system architectures for multi-sensor task allocation
DESCRIPTION
Paper presentation at 4th Annual Conference of the International Technology Alliance (ACITA 2010), Imperial College London, UK September 2010.TRANSCRIPT
ACITA 2010
Fangfei Chen, Tom La Porta
Matthew P. Johnson, Amotz Bar-Noy.
System Architectures forMulti-Sensor Task Allocation
Diego Pizzocaro, Alun Preece
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Why this paper?
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
‣ In our previous work we developed a variety of allocation mechanisms inspired by real scenarios
‣ ...but we decided to imagine what a real interface would look like...
‣ The prototype* interface inspired many discussions about system architectures
‣ The novelty of our architectures consists in a modular approach to solve the allocation problem(integrating a KB module with an allocation mechanism)
* Apple iPhone choice motivated mainly by its popularity and development tool provided.
Outline
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
1. Multi-Sensor Task Allocation (MSTA)
2. Formalizing MSTA problems
3. Conceptual architecture
4. Cetralized, Distributed, Hybrid
5. Discussion & Conclusion
Main problem
is the problem of automatically allocating sensors to tasks satisfying the user’s information requirements.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Multi-Sensor Task Allocation
Sensors (or Sensing assets) Tasks
Simple sensorsUsers create sensing tasks
Platforms
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
TASK 3
Area Surveillance
TASK 4
Area Surveillance
TASK 1
Detectvehicle TASK 2
Identifypeople
Scenario
• A network of heterogeneous sensing assets:
- Support multiple tasks competing for bundles of sensors
TASK 7Localize
Jeep
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
TASK 3
DetectPeople
TASK 5
DetectHelicopter
TASK 1
Identifypeople
TASK 6
IdentifyTank
TASK 4
DetectAircraft
- Sensors are scarce and in high demand.
- Highly dynamic (sensor failures, change of plan)
TASK 2
DetectGround Vehicle
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Multi-Sensor Task Allocation
• Coalition members on the field have usually
no time and no expertise to manually decide the best
sensors for each task.
• We need to automatically allocate sensors to tasks
to satisfy the information requirements of each user.
Various problem settings but the fundamental question:“Which sensor should be allocated to which task?”
We call it the Multi-Sensor Task Allocation (MSTA) problem
Unmanned Sensor
Earthquake Search & Rescue
Detect
(injured) people
Monitor
collapsing building
Formalizing MSTAVarious MSTA instances imply the need for a classification scheme,
helping us to identify the relevant problems.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Framework
Sensors
Task
sA
lloca
tion MSTA
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Taxonomy
• We propose a MSTA taxonomy organized on four main axes:
1. Sensors: Single-task (ST) vs. multi-task (MT).
2. Tasks: Single-sensor (SS) vs. multi-sensor (MS).
3. Assignment: Instantaneous (IA) vs. time-extended (TA).
4. Sensor Network: Homogeneous (HO) vs. heterogeneous (HE).
Given our military/emergency response scenario, our focus is:
• Coalition: Heterogeneous sensor networks (HE)
• Dynamic environment: Instantaneous Allocation (IA)
• Complex sensing requirements: Multi-Sensor tasks (MS)
• Sensor capabilities: both Single-Task (ST) and Multi-Task (MT)
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Single-Task sensors (No sharing)
• Referred to as Disjoint Coalition Formation problem in multi-agent community.
• It can be modelled as a Set Partitioning Problem (SPP)
• Set of sensors S
• Families F of acceptable subsets of S
• Each family F is associated with a utility u
• Goal: Find a maximum utility family X of elements in F
such that X is a partition of S.
• NB: here we allow sensors to remain unallocated and tasks to be dropped.
T2
T1
xUnmanned Sensor
Earthquake Search & Rescue
Detect (injured) people
Monitor collapsing building
MSTA instance:
ST - MS - IA - HE
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Multi-Task sensors (Sharing)
• Referred to as Overlapping Coalition Formation problem in multi-agent community.
• It can be modelled as a Set Covering Problem (SPP)
• Set of sensors S
• Families F of acceptable subsets of S - (representing overlapping coalitions)
• Each family F is associated with a utility u
• Goal: Find a maximum utility family X of elements in F
such that X is a cover of S.
• NB: also here we allow sensors to remain unallocated and tasks to be dropped.
T2
T1
Unmanned Sensor
Earthquake Search & Rescue
Detect (injured) people
Monitor collapsing building
MSTA instance:
MT - MS - IA - HE
Conceptual Architecturevalid for both problem instances (Single and Multi Task sensors)
which highlights the steps necessary to find a solution.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Conceptual Architecture
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Conceptual architecture
KBBundle
Generator
Allocationmechanism
SensorNetwork
1Mobile user creates a sensing task.
2Recommends fit-for-purpose sensor bundles for that task type.
3Finds a solution to eitherST-MS-IA-HE (No sharing) or MT-MS-IA-HE (Sharing).
4Configured accordingly and delivers data to user.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
KB Bundle Generator
• MSTA in Heterogeneous sensor networks requires knowledge of
which sensor types are applicable to which kinds of task.
• Two issues:
(1) Can a type of sensor (or combination) satisfy a task type?
(2) How well a particular sensor instance might perform?
• Addressing these issues requires knowledge from literature, which we encode in a
Knowledge Base.
• The KB stores:
(1) each type of sensor (or combination) that can theoretically satisfy the task
(2) a joint utility function to estimate the utility of sensor instances for that task
KBBundle
Generator
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Reasoning procedure
Task Type
CapabilitiesRequired
Joint Utility FunctionBundle Type
Using sensor & task ontologies& OWL reasoner.
Sensor types to choose.
= Recommendations
Localize vehicle
1) Acoustic intelligence2) Imagery intelligence
BT1 = {AcousticArray}BT2 = {UAV, Camera}
JUF1 = 2D-LocalizationJUF2 = Detection Probability
compatible?
(BT1, JUF1), (BT1, JUF2), (BT2, JUF2)
Which functions can be used to estimate the performances?
Allocation flexibility
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Lightweight KB
• The original implementation of the reasoning process is computationally expensive
• The reasoner uses an exponential-time classification.
• On a mobile device is not recommended.
• Precompute the results of the reasoner and store themin a look-up table
• Task types and sensor types are “stable”
• Reasoning operations are O(1)
• Assumption: the device has sufficiently large store capacity.
• 4000 different task types, 5 different recommendation per task type (BT+JUF)
• ~ 12 MB of storage required, ~ 20 ms (increasing logarithmically)
Task Type Recommendation
1 (BT1 + JUF1)
1 (BT2 + JUF1)
1 (BT2 + JUF2)
2 (BT3 + JUF1)
2 (BT2 + JUF1)
2 (BT2 + JUF2)
... ...
N (BT5 + JUM1)
Allocation Mechanisms
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Architectures
Allocationmechanism
unlike the KB bundle generatorthe allocation mechanism varies greatly depending
on the architecture
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Centralized
Sensor Network
CentralAllocationalgorithm
Hub
CentralKB - Bundle Generator
Base Station(s)
Users
• Currently many mobile apps are cloud based with a powerful central server.
• Mobile devices are just the interface into the system (to create tasks)
‣ Task posted to the
central engine
(KB and Allocation alg).
‣ When connectivity to the
base station is down, can use
the sensor network to
support communication.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Centralized allocation algorithms
ST-MS-IA-HE: Single Task sensors (No sharing)
• This can be seen as a combinatorial auction
‣ bidders: tasks
‣ items: sensors
‣ tasks bids for bundles of sensors
• Many algorithms have been proposed:
we use CASS (Combinatorial Auction Structured Search)
MT-MS-IA-HE: Multi Task sensors (Sharing)
• This can be solved with a Set Covering Problem (SCP)approximation algorithms:
‣ Poor performances if the number of allowed sensors in a bundle is large.
‣ The KB bundle generator limits the number of sensors that can be chosen.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Features
• Pros:
• Global overview of the situation
• Better overall allocation
• No “heavy” computation on the mobile device or on sensor nodes
Sensor Network
CentralAllocationalgorithm
Hub
CentralKB - Bundle Generator
Base Station(s)
Users
• Cons:
• Periodic collection of data and allocation decisions every constant interval
• If a certain area becomes “hot” (e.g. explosion or building collapses):
➡ Suddenly many mobile users occupies the same area,
➡ Overload of the mobile telecommunication network.
• Sending the data over sensor network as a backup
➡ only shifts the overload on the sensor network (and decreases the lifetime of sensors)
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Distributed
Sensor Network
Allocationprotocol
Allocationprotocol
KB Bundle
Generator
Allocationprotocol
KB Bundle
Generator
Allocationprotocol
KB Bundle
Generator
• Moves the KB bundle gen. on the user device (Lightweight KB bundle generator)
• Moves the allocation mechanism on the sensors (distributed protocols)
‣ User communicates directly the task created to the sensors
‣ The sensors autonomously negotiate the best task to serve as part of a bundle
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Protocols & Features
• Well known efficient allocation protocols have been proposed for the disjoint and overlapping
coalition formation problems.
• We have adapted these to solve the ST-MS-IA-HE (No sharing) and MT-MS-IA-HE (Sharing)
• Pros:
• There is no need to periodically collect data or synchronize allocation
• If new task type introduced, there is no need for wireless reprogramming
sensors. We would just update all the KBs on the user devices.
• Cons
• No central base station offering an overview
• Quality of the solution is worse than centralized (hypothesized on past results - verifying)
• Less network traffic (hypothesized on past results - verifying)
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Hybrid
Sensor Network
CentralAllocationalgorithm
Hub
CentralKB - Bundle Generator
Base Station(s)
Users
Allocationprotocol
Syncprotocol
Allocationprotocol
Syncprotocol
KB Bundle
Generator
• Combine aspects of both mechanisms into a hybrid architecture.
• Distributed architecture connected with a base station central allocation engine.
‣ Task posted to the
central engine when user
has connection.
‣ When connectivity to the
base station is down,
the system is able to
perform autonomously
‣ We need a sync protocol,
need to know if the local
protocol has started before
running the central algorithm.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Hybrid with data logs
Sensor Network
Base Stations
DataCenter
Updateprotocol
Allocationprotocol
Allocationprotocol
Updateprotocol
KB Bundle
Generator
• In the previous architecture:
‣ A better solution is not guaranteed, the two mechanisms might work in potential conflict.
‣ The synchronisation protocol will increase the network traffic
‣ Provide the big picture
to base station leaving the
allocation to the
distributed protocol.
‣ Update protocol is used
by sensors and users when an
something changes in the
network.
http://users.cs.cf.ac.uk/D.Pizzocaro [email protected]
Conclusion & Future
• These architectures may be considered in terms of their fit for coalition command and control
structures.
• Central - fits well with high echelon command centre (decisions needs overview)
• Distributed - resembles the case of local commanders (no time and no expertise)
• Hybrid with data logs - might fit both
• We are currently conducting experiments to
• compare solution quality, measure network traffic, sensor/user device performances
• We have presented and analyzed 4 different architectures to solve the MSTA problem
supporting coalition members on the field.