Using Virtual Markets to Program Global Behavior in Sensor Networks

1
Using Virtual Markets to Program Global
Behaviorin Sensor Networks

• Prof. Matt Welsh
• Division of Engineering and Applied Sciences
• Harvard University

Review by Paskorn C. DSSG_at_UMB May 6 2005
2
Market-based macroprogramming

• market-based macroprogramming (MBM), a new
  paradigm for achieving globally efficient
  behavior in sensor networks.
• The goal of market-based macroprogramming is to
  achieve an efficient allocation of network
  resource to optimizee specific objectives

3
Why might this be attractive?

• Markets operate in a (mostly) decentralized
  fashion
• Individual agents make locally greedy
  (profit-making) decisions

4
example

• Nodes closer to target might need to sample more
  frequently
• Ideally, nodes should self-schedule based on
  their local state
• Node should decide locally which operations to
  perform and how often
• Driven by interaction with environment
• Node operation should adapt to changing
  conditions
• e.g., If interesting event happens nearby, node
  might increase sampling rate

5
Basic model

• Nodes act as agents that sell goods (such as
  sensor readings or routed msgs)
• Each good is produced by an associated action
  that produces it
• Nodes attempt to maximize their profit, subject
  to energy constraints
• Each good has an associated price
• Network is programmed by setting prices for
  each good
• Each action has an associated energy cost
• e.g., Cost to sample a sensor/ Cost to transmit a
  radio message

6
Goods and Action

• Agent performs action to produce goods in return
  for payment
• Action
• Sample (S) a local sensor reading
• Send (T) send data towards to base station
• Listen (L) listen for incoming messages
• Aggregate (A) multiple sensor readings

7

• Taking an action may/may not produce a good of
  value to the sensor network application.
• Not all actions will receive a payment

8
Price

• The price on a good defines the payment that an
  agent can expect to get for selling a unit of a
  good.
• In first step Flood prices for each good to the
  network

9
Utility

• Nodes continuously select the action with the
  greatest utility
• The utility for an action is a function of
• Price
• Advertised by base station
• Energy availability
• Taking an action must stay within energy budget

10
Model Objective

• Maximize utility
• Subject to
• Energy budget gt 0

11

• The utility function is the expected profit for
  taking an action.
• B(a) is the estimated probability of payment for
  action a
• An action is only profitable if it is useful
• e.g., Only get paid to transmit if another node
  is listening
• B(a) is learned using an exponentially weighted
  moving average (EWMA) based on whether or not a
  node received payment for action.
• Nodes are conditionally paid for the listen and
  sample actions based on whether or not a message
  was heard and whether or not a sample above a
  threshold was acquired, respectively the a value
  for other actions is always 1
• Action selection uses
• Node usually takes action with highest expected
  profit

12
Constrain

• Energy (real energy --battery)

13
Energy Consume
14
Application Example Object Tracking
Utility functions vary depending on node's
position in network Nodes near target have high
utility for sampling Nodes along routing path
have high utilities for listening and sending
15
Price Updating

• 1 User adjust the price as the market runs in
  response to change in quality rate, latency of
  data.
• 2. Determine prices empirically based on an
  observation of the networks behavior.

16
Simulation model
Simulation based on TinyOS environment Target
travels in circular path Routing using GPSR to
base station
17
Node 23 Observation
18
Node 31
19
Why nodes does sample all times

• This is because the node has enough energy to
  perform these actions, and the -greedy action
  selection policy dictates that it will explore
  among these alternatives despite negligible
  utility.

20
What is the Red Line

• The energy token budget which is refilled in the
  rate of r J/day continuously.
• We can see that at time 350-380 the value of red
  line is very high because it decreases the listen
  action rate.
• At time 380-450, the value is very low because it
  increases the send action rate.

21
Error Listen Price
Increasing the price for the listen action
degrades the trackers accuracy, since fewer
nodes are taking samples
22
Energy

• Low 400J/day (refill)
• High 3000J/day (refill)

23
(No Transcript)
24
Future Work

• Complex price setting
• Price different in area
• Price from sequence of actions not 1 action
• Multiple demand
• Reinforcement Learning (improve Beta
  mechanism)

25
Related Work

• Newest paper
• Decentralized Adaptive Resource Allocation for
  Sensor Networks (SORA)
• Geoff Mainland, David C. Parkes, and Matt Welsh.
  To appear in Proceedings of the 2nd USENIX/ACM
  Symposium on Networked Systems Design and
  Implementation (NSDI 2005), May 2005

26
Related work Sensor Network Resource Allocation
Scheduling

• Static node scheduling
• TAG (26) (database approach, fixed sampling
  periods)
• Cougar(46) (database approach)
• Directed diffusion (17) (database approach,
  data is cached at intermediate nodes, trying
  to identify best route for two nodes, fixed
  sampling periods )
• Dynamic Scheduling
• Hoods tracker (43)
• LEACH (13) (Low Energy Adaptive Clustering
  Hierarchy)(Cluster based Approach, only cluster
  head send the data)

27
Comparative Analysis

• SORA VS
• Static Scheduling
• Dynamic Scheduling
• Hoods Tracker

28
Static scheduling

• Schedule is computed based on the energy budget
• Example (if there are only two actions)
• Energy 100 J /day
• If listen action use 8 J
• If sample action use 2 J
• Thus, in 1 day, the sensor can do listen action
  and sample action 100/(82) 10 times (listen
  action 10 times , sample action 10 times)

29
Static scheduling(TAG, directed diffusion, etc)

• Easy
• Have to know the amount of resource at first
• Sensor performs every action (sample, listen,
  send, aggregate) in the same rate.
• Nodes, which are far from the moving target, lost
  energy the same as working nodes.
• No learning algorithm

30
Dynamic Scheduling

• Nodes continuously adjust their processing rate
  based on their current energy budget
• For example, node that has very low energy can
  perform sample action more than listen action

31
Hoods

• Nodes calculate the target location
• The node that detects the vehicle broadcasts its
  data to its neighbors
• Every node that get the sample data check if it
  is center
• The center node is the only node that send data
  to base station

0
4
10
8
32
Hood A Neighborhood Abstraction for Sensor
Networks (Berkeley)
33
Tracking Accuracy

• Different between estimated and true vehicle
  position at the time of estimated data is
  received
• delay increase error

34
Tracking Accuracy
The static and dynamic schedulers are the most
accurate,since they operate periodically, while
SORA has slightly highererror due to its
probabilistic operation. Disabling aggregation in
SORA causes accuracy to suffer since more
readings are delivered to the base station.
35
(No Transcript)
36
Energy Efficiency
37
Energy Efficiency

• sum of (useful energy) in each hop /total
  energy
• Nodes that in the route to send data to based
  station is useful hop
• In the perfect system, no waste energy

38
The End
39
More EWMA

• Exploration probability
• Adjust waste energy for random action
• Weight value in EWMA
• Adjust more or less quick