based on the commands, direction, and guidance given by appropriate authorities. Its function is to be prepared for events and contingencies that cannot be anticipated in the planning process so operators can absorb perturbations and achieve as many of the planned results as possible. There are multiple issues at hand that make the current processes difficult to execute. For example, incompatible data formats (mission plans, chat data, and sensor observations (state data)), outdated data, and incomplete information makes it difficult to fuse, and identify relevant (and potentially time critical) information. More information is not always the answer, need to provide salient fused data.

Potential Solutions: Of interest are AI/ML applications supporting battle management functions during operational execution. Potential application areas include: 1) supporting information exchange between the Air Battle Manager (ABM), pilots, and planners, and data translations between systems (TACS, AWACS, JSTARS, CRC) could possibly benefit from recent advancements in NLP (deep learning, transformer models for solving challenging sequence-to-sequence translation), 2) identification of mission critical data, correlations and pattern analysis (with respect to specific mission) , 3) identifying and supporting management of mission flows , and 4) developing capabilities for handling deconfliction in/near real-time, and performing dynamic changes to the air tasking cycle, including interpretation of commander’s intent. Other potential ML applications can include operator training and intelligence modeling (Red behavior is a latent requirement behind many other applications).

Possible Input Data: Operational mission requirements derived from Air Combat Planning products (MAAP, AOD, ATO), tactical chat data, data on available airfields, aircraft, and delivery routes, analysis of mission performance and complexity (on time, resource utilization, amount of re-planning), Potential Evaluation Criteria:

Identify courses of actions for replanning that specific mission type

Develop Course of Action (COA) for pop-up targets

Develop deconfliction models

Develop Red intelligent agent, Red TTP, capabilities

BMC2 – Execution

Tactical Chat Problem: Tactical chat plays a central role supporting the execution of battle management functions that identify and assess changes in the threat environment and dynamically adjust missions in reaction to those changes. BM enables the communication between the operators and systems across space and time. Today’s tactical chat systems offer only rudimentary functionality such as chat rooms and rudimentary search, but they do not take advantage of advancements in AI/ML. Mission critical information is often entangled within and across multiple conversations about various operational threads occurring simultaneously and at different points in the coordination process – thereby relying on the operator’s innate ability to identify problems and act accordingly. Current processes often involve operators relying on sticky notes, and sneaker-net solutions to pass mission critical information such as coordination and other data between systems. In addition to the operator’s limitations such as limited attention span or other distractions, an increase in operational tempo can result in missing time-critical information.

Potential Solutions: Some solutions to make sense of the various tactical chat sessions include identifying: one-on-one mission-specific discussions, hidden external relationships among missions among conversations, and finding contextual, mission specific storylines. This particular challenge can possibly leverage the ideas introduced in DeepTriage to match a chat snippet to a mission vignette or an action frame (defined below), where the authors used an attention based, deep bidirectional neural net to find a best match between a bug report and the appropriate developer that can fix that specific bug. Potential workflow to identify and stich related text from chat data may include: 1) develop vignettes for the most common mission threads as performed by the operators using MIRC chat tools; these missions can be any of the following: dynamic target, resource allocation of sensors, blue force overwatch and force protection, SEAD, and automated patterns of life development, 2) develop data sets at sufficient quantity and quality to train AI/ML classification models, 3) use AI/ML trained models to identify mission pertinent information such as person, places, things, beginning of a mission, end of a mission, etc., and 4) map the identified information (person, places, things, time) to a mission specific vignette. This mapping may need to be further broken down and mapped to a subset of a mission, which we refer to as “action frames.” Key activity frames can be derived from chat (in real-time or forensically from chat logs). Capability to identify, classify and stich related information is critical to supporting planning and re-planning BM functions at/or near real-time.

Possible Input Data: Tactical Chat data, operational mission requirements derived from Air Combat Planning products (MAAP, AOD, ATO), roles of the operators.

Potential Evaluation Criteria:

Algorithmically identify key / pertinent information (person, places, things, time, …)

Map identified information to mission and/or vignette

Identify goals / objectives of mission

Air Combat Ops – Planning

MAAP-ATO-Package Planning

Problem: In support of the Joint Force Commander (JFC), the application of air power must be responsive to a dynamic operational environment driven by a flexible planning process that maximizes the effects of many distinct parts. The resultant air battle plan must be produced in a timely manner and capture a battle rhythm that synchronizes personnel, systems and capabilities in the exact time and space necessary to carry out the JFC’s guidance with overwhelming force. A well-orchestrated air campaign requires the input of many expert planners that have applied years of experience and operational art in generating intuitive and effective plans. Sometimes this interaction occurs through an interrelated series of information exchanges of concepts and ideas as well as active involvement in the iterative refinement of high-priority mission packages. At the end of this process, air combat planners will have produced a time-phased air operations scheme of maneuver (MAAP) as the foundation for the detailed tasks (ATO) composed of multi-component missions (package planning).

Potential Solutions: The current state of the art for addressing this problem set is reimplementing large legacy applications as smaller apps using a DevOps approach. As this is occurring, there are likely opportunities to take the evolving datasets and individual apps and apply supervised and unsupervised learning techniques to enhance these capabilities under Task 1. Another option is to anticipate the end state of this development process and consider a reinforcement learning approach that would utilize DevOps auto-planning features as policy subject to evaluation in a simulated environment under Task 2.

Possible Input Data: Commander’s Guidance and Air Operations Directive in various formats. Joint Integrated Prioritized Target Lists and weaponeering options. Blue force allocation parameters, status and unit contracts. Airspace Coordination Order. Mission feasibility constraints such as fuel burn or sortie generation rates.

Potential Evaluation Criteria:

Generate an effective grouping of targets to maximize mission success

Quality of the generated product evaluated by a SME (mission objectives, resource utilization, level of risk, support for/by relationships represented)

Air Combat Ops - Execution

Dynamic Targeting/Retasking Problem: The Dynamic Targeting (DT) has often been called F2T2EA (Find, Fix, Track, Target, Engage, and Assess) or simply the “kill chain.” This process has been used for engaging Time Sensitive Targets (TSTs). Its’ applicability extends to all targets whether developed during deliberate targeting or dynamic targeting during mission execution. Targets of opportunity have been the traditional focus of dynamic targeting because decisions on whether and how to engage must be made quickly. Planned targets (deliberate) are also covered during this phase but the steps are simply to confirm, verify, and validate previous decisions (in some cases requiring changes or cancellation). The find, fix, track, and assess steps tend to be ISR-intensive, while the: target and engage steps are typically labor-, force-, and decision making- intensive. The robustness of an air battle plan is defined as the ability of a plan to handle contingency events during dynamic battle management with minimal need for replanning or retasking. Such contingencies may include dynamic targets or dynamic threats, disrupted or denied environments, or variability in mission performance. The challenge here, is to forecast the right set of provisions that will minimize the need for replanning of the deliberately planned mission.

Potential Solutions: Solutions to this problem could potentially leverage ideas from supply chain planning & logistics to forecast provisions during planning processes for coping with contingency events during dynamic battle management. In that paper, authors were able to map the supply chain planning problem to the machine learning domain (deep learning, & support vector regression). They used PCA to reduce the feature space to a meaningful number of salient features. Other viable approaches for feature reduction may include auto-encoders. If the Dynamic Targeting problem could be cast as the supply chain problem then provisional planning will minimize the need for replanning or retasking thereby reducing the negative impact to the primary mission. Contingency planning for the execution of dynamic targets must also balance a complex set of factors including priority, threat avoidance/neutralization, geospatial proximity, and resource availability.

Possible Input Data: Commanders guidance, operational mission requirements derived from Air Combat Planning products (MAAP, AOD, ATO), data on available aircraft, supporting aircraft, delivery routes and available airfields, and historical data such as planned (ATO’s) and the chat data that supported its’ execution.

Potential Evaluation Criteria:

Rebalancing of the next-day’s strategy assessments based on the results of dynamically tasked missions (measure the increase or decrease in the amount of planning required due to retasking, e.g.: less JIPTL targets, increased coordination, etc.)

AOC Productivity - Planning

Human-Machine Logistics Teaming Problem: Current plan execution processes push operators and systems to their limit, especially when combat operators are required to prosecute dynamic targets in real time. This area’s focus is to relieve some of the intense pressure placed on combat operators by pairing them with AI/ML-based agents that form a highly efficient human-machine team that can minimize perturbations and maximize effects. The goal is to exploit the relative strengths and capabilities of human and machine agents cooperating in teams and performing a variety of necessary logistics tasks. This particular effort can apply to the ATO follower known as the “football carrier.” The ATO follower is an operator who monitors missions being executed today that were planned yesterday. This operator carries around a collection of documents and information that is called the “football” pertaining to the ATO being executed. Some say that this football is the “sum of all knowledge” because it contains the guidance, requests, and rationale for why things are being done. Existing processes rely on the operator’s innate ability to keep track of the complex set of inputs, events, and to provide the support (intent and reasoning) for the execution of current missions.

Potential Solutions: Solutions to this problem could potentially leverage ideas from retail and supply chain optimization to keep track of the mission workflows by following an ATO along with any real-time updates to the mission. The operator’s function would be to ensure that workflows are correct and allow for any adjustments, and/or corrections earlier on in the process, resulting in built-in mission flexibility.

Possible Input Data: Commanders guidance, operational mission requirements derived from Air Combat Planning products (MAAP, AOD, ATO).

Potential Evaluation Criteria:

Design structure for capturing and storing provenance as an initial digital version of the “football”

AOC Productivity -Execution

Operator Workload Balancing Problem: Current AOC processes are operator intensive and span across multiple legacy systems. Often operators must use an out-of-band process or tool and rely on sticky notes and sneaker-net to pass important, mission critical information (e.g., target coordinates) between systems. Today’s AOC processes are highly sequential, leading to bottlenecks and potentially inflexible plans. Recent R&D has focused on DevOps and automated planning to help speed the process, but there is little emphasis on improving the efficiency of the human-machine teaming aspects at the task level that could pay huge dividends in the development of high-quality plans. The intent of this area is to leverage AI/ML to dynamically allocate mission planning tasks among AOC staff to balance workload, improve efficiency, and speed up development of planning artifacts. Currently there are no tools / processes for explicit tracking of user activities, task backlog, or overall productivity. As a result, a subset of users may experience high stress, leading to mistakes that impact the plan or lengthen planning time due to corrections. This problem area could be a complementary activity with the MAAP-ATO-Package Planning area or could address another opportunity to pair an AOC duty position with an AI/ML agent.

Potential Solutions: Solutions to this problem could potentially leverage ideas from optimization of supply chain logistics . These new techniques include end-to-end workflow visibility, agile decision making, and performance management. In addition to the aforementioned, AI/ML may be used to forecast daily tasks based on trends and personalize additional workloads based on current tasks and past performance.

Possible Input Data: All of the same data described in the MAAP-ATO-Package Planning area, with a particular emphasis on the strategy-to-task linkages created by the planners as they step through the planning process that traces all the way back to Commander’s Guidance. In addition, historical data from past AOC exercises or experiments could be distilled to capture workflows and match tasks with the staff that perform them.

Potential Evaluation Criteria:

Percentage reduction of chat messages and sticky notes as a team vs. no automated support (Breadth of Synchronization)

SME review of support for and support by plan elements present with teaming vs. without (Plan Quality)

� � HYPERLINK "https://medium.com/opex-analytics/machine-learning-using-mixed-integer-programming-af95e4d56863" �https://medium.com/opex-analytics/machine-learning-using-mixed-integer-programming-af95e4d56863�

� � HYPERLINK "https://www.kaggle.com/fabiendaniel/predicting-flight-delays-tutorial" �https://www.kaggle.com/fabiendaniel/predicting-flight-delays-tutorial�

� � HYPERLINK "https://www.kaggle.com/dongxu027/airline-delays-eda-deep-dive-lessons-learned" �https://www.kaggle.com/dongxu027/airline-delays-eda-deep-dive-lessons-learned�

� � HYPERLINK "https://arxiv.org/abs/1801.01275" �https://arxiv.org/abs/1801.01275�

� Basu, Mitra, and Tin Kam Ho, eds. Data complexity in pattern recognition. Springer Science & Business Media, 2006.

� Integration and Beyond: Linking Information from Disparate Sources and into Workflow, Journal of the American Medical Informatics Association, Volume 7, Issue 2, March 2000, Pages 135–145, https://doi.org/10.1136/jamia.2000.0070135

� � HYPERLINK "https://arxiv.org/abs/1801.01275" �https://arxiv.org/abs/1801.01275�

� https://doi.org/10.1155/2019/9067367

� https://doi.org/10.1155/2019/9067367