OpML_Industry_Day_Presentation.pptx

Aug 2, 2019 Gennady Staskevich Information Directorate Air Force Research Laboratory

FOR OFFICIAL USE ONLY

Industry Day for Operationalizing Machine Learning for C2 Program

Machine Learning for MDC2 Operational Decision Making

Complexity of Multi Doman (MD) operations will quickly exceed C2 operator capacity to make informed decisions Hypothesis: Applying AI and ML can reduce complexity, increase ops tempo while maintaining manpower levels Problem

C2 Processes

Mapping not well-defined

Highly interactive human processes Non-traditional AI/ML data sources Structured, unstructured, chat, enterprise DBs Dynamic environments and models

OP-ML Problem Space Operational C2 Attributes No clear mapping of AI/ML to Operational C2 processes

Applications of AI/ML are being actively being applied to ISR problems:

MDC2 ECCT objective: explore AI/ML technology in support of enhancing C2 processes & critical functions:

Motivation

Problems: Classification/Detection Algorithms: Google Deep Resnet, etc.

Data: Overhead images, FMV, etc.

Metrics: Classification accuracy, Precision, Recall

Problems: Planning/Learning: Supervised? Reinforcement? Unsupervised? Optimization?

Algorithms: Limited commercial usage Data: Enterprise DBs, Unstructured, etc.

Metrics: Ops tempo, etc.

Extra Motivation Machine Learning / AI “Empty Box Syndrome”

No clear way-ahead to apply AI and ML

Multiple briefs/reports show future application of AI/ML to MDC2, None show way ahead Proposed Mitre Functional Architecture for Multi-Domain C2 defines a “blank” AI/ML component, 2018 Source: National Security Engineering Center 2018 Spring Report U.S. Air Force 2030 Science and Technology Strategy AF Future operation concept AFCEA New Horizons 2018 MDC2 Clean Sheet

MDC2 ECCT (ACC)

Many have identified the Need

Analyze, develop and evaluate the use of AI/ML technologies within operational level C2 processes Analyze and create mapping of AI/ML to operational level C2 processes Develop AI/ML initially to well understood Air C2 domain; expand to support MDO vignettes Evaluate hypothesis in Mod & Sim-based MDC2 environment

End Goal:

Develop AI/ML augmented prototypes supporting operational level C2 processes Results of empirical evaluation of AI/ML augmented processes against non-enhanced baselines Program Objective OP-ML program will address the “Empty Box Syndrome”

State of the Art (Kessel Run: AOC Pathfinder Modernization) AOC Pathfinder Follow on to the Canceled AOC 10.2 WS

AOC modernization approach modern hardware, software (java, ..)

Agile software development “Appify” existing AOC processes Lean startup Procure operational software faster, with higher quality and reduced risk From: Agile DevOps To: Planning with Machine Learning ‘Leapfrog’

AOC 10.1 WS

Legacy suite of tools Significant workload on the operator Operator is responsible for data fusion Existing Approach

(Canceled) AOC 10.2 WS Cost overruns Minimal technical progress

No AI/ML Focus OpML Rapid Prototype Development, Big emphases on ML applications for Planning

State of Art

OP-ML Program Structure Project 1:

AI/ML Prototype Development OP-ML for C2 Project 3:

Multi-Domain Experimentation Project 2:

AI/ML Prototype Dev. & Evaluation Environment Develop AI/ML C2 prototypes based on ranked study outputs Develop input training data sets and algorithms Evaluate quality of the output models Leverage AFRL’s StreamLined ML framework and ecosystem.

Integrate into FlyLeaf

Integration Operational Studies Reinforcement Learning prototype for C2 planning areas identified in studies Develop input training data sets and algorithms Develop simulation environment with sufficient fidelity and speed to refine models at ML scale Evaluate quality of output models Air operational solutions ported to multi-domain environments Develop multi-domain input training data sets and algorithms Apply multi-domain metrics and evaluate models Uses FlyLeaf DevOps/Sim/White Cell approach

OP-ML Program Structure Project 1:

AI/ML Prototype Development OP-ML for C2 Project 3:

Multi-Domain Experimentation Project 2:

AI/ML Prototype Dev. & Evaluation Environment Develop AI/ML C2 prototypes based on ranked study outputs Develop input training data sets and algorithms Evaluate quality of the output models Leverage AFRL’s StreamLined ML framework and ecosystem.

Integrate into FlyLeaf

Integration Operational Studies Reinforcement Learning prototype for C2 planning areas identified in studies Develop input training data sets and algorithms Develop simulation environment with sufficient fidelity and speed to refine models at ML scale Evaluate quality of output models Air operational solutions ported to multi-domain environments Develop multi-domain input training data sets and algorithms Apply multi-domain metrics and evaluate models Uses FlyLeaf DevOps/Sim/White Cell approach

Task 1 Task 1 Option Task 2

2 Collaborative Studies:

Approach:

Identify operational C2 processes with applicable AI/ML attributes: repetitive tasks, machine-to-machine data translation, pattern detection, regression, … Processes with these attributes have the potential to benefit from AI/ML

KEY IDEA:

Operational Studies: Approach

Planning Operational execution management Air Combat Ops AOC Productivity Air Mobility Ops Battle Management C2 – Operation Execution

Operational & tactical level decision making Task name & Description of existing task Define Evaluation criteria Processes with applicable ML Attributes Candidate ML application(s) & Data Req.

Baseball cards:

Key ingredients

Air Combat Ops

AOC Productivity Operational Studies: Results (FY19)

Air Mobility Ops Battle Management C2

Operational Studies Example Air Mobility Ops - Planning Ground MOG planning & Scheduling – Maximum-on-ground (MOG) is an “in-theater” airfield capacity (and infrastructure) to both: park (parking MOG) and process airlift missions (working MOG), with the latter being the harder problem to solve.

Candidate applications for AI/ML Measuring and forecasting working MOG – need to aggregate several services into a single value

Logistics supply chain optimization. Robust air mobility plans will need to balance efficiency of logistics supply chains while smoothly adjusting to meet unexpected increases in demand and mitigate supply chain disruptions.

Operational Studies Example Air Mobility Ops - Execution Flight Monitor Load Balancing - Flight monitors are responsible for monitoring, coordination, and management of currently executing air missions (aeromedical evacuation, air mobility support, air refueling, and airlift). Their oversight persists until the mission has landed.

Candidate applications for AI/ML Predictive load monitoring - Monitor mission management workload and surge cycles, estimate completion-time and task metrics. Use ML to predict workload surges Situational skill levels optimization - Assign users to manage missions that best meet their skills/experiences. Categorize mission situations/scenarios and measure individual skill levels wrt those scenarios. Identify conditions and driving functions for optimal assignment of users to missions based on their skills.

Operational Studies Example Battle Management C2 - Planning Operations Execution Management - Battle Management (BM) is the management of activities within the operational environment based on the commands, direction, and guidance given by appropriate authority. The function of Battle Management is to be prepared for events that can not anticipated so we can recover and achieve as many of the planned results as is possible.

Candidate applications for AI/ML Data translation between systems – (TACS, AWACS, JSTARS, CRC), identification of mission critical data, correlations and pattern analysis, supporting management of mission flows and capabilities for handling deconfliction in/near real-time.

Applications supporting battle management functions during operation execution - Air Support Operations Center (ASOC), Theater Air Control System (TACS), Joint Surveillance Target Attack Radar System (JSTARS), Airborne Warning and Control System (AWACS) and Control and Reporting Center (CRC) functions

AWACS/E3A, JSTARS, Rivet Joint, CRC

Operational Studies Example Battle Management C2 - Execution Tactical Chat - Tactical chat plays a central role in the execution of battle management functions. It 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 basic attribute detection), they do not take any advantages of advancements in AI/ML.

Candidate applications for AI/ML Distill tactical chat data to:

Derive mission workflows from tactical chat data Identify current mission state based on chat data Identify mission critical and time sensitive data Anticipate next steps, resource requirements, and suggest candidate courses of actions

Operational Studies Example Air Combat Ops - Planning MAAP-ATO-Package Planning – Producing an air battle plan that is responsive to a dynamic operational requirement. This is a continuous, iterative, and highly structured process for force management to meet commander’s requirements across range of military options.

Candidate applications for AI/ML Generate skeletal mission packages with coarse-grained target clusters Propose good starting solutions for MAAP templates (package information, weaponeering, smarter resource utilization, and schedule) Balance a complex set of factors including priority, threat avoidance, neutralization, geospatial proximity, and availability of resources.

Optimize existing ATO plans based on time, cost, risk, etc

Identify mission gaps.

Operational Studies Example Air Combat Ops - Execution Dynamic Targeting / Retasking - The Dynamic Targeting (DT), knows as F2T2EA (Find, Fix, Track, Target, Engage, and Assess) or simply the “kill chain.” This process has been used for engaging Time Sensitive Targets (TSTs). It’s applicability extends to all targets whether developed during deliberate targeting or dynamic targeting during mission execution.

Candidate applications for AI/ML Provisional planning for Dynamic Targeting – Use ML to preemptively forecast provisions during planning process for coming up with flexible “policies” that will be able to accommodate contingency events during dynamic battle management that minimize the need for replanning or retasking.

Use ML to balance a complex set of factors including priority, threat avoidance/neutralization, geospatial proximity, and resource availability for the execution of dynamic targets.

Operational Studies Example AOC Productivity - Planning Human-Machine Logistics Teaming - The goal is to optimize the relative strengths and capabilities of human and machine agents. This particular effort is focused on the functions of the ATO follower a.k.a. “football carrier”. The ATO follower is a person who monitors the execution of a mission. 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, which is critical for effective re-planning during plan execution.

Candidate applications for AI/ML ML assisted ATO follower – Leverage ML to keep track of the mission workflow by following an ATO along with any real-time updates to the mission.

Train models using historical data to rapidly identify relevant information (even across historical data) Learn and alert operator of any concerns experienced in similar missions, alert for missing information, and capability Allow for any adjustments, and/or corrections earlier on in the process, allowing for more mission flexibility.

Operational Studies Example AOC Productivity - Planning Operator Workload Balancing - Current AOC tools and processes are highly sequential, operator intensive tasks that span across multiple systems leading to planning bottlenecks. Recent R&D has focused on automated planning to help speed the process, but little focus on improving the efficiency of the human teaming aspects. The intent is to leverage AI/ML to dynamically allocate mission planning tasks among AOC staff to balance workload, to improve efficiency and timeliness of ATO development.

Candidate applications for AI/ML Dynamic allocation and reallocation of mission tasks - Potential application can be used to identify and track relevant data; develop tools & processes that dynamically allocate mission planning tasks among AOC staff to balance workload, improve efficiency and timeliness of ATO development.

Task 1: AI/ML Prototype Development and Experimentation

(Project 1) Identify, and develop AI/ML prototypes that enhance operational C2 plans and processes Reduce uncertainty, reduce complexity of data-to-decision process, and increase operational tempo

Single domain focused (Air, Cyber, and Space) Short duration efforts; 15M efforts (9M prototype + 6M concept refinement) (Midway through effort) identify a candidate multi-domain use-case for your approach

(Project 3) Multi-Domain Experimentation – Executable Option Extend and build upon the identified multi-domain use-case Generate / acquire additional data, services supporting MD use-case Leverage Flyleaf architecture Duration 12M

Objective: Develop supporting environment for RL to enable Operational-level planning Challenge: Existing C2 modeling environments not compatible with modern AI/ML approaches Does not scale – Run at real-time Limited by human factors (experience, SA, fatigue, sticky-notes/sneaker-net) Approach: Develop C2 environment to support AI/ML training/execution

Task 2: RL Prototype Evaluation Environment Development Environment for development of general reinforcement learning algorithms Multi-layer Planning Feedback Learning Loop Architecture

FY20

Common Shared Interface Test Problems/Environments Standardized Benchmarks

Schedule

Current Status (TRL 2)

Identified ML applications for operational C2 Soliciting white-papers for operationally focused prototypes Prototypes will answer our hypothesis reduce complexity increase ops tempo not increasing manpower levels

Summary

Reminder:

Please submit your questions on 3x5 cards image4.jpeg image5.png image6.png image7.emf image8.png image9.jpg image10.png image11.jpeg image12.png image13.png image14.png image15.png image20.png image16.png image17.png image18.png image19.png image21.jpeg image22.jpeg image23.jpg image24.png image25.jpeg image26.jpg image27.png image28.png image29.png image30.png image31.png image32.jpeg image33.png image34.png image35.png image36.jpeg image37.png image38.png image39.emf image1.png image2.jpeg image3.png