autonomous market-based approach for resource allocation in a cluster-based sensor network
DESCRIPTION
Autonomous Market-Based Approach for Resource Allocation in A Cluster-Based Sensor Network Wei Chen , Heh Miao Department of Computer Science Center of Excellence for Battlefield Sensor Fusion Tennessee State University, United States Koichi Wada Nagoya Institute of Technology, Japan. - PowerPoint PPT PresentationTRANSCRIPT
Autonomous Market-Based Approach for Resource Allocation
in A Cluster-Based Sensor Network
Wei Chen, Heh Miao Department of Computer Science
Center of Excellence for Battlefield Sensor FusionTennessee State University, United States
Koichi Wada Nagoya Institute of Technology, Japan
IEE MCDM 2009
IEEE Symposium on Computational Intelligence in Multicriteria Decision-Making, 2009
Introduction: Sensor network, Fusion, Resource Allocation
Problem Statement
Review of Market-Based Resource Allocation: Centralized vs. Decentralized Approaches
Proposed Market-Based Resource Allocation Approach for Cluster-based Sensor Networks
Implementation and Experiment Results
Future work
Presentation OutlinePresentation Outline
IEE MCDM 2009
sink
Sensor Network & Sensor Fusion
Return back sensed/fused data
Ask for data/information
Fusion missions: Target tracks, target identification, environment monitoring …
Upper-level fusion Base Station
Lower-level fusion
Sensor Network
IntroductionIntroduction
How to assign the resources for achieving the requested data with smallest delay while keeping the network alive as long as possible?
Resource Allocation
IntroductionIntroduction
Given a task or tasks, how to assign sensors and network resources for fulfilling the task/tasks with the goal of less delay, high QoS, and long network lifetime?
For example, a task of mobile target tracking can be fulfilled by a sequence of node actions: sampling, listening, transmitting, aggregation, sleeping, and each action uses some resources. What action each node should take at each timeslot to fulfill the task that best matches the above goal?
Problem StatementProblem Statement
Centralized Resource Allocation (CRA) (Dr. T. Mullen and others, Penn State Univ.)
Using an auction mechanism for a single-platform or single-hop sensor network A winner has to be decided from resource bids during each round of scheduling according to the current status of all resources and requirements of given tasks.
Computation intensive
Central Sensor manager
Base Station (Clients, Consumers)
Single-platform or one-hop Sensor Network
Not suitable to a multi-hop sensor network, where communication cost of relaying data are the dominant cost.
Review of Market-Based Approaches
Review of Market-Based Approaches
IRM
IRM
IRM IRM
IRM
IRM
IRM
Sensor Network
Base Station(Clients, Consumers)
Infrequently central control
Decentralized Resource Allocation (DRA) (G. Mainland & others, Harvard Univ.) At each timeslot, the IRM at each node locally selects an action that can maximize the utility function.
Tuning node behavior: when action is “successful,” the utility function receives a reward. Nodes can determine locally which actions were “successful”.Central control: adjusting the price of resource infrequently
No control points, hardly achieving optimal resource allocationOverlap on sensing, computation, and networking
Individual Resource Manager
otherwise 0
available is action theif )()()(
aapricea
au
otherwise 0
available is action theif )()()(
aapricea
au
Review of Market-Based Approaches
Review of Market-Based Approaches
• Local Resource Manager (LRM) at cluster-head nodes is local centralized
• Individual Resource Manager (IRM) at cluster-member nodes is decentralized.
• Simple central control by adjusting the price of resource infrequently
• Using the routing protocols and reconfiguration functions of the underlying cluster-based sensor network
Goal: (1) providing promise solution of resource
allocation for given tasks with less delay and high QoS; and
(2) extending network lifetime
Hierarchical Resource Allocation (HRA) in Cluster-Based Sensor Networks
Cluster head
IRM
IRM
LRM
LRM
IRM
LRMCluster
Sensor Network
IRM
Base Station(Clients, Consumers)
Infrequently central control
Proposed Approach- Framework
Proposed Approach- Framework
Underlying sensor network: cluster-based sensor network
Most sensor networks nowadays are built with hierarchical and reconfigurable structures that introduce efficient sensing, computing and networking, and long network lifetime. One of the most well used hierarchical structures is cluster-based structure.
Market-Based Approach
Instead of low-level sensor programming that manually tunes sensor and other resource usage, we use a market-based approach for dynamic allocation of system resources.
Proposed Approach – Assumptions
Proposed Approach – Assumptions
Goods and ActionsIn the HRA approach, the actions that sensor nodes take depend on the task, but typically can include sampling a sensor, aggregating multiple sensor readings. An available action set is decided at each timeslot. Production of one good may have dependencies on the availability of others. For example, a node cannot aggregate sensor readings until it has acquired multiple readings. Taking an action may or may not produce a good of value to the sensor network as a whole. For example, listening for incoming radio messages is only valuable if a node hears a transmission from another node. We suppose that nodes can determine locally whether a given action deserves a payment.
Resource Constraints There are tradeoffs between the network resources and the quality of the service. Especially, a node’s energy constrains the actions that it can take. In the IRM, a payment-possibility threshold is used. When the estimated probability of payment from an action is smaller than the threshold, the action is not scheduled for the current timeslot. It is expected that the energy can be saved by reducing unnecessary actions and the quality of the service can be maintained by giving no energy constraint to useful actions.
Proposed Approach – Principles
Autonomous Scheduling1. Rather than static scheduling, individual nodes tune their schedules
over time2. Cluster-heads do local optimization in their clusters 3. Nodes avoid wasting energy4. Using the feedback to tune node behavior: nodes receive rewards
when they take useful actions5. Reinforcement learning to select best actions
Action model at nodes:1. Nodes can select an action among a set of actions2. Each action has an associated energy cost3. When an action is “successful,” the node earns a reward
Examples of actions: Sample a sensor, Listen for incoming radio messages, Transmit a radio message, Aggregate multiple sensor readings into a single value
4. Each node attempts to maximize its reward5. Nodes can determine locally which actions were useful
Proposed Approach – Design Details
Algorithm of the IRM at a node r for each timeslot (scheduling cycle) do (1) with 1-ε probability select an action a from the available action set which has largest utility value; (2) with ε probability randomly select an action a from the action set //exploring action space to avoid falling to local minima// (3) if β(a) < payment-possibility threshold then node r goes to sleep //saving energy// else begin node r takes action a; if action a receives a payment then β(a) =α+(1- α)β(a) //estimated probability of success gets larger // else β(a) =(1- α)β(a); //estimated probability of success gets smaller // end; (4) if node r runs out of the energy then call the network reconfiguration functions;
otherwise 0
available is action theif )()()(
aapricea
au
Utility function
G. Mainland’s algorithm: An energy budget is used for each fixed period. Nodes take the actions that can maximize the utility function even the profit is very small when the budget is allowed.
Proposed Approach –Design Details
Algorithm of the LRM at a cluster-head for each timeslot (scheduling cycle) do begin (1) collect status of each member node in the cluster; (2)determine the optimal resource allocation according to the current
status in the cluster and the given tasks; (3) inform the decision to the cluster member nodes; (4) if the head runs out of the energy then call the network reconfiguration functions; end;
Price Selection and Adjustment at the Central Controller • Prices are propagated to sensor nodes from the GRM through data dissemination algorithm. • The client can adjust prices to affect coarse changes in system activity.
Routing ProtocolsBroadcast protocol and data gathering protocol for the underlying cluster-based sensor network are used.
Reconfigurable FunctionWhen a node runs out of battery, the network will be self-reconfigured.
Proposed Approach – Design Details
A Flat WSN level
Hierarchical Architecture level cluster
Underlying Networking Architecture : cluster-based hierarchical networking architecture for supporting hierarchical routing and resource allocation.
Data Dissemination/Collection Algorithms: distributed routing algorithm for time and energy efficient broadcast, multicast, unicast and data gathering
Network Self-Organization Functions: network self-construction/reconfiguration
backbone
A group of specified nodes
Data fusion on a group via routing
sink
Management servicesSynchronization LocalizationNode and event failure detectionArchitecture reconfiguration
Configurable Service level
Networking SericesData query and disseminationData collection and integration Data fusion via routing
Proposed Approach – Underlying Cluster-Based Sensor Network
Clustering-based Network Architecture : Combining the centralized control in local with the decentralized control in global
Efficient Routing Algorithms for Broadcast/Multicast, and Data Gathering
BroadcastingFlat (unstructured) Network
Clustering-based (structured) Network
a
bc
d
e
Euclid circuit traveling
Network Self-Organization for maximizing network lifetime: head rotation, node move-in and move out – Physical layer dependent
Proposed Approach – Underlying Cluster-Based Sensor Network
Implementation and Simulation Application: Tracking Mobile Targets
Field: 105m×105m Nodes: 800 MICA2/Crossbow motesResource: (1) Radio: member – 15 m, head – 30 m; (2) Magnet sensor: sensing range – 11m; (3) Processor Buffer: 2 buffers (2256 byte) with totally 14 packages Sample reading: 29 byte (one buffer can save 17 samples) Moving target: one or two with speed 1.5 m/s or 3 m/s moving on random straight routes Packet size: 35 byte (payload 29 byte with header 6 byte) Data rate: 38.4 kbps Timeslot for an action: 0.25 second Initial energy at each node: e = 3.88 J (energy in an Nickel Cadmium AA battery = 4320 J)MAC protocol: CAMA/CALocal optimization at LRM: cluster-head select the best radio messages (most accurate message) when it receives multiple overlap messages from its member nodesRouting protocols: data dissemination – broadcast protocol by using the backbone tree, message collection – data gathering protocol which relays data back to the base station from sensor nodes by using the backbone tree from children to the parent
Energy consumption for actions at each time slot Action 1: Sending, 2.33 mJ, Action 2: Listening, 6.56 mJ, Action 3: Sampling, 84.1 uJAction 4: Aggregation, 84.1 mJ, (Action 5): sleeping, 12 uJ
Experimental Results
Flat Sensor Network
sink
Experimental Results
Cluster-based Sensor Networks
sink
Experimental Results
Latency (one mobile target) In 20 seconds, DRA received 77 messages, HRA received 119 messages
DRA (Without Local Optimization) HRA (With Local Optimization)
Latency of Messages (One Target, OPT)
46; 39%
45; 37%
19; 16%
9; 7% 2; 1%
0 - 5 sec
5 - 10 sec
10 - 15 sec
15 - 20 sec
>20 sec
Latency of Messages (One Target, NOPT)
16; 3%11; 2% 24; 4%26; 5%
458; 86%
0 - 5 sec
5 - 10 sec
10 - 15 sec
15 - 20 sec
>20 sec
Test field Test field
DRA (Without Local Optimization) HRA (With Local Optimization)
Latency of Messages (Two Targets, OPT)
9; 5%
9; 5%
123; 65%
45; 24%
0 - 5 sec
5 - 10 sec
10 - 15 sec
15 - 20 sec
>20 sec
Latency of Messages (Two Targets, NOPT)
134; 37%
40; 11%25; 7%20; 5%
149; 40% 0 - 5 sec
5 - 10 sec
10 - 15 sec
15 - 20 sec
>20 sec
Latency (two mobile targets)
Test fieldTest field
Experimental Results
After tracking a mobile target 200 seconds
Experimental Results
Experimental Results
The closer to the target, the more accurate sensor readings a sensor node can get
Experimental Results
Experimental Results
Observation: change the price of sending only may not work well.
Future Work
Return back sensed/fused data
Ask for data/information
Fusion Service Level
Fusion missions: Target tracks, target identification, …
Task and sensor management identifying network service, specifying resource and service quality
Mission management: decomposing mission, assigning priority, allocating task, …
Upper-level fusion
Customer/Base Station
A Flat WSN level
Hierarchical Architecture level clusterbackbone
A group of specified nodes
Data fusion on a group via routing
sink
Management servicesSynchronization LocalizationNode and event failure detectionArchitecture reconfiguration
Configurable Service level
Networking SericesData query and disseminationData collection and integration Data fusion via routing
Homework and assignment
1. Discuss the tradeoff between DRA and HRA on latency, energy consumption, and network maintenance, respectively.
2. Who adjusts the prices of actions? Is it centralized control or distributed control? How to make the HRA more efficient by adjusting the price of actions?