Emergent Behavior in Biological Swarms Stephen Motter.

23
Emergent Behavior in Biological Swarms Stephen Motter

Transcript of Emergent Behavior in Biological Swarms Stephen Motter.

Emergent Behavior in Biological Swarms

Stephen Motter

The Papers

• The Self-Organizing Exploratory Pattern of the Argentine Ant

• Study of Group Food Retrieval by Ants as a Model for Multi-Robot Collective Transport

1

Paper 1

• Authors:– J. L. Deneubourg– S. Aron– S. Goss– J. M. Pasteels

• Appeared in the Journal of Insect Behavior, 1990

The Self-Organizing Exploratory Pattern of the Argentine Ant

2

Problem

• How do ants explore?– Rather, how is a single ant’s exploration

affected by the previous ants? Is it a function? Can we model it?

• Homogenous/heterogeneous agents?

Paper 1 CritiqueResultsExperime

ntApproachInsights

Problem Statement

3

InsightsProblem

• Ants explore with no fixed destination.

• They do this at night (so no visual cues).

• The Argentine ant lays her pheromone continuously (not just on return).

Paper 1 CritiqueResultsExperime

ntApproach

Insights

4

ApproachProblem

1) Observe the exploratory pattern.2) Reduce to a binary choice (diamond

bridge).3) Generate a model from observed data.– Does a Monte-Carlo model fit?– General choice function:

Paper 1 CritiqueResultsExperime

ntInsights

Approach

5

Experiment

Problem

• This experiment has two parts.

• Empty arena (no food or debris).

• Automatically photographed every 60 seconds.

• Sand periodically replaced.

Paper 1 CritiqueResultsApproachInsights

Experiment (Open Arena)

Note: This is an artist’s rendition of the experiment, as no image of the arena was provided by the authors. 6

Experiment

Problem

• The second part is more controlled.• Ants crossing bridge counted every

3-minutes.• Ants prevented from doubling back.

Paper 1 CritiqueResultsApproachInsights

Experiment (Diamond Bridge)

7

ResultsProblem

• Ants explore close to the nest first.

• The front advances, but leaves a trail.

• Number of explorers grows logistically.

• Picking out returning explorers halts exploration development.

• Ants will not ‘re-explore’ a well-explored area.

Paper 1 CritiqueExperime

ntApproachInsights

Results (Open Arena)

8

ResultsProblemPaper 1 CritiqueExperime

ntApproachInsights

Results (Open Arena)

9

ResultsProblem

• Both branches chosen equally at first.

• Positive feedback rapidly makes one path preferable.

• Ants act reactively (as a function of # ant passages).

Paper 1 CritiqueExperime

ntApproachInsights

Results (Diamond Bridge)

10

ResultsProblemPaper 1 CritiqueExperime

ntApproachInsights

Results (Diamond Bridge)

11

(Note: The axes on these graphs are not the same)

CritiqueProblem

• The model fits, but a lot of simplifications are required.

• Pheromone quantity estimated by number of ants passing (ignores evaporation, assumes each ant lays equal amount of pheromone).

• The ‘separated ants’ appear more dispersed in experiments than model.

Paper 1 ResultsExperime

ntApproachInsights

Critique

12

Paper 2Study of Group Food Retrieval by Ants as a Model for Multi-Robot

Collective Transport• Authors:– S. Berman– Q. Lindsey– V. Kumar–M. S. Sakar– S. C. Pratt

• Appeared in the Proceedings of the IEEE, 2011

13

Problem

• What is the role of each ant in collective transport? Rules that govern their actions?

• Can we apply this to robots who, like ants. have limited sensing, communication, and computation capabilities?

Paper 2 CritiqueResultsExperime

ntApproachInsights

Problem Statement

14

InsightsProblem

• Ants grab stuff in groups (better than robots do).

• The ant approach is decentralized, scalable, a requires no a priori information.

• Therefore, ants are more flexible and more robust than centralized approach.

• Prey transport teams are superefficient.

Paper 2 CritiqueResultsExperime

ntApproach

Insights

15

ApproachProblem

1) Observe ants in a controlled environment.

2) Develop a behavior model.3) Run a simulation to see if the model

matches.

Paper 2 CritiqueResultsExperime

ntInsights

Approach

16

Experiment

Problem

• Fabricate fake food (out of springs and fig paste) and measure the forces and deformations as ants carry it back to the nest (about 1 meter).

• 27 Trials

Paper 2 CritiqueResultsApproachInsights

Experiment

17

18

ResultsProblemPaper 2 CritiqueExperime

ntApproachInsights

Results (Observation)

• Quasi-static motion• More ants is better

(faster)• Load speed saturation

with increased group size

ResultsProblem

• Hybrid system with probabilistic transitions between two task modes:– search for grasp point– transport

• Start from uniformly randomly distributed positions and orientations

Paper 2 CritiqueExperime

ntApproachInsights

Results (Simulation)

19

CritiqueProblem

• Friction is a major factor which throws the deformation measures off.

• They even observe “stick-slip” motion.

Paper 2 ResultsExperime

ntApproachInsights

Critique

20

Closing Thoughts

• Both use ants as a model of homogenous agents and minimal communication.

• Both attempt to apply lessons from ants to distributed robotics.

• Both simulations use very simple models, while still being reasonably accurate.

21

Questions?

• The Self-Organizing Exploratory Pattern of the Argentine Ant

• Study of Group Food Retrieval by Ants as a Model for Multi-Robot Collective Transport

22