8 September 2016

Capturing Flying Insects: A Machine Learning Approach to Targeting

September 6, 2016

A small insect darts from the security of the branch upon which it sits. It flutters through the leaves providing cover, revealing itself for a fleeting moment to the area just beyond leafy cover. Flying erratically, the insect believes it is safe for the brief moment of exposure — and it would be, from most animals. In that same fleeting moment, a nearby brown bat flies in and eats it. The bat’s ability to “detect, localize, and capture its prey” demonstrates an evolution from how most animals capture prey through constant bearing (same heading and velocity) to a method that succeeds in catching insects flying inconsistent patterns. Failure to adapt would result in extinction, so the bat uses sonar to improve its ability to see, track, and engage its prey.

Today’s asymmetric battlefield presents similar requirements for identifying, tracking, and engaging targets due to the increased data flow and speed of conflict. Current military structures struggle with these new data flows because they are self-contained vertical organizations that rely on hierarchical chains of command for decisions — primarily a function of post-World War II industrial age design. From the counter-insurgency environment in which enemies blend in with local populaces, only briefly exposing themselves for movement or communication, to large-scale conflicts on land, sea, air, and cyberspace that will occur at a velocity never before seen, military decision-makers must process more data in shorter periods of time to achieve success. The changing character of war, fleeting nature of targets, and glut of big data requires the military to integrate machine learning into its targeting process to win wars. Failure to do so will put the U.S. military a step behind its adversaries. We need a targeting revolution driven by machine learning.

Targeting

The process of selecting, prioritizing, and choosing weapons to engage objectives is targeting. Targets include facilities, individuals, virtual components, equipment, or organizations and are characterized by five broad groups (physical, functional, cognitive, environmental, temporal) and, believe it or not, 39 sub-characteristics. Systematic at all levels of command, the targeting cycle is “rational and iterative,” methodically analyzing, prioritizing, and assigning assets against targets or identifying, localizing, and capturing. Involving staff members from various services and functions, numerous lanes of expertise provide intelligence and provide possible solutions. Yet the decision about which targets to engage and how to engage them typically comes down to a single individual, the commander, who doctrinally is only fully engaged in two key steps and is somewhat removed from the analytical elements of the process.

The joint targeting cycle consists of six major steps: (1) end state and commander’s objectives, (2) target development and prioritizing, (3) capabilities analysis, (4) commander’s decision and force assignment, (5) mission planning and force execution, and (6) assessment. Each of these steps contains its own internal processes, such as a target system analysis that evaluates which of the targeted systems impact the operation the most. These processes work well in a conventional war in which operations are defined clearly by time and space. Yet even in this environment, intelligence analysts and targeteers must collect, compile, and evaluate enough intelligence to make the best decision to empower their commander. The amount of big data infusing the process today can overwhelm it. No matter how perfectly a staff conducts the process, individual decision-makers still hold overwhelming influence in the hierarchical structure of the military.

The Need for a Targeting Evolution

Recent conflicts demonstrated the pressures big data puts on the targeting process. Established for the military’s hierarchical approach, targeting led to both tactical and strategic failures in the initial phases of both Afghanistan and Iraq. As U.S. forces and supported militias closed in on Al Qaeda leadership in Tora Bora, human intelligence from locals combined with the knowledge of commanders on the ground determined both a general location and plan to inhibit the group’s escape into Pakistan. Secretary of Defense Donald Rumsfeld and Gen. Tommy Franks prevented U.S. forces and their partners from taking ground action, instead continuing to bomb the area in hopes for a surrender or, better, the collapse of a cave on Osama Bin Laden’s head. As we now know, Bin Laden escaped. A subsequent Senate Foreign Relations Committee investigation determined numerous failures, many involving human bias and slow decision-making. Fought against an enemy that no longer moved in an anticipated fashion, Tora Bora started to show the weakness of how the military implemented the targeting process.

Over the last decade, as the enemy shifted tactics, the intelligence requirements for U.S. forces grew almost exponentially, to the point of testing human processing power. As Gen. Stanley McChrystal stated, we “initially saw our enemy as we viewed ourselves,” making identifying the right target difficult. Relationships, reputation, popularity, money, and communications methods all played a role in insurgent networks, but these individual elements were difficult to consolidate in current target systems analyses. Moreover, insurgent networks could re-form or grow quickly, adapting to changes in American tactics. McChrystal’s organization, Joint Special Operations Command (JSOC), sought to reduce the “blinks”, or the gap between when an organization receives and acts upon information. JSOC improved intelligence collection by networking various agencies and types of intelligence. While an analyst alone may not understand the significance of a report, “crowdsourcing” intelligence through JSOC networking allowed the collective to see trends and predict enemy actions, alleviating the pressure on small intelligence sections. Moreover, JSOC could use immediate feedback from one target to locate another in Iraq. The bureaucratic barrier-breaking and information sharing by JSOC allowed humans to learn and adapt quickly. But success still required an organizational construct inherently difficult and costly to scale due to the training and experience required.

Beyond the asymmetry of insurgency and terror techniques, near-peer actors also present issues in the modern operating environment for the antiquated targeting process by, in the words of Frank Hoffman, blending “all forms of war” to “blur” the lines between regular and irregular warfare. According to Nadia Schadlow, these blurred lines of hybrid warfare challenge the way the American military fights and renders it difficult to understand efforts undertaken by one state. Current targeting processes work well for one method of war, but may not effectively analyze the “unfinished pages” of these hybrid conflicts and the nebulous activities that surround them, such as Russia’s efforts in Ukraine. As part of its incursion into Ukraine, Russia used methods to shape local and global perceptions, affecting decision-makers in their efforts to understand potential actions. Using psychological and information warfare combined with top-tier troops, Russia influenced actions of its adversaries while hiding its true intentions.

Moreover, swarming — when “several units conduct a convergent attack on a target from multiple axes” — may further confuse analysts and decision-makers as to the true threat and aim. Paul Scharre wrote that “decentralized execution (of swarms) can be devilishly hard to defeat” owing to its faster reaction rate, unpredictability, and agility. The overwhelming nature of a swarm may complicate the ability to determine the best targets to engage, especially when added with the blurriness of a hybrid conflict.

Machine Learning Evolution

Unlike current targeting methods, targeting with machine learning by definition (and like the bat) improves with experience. Using algorithms to help identify specific knowledge from data and experience “based on sound statistical and computational principles,” machine learning uses computers to “find hidden insights without being explicitly programmed where to look.” The evolution of machine learning over the past 50 years shows various sectors shifting to preempt the growing amount of data dominating their daily lives. Using a simple game of checkers, Arthur L. Samuel showed how machine learning succeeded with defined rules and a clear end state in 1959. Adding complexity, the Deep Blue computer defeated a human at chess in 1997. Programmed with all possible outcomes, Deep Blue worked her way through all future moves in a speed impossible for any human to beat. Matthew Lai of the Imperial College of London created a chess program named Giraffe that teaches itself to play chess. Instead of requiring coding of thousands of possible moves, Giraffe learns at it goes, spotting patterns in chess with generally successful outcomes. And today, combining a human with a machine in chess creates an unbeatable team.

Like targeting, the chess board offers thousands of moves, but any evolution in targeting must incorporate murkier data than that of a game board. High-frequency trading (HFT) and its use of machine learning is the closest example to that of military decision-making. There are consistencies in terms of the amount of data, variables that can change that data, constraints required, role of humans, and the effectiveness of reinforced learning between the military and high-frequency trading algorithms. These programs gather real time data across numerous variables from around the world and make decisions in micro-seconds — one millionth of a second — outdoing even the brown bat. Moreover, each maintains its own trading strategy in the same way a commander provides guidance on his or her strategy, allowing for the adjustment of algorithms. Establishing constraints in terms of profit and loss limits the computer in the same manner that rules of engagement would in targeting.

Some might argue that machine learning in war is much different than in the financial sector, and it is. However, comparing high-frequency trading results to those of the average stock picker demonstrates that humans are slower, less accurate, and cannot process all of the data efficiently. In the same manner, machine learning allows for faster and more accurate targeting. The computer can identify a targeted building as a hospital instead of an insurgent stronghold by querying a non-governmental organization (NGO) database instead of hoping a commander is aware of that information. As the nature of war shifts from insurgency to counter-terrorism to hybrid and back, capturing the enemy may rely more on social physics than on the vehicle identification cards on which many of us trained. The data arrives not in the form of a picture of a tank formation, but in ones and zeroes, requiring computer code to consolidate and understand the intelligence. Moreover, when machine learning combines with autonomous weapons, as Paul Scharre writes in War on the Rocks, “those who master a new technology and its associated concepts of operation can gain game-changing advantages.”

But what about those things that makes the military great — its people and their ability to lead? I do not talk about removing people from the fight. In fact, successful high-frequency trading algorithms rely on individuals to monitor the outcomes of their trades, entrusting that person with the power to stop trading altogether if something seems awry. For targeting and machine learning, commanders must dictate the strategy behind the machine algorithms for targeting. Targeteers input constraints based on the rules of engagement and other operational variables. And through supervised or semi-supervised learning wherein algorithms use labeled examples to assist in learning, the military will control how the system learns. Moreover, the commander can still pull the trigger. Machine learning just assists in identifying, locating, and determining the best method of engagement for a target. Also, the human at the end of the process mitigates concerns of a “flash crash.” In the same way an artillery officer confirms the digital firing calculations before a fire mission, commanders of each organization will confirm orders before sent, unlike the ripple effects of the crash of 2010. All of these human injections into the algorithm push past the concerns about dividing ethics and machines presented by Brecher, Niemi, and Hill. The military will remain true to its ethical standards while innovating its processes faster than adversaries.

McChrystal’s JSOC demonstrated how innovation in data collection and processing can improve the targeting process. A decade later, the U.S. military must continue to improve upon the targeting process, as big data is difficult for humans to consistently and correctly consume, comprehend, and analyze. We must recognize our own fallibility and apply technologies relied upon for decades by other industries. Machine learning used to consolidate big data, apply that data to a strategy, and make decisions in one-millionth of a second transforms the military’s ability to target from an antiquated approach suitable to only capturing through constant bearing to one that is adaptable to different enemies and methods of fighting war. Instead of relying on archaic industrial age structures that see innovation only in terms of product, the military can use machine learning to step forward into the information age to fight and win this nation’s wars.

Charlie Lewis is a Cyber Operations Officer in the United States Army. A graduate of the U.S. Military Academy and the Harvard Kennedy School, he is currently a Madison Policy Forum Military-Business Cybersecurity Fellow and Council on Foreign Relations Term Member. The views expressed are those of the author and not necessarily those of the Department of the Army, Department of Defense, or any agency of the U.S. government.

No comments: