Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Distributed Binary Consensus in Networks with Disturbances

ALEXANDER GOGOLEV and NIKOLAJ MARCHENKO, University of Klagenfurt


LUCIO MARCENARO, University of Genoa
CHRISTIAN BETTSTETTER, University of Klagenfurt, Lakeside Labs GmbH

This article evaluates convergence rates of binary majority consensus algorithms in networks with different
types of disturbances and studies the potential capacity of randomization to foster convergence. Simulation
results show that (a) additive noise, topology randomness, and stochastic message loss may improve the
convergence rate; (b) presence of faulty nodes degrades the convergence rate; and (c) explicit randomization
of consensus algorithms can be exploited to improve the convergence rate. Watts-Strogatz and Waxman
graphs are used as underlying network topologies. A consensus algorithm is proposed that exchanges state
information with dynamically randomly selected neighbors and, through this randomization, achieves almost
sure convergence in some scenarios.
Categories and Subject Descriptors: C.2.4 [Computer-Communication Networks]: Distributed Systems;
C.4 [Performance of Systems]: Fault Tolerance
General Terms: Algorithms, Performance, Fault Tolerance
Additional Key Words and Phrases: Consensus algorithms, randomization, faulty nodes, Byzantine fault
tolerance, delayed system, self-organization, density classification, majority sorting
ACM Reference Format:
Alexander Gogolev, Nikolaj Marchenko, Lucio Marcenaro, and Christian Bettstetter. 2015. Distributed binary
consensus in networks with disturbances. ACM Trans. Auton. Adapt. Syst. 10, 3, Article 19 (August 2015),
17 pages.
DOI: http://dx.doi.org/10.1145/2746347

1. INTRODUCTION
When centralized decision making in networked systems is difficult or impossible,
distributed consensus algorithms can be used to steer the system to a required state. 19
Such tasks arise, for example, in technical systems for sensor fusion [Olfati-Saber and
Shamma 2005], object tracking [Gee and Cipolla 1996], localization [Simonetto et al.
2010], and mission planning [Alighanbari and How 2006]. Collective and distributed
decision making is also important in natural and social networks [Cao et al. 2008;
Nabet et al. 2009].

This work was performed within the Erasmus Mundus Joint Doctorate in “Interactive and Cognitive Envi-
ronments,’ which is funded by the Education Audiovisual and Culture Executive Agency of the European
Commission under Erasmus Mundus Joint Doctorate Interactive and Cognitive Environments Framework
Partnership Agreement n 2010-0012. The work of A. Gogolev is supported by Lakeside Labs, Klagenfurt,
with funding from the European Regional Development Fund, Kärntner Wirtschaftsf örderungs Fonds, and
the state of Austria under grant 20214/21530/32606.
Authors’ addresses: A. Gogolev, N. Marchenko, and C. Bettstetter, Institute of Networked and Embedded Sys-
tems, Klagenfurt University; emails: {alexander.gogolev, Nikolaj.Marchenko, Christian.Bettstetter}@aau.at;
L. Marcenaro, Department of Naval, Electric, Electronic and Telecommunication Engineering, University of
Genoa, Genoa, Italy; email: Lucio.Marcenaro@unige.it.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for
components of this work owned by others than ACM must be honored. Abstracting with credit is permitted.
To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this
work in other works requires prior specific permission and/or a fee. Permissions may be requested from
Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or permissions@acm.org.
c 2015 ACM 1556-4665/2015/08-ART19 $15.00
DOI: http://dx.doi.org/10.1145/2746347

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:2 A. Gogolev et al.

Every node in a distributed consensus system follows simple rules, utilizing local
information and node-to-node communication to assist global system coordination.
Algorithms performing such tasks should be robust toward different types of distur-
bances. Studies on robustness of consensus consider faulty nodes [Pease et al. 1980],
noise and communication errors [Kar and Moura 2007; Mendonçia 2011], topological
changes [Saks and Zaharoglou 1993; Kingston and Beard 2006], and delays or lack of
synchrony [Bracha and Toueg 1985; Burman et al. 2006]. Out of these impediments,
faulty node behavior appears to be the most severe. Early investigations by Pease
et al. [1980] show that for synchronous systems in which M out of N nodes are faulty,
agreement is possible only if M < N+1 3
. Later, Fischer et al. [1985] strengthen this
condition for asynchronous systems showing that already one faulty node can prevent
an agreement, which has become known as the “FLP impossibility” (named according
to the first letters of the authors’ names).
Scholars approach the faulty-node problem with fault detection [Chandra et al. 1996;
Chandra and Toueg 1996], which generally leads to higher complexity of consensus
algorithms [Chandra and Toueg 1996], or by weakening the system requirements,
for example, on synchrony [Dolev et al. 1987; Dwork et al. 1988]. Another promising
approach to increase the fault tolerance is randomization [Aspnes 2003]. It exploits
the positive impact of random disturbances and is typically introduced in two ways
[Aspnes 2002]. The first approach introduces random components to the environment,
for example, noisy scheduling initially presented by Bracha and Toueg [1985] and later
developed by Aspnes [2000]. The second approach assumes randomization within an
algorithm itself (see Ben-Or [1983] and Bracha and Rachman [1992]). Randomization
can utilize random components of real-life systems, which are usually considered as
threats, and randomization is typically simpler than fault detection.
Research on randomization has mainly focused on multistep algorithms. At the same
time, simple randomization techniques are vital for heuristic one-step algorithms for
which sophisticated fault detectors are not applicable. Such simple consensus algo-
rithms rarely can guarantee the agreement and therefore are often evaluated with
their convergence rate, that is, the fraction of initial network configurations on which
an algorithm achieves an agreement among all nodes. This article focuses on a certain
class of consensus, namely, binary majority consensus, as used for density classification
(see, e.g., Mitchell et al. [1993]). Some one-step heuristic algorithms in this algorithm
class, such as Gacs-Kurdyumov-Levin (GKL) [Gacs et al. 1978], show high convergence
rates in noiseless and error-free ordered networks, but their convergence rate signifi-
cantly decreases with noise [Gacs et al. 1978] or errors [Moreira et al. 2004]. This weak
robustness can explain why simple algorithms were not thoroughly investigated un-
til recently. Some simple algorithms, however, can increase the convergence rate with
disturbances. Moreira et al. [2004] show that Simple Majority (SM) consensus may out-
perform GKL in random networks with communication errors. This can be partially
explained by the randomization provided by communication errors and random topol-
ogy, since some algorithms gain performance with randomization [Saks et al. 1991;
Aspnes 2002]. This motivated us to study the robustness of consensus in a larger range
of system parameters (also see Gogolev and Bettstetter [2012]; Gogolev and Marcenaro
[2014a, 2014b]).
The article at hand studies binary majority consensus (BMC) in synchronous and
asynchronous networks with message loss, random topology, noise, and faulty nodes,
focusing on the impact of these “disturbances” on the convergence rate. We show that
additive noise, topology randomness, and message loss can increase the convergence
rate, while presence of faulty nodes inhibits the consensus. Furthermore, we propose a
new consensus rule, called Random Neighbors Majority (RM), which builds on SM but
embeds randomization directly into the consensus algorithm, thus yielding a better

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:3

convergence rate and robustness toward faulty nodes. It reaches almost sure conver-
gence in some scenarios.
The article is organized as follows. Section 2 describes the algorithms under investi-
gation. Section 3 describes the system model and simulation setup. Section 4 presents
and discusses simulations results. Section 5 contains our conclusions.
2. DISTRIBUTED BINARY MAJORITY CONSENSUS
The algorithms under consideration perform binary majority consensus (BMC) in a
network with N nodes. Each node has an index i ∈ {1, 2, . . . , N}. Some node pairs are
connected by a link. For a given node i, the nodes having a link with i build the set of
neighbors Ni of node i. Nodes with indices j < i are called left-side neighbors. They
build a vector Li sorted in ascending order of |i − j|. Similarly, vector Ri is built with
right-side neighbors j > i. These vectors are further used by algorithms to access their
neighboring nodes.
Such left–right notation is especially meaningful if a physical location is associated
with each node and reflected in the index. It also makes sense in non-geometric graphs,
for example, when the index stands for a scalar node property, such as temperature or
birthday, in which case left–right could also mean up–down.
The binary state of a node i at time t is denoted by si [t] ∈ {−1, 1}, where si = 1
is illustrated in white and −1 in black. The state information that node i receives
from node j is si, j [t]. At starting time t = 0, each node obtains a randomly assigned
state si [0] of 1 or −1 as an outcome of a coin-flipping experiment, that is, both values
have the same probability. A network reaches agreement if all nodes come to the same
state within a certain time period T and this final system state corresponds to the
initial
 majority of values [Gacs et al. 1978]. In other words, the system  converges if
(a) i si [0] < 0 and there exists t ≤ T so that si [t] = −1, ∀ i, or if (b) i si [0] > 0 and
there exists t ≤ T so that si [t] = 1, ∀ i. Algorithms are always terminated after time T
independent of whether agreement has been reached or not (wait-free consensus).
Time is divided into steps of equal size . Without loss of generality, we assume that
the duration of one time step  equals to one unit of time. The time steps are indexed
by t = {0, 1, 2, . . . , T } for simplicity of notation.
Figure 1 visualizes examples of a consensus process. At every time step t, each node
updates its state (black or white) following a given algorithm based on its own state and
state information received from its neighbors in the network. In the given examples,
a network consists of N = 100 nodes, and the system is observed over T = 200
time steps. The figures qualitatively resemble the results obtained by Moreira et al.
[2004]; also see Gogolev and Bettstetter [2012]. Figure 1(a) illustrates a network that
converges to a single value within t < T steps. Figure 1(b) shows a network that
becomes clustered and cannot converge. Figure 1(c) illustrates how a low noise level
can promote consensus. Figure 1(d) shows how a higher noise level prevents consensus.
It may appear that BMC is not a very complex problem. Several decades of research,
however, resulted in algorithms with convergence rates far below the targeted 100%.
Meanwhile some more general consensus problems have been solved with guaranteed
convergence, even in the presence of faulty nodes [Ben-Or 1983; Aguilera and Toueg
2012]. The seeming controversy of simplicity and low convergence rates in BMC is
rooted in differences in the termination requirements [Aspnes 2003] and the used ref-
erence network model. Works on BMC specify that an agreement is expected within
a limited period of time (wait-free consensus) and that the agreed state should corre-
spond to the initial majority of state distribution (majority consensus). Additionally,
convergence rates of BMC algorithms are compared in a 6-connected ring lattice of 149
nodes, rather than in the eventually fully connected networks assumed for Ben-Or’s
algorithm [Ben-Or 1983], or the setup used by Pease et al. [1980] and Fischer et al.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:4 A. Gogolev et al.

Fig. 1. Simple Majority consensus in an ordered grid of N = 100 nodes, each with K = 3 neighbors, observed
over T = 200 time steps.

[1985]. These restrictions cumulatively restrict the convergence rate results of BMC
algorithms in the literature.
2.1. Algorithms
2.1.1. Simple Majority (SM) Consensus. A node calculates its new state on the basis of
information of its current state si,i and the states of all its neighbors [Margush and
McMorris 1981; Moreira et al. 2004]:
⎛ ⎞

si [t + 1] = G ⎝si,i [t] + si, j [t]⎠ . (1)
j∈Ni

Note that the received state information can be affected by noise or message loss. The
update function G(x) is defined as [Moreira et al. 2004]:

−1 for x < 0,
G(x) = (2)
+1 for x > 0.
2.1.2. Gacs-Kurdyumov-Levin (GKL) Consensus. Using GKL, a node i computes its new
state si [t + 1] using information about its current state si,i [t] and about the current
states of the first and third nearest neighbors to the left or right from node i (the “left”
and “right” notation is given earlier). The current state of the node si,i defines the side
from which the neighbors are chosen. We implement the rule

G(si,i [t] + si,l1 [t] + si,l3 [t]) for si,i [t] < 0,
si [t + 1] = (3)
G(si,i [t] + si,r1 [t] + si,r3 [t]) for si,i [t] > 0,
initially proposed in Gacs et al. [1978]. The nodes l1 , l3 ∈ Li and r1 , r3 ∈ Ri are the first
and third neighbors to the left and to the right, respectively.
GKL is a modification of SM with a value-dependent direction bias: depending on
its own value, a node gets information from the opposite directions of, for example,
a lattice. This bias counters the network clustering and provides a high convergence

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:5

rate in ordered grids [Gacs et al. 1978]. Note that GKL was not designed for random
or asynchronous networks and may perform worse under perturbations [Moreira et al.
2004]. We include GKL in our analysis in order to set a reference for convergence rate
and its dynamics under such conditions.
2.1.3. Random Neighbors Majority (RM) Consensus. To evaluate whether randomization
can be beneficially embedded into consensus algorithms, we propose the following
modification to SM: The new state si [t + 1] of a node i is no longer computed using state
information from all its neighbors but using state information from some randomly
selected neighbors. A key feature is that neighbor selection is dynamic, that is, in
general, a node chooses different neighbors in each step t. The resulting consensus
algorithm is called Random Neighbors Majority (RM). Each node running RM uses its
own state information and state information from C random neighbors to update its
state:
⎛ ⎞

C
si [t + 1] = G ⎝si,i [t] + si,n j [t]⎠. (4)
j=1

To select the neighbors n j , we sample from a discrete uniform distribution with repe-
tition, that is, the same neighbor can be queried more than once to transmit its state
information to node i.

2.2. Update Synchrony


Two common state update schemes are synchronous and asynchronous [Dolev et al.
1987; Gershenson 2004]. In a synchronous mode, all nodes update their states simul-
taneously. In the asynchronous mode, nodes are updated sequentially, one after the
other. The update sequence is generated once at the beginning of the simulation run.
To update its state, a node uses the latest available states of its neighbors. This is
an asynchronous regime with bounded delay (see Section 2) as defined in Tanenbaum
and Steen [2010]. Such distributed state updates can provide system randomization in
time.

2.3. Initial Configuration of Node States


The random state values si [0] ∈ {−1, 1} assigned  to all nodes at time t = 0
form the initial configuration I. The sum ρ = i si [0] is the initial density, where
ρ ∈ {−N, . . . , N} [Mitchell et al. 1994]. We initiate all nodes using a flip-coin proce-
dure [Moreira et al. 2004], that is, each node obtains its state by a Bernoulli trial in
which each of the two states si ∈ {−1, 1} is chosen with equal probability. Hence, the
initial density ρ approximates the shape of a Gaussian distribution.

3. MODELING ASSUMPTIONS
3.1. Network Topology
To compare the convergence rates of different BMC algorithms, we simulate these algo-
rithms in three commonly used types of networks: ring lattices [Gacs et al. 1978], Watts-
Strogatz graphs [Watts and Strogatz 1998], and Waxman graphs [Waxman 1988]. These
graphs are models for real-world networks, such as social [Török et al. 2013; Cao et al.
2010; Leonard et al. 2012] and technical systems [Mostofi and Murray 2009]. The three
networks are generated as follows:
(1) Ring Lattices. A ring lattice is initially created as a cellular automaton, in which
each node is connected to K following nodes. Such an automaton is then closed to

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:6 A. Gogolev et al.

a ring to avoid boundary effects. This yields a regular ring lattice in which each
node has |Ni | = 2K neighbors. One- and two-dimensional cellular automata models
are widely used in consensus studies [Tanaka-Yamawaki et al. 1996; Wolz and De
Oliveira 2008; Marques-Pita and Rocha 2008].
(2) Watts-Strogatz Graphs. A Watts-Strogatz (WS) graph can represent networks that
incorporate the small-world phenomenon and the properties of scale-free topolo-
gies [Barabási and Reka 1999]. It is often employed to model social and natu-
ral networks [Watts and Strogatz 1998; Yang 2001]. Initially, a WS graph is cre-
ated as a ring lattice of N nodes, each having 2K neighbors. Next, every link
of node i is rewired with rewiring probability P, that is, it is deleted and reas-
signed to a random node j ∈ / {i − K, . . . , i + K}. Varying P from 0 to 1 results in
the transition from a regular ring lattice to a random graph: At P = 0, the net-
work is a regular 2K-connected ring lattice. At P ≥ 0, it can be represented as a
small-world with N K P shortcuts (further details can be found in Barabási and
Reka [1999]). At P = 1, all links are random and the network is a pure random
graph.
(3) Waxman Graphs. Graphs proposed by Waxman [1988] are used to model human-
designed networks, such as the Internet [Van Mieghem 2001]. A graph is built as
follows. First, for each pair of nodes i, j ∈ {1, 2, . . . , N}, i = j, the distance d is
randomly uniformly chosen from the interval (0, 1]. Next, each node pair is linked
with probability

d
α exp − , (5)
β
with parameters α, β ∈ (0, 1]. An increasing α yields an increasing link probability,
thus increasing the average node degree. An increasing β increases the number of
long links compared to short links, thus increasing the average link length in the
network. The described model with d ∈ (0, 1] is often referred to as the Waxman
network of the second type. Such networks are easy to model and still capture
the important features of real-life random networks. For comparison purposes,
we model sparsely connected Waxman graphs with fixed α = 0.05 and varying
β ∈ [0.01, 0.4]. Within the given parameter range of β, we limit the average node
degree and average link length to match the WS model.

The reasons for having chosen these types of network topologies are as follows. First,
the use of ring lattices in the study of BMC stems from the work by Gacs et al. [1978]
(with N = 149 nodes, each with 2K = 6 neighbors) and since then has been used by
scholars in performance studies of consensus algorithms. It is like a reference setup.
Second, the use of WS networks enables us to study the behavior in networks spanning
from ring lattices to fully random graphs, and thus assess the impact of the increasing
topology randomness. Third, Waxman graphs are random geometric networks in which
the link probability depends on the distance between nodes. This characteristic appears
in many real-world networks, for example, wireless networks. In summary, the three
topologies enable us to compare the convergence rate of BMC algorithms over a variety
of topology characteristics.
Note that GKL explicitly states a minimum required network connectivity of K ≥ 3
neighbors for each node. In our simulations, we thus artificially limit the connectivity
of the simulated networks at K ≤ 3 for all algorithms but GKL. This allows us to
exhibit whether randomization can be beneficial in conditions that are similar or worse
than those of GKL and corresponding reference setup (ring lattice of N = 149 and
K = 3).

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:7

3.2. Disturbances
3.2.1. Faulty Nodes. After initial network configuration, M faulty nodes are added to
the N normal nodes. These faulty nodes are assigned a state value sM opposite to the
initial majority of state values of the normal nodes: if ρ ≤ 0, sM = 1 and if ρ ≥ 0,
sM = −1. The network topology is then created with N + M nodes. All faulty nodes
broadcast their state values to neighbors but never change their own states.1 A network
is considered to be converged if all N nonfaulty nodes reach an agreement within T
steps.
3.2.2. Additive Noise. Each received state information can be disturbed by a random
noise value φ j . We focus here on additive noise by applying the replacement
si, j [t] → si, j [t] + φ j (6)
in Equations (1), (3), and (4). As an example, the new state of a node running RM in a
network with noise is given by:
⎛ ⎞

C
si [t + 1] = G ⎝si,i [t] + φi + (si,n j [t] + φ j )⎠. (7)
j=1

We consider two types of noise: Additive White Gaussian Noise (AWGN), φ j ∼ N (0, ( 3A )2 )
and Additive White Uniform Noise (AWUN), φ j ∼ U(−A, A), with the magnitude A.
Previous studies on the influence of additive noise on consensus [Ruan and Mostofi
2008; Huang and Manton 2007] mostly consider AWGN as the most common noise type
in real environments. AWUN is generally used to model frequency response of some
amplifiers and distributed filters [Vladimirov and Diamond 2002].
3.2.3. Message Loss. A loss of messages with state information can severely decrease
the performance of the studied algorithms, since a node decision is based  on less
state information. If a message is lost, a node can come to a state when Ni si, j = 0
and the state of the node stays unchanged. In our model, a message is lost with the
message loss probability Ei, j ∈ [0, 1):

0, with probability Ei, j
si, j [t] = . (8)
si, j [t], with probability 1 − Ei, j
To study the impact of noise and message loss, we apply them both independently
from each other and jointly. A node can receive a state message corrupted by noise
or, if the message is lost, only the noise sample. Such randomization also affects a
node’s own state, that is, si = si,i in general. This modeling assumption is motivated
by distributed detection systems, in which decisions are based on unreliable (noisy or
missing) sensory data, and thus, the sensing unit can be decoupled from the decision-
making unit [Galov and Moschevikin 2013; Moschevikin et al. 2013]. Here, the noise
and message loss represent the inevitable impact of the environment on the sensed data
rather than the impact of the internode communication (which often can be mitigated
by noise cancellation or robust coding techniques).

1 This model differs from the Byzantine failure [Fischer et al. 1985; Pease et al. 1980; Aspnes 2003], where
faulty nodes may have arbitrary failures and thus may occasionally promote consensus. In our model, faulty
nodes persistently counter consensus by spreading the information opposite to the initial majority. Even the
best algorithms for binary majority consensus (e.g., GKL) do not reach 100% convergence rate and indicate
little tolerance towards noise and errors [Gacs et al. 1978; Moreira et al. 2004]. This can explain why there
were no studies on the impact of the faulty nodes on BMC. In this article, we introduce faulty nodes to study
the robustness of the gains in convergence rate induced by randomization. We chose persistent faulty nodes
since they hinder BMC stronger than Byzantine faulty nodes [Gogolev and Marcenaro 2014a, 2014b].

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:8 A. Gogolev et al.

Table I. Overview of Consensus Rules for Density Classification, N = 149


Test
Rule Year Design Reference R,% ρ Samples
GKL 1978 Human- Gacs et al. [1978] 81.6 Gaussian 104
designed
Davis 1995 Human- Andre et al. [1996] 81.8 Gaussian 104
designed
Das 1995 Human- Das et al. [1995] 82.2 Gaussian 104
designed
Koza 1996 Genetic Andre et al. [1996] 82.3 Gaussian 104
programming
Juille and 1998 Co-evolution Juille and Pollack [1998] 86.0 Gaussian 104
Pollack
SM 2004 Randomized Moreira et al. [2004] 85.0 Gaussian 103
by errors
Traffic 2013 Randomized Fatés [2013] 90.0 Biased 104
Majority switching

4. PERFORMANCE EVALUATION
4.1. Performance Measure
We study the performance of consensus algorithms in terms of their convergence rate R,
that is, the fraction of initial configurations resulting in successful agreement. Table I
summarizes the convergence rates of state-of-the-art approaches.
Note that R can vary significantly depending on N, K, and the stochastic distribution
of ρ. Care must be taken to simulate enough initial configurations, since test sets with
fewer than 10 000 initial configurations can result in a deviation of several percent in
convergence rate. Our simulations reveal that test sets obtained by a flip-coin procedure
tend to result in lower R for both GKL and SM compared to test sets with uniform ρ
distribution used by some scholars [Mitchell et al. 1993]. This article uses the flip-coin
procedure, as we believe that it is a more intuitive way to assign states to nodes.
Randomized algorithms show a high convergence rate. Thus, SM indicates R
85% being randomized by errors and topology, compared to R below 1% registered
without randomization [Moreira et al. 2004]. A randomized Traffic Majority rule yields
a convergence rate up to 90% [Fatés 2013], which is due to a random switching between
Traffic and Majority rules.

4.2. Simulation Methodology and Parameters


The simulator is built in C++ using Boost libraries (www.boost.org). It generates
N × 300 000 binary state values by coin flipping, leading to 300 000 initial configu-
rations. Three random networks are produced. For a given network and a given set of
algorithm parameters, we simulate a set of 10 000 initial configurations to obtain the
convergence rate R of this setup. This experiment is performed over 30 different sets
of configurations for a given network. The resulting 3 × 30 samples for R are finally
averaged by means of NumPy Python libraries (www.numpy.org). These average values
of R are plotted with 95% probability confidence intervals.
We vary the rewiring probability P ∈ [0, 1], noise magnitude A ∈ [0, 4], and message
loss probability E ∈ [0, 1). Waxman graphs are considered with α = 0.05 and β ∈
[0.01, 0.4].
The system is regarded for T = 2N steps as in Gacs et al. [1978], which is a stricter
condition than N 2 used in some cases described in Moreira et al. [2004]. The system is

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:9

Fig. 2. Convergence rate of SM and GKL in ring lattices (K = 3) of different sizes. No noise, no message
loss, no faulty nodes.

considered to be converged if all nodes have the same state at the end of a simulation
and this state corresponds to the initial majority (majority agreement is reached).2
We analyze the impact of the following on the convergence rates of GKL, SM, and
the introduced RM:
—network size (growing N with fixed K)
—update mode (as randomized/asynchronous state updates)
—number of neighbors (growing K with fixed N)
—topology randomization (growing P and β)
—additive noise (randomizing the information exchange with growing A)
—message loss (randomizing the information exchange with growing E)
—faulty nodes (flooding the network with faulty information)

4.3. Convergence of Standard Algorithms


Let us first study the performance of SM and GKL consensus with and without noise,
message loss, and faulty nodes in lattices, WS graphs, and Waxman graphs.
4.3.1. Reference Scenario (Ring Lattice). We consider regular ring lattices with N ∈
{29, . . . , 149} nodes, where each node has K = 3 neighbors on each of both sides.
Figure 2 indicates that the convergence rate R decreases with an increasing number
of nodes N in all algorithms. We explain this behavior by the fact that systems with
a larger K/N ratio are more prone to clustering [Olfati-Saber and Shamma 2005].
This effect is more evident for SM. A lower R for asynchronous GKL compared to
synchronous GKL, as expected, can be explained by the fact that asynchrony random-
izes state updates and inhibits a direction bias of GKL. Randomization by asynchronous

2 In networks with disturbances, the definition of convergence in this article is stricter (not necessarily better)
than that of Moreira et al. [2004]. That article assumes that consensus “is completed successfully when, at
the end of the evolution, any deviation from the correct classification is caused by these fluctuations, meaning
that if the noise were ‘turned off ’ at that moment, all the units would converge to the correct state in the
next time step.”

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:10 A. Gogolev et al.

Fig. 3. Convergence rate of asynchronous SM in Watts-Strogatz networks with N = 99 nodes, each node
with K neighbors.

state updates can provide temporal direction bias for SM, similar to that of GKL, which
can promote consensus. It is shown in Gacs et al. [1978] and Moreira et al. [2004] that
synchronous GKL decreases performance with noise and topology randomization.
4.3.2. Impact of Topology Randomization (Watts-Strogatz Networks). Figure 3(a) indicates
that randomization of the network topology—that is, starting from a ring lattice that
corresponds to a WS graph with P = 0, then increasing P—improves the convergence
rate of SM for all K ∈ {2, 3, 4}. Growth of P influences the topology of a WS graph in
the following way: it leaves the average node degree intact, while more nodes obtain
long-range links, increasing the average link length. Random long-range links dither
topological clusters [Watts and Strogatz 1998; Barabási and Reka 1999]. We conjecture
that this effect promotes SM consensus. In summary, probabilistic rewiring, even with
few neighbors K, counters clustering and thus promotes SM consensus.
4.3.3. Impact of Noise. Figure 3(b) illustrates that a moderate amplitude A of addi-
tive noise (AWUN and AWGN) improves the convergence of SM compared to no noise
(A = 0). This beneficial effect occurs as noise randomizes information exchange and
destabilizes clusters of nodes with equal states. If, however, the noise amplitude is too
high, consensus will become impossible, for example, if A > 2.7 for AWUN with K = 4.
Figure 4 presents the combined impact of topology randomization and noise for asyn-
chronous SM and GKL. It is possible for both algorithms to superimpose the positive
effects of topology randomization and noise. An optimal combination of disturbances
and their levels can be selected, depending on the system prerequisites. This observa-
tion extends the previous results showing that asynchronous GKL can be beneficially
randomized by errors and topology (while synchronous GKL indicates the opposite
effect) [Moreira et al. 2004; Gacs et al. 1978].
4.3.4. Impact of Faulty Nodes. Figure 5(a) presents the impact of faulty nodes and topol-
ogy randomization on convergence. Figure 5(b) shows the impact of faulty nodes and
noise in pure random networks. Faulty nodes have, in general, a negative influence on
the convergence. Part of this negative impact can be mitigated by topology randomiza-
tion and noise.
4.3.5. Impact of Message Loss. Figure 6 indicates that probabilistic message loss can
positively or negatively influence the convergence rate in noiseless networks. The
convergence rate of GKL degrades by growing E (E ≥ 4%), whereas the convergence
rate of SM improves up to a certain value of E (e.g., up to E = 80% for ring lattices).

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:11

Fig. 4. Convergence rate of asynchronous SM and GKL in Watts-Strogatz networks (N = 99, K = 3) with
AWGN and no faulty nodes.

Fig. 5. Convergence rate of asynchronous SM in Watts-Strogatz networks (N = 99, K = 3) with AWGN and
faulty nodes.

Fig. 6. Convergence rate of asynchronous GKL and SM in Watts-Strogatz networks (N = 99, K = 3) with
message loss. No faulty nodes.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:12 A. Gogolev et al.

Fig. 7. Convergence rate of asynchronous SM in Waxman networks (N = 99, K = 3) with AWGN and
message loss without faulty nodes.

Fig. 8. Convergence rate of asynchronous RM and SM in noiseless networks ( A = 0) with N = 99 nodes each
with six node neighbors (K = 3). Solid line is lattice (WS with P = 0); dashed line is small-world networks
(WS with P = 0.48). RM with C = 2 randomly selected neighbors.

The impact of message loss is twofold: it randomizes and reduces the information ex-
change at the same time. Thus, using SM, message loss promotes consensus until its
positive randomizing effect is outweighed by the negative effect of information loss.
The optimum value of E depends on the algorithm and network parameters.
4.3.6. Waxman Networks. For SM in Waxman networks, Figure 7 presents the combined
influence of the Waxman parameter β and message loss. Figure 7(a) shows that SM in
noiseless networks with β ≥ 0.25 reaches R ≥ 50% and can tolerate high message loss.
With α = 0.05 and β = 0.01, a Waxman graph is not guaranteed to be connected [Waxman
1988]. This can explain low convergence rate at β ≤ 0.25. Figure 7(b), in comparison
Figure 7(a), illustrates that noise can promote consensus in Waxman networks for low
β and few message losses, say, β < 1/4 and E < 1/2 in this example.

4.4. Convergence of Random Neighbors Majority


Let us now study the RM algorithm proposed earlier in comparison with the SM. Recall
that RM receives state information from C randomly selected neighbors that are chosen
anew in each time step.
Figure 8 plots the convergence rates in noiseless lattices and small-world networks.
Figure 8(a) presents simulation results without faulty nodes. It illustrates that RM

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:13

Fig. 9. Convergence rate of asynchronous RM with C ∈ {2, 4} in ring lattices and random networks (N = 99,
K = 3) with noise (AWGN) and message loss.

outperforms SM if the message loss is low. This indicates that randomizing the neigh-
bor selection dynamically might be beneficial despite the fact that only two nodes
provide state information. The bad convergence rate of RM compared to SM for more
frequent message loss can be explained by a lower number of state information mes-
sages received by RM (two messages) compared to SM (six messages). Figure 8(b)
presents similar results: with faulty nodes, RM outperforms SM in small-world net-
works up to a certain message loss. In lattice networks, however, even a small number
of faulty nodes decreases the convergence rate of both algorithms to a very low value.
We conjecture that these observations can be generalized as follows: randomization
of neighbor selection embedded into consensus algorithms can promote consensus in
both lattice and random networks, and increase fault tolerance in random networks.
Let us now study RM in scenarios with AWGN. Figure 9 indicates that consensus
randomized by both neighbor selection and noise can reach convergence rates close
to 100% in both lattices and random networks for certain parameters, outperforming
other algorithms for low to medium message-loss rates. Also, note that increasing the
number of selected neighbors C from 2 to 4 increases robustness towards message loss.
This latter effect can be explained by the fact that message loss acts as a randomizing
agent (positive impact) and inhibits the information exchange (negative impact). This
means that, on one hand, the information loss induced by message loss can be countered
with a higher initial number of received messages C. On the other hand, if message loss
is low and its beneficial randomizing impact is weak, it can be enhanced with a higher
level of added noise A. A convergence rate of almost 100% may seem contradicting
to the restriction by Land and Belew [1995], who show that deterministic algorithms
for binary consensus cannot reach R = 100%, but this restriction does not apply to
randomized algorithms.
Finally, we study RM in Waxman networks with faults. As shown in Figure 10(a), RM
again outperforms SM for low to medium E. Figure 10(b) presents the RM convergence
rate as a function of the Waxman parameters β and E. Comparing this performance
with that of SM in Waxman networks (see Figure 7) indicates that asynchronous RM
can show better convergence rates and fault tolerance than SM for all parameters in
this setting.

4.5. Comparison to Related Work


Table II gives an overview of some studies on BMC with randomization and faulty
nodes in relation to the contributions of this article. Based on this article, the focus of
Gogolev and Marcenaro [2014b] is a deeper investigation of the impact of faulty nodes
on convergence rate in SM and GKL.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:14 A. Gogolev et al.

Fig. 10. Convergence rate of asynchronous RM with C = 4 in Waxman networks (N = 99, K = 3) with noise
(AWGN) and message loss.

Table II. Binary Majority Consensus with Disturbances


Disturbance by
Rule Noise Errors Msg. loss Faulty nodes Topology Reference
GKL yes no no — Ring lattice Gacs et al. [1978]
GKL, SM no yes no — WS Moreira et al. [2004]
TM no no no — Ring lattice Fatés [2013]
GKL, SM yes no yes persistent WS, Waxman Section 4.3
GKL, SM yes no yes persistent, Byzantine WS, Waxman Gogolev et al. [2014a]
RM yes no yes persistent WS, Waxman Section 4.4

5. CONCLUSIONS AND OUTLOOK


This article studied the impact of different kinds of “disturbances” in distributed bi-
nary majority consensus. We have seen that GKL, SM, and the newly proposed RM
algorithm exhibit, in general, different reactions to disturbances. Explicit or implicit
randomization with topology, message loss, and stochastic noise can improve the con-
vergence rate of the studied algorithms, and all these means have a cumulative effect
when combined. In each of the studied setups, an optimal combination of randomization
exists that improves the convergence rate.
If noise and message loss are too low in practical implementations, they may also be
artificially integrated to the consensus algorithms to improve their performance. The
proposed RM rule embeds randomization by neighbor selection. It shows a convergence
rate up to 100% in certain cases.
The presence of persistent faulty nodes degrades the convergence rate. Algorithms
with randomization indicate better tolerance toward the influence of faulty nodes in
most of the studied setups. However, even optimally randomized algorithms can toler-
ate the influence of only a few faulty nodes.
Future work could exploit the insights of this article for the design of adaptive
algorithms that automatically adjust their parameters during operation to improve
performance. For example, the number of randomly selected neighbors in RM can be
adjusted, and different nodes may, in general, have a different number of neighbors
for state information exchange. Furthermore, while this article studied the start-up
phase of binary consensus until agreement is reached, future work can analyze the
robustness against disturbance in the stationary state.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:15

ACKNOWLEDGMENTS
This work was performed at University of Klagenfurt and University of Genoa within the Erasmus Mundus
Joint Doctorate in “Interactive and Cognitive Environments,” which is funded by the EACEA Agency of the
EC under EMJD ICE FPA n 2010-0012. The work of A. Gogolev is supported by Lakeside Labs, Klagenfurt,
with funding from the ERDF, KWF, and the state of Austria under grant 20214/21530/32606.

REFERENCES
Marcos K. Aguilera and Sam Toueg. 2012. The correctness proof of Ben-Ors randomized consensus algorithm.
Distributed Computing 25, 5, 371–381.
Mehdi Alighanbari and Jonathan P. How. 2006. An unbiased Kalman consensus algorithm. In Proceedings
of American Control Conference. 3519–3524.
David Andre, Forrest H. Bennett III, and John R. Koza. 1996. Discovery by genetic programming of a cellular
automata rule that is better than any known rule for the majority classification problem. In Proceedings
of the Annual Conference on Genetic Programming. 3–11.
James Aspnes. 2000. Fast deterministic consensus in a noisy environment. In Proceedings of the ACM
Symposium on Principles of Distributed Computing. 299–308.
James Aspnes. 2002. Fast deterministic consensus in a noisy environment. Journal of Algorithms 45, 1,
16–39.
James Aspnes. 2003. Randomized protocols for asynchronous consensus. Distributed Computing 16, 2–3,
165–175.
Albert-Laszlo Barabási and Albert Reka. 1999. Emergence of scaling in random networks. Science 286, 5439,
509–512.
Michael Ben-Or. 1983. Another advantage of free choice: Completely asynchronous agreement protocols. In
Proceedings of the ACM Symposium on Principles of Distributed Computing. 27–30.
Gabriel Bracha and Ophir Rachman. 1992. Randomized consensus in expected O(n2log n) operations. In
Distributed Algorithms. Lecture Notes in Computer Science, Vol. 579. Springer, 143–150.
Gabriel Bracha and Sam Toueg. 1985. Asynchronous consensus and broadcast protocols. Journal of the ACM
32, 4, 824–840.
Janna Burman, Ted Herman, Shay Kutten, and Boaz Patt-Shamir. 2006. Asynchronous and fully self-
stabilizing time-adaptive majority consensus. In Principles of Distributed Systems. Lecture Notes in
Computer Science, Vol. 3974. Springer, 146–160.
Ming Cao, Andrew Stewart, and Naomi E. Leonard. 2008. Integrating human and robot decision-making
dynamics with feedback: Models and convergence analysis. In 47th IEEE Conference on Decision and
Control. 1127–1132.
Ming Cao, Andrew Stewart, and Naomi E. Leonard. 2010. Convergence in human decision-making dynamics.
Systems and Control Letters 59, 2, 87–97.
Tushar Deepak Chandra, Vassos Hadzilacos, and Sam Toueg. 1996. The weakest failure detector for solving
consensus. Journal of the ACM 43, 4, 685–722.
Tushar Deepak Chandra and Sam Toueg. 1996. Unreliable failure detectors for reliable distributed systems.
Journal of the ACM 43, 2, 225–267.
Rajarshi Das, James P. Crutchfield, Melanie Mitchell, and James E. Hanson. 1995. Evolving globally syn-
chronized cellular automata. In Proceedings of the International Conference on Genetic Algorithms.
336–343.
Danny Dolev, Cynthia Dwork, and Larry Stockmeyer. 1987. On the minimal synchronism needed for dis-
tributed consensus. Journal of the ACM 34, 1, 77–97.
Cynthia Dwork, Nancy Lynch, and Larry Stockmeyer. 1988. Consensus in the presence of partial synchrony.
Journal of the ACM 35, 2, 288–323.
Nazim Fatés. 2013. Stochastic cellular automata solutions to the density classification problem. Theory of
Computing Systems 53, 2, 223–242.
Michael J. Fischer, Nancy A. Lynch, and Michael S. Paterson. 1985. Impossibility of distributed consensus
with one faulty process. Journal of the ACM 32, 2, 374–382.
Peter Gacs, Georgy L. Kurdyumov, and Leonid A. Levin. 1978. One-dimensional uniform arrays that wash
out finite islands. Problemy Peredachi Informacii 14, 92–98.
Alexander Galov and Alexey Moschevikin. 2013. Bayesian filters for ToF and RSS measurements for indoor
positioning of a mobile object. In Proceedings of 2013 International Conference on Indoor Positioning
and Indoor Navigation (IPIN). 1–8.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
19:16 A. Gogolev et al.

Andrew Gee and Roberto Cipolla. 1996. Fast visual tracking by temporal consensus. Image and Vision
Computing 14, 2, 105–114.
Carlos Gershenson. 2004. Updating schemes in random Boolean networks: Do they really matter? In Pro-
ceedings of the International Conference on the Simulation and Synthesis of Living Systems. 238–243.
Alexander Gogolev and Christian Bettstetter. 2012. Robustness of self-organizing consensus algorithms:
Initial results from a simulation-based study. In Self-Organizing Systems. Lecture Notes in Computer
Science, Vol. 7166. Springer, 104–108.
Alexander Gogolev and Lucio Marcenaro. 2014a. Density classification in asynchronous random networks
with faulty nodes. In Proceedings of International Conference on Parallel, Distributed and Network-Based
Processing. 256–261.
Alexander Gogolev and Lucio Marcenaro. 2014b. Randomized binary consensus with faulty nodes. Entropy
16, 2820–2838.
Minyi Huang and Jonathan H. Manton. 2007. Stochastic double array analysis and convergence of consensus
algorithms with noisy measurements. In Proceedings of American Control Conference. 705–710.
Hugues Juille and Jordan B. Pollack. 1998. Coevolving the “ideal” trainer: Application to the discovery of
cellular automata rules. In Proceedings of the Annual Conference on Genetic Programming. 519–527.
Soummya Kar and Jose M. F. Moura. 2007. Distributed average consensus in sensor networks with random
link failures and communication channel noise. In Proceedings of Asilomar Conference on Signals,
Systems and Computers. 676–680.
Derek B. Kingston and Randall W. Beard. 2006. Discrete-time average consensus under switching network
topologies. In Proceedings of American Control Conference. 3551–3556.
Mark Land and Richard K. Belew. 1995. No perfect two-state cellular automata for density classification
exists. Physical Review Letters E 74, 5148–5150.
Naomi E. Leonard, Tian Shen, Benjamin Nabet, Luca Scardovi, Iain D. Couzin, and Simon A. Levin. 2012.
Decision versus compromise for animal groups in motion. Proceedings of the National Academy of
Sciences 109, 1, 227–232.
Timothy Margush and Fred R. McMorris. 1981. Consensus n-trees. Bulletin of Mathematical Biology 43, 2,
239–244.
Manuel Marques-Pita and Luis Mateus Rocha. 2008. Conceptual structure in cellular automata—the density
classification task. In ALIFE. 390–397.
J. Ricardo G. Mendonçia. 2011. Sensitivity to noise and ergodicity of an assembly line of cellular automata
that classifies density. Physical Review Letters E 83, 3, 31112–31117.
Melanie Mitchell, James P. Crutchfield, and Peter T. Hraber. 1994. Evolving cellular automata to perform
computations: Mechanisms and impediments. Physica D: Nonlinear Phenomena 75, 1–3, 361–391.
Melanie Mitchell, Peter T. Hraber, and James P. Crutchfield. 1993. Revisiting the edge of chaos: Evolving
cellular automata to perform computations. Complex Systems 7, 89–130.
André A. Moreira, Abhishek Mathur, Daniel Diermeier, and Luı́s A. N. Amaral. 2004. Efficient system-wide
coordination in noisy environments. Proceedings of the National Academy of Sciences of the United States
of America 101, 33, 12085–12090.
Alexey Moschevikin, Alexander Galov, Sergey Reginya, Alexander Volkov, and Axel Sikora. 2013. The impact
of NLOS components in Time-of-Flight networks for indoor positioning systems. In Proceedings of IEEE
International Conference on Intelligent Data Acquisition and Advanced Computing Systems (IDAACS),
Vol. 01. 455–460.
Yasamin Mostofi and Richard M. Murray. 2009. Kalman filtering over wireless fading channels—How to
handle packet drop. International Journal of Robust and Nonlinear Control 19, 18, 1993–2015.
Benjamin Nabet, Naomi E. Leonard, Iain D. Couzin, and Simon A. Levin. 2009. Dynamics of decision making
in animal group motion. Journal of Nonlinear Science 19, 4, 399–435.
Reza Olfati-Saber and Jeff S. Shamma. 2005. Consensus filters for sensor networks and distributed sensor
fusion. In Proceedings of IEEE Conference on Decision and Control and European Control Conference.
6698–6703.
Marshall Pease, Robert Shostak, and Leslie Lamport. 1980. Reaching agreement in the presence of faults.
Journal of the ACM 27, 2, 228–234.
Yongxiang Ruan and Yasamin Mostofi. 2008. Binary consensus with soft information processing in coopera-
tive networks. In Proceedings of IEEE Conference on Decision and Control. 3613–3619.
Michael Saks, Nir Shavit, and Heather Woll. 1991. Optimal time randomized consensus-making resilient
algorithms fast in practice. In Proceedings of ACM-SIAM Symposium on Discrete Algorithms. 351–362.
Michael Saks and Fotios Zaharoglou. 1993. Wait-free k-set agreement is impossible: The topology of public
knowledge. In Proceedings of the ACM Symposium on Theory of Computing. 101–110.

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.
Distributed Binary Consensus in Networks with Disturbances 19:17

Andrea Simonetto, Tamas Keviczky, and Robert Babuska. 2010. Distributed nonlinear estimation for robot
localization using weighted consensus. In Proceedings of IEEE International Conference on Robotics and
Automation. 3026–3031.
Mieko Tanaka-Yamawaki, Sachiko Kitamikado, and Toshio Fukuda. 1996. Consensus formation and the
cellular automata. Robotics and Autonomous Systems 19, 1, 15–22.
Andrew S. Tanenbaum and Maarten van Steen. 2010. Distributed Systems: Principles and Paradigms (3rd
ed.). Prentice-Hall, Inc. Upper Saddle River, NJ.
János Török, Gerardo Iñiguez, Taha Yasseri, Maxi San Miguel, Kimmo Kaski, and János Kertész. 2013.
Opinions, conflicts, and consensus: Modelling social dynamics in a collaborative environment. Physical
Review Letters 110, 8, 88701–88708.
Piet Van Mieghem. 2001. Paths in the simple random graph and the Waxman graph. Probability in the
Engineering and Informational Sciences 15, 4, 535–555.
Igor G. Vladimirov and Phil Diamond. 2002. A uniform white-noise model for fixed-point roundoff errors in
digital systems. Automation and Remote Control 63, 5, 753–765.
Duncan J. Watts and Steven H. Strogatz. 1998. Collective dynamics of “small-world” networks. Nature 393,
6684, 440–442.
Bernard M. Waxman. 1988. Routing of multipoint connections. IEEE Journal on Selected Areas in Commu-
nications 6, 9, 1617–1622.
Dietmar Wolz and Pedro De Oliveira. 2008. Very effective evolutionary techniques for searching cellular
automata rule spaces. Journal of Cellular Automata 3, 4, 289–312.
Xin-She Yang. 2001. Small-world networks in geophysics. Geophysical Research Letters 28, 13, 2549–2552.

Received November 2013; revised December 2014; accepted March 2015

ACM Transactions on Autonomous and Adaptive Systems, Vol. 10, No. 3, Article 19, Publication date: August 2015.

You might also like