self-organized task allocation to sequentially...
TRANSCRIPT
![Page 1: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/1.jpg)
Self-Organized Task
Allocation to
Sequentially
Interdependent Tasks
in Swarm Robotics
Jimmy Lee, Sawyer Jager, Brad Steiner
Team Amirite?
![Page 2: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/2.jpg)
Arne Brutschy, Giovanni Pini, Carlo Pinciroli,
Mauro Birattari, and Marco Dorigo (2014).
Self-Organized Task Allocation to
Sequentially Interdependent Tasks in Swarm
Robotics, Autonomous Agents and
Multiagent Systems, 28(1):290-336.
(Brutschyetal2014.pdf)
![Page 3: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/3.jpg)
How can a swarm of robots allocate
themselves to efficiently complete sequentially
interdependent tasks?
Sequentially interdependent tasks:
A set of subtasks that must be completed
one after the other to complete the overall
task
Overview
![Page 4: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/4.jpg)
Setup - Harvesting and Storing
![Page 5: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/5.jpg)
Proposal
● Goal: Maximize number of objects retrieved per time unit
● Robots independently assign themselves to subtasks based on
how long they must wait for a robot from the other task to arrive at
the task interface.
● Increasing probability of a robot switching from one subtask to
another one while waiting at the interface
● Intuitively, a relatively longer wait at the interface indicates that the
other subtask is understaffed.
● End result: rates of arrival at the interface are equalized
![Page 6: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/6.jpg)
![Page 7: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/7.jpg)
Proposal
● Does not rely on global knowledge or
centralized components
● Does not use any communication
● Self organized
![Page 8: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/8.jpg)
Formal Definitions
![Page 9: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/9.jpg)
Formal Definitions - Tasks/Subtasks
● A task T is composed of subtasks τ1 ... τn
● Sequential interdependence means τ1 ... τn
must be completed in a given order
o T can also be called a “task sequence”
● τ1 > τ2 : τ1 must be completed before τ2
o τ1 is a predecessor
o τ2 is a successor
![Page 10: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/10.jpg)
Formal Definitions - Allocation
● gi : A group of robots working on τi
● Ni : Number of robots in gi
● N1 + N2 + … + Nn = N : total number of
robots
o There are no idle robots
![Page 11: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/11.jpg)
Formal Definitions - Process
A robot allocated to τi waiting at the interface for a robot
from τj experiences interface delay denoted dij
(seconds)
Robots keep track of how long they’ve waited at either
end of the interface. Average wait time of dij denoted dˆij
Probability of a robot switching from τi to τj before dij
seconds have elapsed: Pij(dij)
● 1 - Pij(dij) : Probability a robot will wait dij seconds
● Formally, Pij(dij ; dˆij , dˆji)
![Page 12: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/12.jpg)
Probability Function
Sigmoid Curve
Time Waited at Interface
Pro
ba
bil
ity o
f S
wit
ch
ing
![Page 13: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/13.jpg)
Probability Function
Designed so that robots tend to switch to the
subtask with a longer wait time quicker
subtask j is understaffed
subtask i is understaffed
![Page 14: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/14.jpg)
Probability Function
Function is independent of absolute values of
subtask duration and delays
Probability curve
automatically calibrates
itself
If wait times are generally
longer, the probability of
switching doesn’t increase
until more time has passed
![Page 15: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/15.jpg)
Probability Function Formal Def.
Robots evaluate probability function at discrete
points in time (dictated by control cycles).
At each control cycle, the robot switches from
one task to another with a probability of pij
(switching probability, not the same as Pij)
![Page 16: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/16.jpg)
Probability Function Formal Def.
Probability that the robot does not switch subtasks within
the first dij control cycles.
Pij(dij ; dˆij , dˆji) = 1 - Π(1 - pij(q ; dˆij , dˆji))q=1
dij
![Page 17: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/17.jpg)
Probability Function Formal Def.
k = steepness parameter
lower k => steeper curve
m = shift parameter
higher m => delay longer
![Page 18: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/18.jpg)
Steepness and Shift
![Page 19: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/19.jpg)
Probability Function Formal Def.
asdfasdfasdf
asdfasdfasdf
Relates the average interface times
Ensures robots tend to switch to the
understaffed group and not the other way
around.
![Page 20: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/20.jpg)
subtask j is understaffed
subtask i is understaffed
![Page 21: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/21.jpg)
Experiment
![Page 22: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/22.jpg)
Experimental Framework
Shown:
Asymmetric environment (arena ratio = .33)
![Page 23: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/23.jpg)
Experimental Framework
Two subtasks: harvesting and storing
Two environments: symmetric and asymmetric (.67)
Resources must be harvested before they can be stored
Resources cannot be cached at interface area (direct
handoffs only)
Robots that switch subtasks experience a switching cost, cs
Task switching only occurs at the task interface
![Page 24: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/24.jpg)
Metrics
P : Swarm Performance = # of objects
collected by the swarm
R : allocation ratio = fraction of robots allocated
to storage subtask
Stot = total number of switches performed by
robots in a run of the simulation.
![Page 25: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/25.jpg)
Experiments
1. Optimal Allocation
2. Task Switching Cost
3. Parameter Study
4. Scalability
5. Adaptivity
![Page 26: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/26.jpg)
1. Optimal Allocation
Brute force search for the optimal ratio of
allocation (Ropt) for various numbers of robots
Robots are not allowed to switch subtasks
![Page 27: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/27.jpg)
Symmetric Environment
Asymmetric Environment
![Page 28: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/28.jpg)
2. Task switching cost
Examine influence of task switching cost
Experiment with two groups:
Group 1: Each robot required to complete
both subtasks
Group 2: Robots switch using the proposed
method
![Page 29: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/29.jpg)
Proposed method unaffected
by switching costs
N = 18
![Page 30: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/30.jpg)
3. Parameter Study
Explore the effects of m (shift parameter) and
k (steepness parameter) on performance
![Page 31: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/31.jpg)
![Page 32: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/32.jpg)
![Page 33: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/33.jpg)
![Page 34: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/34.jpg)
![Page 35: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/35.jpg)
4. Scalability
Examine how increasing the number of robots
affects performance
Physical space is considered though this is a
virtual simulation
![Page 36: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/36.jpg)
Diminishing returns
due to congestion
![Page 37: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/37.jpg)
5. Adaptivity
Test flexibility of the proposed method
Change the environment halfway through the
simulation
@ t = 30, symmetric => asymmetric
![Page 38: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/38.jpg)
Quick adaptation
![Page 39: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/39.jpg)
Real Simulation
http://iridia.ulb.ac.be/supp/IridiaSupp2011-
002/sbot_experiment_run1_30x.ogv
![Page 40: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/40.jpg)
Conclusions
● Proposed method allows for self-
organization
● Does not require communication
● Achieves near-optimal allocation
● Environment specific factors have very little
influence
● Adaptive
![Page 41: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/41.jpg)
Questions?
![Page 42: Self-Organized Task Allocation to Sequentially ...cse.unl.edu/~lksoh/Classes/CSCE475H_Spring15/... · asdfasdfasdf asdfasdfasdf Relates the average interface times Ensures robots](https://reader036.vdocuments.us/reader036/viewer/2022071214/60420a4330677b35060bffd8/html5/thumbnails/42.jpg)
Then we have questions for you!
Kidding.
No questions?