#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Sensor-Task Assignment with Combinatorial Auctions
1. Sensor-Task Assignment
in Heterogeneous
Sensor Networks
Diego Pizzocaro
D.Pizzocaro@cs.cf.ac.uk
Research Group: KIS Supervisor: Prof. Alun Preece
2. Why sensor-task assignment?
• Heterogeneous Sensor Networks (HSN) introduce new resource allocation
problems in which sensors must be assigned to the tasks they best help
• An already deployed HSN is usually required to support multiple sensing
tasks of different nature to be accomplished simultaneously
• Tasks might compete for the exclusive usage of the same sensing resource
➡ We need schemes to assign individual sensors to tasks
• Research focus: find the right way to model this allocation problem
• we defined increasingly detailed models
• we developed computationally efficient approaches to solve them
3. Example
• Two target identification tasks.
(Tasks may have different priorities)
• Targets are close in the field.
• We only have one video sensor that could identify both.
• Problem: Where do we point the video sensor?
(i.e. To which task do we assign the sensor?)
X
Target 2
X
Target 1
4. Simple model
Sensor-Task Assignment
• Tasks vary in priority and have a different demand Sensors
for sensing resource capabilities. S1
e11
Tasks
e1 T1 (d1, p1)
2
• Each sensor has a different utility for each task, S2
because of:
• Geography & distance S3
T2 (d2, p2)
• Remaining battery life
S4
• Goal: A sensor assignment that maximizes the
e = utility of sensor to a task
utility that the sensor network can provide d = task utility demand
to tasks. p = task priority
This problem is NP-Complete and very hard to approximate:
We developed many heuristic algorithms to solve it (greedy algs)
5. Non additive utility!
• Utilities from multiple sensors do not always combine additively
• Example:
‣ Triangulation tasks
‣ We need two audio sensors for each task
X X
Target 2 Target 2
X X
Target 1 Target 1
Task 1: Utility(S1,S2) = 100 Task 1: Utility(S1) = 0
6. More detailed model
Sensor-Bundle-Task Assignment
• We first want to group sensors into bundles, and then we want to find the
best assignment of bundles to tasks.
• NP-Complete problem: we will use COMBINATORIAL AUCTION techniques.
Sensors
Bundles Tasks
S1
e11
B1 T1 (p1)
S2
e1
2
S3
B2 T2 (p2)
S4
e = joint utility of a bundle to a task
p = task priority
7. Combinatorial auctions
• Combinatorial auction:
‣ It is a silent auction in which bidders can bid on sets of items
(instead of single items).
‣ Each bidder provides sets of items and corresponding prices for each set.
‣ The auctioneer chooses the set of bids that maximizes the payment.
• The Sensor-Bundle-Task Assignment model can be seen as a combinatorial auction.
8. Related work - overview
• Combinatorial auctions have been already applied to sensor-task assignment
problems in scenarios where sensor utility does not combine additively:
‣ J. Ostwald, V. Lesser, and S. Abdallah.
Combinatorial auctions for resource allocation in a distributed sensor
network. In RTSS ’05 (Real Time Systems Symposium), Washington, DC, USA.
• Problem: assign radar sensors to weather monitoring tasks
‣ Sensors have multiple settings (and can be configured)
‣ Joint utility of a bundle is computed with a probabilistic approach
(i.e. non-additively)
‣ A sensor can be shared by multiple tasks
• This problem is a variant of the classic combinatorial auction:
• they modified a pre-existent algorithm.
9. Similarities and differences
• Similarities:
‣ Their problem can be modeled as a variant of classic combinatorial auctions.
‣ Their joint utility is computed using a task dependent joint utility function.
‣ They use heuristic/preprocessing to limit the number of possible bundles and
sensor configurations.
• Differences:
‣ They assume an homogeneous sensor network (only radars) but configurable.
‣ Their tasks are not different by nature (only “monitoring tasks”):
joint utility is easier to compute.
‣ The size of the network and the number of simultaneous tasks
(a few dozens of radar sensors, and ten tasks).
‣ The computational time is not the main focus (time to solve it is 10 seconds!).
10. Learned lessons
• Applying combinatorial auctions to solve our current model seems reasonable.
• We need to modify (again!) our model considering configurable sensors and
resource sharing
‣ To solve it we could adopt a similar approach to modify combinatorial auction.
• Critique to their approach:
‣ they should have modified other algorithms or developed new greedy algorithms
to compare the performances (quality of solution, computational cost).