1 Distributed Adaptive Sampling, Forwarding, and Routing Algorithms for Wireless Visual Sensor Networks Johnsen Kho, Long Tran-Thanh, Alex Rogers, Nicholas

1

Distributed Adaptive Sampling, Forwarding, and Routing Algorithms for

Wireless Visual Sensor Networks

Johnsen Kho, Long Tran-Thanh,Alex Rogers, Nicholas R. Jennings

{jk05r,ltt08r,acr,nrj}@ecs.soton.ac.uk

12th May 2009

Third International Workshop on Agent Technology for Sensor Networks (ATSN-09)

2

Background: Wireless Visual Sensor Network (WVSN). Research Challenge & Aims.

Inter-Related Adaptive S/F and Routing : Problem Description. Information Metric. The Mechanism:

Algorithm with Fixed Routing. Algorithm with Flexible Routing.

Empirical Evaluations. Conclusions & Future Work.

Outline

12th May 2009 ATSN-09

3

Wireless VisualSensor Network WVSN characteristics:

Array of smart camera devices, Basic processing and compression of (usually large) visual data Decentralised control regime A base-station (BS) to fuse and analyse collected data.

WVSNs are increasingly being deployed for: Object tracking, Unattended area surveillance, Other security related applications.

Both pictures are taken from Kleihorst et al., 2006


Inter-Related S/F +Routing in WVSNs

ATSN-09 412th May 2009

Constraints: Heavy energy constraints on the nodes Sampling is relatively expensive (due to large size of data packets) Forward own data vs. relay data from the others

Sampling, Forwarding (and Routing) share the same energy budget Their performance in efficient data collection are inter-related Efficient coordination is needed

Related research works: USAC: Utility-based Sensing and Communication Protocol (Padhy et al.

2006)

Goal of deployment: efficient information collection

Problem: these algorithms are not efficient for maximising information collection in WVSNs due to the myopic decisions of the agents during operation

Research Challenge: Efficient energy-aware coordination between sampling and

routing actions in WVSNs Minimise energy waste on taking useless actions Non-myopic decisions are needed

ATSN-09 5

Research Challenges and Aims

12th May 2009

Research Aims: Information metric to measure the usefulness of data Efficient S/F + routing mechanisms

Small control messages in the coordination phase Energy-awareness

6


Inter-Related Adaptive S/F and Routing: Problem Description. Information Metric. The Mechanism:


Empirical Evaluations. Conclusions & Future Works.

Outline


The goal is to maximise the total information value delivered to BS in each round

7

Model Description


Set of heterogeneous cooperative nodes I={1,…..,n}. Each node i ∈ I :

mi different sampling (or frame) rates

Each data packet p has the information value of

Bi energy budget to:

Sampling data Forwarding data

)( pVi

The BS collects data from the nodes periodically rounds

},...,{ 21 imiiii cccC

Each node’s memory is flushed and reinitialised after each round

Not delivered data is useless for the application

The nodes can choose one sampling rate for each round

Assumptions:

Nodes are rechargeable Bi can be entirely used in each round

8

Information Metric

Several techniques for valuing information: Kalman Filter [Guestrin et al. 2005; Rogers et al. 2006]. Simple Linear Regression [Padhy et al. 2006]. GP Regression Technique [Mackay 1998; Seeger 2004, Stranders et al. 2008, Kho et al. 2009].


In our model, we use a generic information valuation function

)( ii rV non-decreasing function

9

The Algorithms (Overview)

Algorithm with fixed routing: routing tree is already established by a

routing protocol (e.g. AODV) calculate optimal sampling rates


Phase I: each parent node broadcasts

its capacity to child nodes

Phase II: each node transmits maximal

possible contributions to its parent

Phase III: parents allocate packet

forwarding capacities to its children

Algorithm with flexible routing: optimal routing + optimal sampling are

to be determined

Algorithm with fixed routing

0 1 2 ... M-1 M

0 12 20 ... 27 27

0 1 2 ... 20 20

ATSN-09 10

Algorithm with Fixed Routing (Phase II)

12th May 2009

Each node i maintains an array of 3-tuples: iO })(),(,{ * ncnUnO iii

n : number of packets node i sends to its parent

)(nU i : maximal information value node i can contribute with n packets

)(* nci : sampling rate at node i in this case (node i’s own contribution)

If i is a leaf node: Only its own data is considered Filling is straightforwardiO

MJJJ ,..., 21

If i is not a leaf node: Its child nodes: Wait until all has arrived Maintain a table as follows

kJO

iT

:)(0JI OO

ATSN-09 11

Algorithm with Fixed Routing (Phase II) - cont’d

12th May 2009

:0J

O]}[],1[{max],[

0mlOmkTlkT kJlm

Dynamic programming:

0 1 2 ... M-1 M

:1JO

:iT

)( 0JI Own data (similar to the previous case)

1JI Own data + data from J1

0 1 2 ... M-1 M

0 12 20 ... 27 27

0 1 2 ... 20 20

0 1 2 ... M-1 M

0 16 22 ... 58 58

0 1 2 ... 31 31

0 12 20 ... 27 27

0 16 ... 75 75

21 JJI Own data + data from J1 and J2 0 18 30 ... 100 120

MJJI ...1Own data + data from all children

0 ... ... ... ... ...

0 30 40 ... 100 125

28

0 1 2 ... M-1 M

0 30 40 ... 100 125

0 1 2 ... 12 14

:iO

M: the capacity of node i’s parent

12

Algorithm with Fixed Routing (Phase III)

Efficient: it satisfies the data flow

conservation of the network no energy is wasted by

transmitting data that later will not be delivered to BS


When control messages reach the leaf nodes, that node start to transmit data

The BS maintains its own T table Easily detects the contributions

of its children

13

Algorithm with Flexible Routing (Overview)

The data readings from different nodes could be sent through different

routes if there are more than one option to choose from.

Minor restrictions:

Nodes always forward their data toward the BS; that is, they will not forward

data to a node that is further from the BS (in terms of hop count) than

themselves.

Sampled data from a same

node must be sent in bundle


14

Algorithm with Flexible Routing


: set of parents of node i

: set of descendants of node i

bundles to send

combinations

iOnode i has to send -s for all of the combinations

curse of dimensionality!!!

15


Inter-Related Adaptive S/F and Routing: Problem Description. Information Metric. The Mechanism:


Empirical Evaluations. Conclusions & Future Work.

Outline


16

Algorithm with flexible routing: deliver more information, but has greater computational and communicational cost

Algorithm with fixed routing can be applied on an efficiently chosen spanning tree Fast, but sub-optimal result

A trade-off between the loss in information and the saving in resources

Empirical Evaluation


Benchmark Algorithm: The Uniform Non-Adaptive S/F and Routing

each sensor divides its energy budget equally

Linear information valuation function

Empirical Evaluation (cont’d)

ATSN-09 1712th May 2009

Algorithm with flexible routing is used here Random tree is generated, algorithm with fixed routing is used on this tree

18

Empirical Result I

Algorithm with flexible routing produces optimal performance

Uniform non-adaptive algorithm has the worst performance

By choosing a spanning tree efficiently, fixed routing can achieve near-optimal performance


19

Empirical Result II


20

Conclusion & Future Work

Two novel and optimal decentralised algorithm: Algorithm with fixed routing: calculates the optimal sampling actions

Algorithm flexible routing : optimal in both sampling and routing

Algorithm with flexible routing is optimal, but has higher communication and computational cost

Algorithm with fixed routing can achieve near-optimal result on an efficiently chosen spanning tree

Future work: Develop an efficient way to choose the best spanning tree (e.g. using learning

approach)

Relax the assumptions (topology hierarchy, flexible sampling rate)

Take more rounds into account (long-term data collection)


Thank you (Any Questions?)

Documents

1 Distributed Adaptive Sampling, Forwarding, and Routing Algorithms for Wireless Visual Sensor Networks Johnsen Kho, Long Tran-Thanh, Alex Rogers, Nicholas