[in the SPOTLIGHT]
continued
GOING BEYOND TRADITIONAL NETWORKS: SMART GRID, INTELLIGENT INFRASTRUCTURE, AND SOCIAL NETWORKS SPCOM is evolving to incorporate new networking concepts. The smart grid adds communication networking capability to an infrastructure network like the power grid or water grid. For example, the power grid can deliver power much more efficiently when real-time information about the grid state is available through networked meters and power infrastructure. Challenges in the smart power grid include developing machine-to-machine network protocols for reporting measurements, better techniques for large-scale network state estimation, and robustness to cyber attacks. The smart grid is just one example of a more general trend of networks-of-networks where different networking concepts are used to make infrastructure more intelligent. More examples include intelligent transportation systems, which use vehicle-to-vehicle networks to improve transportation network safety and efficiency. Mathematical tools from SPCOM are also being used to
understand noncommunication networks like social networks. COMPRESSIVE SENSING IN SPCOM Compressive sensing (CS) refers to efficient compression and reconstruction of analog signals that are sparse in some domain, e.g., space, time, or frequency. CS is a component of many different technical committees. In SPCOM, CS has been applied to the detection of impulse radio ultrawideband (exploits time-domain sparsity), radar (exploits sparsity in angle, Doppler, and/or range domain), and spectrum sensing (exploits sparsity in the spectrum). There are many applications of CS remaining in SPCOM, including localization and tracking through radar, or better navigation through global network satellite systems. Challenges remain, especially in evaluating the viability of CS versus non-CS techniques. LOCALIZATION Determining location is receiving renewed interest in SPCOM coupled
with applications such as sensor networks for telemetry and satellite navigation. For low-energy sensor applications, range-based localization is receiving attention where distance measurements between sensors and beacons and/or among the sensors themselves are exploited to compute the location of the sensors. Assisted satellite navigation is also likely to become more important, where signals of opportunity are exploited. New mathematical tools that are being exploited in localization include CS and multidimensional scaling. AUTHORS Shuguang (Robert) Cui (
[email protected]. edu) is an associate professor at Texas A&M University. Robert W. Heath Jr. (rheath@ece. utexas.edu) is an associate professor at The University of Texas Slides at Austin. Geert Leus (G.J.T.Leus @tudelft.nl) is an associate professor at the Delft University of Technology.
A.M. Zoubir, V. Krishnamurthy, and A.H. Sayed
Signal Processing Theory and Methods
T
he scope of the IEEE Signal Processing Theory and Methods (SPTM) Technical Committee has a broad span, ranging from digital filtering and adaptive signal processing to statistical signal analysis, estimation, and detection. There have also been significant advances in the estimation of sparse systems. These areas continue to play a key role in classical and timely applications. Digital Object Identifier 10.1109/MSP.2011.941987 Date of publication: 22 August 2011
Under the unifying theme “how simple local behavior generates rational global behavior,” an SPTM expert session was organized by the authors during ICASSP 2011 in Prague. This article summarizes the session and raises challenging questions for future research. It is by no means representative of all emerging topics in the areas of SPTM, but it includes trends and challenges that, in our opinion, will become important activities in SPTM in the coming years. The bibliography is not exhaustive due to space limitations; it
only gives some representative references the readers may want to consult. IN-NETWORK PROCESSING, LEARNING, AND ADAPTATION Cognitive or adaptive networks are composed of spatially distributed agents that share information over a graph. The topology of the graph may evolve dynamically over time due to movement of the agents or because agents wish to collaborate with other agents and form coalitions (see [1] and the references therein). Each agent possesses adaptation and learning abilities
IEEE SIGNAL PROCESSING MAGAZINE [152] SEPTEMBER 2011
1053-5888/11/$26.00©2011IEEE
[in the SPOTLIGHT]
continued
[2], [3]; for example, each agent can have the capability of running an adaptive or a Bayesian signal processing algorithm based on local data and also information from other nodes. There are at least two questions one can pose: 1) If each agent possesses limited capabilities, can the global behavior of the network be sophisticated? This can be viewed as an analysis problem—how do simple algorithms interact resulting in sophisticated behavior. 2) A related question is: what algorithm should each agent run for the global behavior of the network to achieve a particular objective? This can be viewed as a synthesis problem, since one is interested in designing distributed algorithms. The combination of in-network processing and adaptive cooperation leads to the emergence of learning and self-organization features across the network. Nature provides an abundance of examples of selforganization over biological networks consisting of mobile agents. While individual agents tend to exhibit limited cognitive abilities, it is the coordinated behavior among the agents that leads to the manifestation of decentralized intelligence and enables the agents to perform sophisticated maneuvers to evade predators. In many biological networks, no single agent is in command and yet complex patterns of formation are evident. Examples include fish joining together in schools [1], birds flying in formation [4], bees seeking a new hive, and bacteria foraging for food. A close synergy is emerging between studies on self-organization in the biological [5] and social sciences and studies on cognitive networks in signal processing and communications. There are ample opportunities for cross-disciplinary research that seeks to understand and reverse-engineer the decentralized intelligence encountered in socio-economic-biological networks, by exploring connections with adaptation over networks and by using enhanced signal processing techniques. Adaptive diffusion methods [1] and game theoretic methods [6], [7] are ideal tools for the synthesis and analysis of cognitive networks with varied capabilities. By spreading intelligence throughout the sys-
tem, such methods eliminate the need to transport information to and from a central point, while still allowing local information exchange to any desired degree. ROLE OF ADAPTATION THEORY An important feature of cognitive networks is that the individual nodes are not expected to rely mainly on information fed from their neighbors. Such cooperation among the nodes is only one factor in the learning process. The individual nodes should also possess local adaptive processing abilities that enable them to assess and react to the quality of the information received from their neighbors against their own personal beliefs [1], [4]. For this reason, cognitive networks do not expect all nodes to reach global agreement over the state of the environment, as is common in some useful consensus seeking strategies [8], [9]; nodes in a cognitive network do not need
SOLVING ESTIMATION AND TRACKING PROBLEMS OVER COGNITIVE NETWORKS GENERALLY REQUIRE OPTIMIZING CERTAIN GLOBAL COST FUNCTIONS IN A DISTRIBUTED MANNER. to converge to the same global value [2]. Actually, such variations in the individual levels of performance across a network are commonly observed in nature. Animals in a group do not act in absolute synchrony. There are variations in their patterns of motion and in their individual reactions to obstacles in the environment [1]. The same phenomenon is observed even in agent-based models of macroeconomies: the nodes (such as sellers and buyers) do not need to converge to the same equilibrium state. Instead, their state (and behavior) can fluctuate depending on their individual beliefs and preferences. Solving estimation and tracking problems over cognitive networks generally require optimizing certain global cost functions in a distributed manner. Consensus-based techniques are useful in enabling networks to evaluate average val-
ues across the network. Adaptive diffusion techniques, on the other hand, allow networks to more generally optimize global cost functions and to perform real-time adaptation and learning [2]. This level of generality is useful in modeling mobile adaptive networks, which serve as good models for various patterns of motion observed in biological networks. The analysis of such learning algorithms poses several challenges. What assumptions on graph connectivity, information patterns, rate of information sharing, adaptation rules, and learning dynamics are needed to ensure convergence to within acceptable levels of performance? ROLE OF GAME THEORY Game-theoretic methods [6], [7] can also derive rich dynamics through the interaction of simple components and can be used either as descriptive tools, to predict the outcome of complex interactions, or as prescriptive tools, to design systems around given interaction rules. Game theory is a complexity-based theory, along with percolation theory, cellular automata, and ecology modeling. Game-theoretic learning algorithms [6], [10] can allow individual agents to perform simple algorithms and, under suitable assumptions, ensure that the global performance converges to desired equilibrium sets. The game theoretic approach has also several appealing features for system design and analysis. Simple devices with limited awareness can be equipped with preconfigured or configurable utility functions and routines for maximizing their utility in an interactive, even unknown environment. Such devices can then be deployed to organize themselves effectively in a dynamic and unknown environment. As long as the utility functions are properly specified, these networked devices can be made to exhibit many desired behaviors [7]. The game theoretic design approach echoes natural systems in a form of biomimicry; biological agents such as insects have long since evolved the proper procedures and utility preferences to accomplish collective tasks. The same can be said of humans, who orient their utilities according to economics; the proper specification of utilities in this case are
IEEE SIGNAL PROCESSING MAGAZINE [154] SEPTEMBER 2011
[in the SPOTLIGHT]
continued
dictated by the structure of the economic system and realized through such mechanisms as pricing. The self-organizing feature of game-theoretic (decentralized) systems results in a specific set of benefits and challenges. APPLICATIONS Cognitive networks can be designed to perform a variety of tasks. Examples of applications include environmental monitoring, event detection, resource monitoring, target tracking, communications over cognitive radio channels, processing and control over smart power grids, analysis of swarm and animal flocking behavior [1], [4], design of multiagent systems, and analysis of collective decision making. While it is generally possible to find centralized or hierarchical processing mechanisms that are faster or more accurate in performing a given task, cognitive networks are generally more scalable, adaptable, and resilient. Cognitive networks can also be used to model herding behavior in macroeconomic systems. In agent-based economic models, the individual agents such as buyers, sellers, traders, brokers, and dealers, are capable of behavioral adaptation. The nodes are embodied with goals and beliefs related to patterns they see in pricing and profitability, and they react according to certain behavioral rules. Agent-based models are also prevalent in social networks, where they are used to model social interactions and the spread of disease or information. Extensive studies in computer science and graph theory have been devoted to understanding the structure of social networks in terms of properties such as their centrality (a measure of the influence of a node), closeness (how close individuals are on a social network), and clustering, and network degree (how many connections a node has). BIOMIMICRY Video
There has already been extensive work in the literature on exploiting naturally occurring phenomena in the development of biolog-
ically inspired techniques for application in various domains such as robotics and optimization. For
INTERESTINGLY, THERE IS EVIDENCE TO SUGGEST THAT CERTAIN PATTERNS OF BEHAVIOR MAY BE INDEPENDENT OF THE POPULATION. example, the ant colony optimization (ACO) procedure is based on how ants find the shortest path to food, and the particle swarm optimization (PSO) procedure is based on how birds flock to find food. Other research efforts have focused instead on rules that emulate the emergence of organized behavior in animal colonies. For example, in consensus-seeking models, the individual members in a colony adjust their velocities according to the average velocity of their neighbors. While consensus methods can be effective in emulating the coordinated motion of (animal) agents, they are nevertheless limited in their ability to model the remarkable adaptation, learning, and tracking capabilities that moving (animal) networks exhibit, especially when traversing an environment with unpredictable obstacles and predators. Adaptive diffusion methods provide effective modeling tools in these situations [1], [4]. Research efforts are needed to address broader questions such as understanding how and why complex patterns of behavior arise in biological networks under highly dynamic conditions. How do mobility and the changing topologies influence learning and cognitive abilities? How does information flow through a cognitive network? Are there similarities across different domains? Interestingly, there is evidence to suggest that certain patterns of behavior may be independent of the population. For example, when faced with two identical food sources, ants have been observed to focus on one of these sources for some time before switching to the other source. The same behavior
has been observed in humans choosing between two restaurants—this is modeled by social learning where agents learn from the actions of other agents. A related question is: how can a decision maker make global decisions based on local decisions made by selfish individual agents? It can be shown that even for elementary sequential detection problems, the optimal decision policy no longer has a threshold behavior [11]. AUTHORS A.M. Zoubir (
[email protected]. de) is a professor with Technische Universität Darmstadt, Germany. V. Krishnamurthy (
[email protected]. ca) is a professor with The University of British Columbia, Canada. A.H. Sayed (
[email protected]) is a professor with the University of California, Los Angeles, United States. REFERENCES
[1] S.-Y. Tu and A. H. Sayed, “Mobile adaptive networks,” IEEE J. Select. Topics Signal Processing, vol. 5, 2011. [2] F. Cattivelli and A. H. Sayed, “Diffusion LMS strategies for distributed estimation,” IEEE Trans. Signal Processing, vol. 58, no. 3, pp. 1035–1048, Mar. 2010. [3] C. G. Lopes and A. H. Sayed, “Diffusion least-mean squares over adaptive networks: Formulation and performance analysis,” IEEE Trans. Signal Processing, vol. 56, no. 7, pp. 3122–3136, July 2008. [4] F. Cattivelli and A. H. Sayed, “Modeling bird flight formations using diffusion adaptation,” IEEE Trans. Signal Processing, vol. 59, no. 5, pp. 2038–2051, 2011. [5] S. Camazine, J. L. Deneubourg, N. R. Franks, J. Sneyd, G. Theraulaz, and E. Bonabeau, SelfOrganization in Biological Systems. Princeton, NJ: Princeton Univ. Press, 2003. [6] S. Hart, “Adaptive heuristics,” Econometrica, vol. 73, no. 5, pp. 1401–1430, 2005. [7] M. Maskery, V. Krishnamurthy, and Q. Zhang, “Decentralized dynamic spectrum access for cognitive radios: Cooperative design of a non-cooperative game,” IEEE Trans. Commun., vol. 57, pp. 459–469, Feb. 2008. [8] S. Kar and J. Moura, “Sensor networks with random links: Topology design for distributed consensus,” IEEE Trans. Signal Processing, vol. 56, no. 7, pp. 3315–3326, July 2008. [9] V. Krishnamurthy K Topley, and G. Yin, “Consensus formation in a two-time-scale Markovian system,” SIAM J. Multiscale Model. Simul., vol. 7, no. 4, pp. 1898–1927, 2009. [10] M. Benaim, J. Hofbauer, and S. Sorin, “Stochastic approximations and differential inclusions,” SIAM J. Control Optim., vol. 44, pp. 328–348, Jan. 2005. [11] V. Krishnamurthy. (2011). Bayesian sequential detection with phase-distributed change time and nonlinear penalty—A lattice programming approach. IEEE Trans. Inform. Theory [Online]. Available: http://arxiv.org/ abs/1011.5298
IEEE SIGNAL PROCESSING MAGAZINE [156] SEPTEMBER 2011
Slides
