Collective intelligence: From animal societies to human groups · collective intelligence Basic ingredients of self-organization Self-organization processes in the Argentine antPositive

1

Collective intelligence: From animal societies to human groups

Guy Theraulaz Centre de Biologie Intégrative

Centre de Recherches sur la Cognition Animale CNRS, UMR 5169, Toulouse, France

The Graduate Center City University of New-York

February 28, 2019

Collective behavior in animal societies to human groups

Camazine, S. et al., Self-organization in Biological Systems, Princeton University Press (2001); Sumpter, D.J.T., Collective Animal behavior, Princeton University Press ((2010)

2

Collective behavior in human crowds

Gustave Le Bon (1841-1931)

Protesters march during a demonstration in Toulouse, France on December 29, 2018 © P. Pavani

Le Bon, G., The Crowd: A Study of the Popular Mind (French: Psychologie des Foules), Sparkling Books

edition. Sparkling Books (1896)

Collective motion in animal groups

Procaccini, A. et al., Anim. Behav. (2011); Ballerini, M. et al., PNAS (2008) Selous, E., Thought Transference ( or What? ) in Birds, (1931)

Sturnus vulgaris ( © M. Briola )

How do starlings coordinate their aerial choreography?

Edmund Selous (1857-1934)

3

How do social insects coordinate the construction of their nests?

Coordination of individuals activities

Apicotermes sp. Macrotermes bellicosus

Garnier, S. et al., Swarm Intell. (2007); Perna, A. & Theraulaz, G. , J. Exp. Biol. (2017) Singh, K. et al., Sci. Adv. (2019)

Procornitermes araujoi

The « spirit of the hive »

Maurice Maeterlinck (1862-1949)

1901 The life of the bee

© Ingo Arndt

4

The collective intelligence of super-organisms Complex systems with emergent properties

Sturnus vulgaris Macrotermes bellicosus

© Stéphane De Greef © Nick Dunlop

The collective intelligence of super-organisms Complex systems with emergent properties

5

Mathematical model

Collective level

Individual level

Quantifying and modeling individual behavior and social interactions

How do properties emerge at the group level?

Dell, A.I. et al., Trends Ecol. Evol (2014)

© Institute of Evolutionary Genetics, Düsseldorf

Inactive bee Active bee

QR code tags

Pérez-Escudero, A. et al., Nat. Methods (2014) Lecheval, V. et al., Proc. R. Soc. B (2018)

Hemigrammus rhodostomus

Automated tracking of animal movements

How do properties emerge at the group level?

6

Collective behavior in schooling fish

Analysis of social interactions between fish involved in the coordination of swimming

1 cm Hemigrammus rhodostomus

Calovi, D.S. et al., PloS Comp. Biol. (2018)

Attraction and alignment

Measurement and modeling of social interactions between pairs of fish


Distance Angular position Relative heading

7


Orientation relative Distance Angular position Relative heading

Attraction and alignment

Measurement and modeling of social interactions between pairs of fish


Modeling interactions between fish Comparisons between model predictions and experimental data

Model Experiment

Distance to the wall and distance between the two fish

Relative angle to the wall

Leader Follower

Experiment Model

Experiment Model

■  The model qualitatively and quantitatively reproduces the key features of the motion and spatial distributions of fish observed in experiments

8

Influence of the relative intensities of attraction and alignment on collective motion patterns

Phase 1 (« swarming »)

Calovi, D. et al., New J. Phys. (2014)

Modeling interactions in fish schools

Phase diagram

Phase 2 (« schooling »)


Modeling interactions in fish schools Influence of the relative intensities of attraction and alignment on collective motion patterns

Phase diagram

9

Phase 3 (« milling »)


Modeling interactions in fish schools Influence of the relative intensities of attraction and alignment on collective motion patterns

Phase diagram

Combination of interactions driving a fish school towards a “critical” state




Calovi, D. et al., J. R. Soc. Interface (2015)


Phase diagram

10




Calovi, D. et al., J. R. Soc. Interface (2015)

Combination of interactions driving a fish school towards a “critical” state


Phase diagram

Chartergus chartarius

Apis mellifera

Cubitermes fungifaber

Nasutitermes triodiae

Very elaborate nest architectures Collective nest building in social insects

11

Pierre-Paul Grassé (1895-1985)

1959

“An insect does not control his own work. But its ongoing activity is guided by the by-product of its work”

Grassé, P.-P., Insectes Sociaux (1959)

The stigmergy

■  Stigmergy occurs when insect’s actions are determined or influenced by the consequences of another insect’s previous action

■  This is a form of indirect communication that makes possible the coordination and regulation of insects activities

Response!

Stimulus! S1 S2 S3 S4 S5 Stop

time ■  This process leads to the coordination of the collective work and a

colony seems to follow a pre-defined plan

■  To understand how construction works one has to identify all indirect interactions that control the growth and shape of a nest

Theraulaz & Bonabeau, Artificial Life (1999)

The stigmergy Indirect communication

12

■  Wasp nests are built with wood pulp and plant fibers

■  Colored blotting paper used as building material makes it possible the visualization of successive building steps

■  Individual construction behavior can be studied in great details such as the wasps decisions to build a new cell in particular locations on the comb

Nest construction in Polistinae wasps

Nest building in social wasps A stigmergic behavior

The first construction steps in Polistes dominulus

Nest building in social wasps

13





14



Potential building sites on a comb

The control of nest building in wasps

■  The nest structure controls the organization of building activities

■  To decide where to build a new cell, wasps make use of the information provided by the local arrangement of cells on the comb

Karsaï, I & Theraulaz, G., Sociobiology (1995)

15

Construction rules in Polistes wasps Probability to build a new cell on the comb

■  Wasps have a greater probability to add new cells to a corner area (3 or 4 adjacent walls) than to initiate a new row by adding a cell on the side of an existing row (2 adjacent walls)

■  Wasps are modeled by asynchronous automata with a stimulus-response behavior

■  Virtual wasps move randomly in a 3-D discrete hexagonal lattice

■  Virtual wasps only have a local perception of their environment (the first 26 neighboring cells close the cell occupied by the wasp)

■ … and do not have any representation of the global architecture they build

Behavior of the virtual wasps

Modeling nest building in social wasps

Local neighborhood of the virtual wasp

Theraulaz, G. & Bonabeau, E., Science (1995)

16

1.00

1.00

1.00

1.00

0.05

0.50

1.00

Local construction rules

■  Some configurations of cells trigger the construction of a new cell

■  Construction rules are stochastic


Simulation results of the model

Polistes dominulus


© Alex Wild

Theraulaz, G. & Bonabeau, E., J. Theor. Biol. (1995)

17

0.05 0.5 1.0 0

Parapolybia varia ■  Encoding economy by

changing the execution probability of the building rules

1.0 1.0 1.0 0

Simulation results of the model


Parachartergus fraternus

Vespa crabro Chartergus chartarius

Collective nest building in social wasps

Nest architectures obtained by simulation of the model

18

From stigmergy to self-organization Trail recruitment in ants

■  To coordinate their movements when exploring a new territory or when foraging for food ants use chemical signals (pheromone trails)

■  The stigmergic interactions between ants lead to Amplification processes through positive feedback mechanisms

Bonabeau et al.,Trends Ecol. Evol. (1997)

■  Self-organizing processes allow social insects to develop a form of collective intelligence

Basic ingredients of self-organization

Self-organization processes

■  Positive feedbacks (amplification) promote the creation of structures

■  Negative feedbacks conterbalances positive feedbacks and helps to stabilize the collective pattern

Linepithema humile

Formation of an exploration network in the Argentine ant

Bonabeau et al.,Trends Ecol. Evol. (1997)

19

Goss et al., Naturwissenchaften (1989) Linepithema humile

Selection of the shortest route toward a food source

Collective decision-making in ants

■  Ants first use both paths in equal numbers, laying down pheromones as they move

■  Ants taking the shorter path return to the nest faster

■  The shorter path will then be doubly marked with pheromone and will thus be more attractive

■  Geometrical constraints play a key role in the collective decision-making processes that emerge at the colony level

Selection of the shortest route toward a food source

Collective decision-making in ants

Camazine, S. et al. Self-Organization in Biological Systems. Princeton University Press (2001)

20

Lasius niger

Nest construction in ants Nest architecture of the black garden ant Lasius niger

10 cm

~ 5.10 3 to 10 4 ants

Architectures without architect

Lasius niger

3D structure of a Lasius niger ant nest

21

1 cm Lasius niger

Construction dynamics

Real duration : 36 hours

Nest construction in ants

Rule 3 Probability of adding

a soil pellet on the side

Height of the pillar (mm)

Khuong et al., PNAS (2016)

Rule 2 Probability

of depositing a soil pellet

Nombre de boulettes déposées (n)

Probability of picking up a soil pellet

Rule 1

Number of deposited pellets (n)

Construction rules: picking up and dropping behaviors


Number of deposited pellets (n)

22

3D computational model of ant nest construction


Topochemical landscape


■  The pheromone is not homogeneously present onto the surface of the built structures

■  The spatial organization of pellets creates a topochemical landscape that determines the places at which ants concentrate their building activity

0 low high Pheromone density


23

Comparison between experiments and simulations

10 6 Time steps 5.10 2 Ants 2.10 5 Building

particles

η = 8.10 -4


Number of pillars

Model Experimental data



Inter-pillar distance

Distribution at 96h


Phenotypic plasticity of nest architecture

Dry environment Wet environment


24

Collective level

Individual level

Complex structures

Simple behavioral rules ■  Self-organization

processes simplify the encoding of behavioral mechanisms that make it possible the coordination of building activities

Properties

Self-organization processes

Camazine, S. et al., Self-organization in Biological Systems, Princeton University Press (2001)

Milgram, S., Bickman, L. & Berkowitz L. J. Pers. Soc. Psychol.(1969)

Social influence in human groups Effect of crowd size on Individual behavior

Size of stimulus crowd

Per

cent

of p

asse

rs b

y

Who look up

Who stop

Conformity and Independence, 1975 © Harper & Row

25

Milgram, S., Bickman, L. & Berkowitz L. J. Pers. Soc. Psychol.(1969)

Social influence in human groups Effect of crowd size on Individual behavior

■  Pedestrians assume that if lots of people are doing something, there must be a good reason why

■  A crowd becomes more influential as it becomes bigger

■  When things are uncertain, the best thing to do is just to follow along

Collective estimation tasks in human groups

■  Measuring the impact of social information on collective performance in estimation tasks

■  Groups of subjects are asked to estimate quantities about which they have low prior knowledge (186 subjects in France, 180 subjects in Japan)

■  Low demonstrability questions (number of stars in the galaxy, size of the population of Madrid, number of marbles in a jar, number of books in books does the American Congress library…)

Experimental design

Jayles, B. et al., PNAS (2017)

26

Estimation without social information

Collective estimation tasks in human groups Experimental design


Ep

Estimation after social influence

E

How many stars does the Milky way hold (in millions of stars)?

Ep1

Ep2

Ep3

■  Subjects had to first provide their personal estimate Ep in a limited amount of time (25s)

■  Then, after receiving the social information (geometric mean of some previous estimates) subjects were asked to give a new estimate E

Ep1

Ep2

T

T

Ep1

T

T

T

T

Estimation without social information



■  Social information was controlled by introducing a controlled number of « virtual experts » (giving the right answer), without the subjects being aware of it

Ep

Estimation after social influence

E

T : true value

27

■  Precise control of the quantity and quality of the social information exchanged between subjects

■  Human groups are often composed of individuals with heterogeneous expertise



■  To compare the quantities that can differ by several orders of magnitude, each estimate E is normalized by the true answer T to the question and we used the log-transformed estimate X = log (E/T)

Collective estimation tasks in human groups Distribution of individual estimates

Galton, F. Nature (1907); Jayles, B. et al., PNAS (2017)

Before SI After SI

Francis Galton (1822-1911)

28

■  To compare the quantities that can differ by several orders of magnitude, each estimate E is normalized by the true answer T to the question and we used the log-transformed estimate X = log (E/T)

■  Social information allows a group to collectively improve its performance and the accuracy of its estimates

■  When there is no virtual expert the group is able to improve its collective performance after social influence

Collective estimation tasks in human groups Distribution of individual estimates


Before SI After SI

Without virtual

experts Before SI After SI

■  The sensitivity to social influence Si can be uniquely defined by:



Distribution of individual sensitivities to social influence

New estimate

Personal estimate

Social information

■  The influence of the social information on individuals’ decisions is not uniform

29

■  5 types behavioral responses:



Distribution of individual sensitivities to social influence

u  keeping the initial estimate (S = 0)

u  adopting that of others (S = 1)

u  making a compromise between one’s opinion and that of the group’s (0 < S < 1)

u  amplifying the social information (S > 1)

u  contradicting the social information (S < 0)

■  The sensitivity of subjects to social influence increases with the difference (D) between their (log transformed) personal estimates and the social information (M) they receive

■  The subjects that are the least sensitive to social information are also those whose estimates are closest to the true values



Impact of distance between personal’s and group’s opinion on sensitivity to social influence

Poor High

30

■  A model built and calibrated from the experimental observations quantitatively predicts the improvement in collective performance and collective accuracy as the amount of good information provided by virtual experts to the group increases



Impact of the proportion of virtual experts on collective performance and accuracy

Collective accuracy median(⎟Xi⎟)

Collective performance ⎟median(Xi)⎟

Experience Model

Before social influence After social influence

Proportion of experts

Experience Model

Before social influence After social influence

Proportion of experts

Low performance

High performance

Low accuracy

High accuracy

■  Knowing how individuals respond to social information in animals societies and human groups allows us to understand how this information helps a group of individuals to coordinate their actions and improve their collective performance

Collective intelligence in animal and human groups Conclusions

31

■  Understanding how social information influences individuals’ decisions can help us to determine the conditions under which collective opinions can come closer to the true values

■  In human groups understanding these behavioral mechanisms can help us to develop algorithms to adapt the information delivered to individuals to make optimal collective choices

■  Combining human and machine intelligence to enhance collective intelligence in crowds

Conclusions Usingsocialinforma.ontomakerecommenda.ons

Social Information

Collective intelligence in animal and human groups

Centre de Recherches sur la Cognition Animale, Centre de Biologie Intégrative, CNRS,

Université Paul Sabatier, Toulouse

Service de Physique de l’Etat Condensé, CEA, Saclay

Laboratoire de Physique Théorique Université Paul Sabatier, Toulouse

Acknowledgements

D.S. Calovi V. Lecheval A. Perna J. Gautrais C. Jost A. Khuong R. Escobedo

H. Chaté C. Sire B. Jayles H.-R. Kim T. Kameda A. Blanchet

Department of Behavioral Science, Hokkaido University, Sapporo, Japan

Department of Social Psychology, The University of Tokyo, Japan

Toulouse School of Economics, France

Collective intelligence: From animal societies to human groups · collective intelligence Basic ingredients of self-organization Self-organization processes in the Argentine antPositive

Documents