The Information Sciences Directorate describes its programs in network science as follows: Programs in the Network Science Division pursue “discovery and understanding of robust mathematical principles and laws that govern a broad variety of networks including organic, social, and electronic. These principles and laws serve as the foundation for the creation of algorithms which may be leveraged for autonomous system reasoning.” The programs are Multi-Agent Network Control, Wireless and Hybrid Communication Networks, Social and Cognitive Networks, Communications and Human Networks, Intelligent Information Networks, and Network Science and Intelligent Systems (an international program).1
Four of the Network Science programs were presented for review: Multi-Agent Network Control, Social and Cognitive Networks, Communications and Hybrid Networks, and Intelligent Information Networks. The division’s $58.7 million budget, including $8.3 million in core funding and $50.4 million in leveraged funding, supports 236 current projects.
Scientific Quality and Degree of Innovation
The projects presented were strong, and although consistent with the scientific objectives, covered a very narrow selection of them. The scientific objectives were stated differently in various presentations but are interpreted to be as follows: distributed control (i.e., to design new frameworks for distributed control of multi-agent systems with nonlinear behavior); co-evolving networks (stochastic dynamical systems of interacting agents—i.e., to design new frameworks for describing and analyzing dynamics of asymmetric network interactions between heterogeneous agents); and collective information processing (i.e., to create distributed information collection and processing systems for inference, prediction, and control of system-level dynamics, with special attention paid to emerging properties in interactive learning in distributed state estimation).
This program has been in existence for a number of years and is currently in transition, as a new program manager is expected to arrive within a few months of the August review. A new program manager can be expected to bring a welcome injection of new energy and ideas to this program to drive a new wave of innovation and contributions, and to increase the coherence of the portfolio of research, as the vision and agenda are manifest in the addition and elision of program elements.
This is one of the longest running programs in the Network Science Division. It targets a set of areas that is relatively mature. Careful management will be needed to continue to find new opportunities for significant progress here. A potentially productive activity for the incoming program manager would be to hold one or more brainstorming workshops to identify a set of opportunities that could impact the Army in the coming decades, reflecting the long look ahead inherent in the Army Research Office’s (ARO)’s basic research mission.
While examples of coevolving networks were presented, there could have been more substance to the specific research issues associated with them. For example, given the emergence of cyberwarfare and its importance to the Army, using the mathematical principles developed might provide insight into the mathematics of coevolution of cyberdefense and cyberoffense. A mathematical treatment of software coevolution in a world of rapid releases, software vulnerability marketplaces and advances in cybervulnerability discovery—for example, as demonstrated in the Defense Advanced Research Projects Agency (DARPA) Cyber Grand Challenge—could form a basis for next-generation software development practices for the Army. The topic of emerging properties in interactive learning in distributed state estimation is timely and interesting.
In the area of multi-agent network control, as applied to fields such as biology, an application (particularly a potentially reachable visionary application that is Army-centric) might include self-forming physical objects, or physical objects that are transformable. For example, a parachute, rather than buried to conceal a Ranger’s presence, might be reshaped on command into clothing if decentralized control were used. Another possibility is self-fitting adhesive wound-dressing for battle space wound treatment. For both, polymers would be well-suited; it is the distributed control that the body exhibits with clotting and scabbing that would allow the basic research developed here to save soldier lives when applied.
Several impressive accomplishments were identified: new frameworks for distributed control, based on reduction to optimization and proof that it works for linear time-invariant systems; a new framework to account for global temporal constraints by mixing discrete and continuous control under uncertainty; experimentation with swarm behavior of insects, deriving global principles; and evolutionary game theory-based formalization of interacting networks in nature.
Many Army problems map into a multi-agent network framework, so the potential for transition of significant results should, in principle, be high. The reported transitions demonstrate interest both inside and outside the Army. For example, one project developed methods incorporated into a Small Business Technology Transfer project sponsored by the Army’s Aviation and Missile Research, Development, and Engineering Center (AMRDEC). This project could both address an Army requirement at AMRDEC and lead to commercialization through the small business partner. The list of transitions for this program was shorter and less impressive than for other programs in the Network Science Division, perhaps as a result of the lack of a consistent program manager in recent years.
Scientific Quality and Degree of Innovation
The Social and Cognitive Networks Program has a strong coherence in pursuit of its strategic aims in team science and computational social science. Within the team science space, there is a clearly defined pursuit of research investigating the fundamental principles of team effectiveness and most specifically on quantification of cognitive dimensions of teamwork and team design and assembly. Within the computational social science domain, there is a clear articulation of focus on understanding online/in-person behaviors and how to meaningfully make sense of and use high-flow rate social media data. These programs align well with the Army missions and have potential to meaningfully contribute to proximal and far-term Army needs. Across the program there are multiple investigators on separate awards pursuing different, sometimes complementary, approaches to the same scientific topic or problem. This approach increases the likelihood of overall program success and helps the program manager avoid inadvertent immersion and wedding to a single approach.
The Social and Cognitive Networks Program capitalizes on recent scientific developments in unobtrusive measurement techniques to extend and deepen the models of cognitive dimensions of teamwork, with a particular emphasis on transactive memory systems. The program is moving this area of research in new directions to explore a wider array of knowledge-sharing behaviors as well as mechanisms to track these and other behaviors within teams in near-real time. Additionally, the program appears to be capitalizing wisely on developments in machine learning applications from computational social science and other advanced analytic techniques to improve the science and productivity in the team effectiveness.
The research on deviant and socially destructive behavior in social media was interesting and appears to have strong potential to make both significant accomplishments as well as transitions to national security organizations. It is already garnering interest from several agencies and producing early transitions.
The idea of studying the impact of social networks on the brain and vice versa, involving functional magnetic resonance imaging (fMRI) data, is an ambitious, high-risk endeavor, because it is a priori unclear what insights could be gained from such end-to-end study. Additionally, it would be beneficial to identify and understand the boundary conditions of the scientific topics and findings within the program, for appropriately managing the transitions and to help drive the innovations by pushing against those boundaries when possible. The two basic research objectives—namely, the science of team formation and computation social science—were articulated very clearly. The specific research challenges (scientific barriers), approaches for solutions, and payoffs were clearly defined.
Two research objectives were given the highest priority: the discovery of fundamental principles of human team formation and the modeling of networked human behavior. In both areas the specific project goals, scientific barriers, and proposed approaches provide a reasonable basis to believe that the objectives will be met.
The preliminary accomplishments have provided new and interesting insights with respect to factors affecting performance of human teams, and the formation of deviant cyber flash mobs. The key performance metrics currently applied by ARO were described, and preliminary results indicate high
principal investigator (PI) productivity and outside recognition. The appendix of this report lists a broad set of metrics that ARO could consider for assessment of its programs.
Research results have been transitioned to a variety of stakeholders, including a system to improve team formation processes, transitioned to the Army Research Laboratory (ARL); research methodologies used to identify and study deviant groups, transitioned to the Federal Bureau of Investigation and the U.S. Cyber Command; and cyberforensic methodologies transitioned to the North Atlantic Treaty Organization, the Office of Naval Research, and DARPA.
Scientific Quality and Degree of Innovation
The three basic research objectives—finding fundamental (e.g., capacity) bounds of the physical layer in networks and how to achieve them; mathematical models of network performance; and network coevolution—were articulated very clearly. The specific research challenges (scientific barriers), approaches for solutions, and payoffs were clearly defined. Overall, there is a healthy mixture of work in traditional areas of wireless mobile networks and innovative use of new techniques in distributed learning and optimization.
In all three areas the specific project goals, scientific barriers, and proposed approaches provide a reasonable basis to believe that the objectives will be met.
Preliminary results indicate high PI productivity and outside recognition. Examples included prize medals, best paper awards, and elevations to fellow grade in professional societies. The preliminary accomplishments included interesting new insights in multiple-input and multiple-output systems (e.g., the study of robustness of the interference alignment technique), optimal real-time network traffic scheduling with tight deadlines, and autonomous sharing of a limited quality channel. Potential applications include those related to social networking and the ability to network in ground warfare across service lines.
The accomplishments represent significant scientific progress. For example, a major challenge in wireless networks is management of interference. This program has funded multiple efforts aimed at identifying methods for aligning interfering signals, as well as finding bounds on the performance of these methods. These projects have delivered results such as generalizing the degree of freedom metric to contend with channel uncertainty and identifying the total transmission capacity in millimeter wave networks. The latter result is an example of work that is of greater interest to the Army than to private industry, because millimeter wave technology is potentially better suited for military applications. These results directly address the scientific objectives for this program.
The transitions of the research results have been to a variety of stakeholders within the U.S. Army (e.g., ARL, Communications and Electronics Research, Development, and Engineering Center). Transitions to industry (NXP Corp.) and other government agencies (Office of Naval Research and Naval Research Laboratory) indicate the high degree of interest in this area. There are additional opportunities to transition novel approaches for controlling end-to-end delay bounds—for example, in DARPA programs where backpressure-based scheduling policies have been used due to throughput optimality in the wide area—while for tactical wireless networks used by deployed Army warfighters, the channel state information is often uncertain or unknown, and the new work supported by the ARO may lead to better performance of DARPA approaches in this specific Army network scenario.
Scientific Quality and Degree of Innovation
Three of the four basic research objectives—algorithmic game theory, reasoning about crowds, and algorithms for network inference—were articulated very clearly. The fourth objective, natural language processing, seemed less naturally connected to network science than these three, although a deep mathematical connection exists between graph construction and the often inconsistent observables in textual and visual data. The specific research challenges (scientific barriers), approaches for solutions, and payoffs were clearly defined. The novelty of the specific problems was addressed both independently and in the context of their relevance of specific U.S. Army concerns.
In each of the first three research objectives—game theory, crowds, and inference—the specific project goals, scientific barriers, and proposed approaches provide a reasonable basis to believe that the objectives will be met.
The preliminary accomplishments have provided new and interesting insights. Examples are game theory applications in learning in the context of partial observations and bounded rationality. In particular, the applications to attacker-defender games, repeated games and strategy updates, and application to real-life scenarios were interesting. For example, an analysis of risk-taking biases of attackers rooted in the behavioral economics work was validated in a series of experiments and led to an optimal strategy for security games. A notable characteristic of the game theory work was the diversity and relevance of the applications, including HIV prevention strategies for the large homeless population in Los Angeles, U.S. government agency impacts (Transportation Security Administration and the U.S. Coast Guard), and international impacts such as poacher behavior prediction to optimize placement of patrols in a Ugandan national park.
Key performance metrics (peer-reviewed publications, manuscripts, graduate students, postdoctoral researchers) were described, and preliminary results indicate high PI productivity and outside recognition, such as endowed chairs, elevations to fellow in professional societies, honorary degrees, and best paper awards. The appendix of this report lists a broad set of metrics that ARO could consider for assessment of its programs.
The transitions of the research results have been to a variety of stakeholders within the Army (e.g., application of Bayesian reasoning about societies, transitioned to the ARL) and to industry (e.g., GraphLab, a framework for developing AI algorithms, transitioned to Apple; and game theoretic algorithms for security, transitioned to Avata Intelligence), indicating the high degree of interest in this area. An opportunity for a different dimension of research is designing distributed systems (team missions) for resilience in spite of misbehaving participants (whether due to conscious malice or incompetence). It is not always possible to identify that there is a bad actor or to identify which actor is behaving counterproductively. Particularly in large anonymous teams, it is likely that the team will include badly behaving participants, and designing a collaborative algorithm that will make positive progress despite bad actors can be challenging. Given the applications of this research direction in the cyberdomain, the strength of the Network Science Division in identifying and funding principled and mathematically based foundational approaches to such problems may have significant promise for transition and impact on the Army. For example, the Iterated Feature Boosted Decision Tree scheme used to improve patrol placement to deter poachers in Uganda is applicable to similar Army roles such as peacekeeping in urban areas.
Scientific Quality and Innovation
The Network Science Division has, despite its small size, supported a broad swath of strong basic science that in spite of a long time horizon is directed toward areas of anticipated future needs of the U.S. Army. The division interprets the term “network” broadly as spanning the science to include social science (where a social network might arise) as well as traditional telecommunications networking. Telecommunications science research crosses technologies ranging from mobile networks to emerging quantum communications capabilities.
Across the division, the overall scientific quality is high, although some specific programs and investment areas are stronger than others. New areas identified for investment are unique and promising, with strong possibilities for contributing to the Army science and technology (S&T) mission.
Each program manager presented a unifying scientific vision defining the program area. The Social and Cognitive Networks Program is a particularly good example of this. Scientific objectives were given for fulfilling this vision. Thrust areas were defined to achieve these program objectives.
The Network Science Division’s program managers evinced a high level of engagement in community building and discipline building, in venues such as disciplinary meetings and academic institutions. They showed a strong sense of stewardship for these communities, particularly where the division pursued a distinct strategy, as in the social and cognitive networking area.
Across the division, program managers are actively seeking emerging developments in relevant scientific fields that can help move their programs forward in useful ways, as well as working to capitalize on recent major scientific advances. The challenge is for the program managers to identify those opportunities that will allow them to make a unique and meaningful contribution to the advancement of science in alignment with the Army S&T mission. It is important that the program managers continue to seek ripe opportunities at the gaps between major fields (e.g., self-organizing biological network structures for applications in medicine) and continue to collaborate across disciplines.
The programs in the Network Science Division are relatively new in comparison to those of the other divisions in the Information Sciences Directorate. While some of these programs (e.g., multi-agent network control and intelligent information networks) have been in existence in approximately their current form for 8-10 years, others (e.g., social and cognitive networks) are less than 5 years old. As a point of reference, the Mathematical Sciences Division has existed for decades, and many of its programs have 10-20 years of history. Since the Network Science Division is newer, it has fewer accomplishments compared to the older divisions; however, each program in the Network Science Division has a number of substantive accomplishments.
Specific scientific accomplishments were named for only the example projects detailed under each program. No global list of scientific accomplishments analogous to the lists of transitions was provided. The only means for assessment then were the key metrics presented (e.g., number of publications, students, and postdoctoral researchers). The appendix of this report lists a broad set of metrics that ARO could consider for assessment of its programs.
Programs across the division evidenced forethought and focus. As a result, all programs showed strong evidence of recent and ongoing transitions to applied research programs within the Army, as well as several to the broader defense science community. The program managers were very conscious of the need for basic science to remain relevant to the Army mission, and they show the potential to bring new developments to the Army. An indicator of the broader perceived value of the research commercially is the software package Graphlab, which was subsequently purchased by Apple for over $400 million.
The overall scientific quality was strong and, taking into account the magnitude of the budget available in ARO, quite impressive. It would be impossible for the division to support every relevant research area. The programs in this division address the wide variety of network topics beyond communication networks, although that is a central topic that they address. The high-quality research on social, cognitive, and information networks is to be commended.
The programs reviewed all funded strong basic science, but differences existed among the programs in the degrees of coherence and uniqueness within the government research funding ecosystem. To first order, this seemed related to the time that a program manager had shaped the priorities within their portfolio within the larger Network Science Division and ARO more generally.
In the area of quantum networking, considerable benefit could be gained from allying with other research thrusts elsewhere in ARO, ARL, or elsewhere in Department of Defense (DoD) to obtain a combination of basic science (e.g., theoretical analyses or mathematical models) and the best experimental science. For quantum science generally, it makes sense to maintain a strong and constant scientific interchange between models and experiments.
There are key methodological differences between ARO and research funding entities such as the National Science Foundation (NSF) and DARPA. ARO program managers energetically try to create communities with workshops, visits to universities, and talks. ARO program managers engage in an interactive process with proposers to mature ideas into projects through discussion, white papers, and the Short-Term Innovative Research program. ARO program managers are hands-on managers of their projects; this may be a function of the process whereby program managers work with potential proposers (which NSF does not generally do)—the ARO process seems to resemble DARPA’s management style. ARO program managers view as a transition a follow-on effort at another government research funding
agency that stems from one of their projects. This has two important benefits: (1) ARO program managers seek opportunities, in particular with other DoD components such as DARPA, to keep the basic research results alive in the research and development ecosystem; and (2) ARO program managers are encouraged to trace the progress of an idea from its origins in basic research through development, deployment, and use.