evolutionary conditions for the emergence of communication in robots dario floreano, sara mitri,...
TRANSCRIPT
Evolutionary conditions for theemergence of communication in robots
Dario Floreano, Sara Mitri, Stephane Magnenat, and Laurent KellerCurrent Biology, vol. 17, no. 6, pp. 514-519, 2007.
2010. 05. 11Jongwon Yoon
Contents
• Introduction
• Evolution of multiagent systems in robotics
• Overview
• Experimental setup– Robots
– Foraging arena
– Neural controller
– Evolution process
• Data analysis
• Experimental results
• Conclusion
1/13
Introduction
• Information transfer & communication systems– Plays a central role in the biology of most organisms, particularly social
species– Extremely sophisticated in large and complex societies– Key component ensuring the ecological success of highly social species
• Evolution of communication– Efficient communication requires tight coevolution between the signal
emitted and the response elicited– Conditions and paths remain largely unknown
• Contributions of this study– Predict about the evolutionary conditions conductive to the emergence of
communication– Provide guidelines for designing artificial evolutionary systems
2/13
Evolution of multiagent systems in ro-botics
Author Target Year
Team composition Level of selection
Hetero-geneous
Homo-geneous
Individual Team
S. Raik and B. Durnota Behavior 1994 O O
S. Luke and L. Spector Behavior 1996 O O
S. G. Ficicici et al. Behavior 1999 O O
A. S. Wu et al. Behavior 1999 O O
A. Martinoli Behavior 1999 O O
M. Quinn Behavior 2001 O O O
E. Simoes and D. Barone Behavior 200
2 O O
L. Steels Communica-tion
2003 O O
L. Spector et al. Behavior 2005 O O
M. Mirolli and D. Parisi Communica-tion
2005 O O
V. Trianni et al. Communica-tion
2006 O O
3/13
Overview
• Purpose– Studying the evolution of communication
• Consideration of the kin structure of groups (Relatedness)• The scale at which cooperation and competition occur (Level of selection)
• Experiments overview– Colonies of robots forage in an environment
• Containing a food and a poison– Use 100 colonies of 10 robots– Selection experiments over 500 generations
• By using physics-based simulations
4/13
Robots Experimental setup
• Equipments– Two tracks : Independently rotate in both direc-
tions– Translucent ring : Emit blue light– 360 degree vision camera– Infrared ground sensors
• Sensory-motor cycle– Length : 50ms
• Use a neural controller to process visual informa-tion and ground-sensor input
• Set direction and speed of the two tracks• Control the emission of blue light
• Performance unit– Gain one unit : if it detected food– Lost one unit : if it detected poison
• 1 Trial = 1200 sensory-motor cycles * 50ms = 1min 5/13
Foraging arena Experimental setup
• Size : 300cm x 300cm (Robots are placed randomly)• A food and a poison source
– Radius : 10cm– Placed at 100cm from one of two opposite corners– Constantly emit red light– Circular gray and black papers
• Placed under the food and the poison• Robots detect by infrared ground sensors
6/13
Neural controller Experimental setup
• Evolutionary Neural network– Feed-forward neural network– Ten inputs & three outputs
• Genetic encoding– Encoded the synaptic weights of 30 neu-
ral connections– Each weight was encoded in 8bits, giv-
ing 256 values mapped onto the interval [-1, 1]
– Total length : 8bits x 3 inputs x 10 out-puts
= 240 bits
7/13
Evolutionary process Experimental setup
• Population– 100 colonies x 10 robots in each colony = Total 1000 robots– 20 independent selection lines (replicates)
• Selection– Four treatments
• Colony-level / High relatedness• Individual-level / High relatedness• Colony-level / Low relatedness• Individual-level / High relatedness
• Recombination– Crossover rate : 0.05 (5%)– Mutation rate : 0.01 (1%)
8/13
Data analysis
• Performance– Average performance of the 100 colonies over the last 50 generations– Compared with nonparametric (Kruskal-Wallis and Mann-Whitney) tests
• Some of the data did not follow a normal distribution
• Signaling strategy
– NF / NP : Total number of cycles spent near the food / the poison– bF
rn / bPrn : Whether robot r was emmiting light at cycle n near the food or
poison
• Tendency– The tendency of robots to be attracted by light
– ar : Decrease in the distance as attraction
– vr : Increase in the distance as avoidance
9/13
Experimental results
• Performance
• Performance comparison
10/13
Experimental results (cont.)
• Strategy comparison– Produce light in the vicinity of the food : 12 / 20– Produce light in the vicinity of the poison : 8 / 20– The communication strategy where robots signaled near the food re-
sulted in higher performance (259.6 ± 29.5) than the strategy of produc-ing light near the poison (197.0 ± 16.8)
• Signaling near the food while they feed• Food signal can easily be detected by other robots
• Tendency comparison– Attracted to the light : 12 / 12– Repelled by the light : 7 / 8
11/13
Experimental results (cont.)
12/13
Conclusion
• Cooperative communication and deceptive signaling can evolve
• Communication readily evolves when ..– Colonies consist of genetically similar individuals– Selection acts at the colony level
• May constrain the evolution of more efficient communication sys-tem
– Communication between signalers and receivers can be perturbed– Evolved biological systems can be maintained despite their suboptimal
nature
• Evolutionary principles are demonstrated– Can be useful for designing efficient groups of cooperative robots
13/13