The Input-Output Problem: Managing the Military’s Big Data in the Age of AI
Editor’s Note: This article was submitted in response to the call for ideas issued by the co-chairs of the National Security Commission on Artificial Intelligence, Eric Schmidt and Robert Work. It addresses the third question (part b.) on how data should be collected, stored, protected, and shared.
How can you beat an adversary without overpowering them? One strategy is to overwhelm it with information. The Russian government has successfully executed several cyber attacks against European countries in the past decade that did exactly that, most notably against Ukraine as part of a coordinated ground attack to seize Crimea in 2014. Russia used a distributed denial of service attack to cripple communications in preparation for its assault. It overwhelmed a digital system by inundating it with more data than the system could process, creating a digital traffic jam where legitimate communications cannot travel across the system. The outcome of such an attack is chaos.
The problem of being overwhelmed by data in the U.S. military is not an external challenge. Instead, the U.S. military is inflicting information overload on itself. It is generating more data than it is able to process, resulting in a real inability to execute missions. Indeed, the RAND Corporation published a study in 2014 titled, “Big Data: Challenges and Opportunities.” The report projected exponential growth in the “digital universe” to rise above 40 billion terabytes by 2020, from below 500 million terabytes in 2005. The Navy additionally projected exponential growth in data inputs due to improvements in sensor quality and quantity.
Unfortunately, development of technologies that process growing information flows seems to have lagged behind. This capability mismatch between inputs (e.g., full motion video, global positioning systems, network services, etc.) and outputs (e.g., intelligence estimates, mission orders, target lists, etc.) is drowning tactical military organizations in waves of data, promoting increased error and indecision. We refer to this mismatch as the “input-output problem.” The problem is severe but it is not insurmountable. In the face of an unprecedented influx of data, implementing artificial intelligence within military decision-making processes could turn this vulnerability into a strength. By taking advantage of autonomous systems, governments can avoid the onslaught of big data, and wade into the fog of modern war with eyes open. Artificial intelligence can help leaders see unseen trends, bypass bureaucratic structure, and make better decisions faster.
The U.S. military, as it currently exists, is struggling to cope with the vast amounts of data it receives. While staffs make valiant efforts at processing this data, they are nonetheless human with human limits. To understand the input-output problem, there are two major factors to keep in mind: structure and process. Think of information like water flowing through a plumbing system. The time it takes for water to get from one place to another is dictated by total distance and the width of the pipes. The U.S. military now faces a problem in which it has too much water trying to get through pipes that are too long and too narrow.
The military’s structure (its plumbing) is built around a very simple idea — concentrate the capabilities of all branches of the military against an adversary at one place and at one time to overwhelm the adversary’s ability to make decisions. To achieve this, the U.S. military created unified commands made up of all the services, led by a single commander. Despite its benefits — like unity of effort, overwhelming battlefield effects, and flexibility — unified command is a costly and often inefficient enterprise; it contributes to the input-output problem by forcing information to be redundantly processed at each tier of leadership. Unified command, in combination with a tiered formal hierarchy that lends itself to specialization, leads to inefficiency.
The cost of operating jointly manifests itself in the time and effort wasted dealing with structural inefficiencies in the flow of information. In order to leverage joint resources (e.g., a request for Air Force close air support for an Army ground unit, or a request for a Navy logistics vessel to support a newly established Air Force landing strip), commanders must process requests up the chain of command until the request arrives at the resource-controlling commander. A joint capability request often takes much longer than it should. A major feature of hierarchical organizations like tactical military units is that the cost of coordination increases the further down the structure you travel.
In the U.S. Army, the lowest echelon of command that requests joint capability through a formal process is the battalion level. The targeting process is the primary tool military units use to request joint capabilities and generate effect on targets — an effect could take the form of a missile strike, an electronic warfare attack, or a surgical raid. That means that every echelon from the battalion to the joint task force conducts deliberate planning to request scarce joint resources. In addition to generating and processing their own requests, command nodes also evaluate subordinate requests. There are at least six tiers of leadership between a battalion and the joint headquarters. A joint capability request must travel through all six tiers of leadership before arriving at a decision-maker’s desk. When you apply that math to the doctrinally prescribed span of control of between three and five per command tier, there could be hundreds if not thousands of target requests moving through the system at any given time (the number of requests grow exponentially as the organization gets larger). Clearly, not all those requests make it to the joint force commander’s staff. Nevertheless, they must still be processed somewhere in the organization. This data processing demand should not be taken lightly; in the case of the Army, whole units exist (e.g., battle coordination detachments) to process requests for joint capability at the highest levels.
All those requests are currently sorted, prioritized, and allocated by humans. While many targeteers (the soldiers, airmen, sailors, and marines who process requests for joint capability) would say they are up to the challenge, the data processing requirements just described do not account for the time it takes to complete the joint targeting process (a six-step process partially outlined below). After a target is executed, the staff must still determine the effectiveness of target execution. Assessment involves receiving and processing data from maneuver units, reconnaissance units, unmanned systems, electronic/signal intelligence systems, and myriad other sources. That assessment data must be processed within the same structure previously described, with the same associated costs. This raises the question: Who has time to do all this work?
A Process in Time
When it comes down to it, the input-output problem is a problem of time. Scarcity of time forces tactical units into a dilemma. Units either take the time to process inputs and run out of time to submit joint target requests, or they generate an output by limiting input processing time, therefore increasing the risk of error. Regardless, each outcome is suboptimal. To return to our plumbing analogy, if structure is all about the length of pipe that water has to travel, then process is all about the speed the water flows through those pipes.
In terms of process, military organizations are bound by constraints derived from the time it takes to formally request support from the various military branches. As outlined in Joint Publication 3-60, the joint targeting process pulls data from across the joint force as inputs to create a list of agreed upon targets each of the different components thinks is important (e.g., a priority target list). That list is made of targets that commanders seek to affect in some way with either missiles, intelligence, or myriad other tools. The targeting process is generally set to a time horizon of 96 hours from execution of an operation. This 96-hour time horizon is derived from time requirements levied against all the branches of the joint force by the air tasking cycle. The air tasking cycle plays a central role in the targeting process because it aligns and synchronizes a large portion of joint assets against the priority target list.
As a result of the demand for unified command across all military branches, the air tasking cycle and joint targeting process bind all tiers of the joint force from the joint task force commander (a four-star general or admiral at the highest levels) down to the battalion commander (in the case of the Army) to a 96-hour data processing limit. If an organization plans sequentially, it takes too long and misses the window to request joint support. If an organization plans in parallel, it is most certainly cutting corners, making assumptions and increasing the risk of poorly allocating scarce resources. In either case, the inability to process requests and data undercuts the ability of the unit to accomplish its objective.
Technical “Building Blocks” of Autonomous Systems
Artificial intelligence is uniquely positioned to address the dilemma presented by the input-output problem. Its greatest strength is its superior ability to handle vast amounts of data that would overwhelm human military staffs. Indeed, the Defense Science Board’s study suggests that “given the limitations of human abilities to rapidly process the vast amounts of data available today, autonomous systems are now required to find trends and analyze patterns.” This is particularly true for less kinetic applications of artificial intelligence that revolve around analysis and solutions in the virtual realm (autonomy at rest), rather than the unmanned robotics and lethal autonomous weapons systems that abound in existing military literature (autonomy in motion). While several variations of artificial intelligence do exist, autonomous systems all follow the same basic process: They collect data, process data, and generate an action. The two critical building blocks for this discussion are the data collection and data processing mechanisms.
Data Collection Capacity
The most fundamental requirement of a functioning autonomous system is abundantly available data. These data pools — regardless of whether they are shoe purchases or missile strikes — drive the machine-learning process. Assimilating large amounts of labeled data allows programmers to “train” their systems to find minute correlations between inputs and resultant outputs. Examples include speech recognition (think Apple’s Siri or Amazon’s Alexa), natural language processing (think chatbots), or machine vision (think Google’s self-driving car). The exact quantity and type of data required naturally depends on the complexity of the task at hand. However, the magnitude tends to be robust, evidenced by multiple contemporary examples like Google’s self-driving car, which alone requires approximately 1 GB of sensorial data per second, derived from thousands of video streams and images.
Data can be difficult to gather. The lack of available data often poses the most significant barrier to entry for private and public organizations into artificial intelligence applications. More interestingly, however, this mismatch is one of the primary reasons artificial intelligence is uniquely postured for success in the military. By virtue of its reliance on significant amounts of data and iterative design, the operations process is rife with data pools available for use.
Military planning methodologies mirror the basic process autonomous systems follow. When operations officers plan, they are gathering data. “Preparing” this information is similar to “processing it.” Both military organizations and autonomous systems generate actions based off of these information inputs. Finally, both processes assess the outcome, driving adjustment in subsequent iterations (in the military operations realm, examples include a battle damage assessment or an after action review). Indeed, it is no wonder that a report from the Boston Consulting Group claims “operational practices and processes are naturally suited for AI [artificial intelligence]. They often have similar routines or steps, generate a wealth of data, and produce measurable outputs.” While human staff sections engaged in the operations process might perceive this data overabundance as an obstacle, autonomous systems would thrive in it.
Data Processing Speed
The military needs systems that can process data quickly. The speed at which autonomous systems operate distinguishes them most from alternate solutions. Evidence from a variety of sources and practical applications suggests that machines are faster — at least, to some degree — than humans when it comes to data processing. At the furthest extreme, Fujitsu’s “K” supercomputer can generate 8.2 billion “megaflops” (each megaflop is equivalent to a million operations per second), while the human brain can generate only 2.2 billion.
Artificially intelligent systems, by extension, have demonstrated a similar processing advantage on multiple occasions, across multiple industries. For instance, the U.S. credit, debit, and prepaid card industry utilizes automated technology to scan over 1,200 transactions per second. These systems can identify fraudulent transactions within 10 milliseconds, despite monitoring a customer base of over 1.3 billion cards. Similarly, IBM has developed an autonomous system that scans over 12 million pages of medical content, including over 200 medical textbooks and 290 medical journals, ultimately providing oncologists with recommended courses of treatment within a mere 30 seconds.
Autonomous systems could foreseeably condense the phases between target guidance and air tasking order production through automation. As a result, the window of time available for units to submit targets would expand, allowing more units to request joint resources. From a practical perspective, autonomous systems would shorten the targeting process by creating target lists faster than people can. According to the Defense Science Board, the joint air tasking process is currently “heavily manual,” with up to 40–50 people involved in target data input, mission planning, and resource allocation planning.
Artificial intelligence could also be used to condense the time it takes for units to plan operations. For example, it could expedite the internal planning methodologies practiced at each echelon during the joint air tasking process. It is challenging to imagine an operational environment sophisticated enough (and subsequently, a budget large enough to finance it) to supplement echelons as low as the battalion level with access to autonomous systems. However, it’s important to think through how doing so would create positive outcomes derived from expediting planning methodologies, and do so in a way that does not detract from output quality at every echelon. In this case, reducing planning time in each phase, of each echelon, by a factor of two once again results in every echelon gaining access to the joint air tasking target submission process.
Viewed from a practical standpoint, planning methodologies could be condensed in a variety of ways. One example might be the automatic refinement of orders pushed down from higher echelons. As the system currently exists, junior officers are typically charged with combing through repetitive and irrelevant information contained in higher orders, and hand-picking the items to analyze in more depth — a time-consuming and tedious process that occurs at every echelon depicted. Another time-costly process within planning is course of action development. An artificially intelligent application could feasibly facilitate speed by generating generic “baseline courses of action,” overlaid on terrain analysis and associated products from information inputs and preexisting operational courses of action Indeed, most tactical tasks at the lowest levels are routine enough that deviation from standard operating procedures is unnecessary.
Think Faster or Buy More Time
Kinetic applications of artificial intelligence dominate existing military literature on autonomous systems. After all, it’s pretty cool to think about killer robots destroying enemy targets. However, the concept of “autonomy at rest” — the softer and less elegant applications of artificial intelligence that revolve around data compilation, analysis, and a solution in the virtual realm — may be a more important contribution to the future of the U.S. military. The technical foundations of artificial intelligence suggest applications in the latter category possess far more innovative potential — particularly in military decision-making processes.
Speed in warfare has a quality all its own. A major aim of future operations will be the ability to act faster than our adversaries, and our decision-making processes must be optimized to allow this. In order to do so, the Pentagon will need to embrace artificial intelligence as a means to process data.
The U.S. military either needs to make decisions faster or buy more time. Artificial intelligence is a possible solution to the input-output problem. It integrates well with existing structured systems of planning and assessment, and it takes advantage of vast data streams that currently sit untapped. Artificial intelligence also lends itself to those tasks that occupy the most time, freeing people to think critically and creatively about how to win wars.
David Zelaya is a U.S. Army officer commissioned from the University of Maryland’s ROTC program with a B.A. in Government and Politics. He is currently serving a tour in the Indo-Pacific at Fort Shafter, Hawaii. David’s interests focus on the intersection of economics and security.
Nicholas Keeley is a U.S. Army officer commissioned from Princeton University’s ROTC program with a B.A. in East Asian Studies. He is currently serving a tour in the Indo-Pacific and is stationed at Schofield Barracks, Hawaii. He is passionate about emerging technologies, East Asian politics, and data analytics.