Tag Archives: machine learning

Future Tech

Alexa, Write my OPORD: Promise and Pitfalls of Machine Learning for Commanders in Combat

1 Comment

By Jeff Wong

Introduction

Jump into the time machine and fast forward a few years into the future. The United States is at war, things are not going well, and the brass wants to try something new. John McCarthy,1 a Marine lieutenant colonel whose knowledge of technology is limited to the Microsoft Office applications on his molasses-slow government laptop, mulls over his tasks, as specified by his two-star boss, the commander of Joint Task Force 58:

1. Convene a joint planning group to develop a plan for the upcoming counteroffensive. (Check.)

2. Leverage representatives from every staff section and subject-matter experts participating virtually from headquarters in Hawaii or CONUS. (Roger.)

3. Use an experimental machine-learning application to support the planning and execution of the operation. (We’re screwed.)

Nearly 7,000 miles from a home he might never see again, McCarthy considered two aphorisms. The first was from Marcus Aurelius, the second-century Roman emperor and stoic thinker: “Never let the future disturb you. You will meet it, if you have to, with the same weapons of reason which today arm you against the present.” 2 The second was from Mike Tyson, the fearsome boxer, convicted felon, and unlikely philosopher: “Everybody has a plan until they get punched in the mouth.”3

Artificial intelligence (AI), including large-language models (LLMs) such as ChatGPT, is driving a societal revolution that will impact all aspects of life, including how nations wage war and secure their economic prosperity. “The ability to innovate faster and better – the foundation on which military, economic, and cultural power now rest – will determine the outcome of great-power competition between the United States and China,” notes Eric Schmidt, the former chief executive officer of Google and chair of the Special Competitive Studies Project.4 The branch of AI that uses neural networks to mimic human cognition — machine learning (ML) — offers military planners a powerful tool for planning and executing missions with greater accuracy, flexibility, and speed. Senior political and military leaders in China share this view and have made global technological supremacy a national imperative.5

Through analyzing vast amounts of data and applying complex algorithms, ML can enhance situational awareness, anticipate threats, optimize resources, and adapt to generate more successful outcomes. However, as ML becomes more widespread and drives technological competition against China, American military thinkers must develop frameworks to address the role of human judgment and accountability in decision-making and the potential risks and unintended consequences of relying on automated systems in war.

To illustrate the promise and pitfalls of using ML to support decision-making in combat, imagine the pages of a wall calendar flying to some point when America must fight a war against a peer adversary. Taking center stage in this fictional journey are two figures: The first is McCarthy, an officer of average intelligence who only graduated from the Eisenhower School for National Security and Resource Strategy thanks to a miraculous string of B+ papers at the end of the academic year. The second is “MaL,” a multimodal, LLM that is arguably the most capable – yet most immature – member of McCarthy’s planning team. This four-act drama explores how McCarthy and his staff scope the problems they want MaL to help them solve, how they leverage ML to support operational planning, and how they use MaL to support decision-making during execution. The final act concludes by offering recommendations for doctrine, training and education, and leadership and policies to better harness this capability in the wars to come.

Act One: “What Problems Do We Want This Thing to Solve?”

The task force was previously known as “JTF-X,” an experimental formation that blended conventional legacy platforms with AI and autonomous capabilities. As the tides of war turned against the United States and its allies, the Secretary of Defense (SecDef) pressured senior commanders to expand the use of AI-enabled experimental capabilities. Rather than distribute its capabilities across the rest of the joint force, the SecDef ordered JTF-X into duty as a single unit to serve as the theater reserve for a geographic combatant commander. “Necessity breeds innovation… sometimes kicking and screaming,” she said.

 Aboard the JTF command ship in a cramped room full of maps, satellite imagery, and charts thick with unit designators, McCarthy stared at a blinking cursor on a big-screen projection of MaL. Other members of his OPT looked over his shoulder as he impulsively typed out, “Alexa, write my OpOrd.” Undulating dots suggested MaL was formulating a response before MaL responded, “I’m not sure what you’re asking for.”

The JPG chief, an Air Force senior master sergeant, voiced what the humans in the room were thinking: “Sir, what problems do we want this thing to solve?”

The incredible capacity of ML tools to absorb, organize, and generate insights from large volumes of data suggests that they hold great promise to support operational planning. Still, leaders, planners, and ML tool developers must determine how best to harness the capability to solve defined military problems. For instance, the Ukrainian military uses AI to collect and assess intelligence, surveillance, and reconnaissance (ISR) data from numerous sources in the Russia-Ukraine conflict.6 But as Ukrainian planners are probably discovering today, they must do more than throw current ML tools at problem sets. Current LLMs fall short of desired capabilities to help human planners infer and make sense within the operating environment. Military professionals must tailor the problem sets to match the capabilities and limitations of the ML solutions.

Although tools supporting decision advantage comprised a small fraction of the 685 AI-related projects and accounts within the DoD as of 2021, existing efforts align with critical functions such as the collection and fusion of data from multiple domains (akin to the DoD’s vision for Joint All-Domain Command and Control (JADC2)); multidomain decision support for a combatant command headquarters; automated analysis of signals across the electromagnetic spectrum; and location of bots to support defensive cyber operations.7 There are numerous tools with various tasks and functions, but the crux of the problem will be focusing the right tool or set of tools on the appropriate problem sets. Users need to frame planning tasks and precisely define the desired outputs for a more immature LLM capability, such as the fictional MaL.

McCarthy mapped out a task organization for the JPG to align deliverables with available expertise. The team chief scribbled dates and times for upcoming deliverables on a whiteboard, including the confirmation brief for the commander in 24 hours. An Army corporal sat before a desktop computer to serve as the group’s primary interface with MaL. To help the group develop useful queries and troubleshoot, MaL’s developers in Hawaii participated via a secure video teleconference.

MaL was already able to access volumes of existing data – operations and contingency plans, planning artifacts from previous operations, ISR data from sensors ranging from national assets to stand-in forces in theater, fragmentary orders, and mountains of open-source information.

Act Two: MaL Gets Busy to the ‘Left of the Bang’

Some observers believe that ML capabilities can flatten the “orient” and “decide” steps of the late military theorist John Boyd’s observe-orient-decide-act decision (OODA) loop, expanding a commander’s opportunity to understand context, gain an appreciation of the battlespace, frame courses of action, and explore branches and sequels to inform decisions.8 Nevertheless, the greater capacity that ML tools provide does not eliminate the need for leaders to remain intimately involved in the planning process and work with the platform to define decision points as they weigh options, opportunities, risks, and possible outcomes.

Planners should examine frameworks such as the OODA Loop, IPB, and the Joint Operational Planning Process (JOPP) to guide where they could apply ML to maximum effect. To support IPB, ML can automate aspects of collection and processing, such as identifying objects, selecting potential targets for collection, and guiding sensors. ML capabilities for image and audio classification and natural language processing are already in commercial use. They support essential functions for autonomous vehicles, language translation, and transit scheduling. These commercial examples mirror military use cases, as nascent platforms fuse disparate data from multiple sources in all five warfighting domains.9

MaL’s digital library included the most relevant intelligence reports; adversary tactics, techniques, and procedures summaries; videos of possible target locations taken by uncrewed systems; raw signals intelligence; and assessments of the enemy orders of battle and operational tendencies from the early months of the conflict. The corpus of data also included online news stories and social media postings scraped by an all-source intelligence aggregator.

 McCarthy said, “As a first step, let’s have MaL paint the picture for us based on the theater-level operational context, then create an intelligence preparation of the battlespace presentation with no more than 25 PowerPoint slides.” After the clerk entered the query, the graphic of a human hand drumming its fingers on a table appeared.

Two minutes later, MaL saved a PowerPoint file on the computer’s desktop and announced in a metallic voice, “Done, sir, done!” McCarthy and his J-2 reviewed the IPB brief, which precisely assessed the enemy, terrain, weather, and civil considerations. MaL detailed the enemy’s most likely and dangerous courses of action, discussed adversary capabilities and limitations across all domains, and provided a target-value analysis aligning with the most recent intelligence. The J-2 reviewed the product and said, “Not bad.” She added, “Should I worry about losing my job?”

“Maybe I should worry about losing mine,” McCarthy said. “Let’s go through the planning process with MaL and have it generate three COAs based on the commander’s planning guidance and intent statement.”

American military planning frameworks – JOPP and its nearly identical counterparts across the services – are systematic processes that detail the operational environment, the higher commander’s intent, specified and implied tasks, friendly and enemy COAs, and estimates of supportability across all warfighting functions. They benefit the joint force by establishing uniform expectations about the information needed to support a commander’s estimate of the situation. However, current planning frameworks may hinder decision-making because a commander and his staff may become too focused on the process instead of devoting their energies and mental bandwidth to quickly orienting themselves to a situation and making decisions. Milan Vego, a professor of operational art at the U.S. Naval War College, writes of “a great temptation to steadily expand scope and the amount of information in the (commander’s) estimate. All this could be avoided if commanders and staffs are focused on the mental process and making a quick and good decision.”10

An ML-enabled decision-support capability could help planners stay above process minutiae by suggesting options for matching available weapon systems to targets, generating COAs informed by real-time data, and assessing the likelihood of success for various options in a contemporary battlespace, which features space and cyberspace as contested warfighting domains.11

MaL developed three unacceptable COAs which variously exceeded the unit’s authorities as outlined in the JTF’s initiating directive or extended kinetic operations into the adversary’s mainland, risking escalation.

McCarthy rubbed his face and said, “Time for a reboot. Let’s better define constraints and restraints, contain operations in our joint operational area, and have it develop assessments for risks to mission and force.” He added, “And this time, we’ll try not to provoke a nuclear apocalypse.”

The planning team spent several more hours refining their thoughts, submitting prompts, and reviewing the results. Eventually, MaL generated three COAs that were feasible, acceptable, complete, and supportable. MaL tailored battlespace architecture, fire support coordination measures, and a detailed sustainment plan for each COA and mapped critical decision points throughout the operation. MaL also assisted the JTF air officer develop three distinct aviation support plans for each COA. 

The planning team worked with MaL to develop staff estimates for each COA. The logistics and communications representatives were surprised at how quickly MaL produced coherent staff estimates following a few hours of queries. The fires, intelligence, and maneuver representatives similarly worked with MaL to develop initial fire support plans to synchronize with the group’s recommended COA.

McCarthy marveled at MaL’s ability to make sense of large amounts of data, but he was also surprised at the ML platform’s tendency to misinterpret words. For instance, it continually conflated the words “delay,” “disrupt,” and “destroy,” which were distinct tactical tasks with different effects on enemy formations. The planning team reviewed MaL’s COA overviews and edited the platform’s work. The staff estimates were detailed, and insightful, but still required corrections.

During the confirmation brief, the JTF commander debated several details about MaL’s outputs and risk assessments of the planning team’s recommended COA. McCarthy said, “Respectfully, Admiral, this is your call. MaL is one tool in your toolbox. We can tweak the COA to suit your desires. We can also calibrate the level of automation in the kill chain based on your intent.”

After a moment, the admiral said, “I’ll follow the planning team’s recommendation. Notice that I didn’t say MaL’s recommendation because MaL is just one part of your team.”

Act Three: MaL Lends a Hand to the Right of the ‘Bang’

Contemporary military thinkers believe that ML-enabled tools could improve decision-making during the execution of an operation, leveraging better situational awareness to suggest more effective solutions for problems that emerge in a dynamic battlespace. However, critics argue that developing even a nascent ML-enabled capability is impossible because of the inherent limits of ML-enabled platforms to generate human-like reasoning and infer wisdom from incomplete masses of unstructured data emanating from a 21st-century battlefield. Some are also concerned about the joint force’s ability to send and receive data from a secure cloud subject to possible malicious cyber activities or adversarial ML. Prussian military thinker Carl von Clausewitz reminds practitioners of the operational art that “War is the realm of uncertainty; three-quarters of the factors on which action in war is based are wrapped in a fog of greater or lesser uncertainty.”12 Technological solutions such as ML-enabled capabilities can temporarily lift parts of this fog for defined periods. Still, users must understand the best use of these capabilities and be wary of inferring certainty from their outputs.

Emerging capabilities suggest that ML can augment and assist human decision-making in several ways. First, ML systems can enhance situational awareness by establishing and maintaining a real-time common operational picture derived from multiple sensors in multiple domains. Greater battlespace awareness provides context to support better decision-making by commanders and more accurate assessments of events on the battlefield. Second, ML can improve the effectiveness and efficiency of kill-chain analytics by quickly matching available sensors and shooters with high-value or high-payoff targets.13 This efficiency is essential in the contemporary battlespace, where ubiquitous sensors can quickly detect a target based on a unit’s emissions in the electromagnetic spectrum or signatures from previously innocuous activities like an Instagram post or a unit’s financial transaction with a host-nation civilian contractor.

Indeed, some AI observers in the U.S. defense analytic community argue that warfighters must quickly adopt ML to maintain a competitive edge against the People’s Liberation Army, which some observers believe will favor the possible gains in warfighting effectiveness and efficiency over concerns about ethical issues such as implementing human-off-the-loop AI strategies.14 ML-enabled feedback constructs will enhance the control aspects of command and control to employ a large, adaptable, and complex multidomain force.15

It was now D+21. JTF-58 achieved its primary objectives, but the campaign did not unfold as intended. During the shaping phase of the operation, several high-value targets, including mobile anti-ship cruise missile batteries, escaped kinetic targeting efforts, living to fight another day and putting U.S. naval shipping at risk for the latter phases of the campaign. MaL failed to update the developers’ advertised “dynamic operating picture” despite attempts by forward-deployed sensors and reconnaissance teams to report their movement. Incredibly, the DoD still did not have access to data from some weapon systems due to manufacturer stipulations.16

MaL’s developers insisted that the forward-deployed sensors should have had enough computing power to push edge data to the JTF’s secure cloud. The CommO speculated that environmental conditions or adversary jamming could have affected connectivity. McCarthy shook his head and said, “We need to do better.”

MaL performed predictably well in some areas to the right of the bang. The task force commander approved using MaL to run networked force protection systems, including a Patriot battery that successfully intercepted an inbound missile and electronic-warfare (EW) systems that neutralized small unmanned aerial systems targeting a fuel farm. MaL’s use in these scenarios did not stretch anyone’s comfort level since these employment methods were barely different than the automation of systems like the U.S. Navy’s Phalanx close-in weapon system (CIWS), which has detected, evaluated, tracked, engaged, and conducted kill assessments of airborne threats for more than four decades.17

MaL’s communications and logistics staff estimates were precise and valuable for the staff. The CommO adjusted the tactical communications architecture based on MaL’s predictions about enemy EW methods and the effects of weather and terrain on forward maneuver elements. Similarly, the LogO worked with the OpsO to establish additional forward-arming and refueling points (FARPs) based on MaL’s fuel and munitions consumption projections.

In the middle of the operation, the task force commander issued a fragmentary order to take advantage of an unfolding opportunity. MaL leveraged open-source data from host-nation news websites and social media postings by enemy soldiers to inform battle damage assessment of kinetic strikes. Some of that information was fake and skewed the assessment until the intelligence officer corrected it by comparing it with satellite imagery and human intelligence reporting.

As with any emerging capability, commanders and their staffs must consider the risks of integrating ML into the planning, execution, and assessment of operations. One of the risks is inherent in forecasting, as the ML platform artificially closes the feedback loop to a decision-maker sooner than one would expect during real-world operations. Retired U.S. Navy Captain Adam Aycock and Naval War College professor William Glenney IV assert that this lag might make ML outputs moot when a commander makes a decision. “The operational commander might not receive feedback, and the effects of one move might not be recognized until several subsequent moves have been made,” Aycock and Glenney write. “Furthermore, a competent enemy would likely attempt to mask or manipulate this feedback. Under such circumstances … it is difficult to ‘learn’ from a first move in order to craft the second.”18

Another risk is that the data used by ML platforms are, in some way, inaccurate, incomplete, or unstructured. Whether real or training, flawed data will lead to inaccurate outputs and likely foul an ML-enabled tool’s assessment of the environment and COA development. “Unintentional failure modes can result from training data that do not represent the full range of conditions or inputs that a system will face in deployment,” write Wyatt Hoffman and Heeu Millie Kim, researchers with the Center for Security and Emerging Technology at Georgetown University. “The environment can change in ways that cause the data used by the model during deployment to differ substantially from the data used to train the model.”19

The corollary to inaccurate data is adversarial ML, in which an enemy manipulates data to trick an ML system, degrade or disrupt optimal performance, and erode users’ trust in the capability. Adversarial ML tricks can trick an ML model into misidentifying potential targets or mischaracterizing terrain and weather. In one notable demonstration of adversarial ML, researchers at the Chinese technology giant Tencent placed stickers on a road to trick the lane recognition system of a Tesla semi-autonomous car, causing it to swerve into the wrong lane.20 Just the possibility of a so-called “hidden failure mode” could exacerbate fears about the reliability of any ML-enabled system. “Operators and military commanders need to trust that ML systems will operate reliably under the realistic conditions of a conflict,” Hoffman and Kim write. “Ideally, this will militate against a rush to deploy untested systems. However, developers, testers, policymakers, and commanders within and between countries may have very different risk tolerances and understandings of trust in AI.”21

Act Four: Hotwash

McCarthy took advantage of an operational pause to conduct a hotwash. Over lukewarm coffee and Monsters, the conversation gravitated toward how they could use MaL better. The group scribbled a few recommendations concerning integrating ML into doctrine, training and education, and leadership and policies until the ship’s 1-MC sounds general quarters. 

Doctrine: To realize the utility of ML, military leaders should consider two changes to existing doctrine. First, doctrine developers and the operational community should consider the concept of “human command, machine control,” in which ML would use an auction-bid process akin to ride-hailing applications to advertise and fulfill operational requirements across the warfighting functions. Under this construct, a commander publishes or broadcasts tasks, including constraints, priorities, metrics, and objectives. “A distributed ML-enabled control system would award-winning bids to a force package that could execute the tasking and direct the relevant forces to initiate the operation,” write naval theorists Harrison Schramm and Bryan Clark. “Forces or platforms that accept the commander’s bid conducts (or attempts to conduct) the mission and reports the results “when communications availability allows.”22 This concept supports mission-type orders/mission command and allows C2 architectures to flex to instances and areas subject to low-bandwidth constraints.23

Second, doctrine developers should adjust joint, service, and domain-centric planning processes to account for additional planning aids, such as LLMs, modeling and simulation, and digital twins, which can more deeply explore COAs, branches, and sequels and accelerate understanding of threats and the operating environment. Explicitly changing planning doctrine to account for these emerging capabilities will encourage their use and emphasize their criticality to effective planning.

Training and Education: Tactical units must train and continually develop ML technical experts capable of conducting on-the-spot troubleshooting and maintenance of the tool. Meanwhile, the services should develop curricula to train budding junior leaders — corporals, sergeants, lieutenants, ensigns, and captains — that introduce them to machine-learning tools applicable to their warfighting domains, provide best principles for generating productive outputs, and articulate risks – and risk mitigations – due to skewed data and poor problem framing.

Best practices should also be documented and shared across the DoD. Use of ML capabilities should become part of a JPG’s battle drill, featuring a designated individual whose primary duty is to serve as the human interface with a decision-support tool such as MaL. Rather than work from scratch at the start of every planning effort, JPGs should have a list of queries readily available for a range of scenarios that can inform a commander’s estimate of the situation and subsequent planning guidance, and formulation of intent based on an initial understanding of the operating environment. Prompts that solicit ML-enabled recommendations on task organization, force composition and disposition, risks to force or mission, targeting, and other essential decisions should be ready for immediate use to speed the JPG’s planning tempo and, ultimately, a unit’s speed of action as well. The information management officer (IMO), which in some headquarters staffs is relegated to updating SharePoint portals, should be the staff’s subject matter expert for managing ML capabilities. IMOs would be the military equivalent of prompt engineers to coax and guide AI/ML models to generate relevant, coherent, and consistent outputs to support the unit’s mission.24

Leadership and Policies: There are implications for senior leaders for warfighting and policy. Within a warfighting context, senior defense leaders must identify, debate, and develop frameworks for how commanders might use ML to support decision-making in wartime scenarios. It seems intuitive to use a multimodal LLM tool such as the fictitious MaL to support IPB, targeting, and kill chain actions; in the same way, campaign models are used to inform combatant commander planning for crises and contingencies.

However, leaders and their staffs must also understand the limitations of such tools to support a commander’s decision-making. “Do not ask for an AI-enabled solution without first deciding what decision it will influence and what the ‘left and right limits’ of the decision will be,” Schramm and Clark warn.25 Likewise, AI might not be the appropriate tool to solve all tactical and operational problems. “Large data-centric web merchants such as Amazon are very good at drawing inferences on what people may be interested in purchasing on a given evening because they have a functionally infinite sample space of previous actions upon which to build the model,” they write. “This is radically different from military problems where the amount of data on previous interactions is extremely small and the adversary might have tactics and systems that have not been observed previously. Making inference where there is little to no data is the domain of natural intelligence.”26

Meanwhile, future acquisition arrangements with defense contractors must provide the DoD with data rights – particularly data generated by purchased weapon systems and sensors – to optimize the potential of ML architecture in a warfighting environment.27 Such a change would require the DoD to work with firms in the defense industrial base to adjudicate disagreements over the right to use, licensing, and ownership of data – each of which might bear different costs to a purchaser.

Epilogue

Technologists and policy wonks constantly remind the defense community that the Department must “fail fast” to mature emerging technologies and integrate them into the joint force as quickly as possible. The same principle should guide the development of AI/ML-enabled warfighting solutions. Commanders and their staffs must understand that this is a capable tool that, if used wisely, can significantly enhance the joint force’s ability to absorb data from disparate sources, make sense of that information, and close kill chains based on an ML tool’s assessment.

If used unwisely, without a solid understanding of what decisions ML will support, the joint force may be playing a rigged game against a peer adversary. ML-enabled capabilities can absorb large amounts of data, process and organize it, and generate insights for humans who work at a relative snail’s pace. However, these nascent tools cannot reason and interpret words or events as a competent military professional can. As strategic competition between the United States and China intensifies over Taiwan, the South China Sea, the Russian-Ukraine war, and other geopolitical issues, American political and military leaders must develop a better understanding of when and how to use ML to support joint force planning, execution, and assessment in combat, lest U.S. service members pay an ungodly sum of the butcher’s bill.

Lieutenant Colonel Jeff Wong, a U.S. Marine Corps reserve infantry officer, studied artificial intelligence at the Eisenhower School for National Security and Resource Strategy, National Defense University in the 2022-2023 academic year. In his civilian work, he plans wargames and exercises for various clients across the Department of Defense.

The views expressed in this paper are those of the author and do not necessarily reflect the official policy or position of the National Defense University, the Department of Defense, or the U.S. Government.

References

1. The fictional hero of this story, John McCarthy, is named after the Massachusetts Institute of Technology researcher who first coined the term “artificial intelligence.” Gil Press, “A Very Short History of Artificial Intelligence,” Forbes, December 30, 2016, https://www.forbes.com/sites/gilpress/2016/12/30/a-very-short-history-of-artificial-intelligence-ai/?sh=51ea3d156fba.

2. Marcus Aurelius, Meditations, audiobook.

3. Mike Berardino, “Mike Tyson Explains One of His Most Famous Quotes,” South Florida Sun-Sentinel, November 9, 2012, https://www.sun-sentinel.com/sports/fl-xpm-2012-11-09-sfl-mike-tyson-explains-one-of-his-most-famous-quotes-20121109-story.html.

4. Eric Schmidt, “Innovation Power: Why Technology Will Define the Future of Geopolitics,” Foreign Affairs, March/April 2023.

5. “Military-Civil Fusion and the People’s Republic of China,” U.S. Department of State, May 2020.

6. Eric Schmidt, “Innovation Power: Why Technology Will Define the Future of Geopolitics,” Foreign Affairs, March/April 2023.

7. Wyatt Hoffman and Heeu Millie Kim, “Reducing the Risks of Artificial Intelligence for Military Decision Advantage,” Center for Security and Emerging Technology Policy Paper (Washington, D.C.: Georgetown University, March 2023), 12.

8. James Johnson, “Automating the OODA Loop in the Age of AI,” Center for Strategic and International Studies, July 25, 2022, https://nuclearnetwork.csis.org/automating-the-ooda-loop-in-the-age-of-ai/.

9. Hoffman and Kim, “Reducing the Risks of Artificial Intelligence for Military Decision Advantage,” 7.

10. Milan Vego, “The Bureaucratization of the U.S. Military Decision-making Process,” Joint Force Quarterly 88, January 9, 2018, https://ndupress.ndu.edu/Publications/Article/1411771/the-bureaucratization-of-the-us-military-decisionmaking-process/.

11. Hoffman and Kim, 7.

12. Carl von Clausewitz, On War, ed. and trans. Michael Howard and Peter Paret (Princeton: Princeton University Press, 1976), 101.

 13. Hoffman and Kim, 7.

 14. Elsa Kania, “AI Weapons” in China’s Military Innovation, Brookings Institution, April 2020.

 15. Harrison Schramm and Bryan Clark, “Artificial Intelligence and Future Force Design,” in AI at War (Annapolis, Md.: Naval Institute Press, 2021), 240-241.

16. Josh Lospinoso, Testimony on the State of Artificial Intelligence and Machine Learning Applications to Improve Department of Defense Operations before the Subcommittee on Cybersecurity, U.S. Senate Armed Services Committee, April 19, 2023, https://www.armed-services.senate.gov/hearings/to-receive-testimony-on-the-state-of-artificial-intelligence-and-machine-learning-applications-to-improve-department-of-defense-operations. “Today, the Department of Defense does not have anywhere near sufficient access to weapon system data. We do not – and in some cases, due to contractual obligations, the Department cannot — extract this data that feeds and enables the AI capabilities we will need to maintain our competitive edge.”

 17. MK15 – Phalanx Close-In Weapon System (CIWS), U.S. Navy, September 20, 2021, https://www.navy.mil/resources/fact-files/display-factfiles/article/2167831/mk-15-phalanx-close-in-weapon-system-ciws/.

18. Adam Aycock and William Glenney IV, “Trying to Put Mahan in a Box,” in AI at War, 269-270.

19. Hoffman and Kim, CSET Policy Brief, 8.

20. Ibid, 8-9.

21. Ibid, 11.

22. Schramm and Clark, “Artificial Intelligence and Future Force Design,” 239-241.

23. AI at War, 241.

24. Craig S. Smith, “Mom, Dad, I Want To Be A Prompt Engineer,” Forbes, April 5, 2023, https://www.forbes.com/sites/craigsmith/2023/04/05/mom-dad-i-want-to-be-a-prompt-engineer/amp/.

25. AI at War, 247-248.

26. AI at War, 248.

27. Heidi M. Peters, “Intellectual Property and Technical Data in DoD Acquisitions,” Congressional Research Service In-Focus, IF12083, April 22, 2022, https://crsreports.congress.gov/product/pdf/IF/IF12083.

Featured Image: PHILIPPINE SEA (Sept. 22, 2020) Cpl. Clayton A. Phillips, a network administrator with Marine Air Control Group 18 Detachment, 31st Marine Expeditionary Unit (MEU), and a native of Beech Bluff, Tennessee, tests the connectivity of the Networking On-the-Move Airborne (NOTM-A) communications system during flight operations from the amphibious assault ship, USS America (LHA 6). (U.S. Marine Corps photo by Lance Cpl. Brienna Tuck)

Emerging Tech Week

A Roadmap to Successful Sonar AI

2 Comments

Emerging Technologies Topic Week

By LT Andrew Pfau

Even as the private sector and academia have made rapid progress in the field of Artificial Intelligence (AI) and Machine Learning (ML), the Department of Defense (DoD) remains hamstrung by significant technical and policy challenges. Only a fraction of this civilian-driven progress can be applied to the AI and ML models and systems needed by the DoD; the uniquely military operational environments and modes of employment create unique development challenges for these potentially dominant systems. In order for ML systems to be successful once fielded, these issues must be considered now. The problems of dataset curation, data scarcity, updating models, and trust between humans and machines will challenge engineers in their efforts to create accurate, reliable, and relevant AI/ML systems.

Recent studies recognize these structural challenges. A GAO report found that only 38 percent of private sector research and development projects were aligned with DoD needs, while only 12 percent of projects could be categorized as AI or autonomy research.1 The National Security Commission on Artificial Intelligence’s Final Report also recognizes this gap, recommending more federal R&D funding for areas critical to advance technology, especially those that may not receive private sector investment.2 The sea services face particular challenges in adopting AI/ML technologies to their domains because private sector interest and investment in AI and autonomy at sea has been especially limited. One particular area that needs Navy-specific investment is that of ML systems for passive sonar systems, though the approach certainly has application to other ML systems.

Why Sonar is in Particular Need of Investment

Passive sonar systems are a critical component on many naval platforms today. Passive sonar listens for sounds emitted by ships or submarines and is the preferred tool of anti-submarine warfare, particularly for localizing and tracking targets. In contrast to active sonar, no signal is emitted, making it more covert and the method of choice for submarines to locate other vessels at sea. Passive sonar systems are used across the Navy in submarine, surface, and naval air assets, and in constant use during peace and war to locate and track adversary submarines. Because of this widespread use, any ML model for passive sonar systems would have a significant impact across the fleet and use on both manned and unmanned platforms. These models could easily integrate into traditional manned platforms to ease the cognitive load on human operators. They could also increase the autonomy of unmanned platforms, either surfaced or submerged, by giving these platforms the same abilities that manned platforms have to detect, track, and classify targets in passive sonar data.

Passive sonar, unlike technologies such as radar or LIDAR, lacks the dual use appeal that would spur high levels of private sector investment. While radar systems are used across the military and private sector for ground, naval, air, and space platforms, and active sonar has lucrative applications in the oil and gas industry, passive sonar is used almost exclusively by naval assets. This lack of incentive to invest in ML technologies related to sonar systems epitomizes the gap referred to by the NSC AI report. Recently, NORTHCOM has tested AI/ML systems to search through radar data for targets, a project that has received interest and participation from all 11 combatant commands and the DoD as a whole.3 Due to its niche uses, however, passive sonar ML systems cannot match this level of department wide investment and so demands strong advocacy within the Navy.

Dataset Curation

Artificial Intelligence and Machine Learning are often conflated and used interchangeably. Artificial Intelligence refers a field of computer science interested in creating machines that can behave with human-like abilities and can make decisions based on input data. In contrast, Machine Learning, a subset of the AI filed, refers to computer programs and algorithms that learn from repeated exposure to many examples, often millions, instead of operating based on explicit rules programmed by humans.4 The focus in this article is on topics specific to ML models and systems, which will be included as parts in a larger AI or autonomous system. For example, an ML model could classify ships from passive sonar data, this model would then feed information about those ships into an AI system that operates an Unmanned Underwater Vehicle (UUV). The AI would make decisions about how to steer the UUV based on data from the sonar ML model in addition to information about mission objectives, navigation, and other data.

Machine learning models must train on large volumes of data to produce accurate predictions. This data must be collected, labeled, and prepared for processing by the model. Data curation is a labor- and time-intensive task that is often viewed as an extra cost on ML projects since it must occur before any model can be trained, but this process should be seen as an integral part of ML model success. Researchers recently found that one of the most commonly used datasets in computer vision research, ImageNet, has approximately 6 percent of their images mislabeled 5. Another dataset, QuickDraw, had 10 percent of images mislabeled. Once the errors were corrected, model performance on the ImageNet dataset improved by 6 percent over a model trained on the original, uncorrected, dataset.5

For academic researchers, where the stakes of an error in a model are relatively low, this could be called a nuisance. However, ML models deployed on warships face greater consequences than those in research labs. A similar error, of 6 percent, in an ML model to classify warships would be far more consequential. The time and labor costs needed to correctly label data for use in ML model training needs to be factored into ML projects early. In order to make the creation of these datasets cost effective, automatic methods will be required to label data, and methods of expert human verification must ensure quality. Once a large enough dataset has been built up, costs will decrease. However, new data will still have to be continuously added to training datasets to ensure up to date examples are present in the training of models.

A passive acoustic dataset is much more than an audio recording: Where and when the data is collected, along with many other discrete factors, are also important and should be integrated into the dataset. Sonar data collected in one part of the ocean, or during a particular time of year, could be very different than other parts of the ocean or the same point in the ocean at a different time of year. Both the types of vessels encountered and the ocean environment will vary. Researchers at Brigham Young University demonstrated how variations in sound speed profiles can affect machine learning systems that operate on underwater acoustic data. They showed the effects of changing environmental conditions when attempting to classify seabed bottom type from a moving sound source, with variations in the ability of their ML model to provide correct classifications by up to 20 percent.6 Collecting data from all possible operating environments, at various times of the year, and labeling them appropriately will be critical to building robust datasets from which accurate ML models can be trained. Metadata, in the form of environmental conditions, sensor performance, sound propagation, and more must be incorporated during the data collection process. Engineers and researchers will be able to analyze metadata to understand where the data came from and what sensor or environmental conditions could be underrepresented or completely missing.

These challenges must be overcome in a cost-effective way to build datasets representative of real world operating environments and conditions.

Data Scarcity

Another challenge in the field of ML that has salience for sonar data are the challenges associated with very small, but important datasets. For an academic researcher, data scarcity may come about due to the prohibitive cost of experiments or rarity of events to collect data on, such as astronomical observations. For the DoN, these same challenges will occur in addition to DoN specific challenges. Unlike academia or the private sectors, stringent restrictions on classified data will limit who can use this data to train and develop models. How will an ML model be trained to recognize an adversary’s newest ship when there are only a few minutes of acoustic recording? Since machine learning models require large quantities of data, traditional training methods will not work or result in less effective models.

Data augmentation, replicating and modifying original data may be one answer to this problem. In computer vision research, data is augmented by rotating, flipping, or changing the color balance of an image. Since a car is still a car, even if the image of the car is rotated or inverted, a model will learn to recognize a car from many angles and in many environments. In acoustics research, data is augmented by adding in other sounds or changing the time scale or pitch of the original audio. From a few initial examples, a much larger dataset to train on can be created. However, these methods have not been extensively researched on passive sonar data. It is still unknown which methods of data augmentation will produce the best results for sonar models, and which could produce worse models. Further research into the best methods for data augmentation for underwater acoustics is required.

Another method used to generate training data is the use of models to create synthetic data. This method is used to create datasets to train voice recognition models. By using physical models, audio recordings can be simulated in rooms of various dimensions and materials, instead of trying to make recordings in every possible room configuration. Generating synthetic data for underwater audio is not as simple and will require more complex models and more compute power than models used for voice recognition. Researchers have experimented with generated synthetic underwater sounds using the ORCA sound propagation model.6 However, this research only simulated five discrete frequencies used in seabed classification work. A ship model for passive sonar data will require more frequencies, both discrete and broadband, to be simulated in order to produce synthetic acoustic data with enough fidelity to use in model training. The generation of realistic synthetic data will give system designers the ability to add targets with very few examples to a dataset.

The ability to augment existing data and create new data from synthetic models will create larger and more diverse datasets, leading to more robust and accurate ML models.

Building Trust between Humans and Machines

Machine learning models are good at telling a human what they know, which comes from the data they were trained on. They are not good at telling humans that they do not recognize an input or have never seen anything like it in training. This will be an issue if human operators are to develop trust in the ML models they will use. Telling an operator that it does not know, or the degree of confidence a model has in its answer, will be vital to building reliable human-machine teams. One method to building models with the ability to tell human operators that a sample is unknown is the use of Bayesian Neural Networks. Bayesian models can tell an operator how confident they are in a classification and even when the model does not know the answer. This falls under the field of explainable AI, AI systems that can tell a human how the system arrived at the classification or decision that is produced. In order to build trust between human operators and ML systems, a human will need some insight into how and why an ML system arrived at its output.

Ships at sea will encounter new ships, or ships that were not part of the model’s original training dataset. This will be a problem early in the use of these models, as datasets will initially be small and grow with the collection of more data. These models cannot fail easily and quickly, they must be able to distinguish between what is known and what is unknown. The DoN must consider how human operators will interact with these ML models at sea, not just model performance.

Model Updates

To build a great ML system, the models will have to be updated. New data will be collected and added to the training dataset to re-train the model so that it stays relevant. In these models, only certain model parameters are updated, not the design or structure of the model. These updates, like any other digital file can be measured in bytes. An important question for system designers to consider is how these updates will be distributed to fleet units and how often. One established model for this is the Acoustic- Rapid COTS Insertion (ARCI) program used in the US Navy’s Submarine Force. In the ARCI program, new hardware and software for sonar and fire control is built, tested, and deployed on a regular, two-year cycle.7 But two years may be too infrequent for ML systems that are capable of incorporating new data and models rapidly. The software industry employs a system of continuous deployment, in which engineers can push the latest model updates to their cloud-based systems instantly. This may work for some fleet units that have the network bandwidth to support over the air updates or that can return to base for physical transfer. Recognizing this gap, the Navy is currently seeking a system that can simultaneously refuel and transfer data, up to 2 terabytes, from a USV.8 This research proposal highlights the large volume of data will need to be moved, both on and off unmanned vessels. Other units, particularly submarines and UUVs, have far less communications bandwidth. If over-the-air updates to submarines or UUVs are desired, then more restrictions will be placed on model sizes to accommodate limited bandwidth. If models cannot be made small enough, updates will have to be brought to a unit in port and updated from a hard drive or other physical device.

Creating a good system for when and how to update these models will drive other system requirements. Engineers will need these requirements, such as size limitations on the model, ingestible data type, frequency of updates needed by the fleet, and how new data will be incorporated into model training before they start designing ML systems.

Conclusion

As recommended in the NSC AI report, the DoN must be ready to invest in technologies that are critical to future AI systems, but that are currently lacking in private sector interest. ML models for passive sonar, lacking both dual use appeal and broad uses across the DoD, clearly fits into this need. Specific investment is required to address several problems facing sonar ML systems, including dataset curation, data scarcity, model updates, and building trust between operators and systems. These challenges will require a combination of technical and policy solutions to solve them, and they must be solved in order to create successful ML systems. Addressing these challenges now, while projects are in a nascent stage, will lead to the development of more robust systems. These sonar ML systems will be a critical tool across a manned and unmanned fleet in anti-submarine warfare and the hunt for near-peer adversary submarines.

Lieutenant Andrew Pfau, USN, is a submariner serving as an instructor at the U.S. Naval Academy. He is a graduate of the Naval Postgraduate School and a recipient of the Rear Admiral Grace Murray Hopper Computer Science Award. The views and opinions expressed here are his own.

Endnotes

1. DiNapoli, T. J. (2020). Opportunities to Better Integrate Industry Independent Research and Development into DOD Planning. (GAO-20-578). Government Accountability Office.

2. National Security Commission on Artificial Intelligence (2021), Final Report.

3. Hitchens, T. (2021, July 15) NORTHCOM Head To Press DoD Leaders For AI Tools, Breaking Defense, https://breakingdefense.com/2021/07/exclusive-northcom-head-to-press-dod-leaders-for-ai-tools/

4. Denning, P., Lewis, T. Intelligence May Not be Computable. American Scientist. Nov-Dec 2019. http://denninginstitute.com/pjd/PUBS/amsci-2019-ai-hierachy.pdf

5. Hao, K. (2021, April 1) Error-riddled data sets are warping our sense of how good AI really is. MIT Technology Review. https://www.technologyreview.com/2021/04/01/1021619/ai-data-errors-warp-machine-learning-progress/

6. Neilsen et al (2021). Learning location and seabed type from a moving mid-frequency source. Journal of the Acoustical Society of America. (149). 692-705. https://doi.org/10.1121/10.0003361

7. DeLuca, P., Predd, J. B., Nixon, M., Blickstein, I., Button, R. W., Kallimani J. G., and Tierney, S. (2013) Lessons Learned from ARCI and SSDS in Assessing Aegis Program Transition to an Open-Architecture Model, (pp 79-84) RAND Corperation, https://www.jstor.org/stable/pdf/10.7249/j.ctt5hhsmj.15.pdf

8. Office of Naval Research, Automated Offboard Refueling and Data Transfer for Unmanned Surface Vehicles, BAA Announcement # N00014-16-S-BA09, https://www.globalsecurity.org/military/systems/ship/systems/oradts.htm

Featured Image: Sonar Technician (Surface) Seaman Daniel Kline performs passive acoustic analysis in the sonar control room aboard the guided-missile destroyer USS Ramage (DDG 61). (U.S. Navy photo by Mass Communication Specialist 2nd Class Jacob D. Moore/Released)

Future Tech

Hyper-Converged Networks and Artificial Intelligence: Fighting at Machine Speed

10 Comments

By Travis Howard

Lieutenant Stacey Alto sits in the Joint Intelligence Center aboard the Wasp-class Amphibious Assault ship USS ESSEX (LHD 2). As the Force Intelligence Watch Officer (FIWO), her job is to absorb relevant information related to current and future operations of the Essex Amphibious Ready Group, as well as the general intelligence within the operating theater. Her zero-client, virtual desktop environment (VDE) 6-panel display at her watch station allows her a single-pane-of-glass into Unclassified, Secret, Top Secret, and Coalition enclaves through the Consolidated Afloat Networking and Enterprise Services (CANES) network.

One of her watch standers, an Intelligence Specialist Second Class, approaches her desk with new information from the Joint Operations Center (JOC), the nerve center of ARG operations, announcing new orders from the fleet commander to enter the Gulf of Oman, which represents a shift in operating theater from their current position in the Arabian Sea.

Stacey goes to work immediately, enlisting the help of two Intelligence Specialists and one of the Information Systems Technicians standing watch in the Ship’s Signal Exploitation Space (SSES). She queries the onboard widget carousel on her CANES SECRET terminal. Using a combination of mouse, keyboard, and touchscreen, she pulls together several ready-made widgets and snaps them into place, each taking advantage of a pool of “big data” information stored on the ship’s carry-on Distributed Common Ground System-Navy (DCGS-N) and off-ship sources from the intelligence cloud. Her development work gets passed to the next watch team, as they set the application’s variables for data parsing, consolidating inputs, and terrain mapping to put together a relevant, real-time intelligence picture.

By the time Stacey returns to her watch station almost 24 hours later, the IT personnel in SSES have put the new application through the automated cybersecurity testing process and have released it to the onboard “app store,” which Stacey can now install on her virtualized, thin-client desktop within seconds. She calls the JOC, the Marine Landing Force Operations Center (LFOC), and the ship’s Combat Information Center (CIC) announcing the system’s readiness with separate logins at the appropriate classification level for each watch station. By the time ESSEX enters the Gulf of Oman, the application has mapped adversarial positions and capabilities, pulled from several disparate databases afloat and ashore, all at varying levels of classification necessary for operational planning throughout the ship.

Building a More Maneuverable Network Afloat

The above scenario is almost a reality, representing several emergent advances in network technology and application portability (the “mobility” factor) that the Navy will soon capitalize on: a hardware and network-layer software architecture known as hyper converged infrastructure (HCI). The performance and cost efficiencies realized by this architecture will pave the way for disruptive changes to how we maneuver the network across the entire spectrum of operations: as a business system, as a decision support system, and as a warfighting platform.

Hyper-convergence is the integration of several hardware devices through a hypervisor, which acts as an intermediary and resource broker between software and hardware. Independent IT components are no longer siloed but combined, simplifying the entire infrastructure and improving speed and agility of the virtual network.1 The advantages of HCI seem obvious, but the real disruptive effect is how we can build upon it. The opening scenario describes on-demand application development at the tactical edge. This is achievable through HCI efficiency and another emerging network process known as Agile Core Services (ACS), a joint software development initiative being built into several programs throughout the Navy and Air Force, and one that CANES (as the afloat and maritime operations center network provider) is leveraging.

Hyper-Convergence in Network Hardware combines storage and processing power into a single appliance for simplified management, faster deployment, and could even lower acquisition costs ( Helixstorm.com)

ACS allows applications to use a common mix of services at the platform level, reducing cost and time of development but also forcing all applications to “speak the same language.” All that is needed to make on-demand, tactical application delivery a reality is a framework for plug-ins that takes advantage of big data we already have aboard ships and available at both the operational and tactical levels of war.

Previous articles in the United States Naval Institute’s magazine Proceedings have argued for thin-client solutions aboard warships,2 leveraging the CANES network program to ultimately achieve network efficiency that can remove “fat clients” (standard computer desktops) from the architecture to be replaced by thin or zero-clients (user workstation nodes with virtualized desktops and no onboard storage or input devices beyond keyboard and mouse). Removing clients from the equation eases the burden on shipboard technicians, consolidates the information security posture, and overall presents a more efficient network management picture through smart automation that makes better use of available manpower. HCI is the architecture solution that will eventually enable a full-scale, afloat, thin-client solution.

Hyperconverged.org is a website dedicated to delivering the message of advantages that HCI can bring,3 and lists ten compelling advantages that HCI brings to any IT infrastructure, to include:

  • Focus on software-defined data centers to allow faster software modernization and more agile vulnerability patching
  • Use of commercial off the shelf (COTS) commodity hardware that provides failure avoidance without the additional costs
  • Centralized systems and management
  • Enhanced agility in network management, automation, virtualization of operating systems, and shared resources across a common resource manager (such as hypervisor)
  • Improved scalability and efficiency
  • Potentially lower costs (caveat: in the commercial sector this may be truer than in the government sector, but smart contract competitions and vendor choices can drive down costs for the government as well)
  • Consolidated data protection through improved backup and recovery options, more efficient resource utilization, and faster network management tools

The advantages of HCI are numerous, and represent the true next step in IT architecture that will enable future software capabilities. How can we, as warfighters, take advantage of this emerging technology? It cannot be overstated that our current processes for procuring and delivering software-based services and capabilities must be revamped to keep pace with industry and take advantage of the speed and agility that HCI brings.

Faster, More Efficient Application Development is the Next Step

In our current hardware development methodology, programs of record within the Department of Defense (DoD) have little difficulty determining a clear modernization path that fits within the cost, schedule, and performance constraints outlined by the DoD acquisition framework. However, software development is an entirely different story, and is no longer agile enough to suit our needs. If we can iterate hardware infrastructure at near the speed of industry, then software and application development becomes the pacing function that we must address before we can realize the opening scenario of this essay.

The key term when discussing the speed of system development is agility, defined by the Massachusetts Institute of Technology (MIT) as “the speed of operations within an organization and speed in responding to customers…or reduced cycle times.”4 The federal government, DoD in particular, has been struggling with acquisition reform for some time, and with the signing of the National Defense Authorization Act in fiscal year 2010, Congress placed renewed emphasis on the need to transform the acquisition process for information technology. Several programmatic changes to acquisition helped (such as the approval of the “IT Box” programmatic framework in the joint requirements process), but the agility of software development and modernization remains challenged. Ensuring proper testing and evaluation (T&E) methodology, bureaucratic approval processes to ensure affordability, joint interoperability testing, and lengthy proof-in testing are just some of the processes facing software applications prior to gaining approval for full-rate production and fielding to the warfighter.

Matthew Kennedy and Lieutenant Colonel Dan Ward (U.S. Air Force), in a 2012 article for Defense Acquisition University, argued for agility in system development by discussing flaws in the current “agile software development” model.5 Developed in the early 2000s, this model is not as agile as the name would imply, and still defines requirements to be developed in advance, which doesn’t leave room for innovation or rapid, iterative changes to keep pace with the speed of industry. Exciting initiatives are being fielded in the commercial sector, such as cloud-based development and learning models, and mobility technology that many of the services would use to great effect. Innovative prototyping of disruptive technology at the service or component level of DoD, such as the now-disbanded Chief of Naval Operation’s Rapid Innovation Cell (CRIC), proved that there are operational advantages to emerging tech such as wearable mobile devices, if only we could “turn a tighter circle” within our acquisition framework and work with agility to field newer and better versions to the force.

Thankfully, we don’t have to reinvent the wheel when implementing a more agile software development framework; we must take lessons from industry and apply them to the unique needs of each of the DoD components. This may be easier said than done, but Kennedy and Ward, and indeed likely many other acquisition professionals and scholars, would agree that it is entirely possible if leadership demanded it, and the policies, procedures, and resourcing followed suit to support it. Kennedy and Ward offered a common set of software and business aspect practices to support agile practices that would allow a predictable, faster software refresh cycle (not just patches, but cumulative updates) to ensure software remains agile and relevant to the warfighter. Using small teams for incremental development, lean initiatives to shorten timelines, and continuous user involvement with co-located teams are just some of the practices offered.6

Improving our software development and modernization framework to be even more agile than it is now is necessary considering the recent industry shift to software-as-a-service and cloud-based business models. No longer will software versions be deliberate releases, but rather iterative updates such as Microsoft’s “current branch for business” (CBB) model. With this model, Microsoft envisions that Windows 10 could be the last “version” of Windows to be released, which will then be built upon in future “service pack-like” updates every 12-18 months. Organizations that do not update their operating systems to the latest CBB will be left behind with unsupported versions. Not only does such a change demand a rapid speed-to-force update solution for DoD, but it represents a disruptive process change that will ultimately allow us to reach the opening scenario’s on-demand tactical application process, leveraging big data in a way that units at the tactical edge have never done before – and in a way that may never have been imagined by the system’s original developers.

Hyper-convergence infrastructure, together with agility-based application development and modernization, represents a near-term possibility that will enable true innovation at the tactical level of war and put the power of information superiority into the hands of the warfighter. While re-developing the acquisition framework to achieve this may be difficult, it is entirely possible and, many would say, necessary if DoD is to keep pace with emerging threats, take advantage of emerging technology and innovation, and ultimately retain its status as the best equipped and trained force the world has ever known.

Artificial Intelligence: The Next AEGIS Combat System

Now let’s imagine another scenario. USS LYNDON B. JOHNSON (DDG 1002), last of the Zumwalt-class destroyer line and used primarily to test emergent technology prototypes in real-world scenarios, slips silently through the South China Sea in the dead of night. She is the first ship in the U.S. Navy to possess Nelson, a recursively-improving artificial intelligence (RIAI). Utilizing an HCI supercomputer core, Nelson acts as an integrator for the various shipboard combat systems in a similar concept to today’s AEGIS Combat System, except much faster and with machine-speed environmental adaption.

American relations with China have broken down, resulting in a shooting war in the South China Sea that threatens to spill into the Pacific proper, and eventually reach Hawaii. In an effort to change the dynamic, DDG-1002 forward deploys in stealth to collect intelligence on enemy force disposition and, if the opportunity presents itself, offer a first-strike capability to the U.S. Pacific Command. JOHNSON is spotted by a surface action group of three Chinese destroyers, who take immediate action by firing a salvo of anti-ship cruise missiles followed by surface gunnery fire once in range.

At the voice command of the Tactical Action Officer, Nelson goes to work, taking control of the ship’s self-defense system and prioritizing targets in a similar fashion to Aegis, only much faster, while constantly providing voice feedback on system readiness, target status, and battle damage assessments through the internal battle circuit, essentially acting as a member of the CIC team. Nelson’s adaptability as an AI allows it to evolve its tactical recommendations based on the environment and the sensory input from the ship’s 3D and 2D radars, intelligence feeds, and even the voice reports over the battle circuit. Compiling the tactical picture on a large display in CIC, Nelson simultaneously responds to threats against the ship while providing a fused battle management display to the Captain and Tactical Action Officer. The RIAI does much to lift the fog of war, and automates enough of the ship’s defensive and information-gathering functions to allow the humans to focus on tactically employing the ship to stop the threat rather than reacting to it.

While hyper-convergence, coupled with agile and rapidly-developed software innovation, is the emerging technology, recursively-improving artificial intelligence is the ultimate disruptive technology in the near to medium-term and represents the giant leap forward that many research and development efforts are striving towards. AI has often been relegated to the work of science fiction, and while many futurists see it as the inevitable “singularity” to happen as soon as the mid-21st century, it has not quite gained acceptance in the mainstream technical community. What must be focused on from a warfighter’s perspective is the near-term (within the next 30-50 years) prospects of advances in quantum computing, neural networks, robotics, nanotechnology, and hyper-convergence. These advances could put us on a path towards artificial intelligence within the lifetime of generations currently serving or about to serve in the armed forces.

The debate over whether recursively self-improving artificial intelligence is possible continues,5 with some theorists stating that such an AI cannot be achieved because intelligence could be “upper bounded” in a way that transcends processor speed, available memory, and sensor resolution improvements. Others suggest that intelligence “is the ability to find patterns in data”7 and that, regardless of the more fringe theories surrounding AI, transhumanism, and the ontological discussions of the singularity, “a sub-human level system capable of self-improvement can’t be excluded.”8  It is the sub-human AI, capable of adapting to changing data patterns, that makes a combat system AI an exciting near-future prospect. 

Conclusion

This article presented two hypothetical scenarios. In the near-term, a Navy watchstander takes advantage of a hyper-converged infrastructure network environment onboard a U.S. Navy warship to rapidly develop a tactical application to take advantage of disparate databases and cloud data resources, ultimately producing a battle management aid for the ship’s next mission. This scenario took advantage of two emerging technological concepts: hyper-convergence in hardware infrastructure, a reality some major defense acquisition programs such as the Navy’s CANES has already resourced and on-track to field in the coming years, and agile software development in defense acquisition, which is a conceptual framework that must be developed to ensure more rapid and innovative software capabilities are delivered to the force.

The funding for these technological advances must remain stable to deliver HCI to our operating forces as a hardware baseline for future development, and policy makers must continue to find efficiencies in IT acquisition that lead to agile software development to really take advantage of the efficiencies HCI brings. Additionally, DoD IT leaders must think critically and dynamically about how future software updates will be tested and fielded rapidly; our current lengthy testing and evaluation cycle is no longer compatible with either the speed of industry’s vulnerability patching, a fluid content upgrade schedule, or the pace of adversarial threats.

The second scenario describes a near-future incorporation of recursively-improving artificial intelligence within a combat system, which builds upon hyper converged hardware and recursively improving software to deliver a warfighting platform that can defend itself more rapidly and learn from its tactical situation. The simple fact is that technology is changing at a pace no one dared dream as early as 20 years ago, and if we don’t build it, our adversaries will. A recent (2016) article in Reuters, and reported in other media outlets, showcases the People Republic of China’s (PRC) desire to build AI-integrated weapons,9 citing Wang Changqing of China Aerospace and Industry Corp with saying “our future cruise missiles will have a very high level of artificial intelligence and automation.” DoD must adapt its processes to keep pace and remain the world’s leader in incorporating emerging and disruptive technology into its warfighting systems.

Travis Howard is an active duty U.S. Naval Officer assigned to the staff of the Chief of Naval Operations in Washington D.C. He holds advanced degrees and certifications in cybersecurity policy and business administration, and has over 16 years of enlisted and commissioned experience in surface warfare and Navy information systems. The views expressed here are solely those of the author and do not necessarily reflect those of the Department of the Navy, Department of Defense, or the United States Government.

References

1. Scott Morris. “Putting The ‘Hyper’ Into Convergence.” NetworkWorld Asia 12.2 (2015): 44. 28 Jan 2017.

2. Travis Howard, LT, USN. “’The Next Generation’ of Afloat Networking.” Proceedings Magazine, Mar 2015, Vol. 141/3/1,345

3. Hyperconverged.org. “Ten Things Hyperconverged Can Do For You: Leveraging the Benefits of Hyperconverged Infrastructure.” Retrieved Feb 2 2017, http://www.hyperconverged.org/10-things-hyperconvergence-can-do/

4. Matthew Kennedy & Lt Col Dan Ward. “Inserting Agility In System Development.” Defense Acquisition Research Journal: A Publication Of The Defense Acquisition University 19.3 (2012): 249-264. 4 Feb 2017.

5. Ibid

6. Ibid

7. Roman Yampolskiy. “From Seed AI to Technological Singularity via Recursively Self-Improving Software.” Cornell University Library. arXiv:1502.06512 [cs.AI]. 23 Feb 2015.

8. Ibid

9. Ben Blanchard. “China eyes artificial intelligence for new cruise missiles.” Reuters, World News. 19 Aug 2016, http://www.reuters.com/article/us-china-defence-missiles-idUSKCN10U0EM

Featured Image: Electronic Warfare Specialist 2nd Class Sarah Lanoo from South Bend, Ind., operates a Naval Tactical Data System (NTDS) console in the Combat Direction Center (CDC) aboard the USS Abraham Lincoln as it conducts combat operations in support of Operation Southern Watch. (U.S. Navy photo by Photographer’s Mate 3rd Class Patricia Totemeier)