2012 IEEE Wireless Communications and Networking Conference: PHY and Fundamentals

An Adaptive Learning Approach for Fault-Tolerant

Routing in Internet of Things
Sudip Misra P. Venkata Krishna, Harshit Agarwal
School of Information Tech School of Computing Science and Engineering
Indian Institute of Technology VIT University
Kharagpur, India Vellore, India
[email protected] [email protected], [email protected]

Anshima Gupta Mohammad S. Obaidat

School of Information Technology and Engineering Fellow of IEEE & Fellow of SCS
VIT University Dept. of Computer Science
Vellore, India Monmouth University, NJ, USA
[email protected] [email protected]

Abstract—Internet of Things (IOT) is a wireless ad-hoc network Indubitably, IOT is envisioned to have huge influence on a
of everyday objects collaborating and cooperating with one other variety of aspects of our everyday life [1]. It is likely to emerge
in order to accomplish some shared objectives. The envisioned as a new technology capable of playing a prominent role in a
high degrees of association of humans with IOT nodes require wide range of applications ranging from basic domestic
equally high degrees of reliability of the network. In order to assistance to intelligent automated industrial systems. From the
render this reliability to IOT networks, it is necessary to make bird‟s eye view, IOT is a self-configuring wireless network
them tolerant to faults. In this paper, we propose mixed cross- which integrates the physical world with the world of Internet
layered and learning automata (LA)-based fault-tolerant routing [2]. It has the potential to equip humans with the ability to have
protocol for IOTs, which assures successful delivery of packets high degrees of control over physical objects. It will enable
even in the presence of faults between a pair of source and centralized unified control which is extended to almost every
destination nodes. As this work concerns IOT, the algorithm
object in the proximity.
designed should be highly scalable and should be able to deliver
high degrees of performance in a heterogeneous environment. The huge interconnection requirement between the nodes in
The LA and cross-layer concepts adopted in the proposed IOT will require the enabling of huge centralized/decentralized
approach endow this flexibility to the algorithm so that the same database of objects across the globe. It will require a unique
standard can be used across the network. It dynamically adopts addressing system so that each device can be uniquely
itself to the changing environment and, hence, chooses the identified. IOT draws lot of similarity with the current day
optimal action. Since energy is a major concern in IOTs, the Internet, but it has significantly high level of diversity and
algorithm performs energy-aware fault-tolerant routing. To save
enormous number of communicating devices. IOT devices
on energy, all the nodes lying in the unused path are put to sleep.
share information about their states and other vital factors
Again this sleep scheduling is dynamic and adaptive. The
simulation results of the proposed strategy shows an increase in
which imparts them with essential, though unique,
the overall energy-efficiency of the network and decrease in characteristics such as self-management, self-decision making
overhead, as compared to the existing protocols we have and self-governance [3]. The aspects of cooperation endows
considered as benchmarks in this study. them autonomy and intelligence.
An IOT is conceived to be open and self-assimilating
Keywords-IOT; Cross-Layer Design; Learning Automata; internetwork. These kinds of internetworks are highly prone to
Fault-tolerant routing; faults and security threats [4]. These faults, if not handled
properly, may lead to serious network downtime. Faults tend to
I. INTRODUCTION degrade the network performance and affect the network‟s
IOT refers to a nexus of customary and conventional operation time by introducing unnecessary overheads. Faults
objects connected in the form of an internetwork. The primary can occur due to a variety of factors. To generalize, they can be
concept behind this new technology is the pervasive presence classified into two major categories namely hardware-based
of these objects such as RFID around us interacting and and software-based. The more critical the application of
collaborating with each other to attain mutual goals. Of course network would be, the more severe would be the implications
in order to communicate with one another, the network-enabled because of such faults. In case of IOTs, due to their large scale,
objects need common protocols and standards for these faults have magnified repercussions. To make IOT a
communication. robust, reliable and dependent technology, it is necessary to
adopt a strategy to avoid and counter these faults.

U.S. Government work not protected by U.S. copyright 815

 In this paper, we focus on the fault-tolerance aspects in III. LEARNING AUTOMATA
routing. We propose a learning automaton (LA) [5-7] based The theory of LA centers on the notion of an “automaton,”
intelligent fault-tolerant routing algorithm for IOT. We which is a self-operating machine or a mechanism that
introduce the concept of cross-layering [8-9] to optimize the responds to a sequence of instructions in a certain way, so as
energy saving while handling the faults at the same time. LA is to achieve a certain goal. The automaton either responds to a
an intelligent and adaptive approach which takes decision pre-determined set of rules, or adapts to the environmental
based on the feedback from the environment. It reiterates this dynamics in which it operates. The term “learning” refers to
process to choose the most optimized action. This paper is the action of procuring knowledge and modifying one‟s
inspired from the work presented in [5], where Misra et al. behavior based on the experience earned. Thus, the learning
proposed an LA-based fault-tolerant routing algorithm for automata adapt to the responses from the environment through
mobile ad hoc networks (MANETs). We have modified the a series of interactions within them. The automata, then,
existing approach to meet the specifications of IOT-like attempt to learn the best action from a set of possible actions
networks and tested it by simulating it using ns-2. that are offered to them by the random stationary or non-
stationary environment in which they operate. The automata,
II. MOTIVATION thus, act as decision makers to arrive at the best action.
Our literature survey shows that the fault-tolerance aspects
on routing in IOTs require serious attention. As discussed in
Section I, an IOT consists of huge number of devices, most of
which are enabled by RFID or other similar type of devices.
Such kinds of objects have low computational capabilities due
to limited resources and is highly prone to both software and
hardware faults. For this type of network to be functional, it is
essential to have some fault-countering strategy in place. Not
only in respect of computational power, are these devices also
constrained in terms of energy. Therefore, a fault-tolerant
approach which considers energy as the important factor while
taking routing decisions is a crucial necessity. To handle the Fig. 1 The learning automaton
complexities of IOT it is desired to have an intelligent and
adaptive solution. LA can be used in optimization problems, since an
automaton in LA selects that action which is more likely to be
The major challenge involved in designing network awarded by the environment. Over a period of time, LA learns
protocol for IOT is the enforcement of same standards for a from its actions and chooses an optimal solution. A
variety of devices of which IOT is composed of. As these comprehensive overview of LA can be found in the classic
devices vary primarily in their capabilities, following a text by Narendra and Thathachar [6] and in the recent book
common standard for all is likely to cause considerable chapter by Oommen and Misra [7].
performance degradation. But having different communication A. The Automaton
standard for different set of devices is impractical and
unfeasible as well. Firstly, defining so many different The automaton can be represented as a quintuple
standards is unrealistic and, secondly, achieving flawless represented as {Q, A, B, F, H}, where [10]:
interoperability among these protocols is unattainable. So, in  Q is the finite set of internal states Q = {q1, q2, q3… qn}
order to tackle these issues, we devised an approach which is where qn is the state of the automaton at instant n.
adaptive in nature.
 A is a finite set of actions performed by the automaton. A
We employ LA in the proposed solution which constrains = {α1, α2…αn} where αn is the action performed by the
all the devices to follow the same protocol and simultaneously automaton at instant n.
providing customization capabilities for every device. The
proposed algorithm adjusts various parameters based on the  B a finite set of responses from the environment. B = {β1,
environment in which the nodes are operating. Since we are β2, β3… βn} where βn is the response from the
following the same standard, the interoperability issue is environment at an instant n.
obviated. To further optimize the performance, the cross-layer  F is a mapping function. It maps the current state and
model helps by letting the algorithm perform fault-tolerant input to the next state of the automaton. Q × B → Q.
routing while being aware of the energy of nodes lying in the
route, thereby empowering it with the competency to avoid  H is a mapping function. It maps the current state and
faults that might take place due to deprivation of energy. A response from the environment to determine the next
action to be performed.
goodness value is calculated for every possible path between a
source and destination. This value denotes the fitness of a path B. The Environment
or its suitability to carry out communication. One with the
The environment corresponds to the medium in which the
highest goodness values is preferred.
automaton functions. Mathematically, an environment can be

816
abstracted as a triple {A, B, C}. A, B, and C are defined as The vertices are the nodes in the network and the edges are the
follows [10]: wireless links in between the wireless nodes. A path is a set of
vertices connected to each other from a vertex (which can also
 A = {α1, α1,…,αn} represents a finite input set;
be source) to destination (sink). Faults can occur unpredictably
 B = {β1, β2,…,βn} is the output set of the environment; in any node in the network. We assume all links in the
and network to be bidirectional, i.e., if (vi, vi+1) → E, then (vi+1, vi)
→E also exists. Each node „v‟ has two components: a routing
 C = {c1, c2,…,cn} is a set of penalty probabilities, where
element ci Є C corresponds to an input action αi. component and an LA component. Each node‟s LA
component functions independently of others and shares
We now provide a few important definitions used in the updates through an update table maintained at the routing
field of LA. Given an action probability vector P (t) at time t, component which shares LA information through the neighbor
the average penalty is defined as [10] nodes. Apart from network layer inputs it uses the data from
M (t )  E t  | Pt   Pr t   1 | Pt  the physical layer while rewarding or penalizing the path. Fig.
3 depicts the proposed system model outline.
r
  Pr t   1 |  t     Pr t    
i 1
i i

r
  c p t .
i 1
i i
(1)
The average penalty for the “pure-chance” automation is
given by [4]
r
1
M0 
r c .
i 1
i
(2)
As t→∞ if the average penalty M(t) < M0, at least
asymptotically, the automaton is generally considered to be
better than the pure-chance automaton. E [M(t)] is given by Fig. 2. Cross-Layer Design
[10]
EM t   EE t  | Pt   E t . (3)

IV. CROSS LAYER DESIGN

The need to have an energy-aware fault-tolerant routing
motivated us to choose cross-layer design. The traditional OSI
model does not allow the interaction among the different
layers of the network stack. However, to improve performance
or to increase services, this rigid model has been challenged
by the researchers [11-12]. The conventional model fails to
serve all the requirements of wireless network. IOT generally
is composed of small devices such as RFID, which are low in
energy. So, energy limitation is a major design constraint for
any protocol in IOT. Therefore, to address this limitation,
cross-layer design has been employed, which permits the Fig. 3. System-model
access to the energy statistics. As the energy related
information is available with physical layer, this information is VI. ALGORITHM
passed on to the network layer via shared data structure. The The proposed algorithm, named as Cross-Layer-Based
fault tolerant routing algorithm stationed at network layer uses Adaptive Fault-Tolerant Routing Algorithm for IOTs (i-
this energy knowledge to take better decisions hence deliver CLAFTRA) uses multipath for the transmission of data
better performance. Fig. 2 gives the diagrammatic between a pair of communicating devices. During the data
representation of the cross-layer design component. transmission, the goodness value [5] of the various paths
available is calculated using reward/penalty scheme of LA. If
V. SYSTEM MODEL the goodness value of the current path is above the threshold,
then it continues the transmission using the same path
We represent a wireless network using a graph W = (V, E), otherwise the path with highest goodness value is chosen for
where V represents the set of vertices and E the set of edges. further transmission of packets. The remaining nodes that are

817
not lying on any currently being used path are put to sleep to VII. SIMULATIONS
save energy. The goodness value of the path is continuously We have performed simulations of the proposed solution,
and dynamically updated using the underlying reward/penalty i-CLAFTRA, using ns-2 [14], and have compared its
scheme by LA. performance with the corresponding algorithms in ENFAT-
AODV [13] and AODV [15]. The parameters used for ns-2
A. Reward/Penalty Scheme simulation are specified in Table 2.
As described earlier, each node has a learning automaton
stationed in it. Therefore, the automaton keeps a check on the TABLE 1: Simulation parameters
delivery of the packet sent through it. If the packet delivery is
reported to be successful then the stationed LA increments the Parameter Value
value of the goodness value of that particular node. The Number of Nodes 40-240
Speed 0-15 m/s
increment is the sum of a constant R, the reward constant and MAC IEEE 802.11
a function of energy remaining level of the node. The node Traffic type CBR (at TCP-IP
with higher level of energy remaining will get the higher interface)
reward as it is better to select the nodes with the more energy Terrain Dimensions 1200 m x 1000 m
as they will deliver better operation time.
A. Variation in average energy consumption with respect to
To calculate the goodness value of the path, we take the mobility in the network
commulative sum of the goodness values of the nodes lying in
that path. Once the computation of the goodness value is The aim of this experiment was to study the energy
completed, it is then compared with the threshold and consumption by the network with varying percentage of
accordingly the most suitable path is selected. The nodes of mobile nodes.
the remaining unused path areput to sleep. This increases the
network lifetime. Fig. 4 shows the variation in average energy consumption
of i-CLAFTRA, AODV and ENFAT-AODV. It is observed
B. Sleep/Wake up Scheme from the graph that energy consumption by i-CLAFTRA is
less than other protocols. i-CLAFTRA performs well because
As soon as a fault is detected in the currently used path,
of the ability of learning automata adaptability and its sleep
then the path is switched to the alternative path with the next
mechanism. Energy consumption in AODV remains constant
highest path goodness value. Other than this, if the goodness
as all nodes remain active and dissipate nearly same amount of
value of a path falls by 10%, then the source node will check
energy. The average energy consumption by ENFAT-AODV
if the alternate path is good enough (i.e., has goodness value
increases as number of nodes increases as it has to spend more
higher than the currently used path) for data transmission, then
energy in maintaining the alternate path.
it will use alternate path. If the goodness value of any of the
alternate paths is not more than current path, then in such a
case, it will continue using the current path.
15.1
In the given strategy, the nodes undergo sleep and wake-
up, depending upon the goodness value of path which is
Average Engery Consumption (mW)

different from S-MAC [13]. The higher the goodness values of

AODV
a path, the more number of nodes will be active in that path, 10.1 ENFAT-AODV
resulting in the path‟s greater suitability and higher reliability i-CLAFTRA
in transferring data. Since the proposed scheme dynamically
controls the sleep scheduling of the nodes, this algorithm
results in reduced overhead and less energy consumption. 5.1

The algorithm is shown by the pseudo-code below:

Algorithm: i-CLAFTRA
0.1
0 10 20 30 40 50 60 70 80 90 100
1. Initialize the LA parameters. Percentage of Mobility
2. Using route discovery find path between source and
destination. Fig. 4 Graph for average energy consumption versus percentage of mobility
3. Find optimal path between source and destination. A. Variation in packet delivery ratio with respect to
4. Put remaining nodes to sleep. percentage of mobility in the network
5. Packet delivery across the node----reward/penalize by
altering the goodness value of the path. In this experiment, we examined the packet delivery ratio
6. Calculate the updated goodness value of node. while varying the percentage of mobility in the network.
7. For every 10% drop in goodness value. Fig. 5 shows the variation in packet delivery ratio with
8. If(goodness value of current path> goodness value of respect to percentage of mobility where pause time was kept
alternate path)----use current path constant at 300 sec. As depicted by the graph, the performance
9. else---switch to alternate path of the i-CLAFTRA is significantly better than its counterparts.

818
i-CLAFTRA‟s packet delivery ratio is affected by a very ACKNOWLEDGEMENT
minor factor as the percentage of mobility increases. By The work of the first author was supported in part by the
percentage of mobility, we imply the fraction of mobile nodes
Council for Scientific and Industrial Research (CSIR), New
of the total nodes in a network. As we know, IOT can have
both mobile as well as stationary nodes at the same time, we Delhi, India, Grant Ref. No. 22(0477)/09/EMR-II.
have taken this parameter as one of the criteria for analyzing i- REFERENCES
CLAFTRA‟s accomplishment. The increasing gap between
the curve of i-CLAFTRA and other two curves shows the
stability in the execution the proposed protocol has as [1] L. Atzori, A. Iera and G. Morabito, “The Internet of Things: A survey”,
Computer Networks, Vol. 54, No. 15, October 2010, pp. 2787-2805,
compared to AODV and ENFAT-AODV. ISSN 1389-1286, DOI: 10.1016/j.comnet.2010.05.010.
[2] O. Zhu, R. Wang; Q. Chen, Y. Liu and W. Qin, "IOT Gateway:
110
BridgingWireless Sensor Networks into Internet of Things," Embedded
and Ubiquitous Computing (EUC), 2010 IEEE/IFIP 8th International
100 Conference on , vol., no., pp.347-352, 11-13 Dec. 2010
doi: 10.1109/EUC.2010.58.
[3] ITU Internet Reports, The Internet of Things, November 2005.
90
Packet Delivery Ratio

[4] S. Misra, P. V. Krishna, H. Agarwal, A. Saxena and M. S. Obaidat, “A

Learning Automata Based Solution for Preventing Distributed Denial of
80 Service in Internet of Things”, Proceedings of the International
i-CLAFTRA Conference on Internet of Things (iThings 2011), Dalian, China,
AODV October 2011.
70 ENFAT-AODV
[5] S. Misra, P. V. Krishna, A. Bhiwal, A. S. Chawla, B. Wolfinger, and
C.Lee, “A learning automata-based fault-tolerant routing algorithm for
60 mobile ad hoc networks”, Journal of Supercomputing, 2011-07-05,
Springer Netherlands, Doi: 10.1007/s11227-011-0639-8.
[6] K. S. Narendra and M. A. L. Thathachar, Learning Automata, Prentice-
50 Hall, 1989.
0 10 20 30 40 50 60 70 80
Percentage of Mobility [7] M. A. L. Thathachar and P. S. Sastry, “Networks of Learning
Automata”, Kluwer Academic, 2003.
Fig. 5 Graph for percentage mobility versus packet delivery ratio in the [8] V. Srivastava and M. Motani, "Cross-layer design: a survey and the road
network ahead," IEEE Communications Magazine, vol.43, no.12, pp. 112-119,
Dec. 2005, doi: 10.1109/MCOM.2005.1561928.
[9] M. Conti, G. Maselli, G. Turi, S. Giordano, "Cross-Layering in Mobile
VIII. CONCLUSION Ad Hoc Network Design," Computer, vol. 37, no. 2, pp. 48-51, Feb.
The proposed protocol is a fault-tolerant routing protocol 2004, doi:10.1109/MC.2004.1266295.
for IOT. The protocol has been designed using the LA and [10] B. J. Oommen and S. Misra, "A Fault-Tolerant Routing Algorithm for
Mobile Ad Hoc Networks Using a Stochastic Learning-Based Weak
cross-layer design concepts. It has been designed to operate in Estimation Procedure," IEEE International Conference on Wireless and
the IOT like environment where the diversity of devices and Mobile Computing, Networking and Communications, 2006
the figure of devices are huge. The energy saving strategy (WiMob'2006)., pp. 31-37, 19-21 June 2006, doi:
ensures the longer operational lifetime for the network. Cross- 10.1109/WIMOB.2006.1696374
layer design not only equips this algorithm with the ability to [11] S. Shakkottai, T. S. Rappaport and P. C. Karlsson, "Cross-layer design
for wireless networks", IEEE Communications Magazine, vol.41, no.10,
save power but also helps it in avoiding faults taking place due pp. 74- 80, Oct 2003.
to paucity of energy. LA is stationed at each node to select the [12] V. T. Raisinghani and S. Iyer, "Cross-layer feedback architecture for
best path available among the multiple paths based on its mobile device protocol stacks", IEEE Communications Magazine,
goodness value. The goodness value of the path is updated vol.44, no.1, pp. 85- 92, Jan. 2006.
using the reward/penalty scheme of LA. These parameters are [13] Z. Che-Aron, W. F. M. Al-Khateeb and F. Anwar, "ENFAT-AODV:
The fault-tolerant routing protocol for high failure rate Wireless Sensor
well influenced by energy of the node lying in the path. So the Networks," Proceedings of the 2nd International Conference on Future
path that has the least probability of fault occurrence is Computer and Communication (ICFCC), pp.V1-467-V1-471, 21-24
preferred. To summarize, this work provides an effective May, 2010, doi: 10.1109/ICFCC.2010.5497747.
mechanism for fault tolerant routing for IOTs. [14] ns-2: Network Simulator 2, http://www.nsnam.org
In the future, we want to evaluate this solution to assess [15] C. Perkins, E. Belding-Royer and S. Das, “Ad hoc On-Demand Distance
its scalability and usefulness in a wide range of application Vector (AODV) Routing”, IETF, RFC 3561, July 2003.
domains.

819