Performance and design of cycles in consensus networks

https://doi.org/10.1016/j.sysconle.2012.10.014Get rights and content

Abstract

This work explores the role of cycles in consensus seeking networks for analysis and synthesis purposes. Cycles are critical for many reasons including improving the convergence rate of the system, resilience to link failures, and the overall performance of the system. The focus of this work examines how cycles impact the H2 performance of consensus networks. A first contribution shows that the addition of cycles always improves the performance of the system. We provide an analytic characterization of how the addition of edges improves the performance, and show that it is related to the inverse of the cycle lengths and the number of shared edges between independent cycles. These results are then used to consider the design of consensus networks. In this direction we present an 1-relaxation method that leads to a convex program for adding a fixed number of edges to a consensus networks. We also demonstrate how this relaxation can be used to embed additional performance criteria, such as maximization of the algebraic connectivity of the graph.

Introduction

The consensus protocol has recently emerged as a canonical model for studying networked dynamic systems. The simplicity of the model, most often presented as a collection of single integrators interacting over a communication graph, reveals a deep connection between its dynamic behavior and the underlying properties of the graph [1]. The use of consensus models, even beyond its analytical elegance, is its practical relevance in applications ranging from optimization and sensor fusion to problems in formation control and distributed control and estimation [2], [3], [4], [5], [6]. It is precisely the utility and simplicity of this model that has pushed research in this area to pursue a more general control theory for networked dynamic systems.

Within the framework of this setup, the use of graph theory as a tool for analysis is recognized as the correct mathematical abstraction to study these systems. Perhaps the most celebrated result in this direction is the relationship of the algebraic connectivity of a graph, sometimes referred to as the Fiedler eigenvalue [7], to the convergence rate of the dynamic system. The importance of this property has been extended in many directions, including convergence analysis for random graphs [8] and graphs with communication delays and switching topologies [9]. Various modifications to the consensus protocol have also led to more general system theoretic notions such as controllability and observability [10], [11] and input–output properties [12], [13], [14].

While the analysis of consensus systems has matured, work related to the design of consensus protocols and the characterization of system performance beyond that of its convergence rate have not been as deeply studied. As the consensus protocol becomes more integrated into complex real-world networks, the importance of design from an optimality perspective becomes crucial. A fundamental challenge for the design of consensus networks, however, relates to the computational complexity of solving combinatorial problems. A common approach to this problem is to consider optimization over weighted graphs or other convex relaxations [15], [16], [17]. However, most results that deal with the design of networks focus on the optimization of the Fiedler eigenvalue, aligning with the mainstream results related to connectivity and convergence rates [18], [19].

An important extension of these works, therefore, is the introduction of more general notions of system performance in coordination with the design of these networks. A first step in this direction is to include exogenous inputs in the form of noises and disturbances into the consensus protocol. Such models have been considered in [17], [20], [21] where noises were introduced in either the process or measurement of the consensus protocol. This more general model of the consensus protocol better reflects their use in real-world systems. Indeed, multi-agent systems relying on relative sensing, such as formation flying or sensor fusion, will depend on sensors and actuators that are imperfect.

Noises in the consensus protocol lead to a random walk of the agreement value, and attempts to compensate for this includes the design of edge weights [17], or the introduction of a time-varying gain on the control to effectively reject the noise asymptotically [20]. When the consensus protocol is corrupted by noise, the H2 system norm provides a measure of how the noises affect the asymptotic deviation of each node’s state from a consensus state. This performance metric has been used to study, for example, the leader selection problem for consensus networks [22]. The H2 performance of large scale networked systems and fractal graphs were considered in [23], [24]. Other works have considered this metric for consensus networks over directed graphs [25], spanning trees [26], and biochemical networks [27].

However, many of these results do not exploit the underlying combinatorial properties of the graph as it relates to these metrics. We emphasize a distinction between spectral properties of the graph, i.e. the eigenvalues of the Laplacian matrix, and combinatorial properties of a graph such as path lengths and cycles. Indeed, spectral properties introduce a layer of abstraction to the underlying graph that makes more tangible design issues, such as edge costs and distances, less intuitive.

A thorough treatment in this direction was recently given in [28] via the introduction of the Edge Laplacian and its corresponding edge agreement problem. The edge Laplacian is a variant of the graph Laplacian that provides a more transparent understanding of how spanning trees and cycles affect certain algebraic properties of a graph. When the consensus protocol is analyzed using this construction, clear graph theoretic interpretations of the H2 norm of the system can be derived [28].

An unresolved question from [28], however, was the precise role that cycles play in the performance of consensus networks. The main result from [28] showed that the H2 performance is determined by the inverse of a matrix related to the cycle structure of the graph. However, the precise structure of this matrix was not considered, and this work contributes in that regard. It is well known that the addition of cycles in a graph will increase its algebraic connectivity [29],1 however, a similar result has not been found for other performance metrics. A fundamental contribution of this work, therefore, is an analytic characterizations of how cycles affect the H2 performance of consensus networks. In this direction, we provide new interpretations for the role of cycles as related to algebraic properties of the edge Laplacian, and dynamic properties of the consensus protocol. In particular, we show that the H2 norm of the consensus protocol always improves with the addition of cycles. We provide an exact characterization of how the addition of one or two cycles improves the performance. The improvement is proportional to the inverse of the cycle lengths and also related to the number of edges that are shared between cycles. This establishes a strong connection between combinatorial properties of the underlying graph and dynamic properties of the corresponding protocol.

The analytic results are then used to formulate a synthesis problem for consensus networks. The problem considers the task of adding a fixed number of edges to an existing consensus network over a spanning tree that leads to the greatest improvement in performance. A first approach to this problem leads to a mixed-integer program, generally considered a hard problem to solve due to its combinatorial nature. This combinatorial problem can be formulated as an 0-optimization problem. By combining graph theoretic insights with results from compressed sensing [30], [31], we reformulate the problem as a reweighted 1-optimization problem. The 1-optimization problems are well known to achieve sparse solutions [32], [33], [34] and we show that this convex relaxation of the original mixed-integer problem gives very good results. We also highlight how the weighting mechanism used in the 1 formulation provides an important tuning parameter for design. The analytic results of this work provide intuition for appropriate weighting functions, including cycle length weighting and cycle correlation weighting. This formulation allows to consider additional performance criteria including maximizing the algebraic connectivity of the graph. These results are demonstrated via some simulation examples.

This paper is organized as follows. Section 2 reviews the fundamental properties of the edge Laplacian and provides results relating to algebraic properties of this matrix and cycles in this graph. The edge agreement problem is given in Section 3. These results are then applied in Section 4 to derive the H2 performance of the agreement protocol as a function of the edges in the graph. The synthesis problem and 1 formulation is given in Section 5, and numerical simulations are presented in Section 6. Finally, we offer some concluding remarks in Section 7.

The notation used is standard. The set of real numbers is denoted by R. For a vector xRn, its transpose is given by xT and the ith component by xi; the ijth element of a matrix A is denoted [A]ij. The trace of a matrix is denoted tr[]. The null space and range space of a matrix is denoted as N(A) and R(A) respectively. A symmetric matrix is positive definite (semi-definite) if all its eigenvalues are positive (non-negative), and is denoted A>0(A0); the linear matrix inequality A>B(AB) is equivalent to AB>0(AB0). The cardinality of a set M is denoted as |M|. The 1 norm of a vector xRn is defined as x1=i|xi|. The 0-norm of a vector is defined as x0=|{ixi0}|, the number of non-zero elements in the vector x. Note that in fact, the 0-norm is not a true norm, but it is commonly referred to as a norm in the literature and we adopt that convention in this work.

Section snippets

Cycles and the edge Laplacian

As discussed in the introduction, graph theory plays a central role in the analysis of consensus networks. In particular, the consensus protocol is described in terms of a certain algebraic representation of the underlying communication graph known as the (graph) Laplacian. The work presented in here, however, relies on an alternative representation that we term the edge Laplacian [28]. In this section, we review the construction of the edge Laplacian and focus the presentation on the role

The edge agreement problem

The standard consensus model is based on a collection of n single integrator agents that exchange relative state information over a communication graph to generate a control. The model is usually presented as an autonomous system with no noises or disturbances [1], ẋ(t)=L(G)x(t). Here, the vector x(t)Rn is the concatenated state of each agent, and L(G) is the Laplacian matrix.

A two-port interpretation of the consensus protocol provides a framework for considering the presence of exogenous

Performance of cycles

In this work, we consider the H2 performance of the edge agreement problem (10), and in particular focus on how the addition of cycles impacts this measure. Analyzing the consensus protocol with this metric is meaningful in the sense that it provides a measure of how noises can affect the asymptotic deviation of each agent from a consensus configuration; i.e., it studies the effect of noises on the relative states in the consensus protocol. As discussed in [17], [20], noises can lead to a

Design of cycles

The results of Section 4 provide a clear analytic picture of how cycles impact the H2 performance of the edge agreement problem. A trivial conclusion that one may arrive at is to always use the complete graph. The complete graph also, for example, has the largest algebraic connectivity (λ2(G)) and thus is desirable for other performance indicators. However, as is common in real-world engineering applications, it may not be feasible to implement a consensus network with all possible

Simulation example

In this section we demonstrate the design procedure described in Section 5 with a few numerical examples. For each example we will work with the same spanning tree graph on |V|=30 nodes, generated randomly in MATLAB; see Fig. 4. The H2 performance and algebraic connectivity for this graph can be determined as Σ(T)22=43.5 and λ2(G)=0.07 and for the complete graph Σ(Kn)22=29.97 and λ2(Kn)=30. The longest cycle in this graph is determined by its diameter, diam[G]=10.

For this example, there are

Concluding remarks

This work provided a characterization of how cycles impact the H2 performance of consensus networks. This analysis was facilitated by using an edge variant of the graph Laplacian matrix, termed the essential edge Laplacian. This matrix representation of the graph leads to a deeper insight of the role cycles play for certain algebraic properties. When applied to the corresponding edge agreement problem, the role of cycles were related to the performance of the system. In particular, the purely

References (45)

  • M. Fiedler

    Algebraic connectivity of graphs

    Czechoslovak Mathematical Journal

    (1973)
  • Y. Hatano et al.

    Agreement over random networks

    IEEE Transactions on Automatic Control

    (2005)
  • R. Olfati-Saber et al.

    Consensus problems in networks of agents with switching topology and time-delays

    IEEE Transactions on Automatic Control

    (2004)
  • G. Parlangeli et al.

    On the reachability and observability of path and cycle graphs

    IEEE Transactions on Automatic Control

    (2012)
  • A. Rahmani et al.

    Controllability of multiagent systems: from a graph-theoretic perspective

    SIAM Journal on Control and Optimization

    (2008)
  • B. Briegel, D. Zelazo, M. Bürger, F. Allgöwer, On the zeros of consensus networks, in: Proc. 50th IEEE Conference on...
  • S. Tonetti, R.M. Murray, Limits on the network sensitivity function for homogeneous multi-agent systems on a graph, in:...
  • S.P. Boyd, Convex optimization of graph Laplacian eigenvalues, in: International Congress of Mathematicians, Vol. 3,...
  • S.P. Boyd, A. Ghosh, Growing well-connected graphs, in: Proceedings of the 45th IEEE Conference on Decision & Control,...
  • R. Dai et al.

    Optimal topology design for dynamic networks

  • A. Das et al.

    Agreement over noisy networks

    IET Control Theory & Applications

    (2010)
  • Y. Hatano, M. Mesbahi, A. Das, Agreement in presence of noise: pseudogradients on random geometric networks, in: 44th...
  • Cited by (88)

    • Sensitivity-based link addition for robust linear dynamical networks

      2021, Journal of the Franklin Institute
      Citation Excerpt :

      Therefore, modification of an already designed topology by adding links, rather than designing a new topology from the scratch, is an inexpensive and feasible way to improve or enhance a certain desired measure of the designed network. Measures such as algebraic connectivity [9,12], number of spanning trees [19], rigidity [31], controllability [4,25,40], network robustness [13,36], network coherence [10,33,38], have been studied in the context of topology modification. In this paper, similar to [13], we are concerned with adding a set of communication links to a network, such that the network robustness is maximized or improved.

    • Distributed Stabilization of Signed Networks via Self-Loop Compensation

      2023, IEEE Transactions on Network Science and Engineering
    View all citing articles on Scopus

    The authors thank the German Research Foundation (DFG) for financial support within the Cluster of Excellence in Simulation Technology and the Priority Programme 1305 “Control Theory of Digitally Networked Dynamical Systems”.

    View full text