The world is alive with artificial intelligence, machine learning and natural language processing. IBM’s Watson made headlines by besting Jeopardy champions while more recently, Google’s AlphaGo defeated several acknowledged champions in the board game Go. Companies pursue numerous solutions to apply the latest technologies and add to their product lines: Amazon uses its algorithms to give better product recommendations; Facebook is looking to establish (and now create) even more connections among social media users; meanwhile, IBM is advertising Watson as an analyst who ostensibly brings productivity and efficiency to companies that use it.
As in industry, defense is attempting to leverage these same technologies to solve military problems. Computer vision helps imagery analysts recognize potential targets faster; sustainers can project Stryker maintenance needs with machine learning; and the intelligence community uses deep learning to conduct pattern analysis to discover unseen linkages in threat activity. Unfortunately, these efforts fall short of the goal the Mission Command Center of Excellence has in mind when thinking of artificial intelligence (AI)—a true assistant that aids in understanding the operational environment while supporting the operations process.
To achieve this goal, the Army must look further out and attempt to simultaneously unify and integrate multiple disparate tools under the banner of general artificial intelligence.
Cautionary Tales
Over the decades, movies have provided several prominent examples of general AI: HAL 9000 from 2001: A Space Odyssey, Joshua from WarGames and Skynet from the Terminator series. While science fiction, these popular movies offer cautionary tales of AI gone wrong, providing real-world boundaries in AI creation.
Spurred by Deputy Defense Secretary Robert Work’s publication of the Third Offset Strategy, DoD is undergoing multiple concerted AI efforts. Work recently commissioned the Algorithmic Warfare Cross-Functional Team to find and field additional AI applications. However, these limited solutions risk being stovepiped as they solve relatively small problem sets and have narrow goals.
Indeed, the Army envisions a much broader application for AI. The Army Operating Concept describes AI that will “enable the deployment of autonomous and semi-autonomous systems with the ability to learn.” The Robotics and Autonomous Systems Strategy describes applications for indications and warnings, counter-messaging and cyber defense, among other uses. However, a key assumption is that the AI in question has been trained to understand the operational environment so that it can defend a network or give recommendations. Yet there is no effort to help nascent AI understand what we do much less how we do it.
For AI to be useful and effective, it must discriminate what information means in various contexts. For example, if you ask for “ISIS,” it does not know if one is looking for terrorists, a pharmaceutical company or an Egyptian goddess. However, if one asked the question while in a world history class, the most likely answer is Isis, the Egyptian goddess. The world history class adds the necessary context to vector the artificial intelligence to give the best, most probable answer. This is similar to how Amazon recommends books by different authors, but under the same subject or genre. Their curated algorithm takes a search history and matches it with similar users or relative products. Creating and managing these algorithms with relationships of people and products are closely guarded trade secrets.
Realizing the Vision
Through research, teaming and experimentation, Mission Command Battle Lab is helping the Army aggregate its base layer of data so it can also create the algorithms and data structures needed to support decision-making. Both must occur if the Army is to realize its vision of a learning machine to inform or recommend actions to future unseen circumstances.
A first component is training an AI program how operational units perform missions across the range of military operations by echelon. The military decision-making process, for example, is a good start in that there are defined inputs and outputs. Further along, it could learn to disseminate an imagery report of vehicle X (exploited through computer vision) to S3 plans, targeting, operations and S2 personnel for situational awareness and staff estimates. This logic of how we manage information and distribute knowledge products is essential to the algorithm. Much like commercial counterparts, this allows the computer to offer valid recommendations such as “people who looked for vehicle X also searched for …”
The basis for this future general AI must be grounded with four basic items: history, doctrine, theory and lessons learned. While all this may sound mundane, it is the foundation by which all other forms of Army AI must be applied.
After creating a foundational set of Army knowledge, experts of all types in various AI disciplines will be able to add context and knowledge to the algorithm. Much like we must train people to understand military jargon, so too we must assist machines to understand the relative value of words in each context. It is here where humans add value to various forms of AI to bring about a more useful general AI. This is an opportunity to crowdsource algorithm additions or development—where users contribute best practices or location-specific data to the algorithm. But to truly achieve value across the Army, a holistic approach is needed where the AI is trained at multiple echelons and across warfighting functions.
A Look Ahead
As technology evolves and becomes more economical and prolific, the number of sensors and data inputs to processes and activities will continue to grow. As data from these sensors and inputs accumulates, existing staff tools are ineffective to analyze the thousands, even millions, of collected data points. AI that collates and synthesizes the various forms of AI techniques can enable staff to process data into information and knowledge, enabling commanders to maintain situational awareness and enhance effective decision-making.
At present, these tools require large amounts of processing power and memory, placing this capability out of reach for small tactical units. However, the necessary computing horsepower is present at division, corps and theater Army elements. It is likely these echelons have the resources to field some form of general AI where data and algorithm management will occur. Such an approach parallels industry practices with tiers where subordinate units are subscribers to portions or clusters of AI located at different locations. In this way, if one node fails or is destroyed, it fails over to other nodes—similar to the concept of “graceful degradation.” The algorithm could also “slice off” relevant portions of data to be assigned and reassigned to units based on changing mission sets, or to support disconnected operations.
Invest Today
To achieve the vision of more general AI for Armywide application, it is necessary to invest today. The Army must conduct research and experimentation to understand how these new technologies will affect the force from a holistic doctrine, organization, training, materiel, leadership and education, personnel, facilities and policy perspective. The pieces and parts of AI are not set, providing an opportunity to shape the future by experimenting with the tools available now. This will take time and commitment. For example, it took IBM’s Watson and the Memorial Sloan Kettering Cancer Center three years to start making accurate cancer treatment recommendations. As AI technologies improve, it is possible to reduce this time.
To speed the process, the Army should continue adhering to standards as set out in the joint information and common operating environments. But these standards must extend to the various forms of AI and machine learning, plus be compatible with other efforts, such as the newly formed Algorithmic Warfare Cross-Functional Team. These linkages are where the Mission Command Battle Lab will focus its early research and experimentation. With all this innovation and creation, it is likely there will be data fratricide, and some sort of governance will need to be established. Indeed, the Defense Innovation Board recommended establishing an AI Center of Excellence to help ensure AI and machine learning address impacts across formations.
The past 100 years may have been about oil, but the next 50 will be about big data. To effectively manage and wield big data, the Army needs AI to support autonomous systems, intelligence analysis, logistics management and the decision-making process. However, the Army must invest the time and resources now to produce this future tool. Army general AI must be trained properly with doctrine, history, theory and lessons learned to help leaders reduce the time necessary to translate data to information.
An effective future AI will give humans back the time to conduct higher cognitive load functions of gaining understanding and applying knowledge to problem solving. Initially, it is probable that this capability will reside at echelons above brigade; placing the Mission Command Center of Excellence, specifically Mission Command Battle Lab, supported by the Combined Arms Center at Fort Leavenworth, Kan., in the right place at the right time to lead the integration of AI into Mission Command systems and the common operating environment.