case study: better stay connected… or not?

Post on 21-Dec-2014

192 Views

Category:

Documents

2 Downloads

Preview:

Click to see full reader

DESCRIPTION

Benefits and limits of distributed intelligence! wrt. ecological diversity in the environment

TRANSCRIPT

Nicolas Bredeche !Université Pierre et Marie Curie Institut des Systèmes Intelligents et de Robotique ISIR, UMR 7222 Paris, France nicolas.bredeche@upmc.fr

FoCAS summer school (Crete), 23/6/2014

benefits and limits of distributed intelligence!wrt. ecological diversity in the environment

Case study: better stay connected… or not?

nicolas.bredeche@upmc.fr

!2

Question !

What about adaptation to an open environment?

nicolas.bredeche@upmc.fr

Open environments

• behaviors: generalists or specialists ?

• optimizer: centralized or distributed ?

!3

J.M. Turner, 1813

Applications: robots in the real world, video games, simulation, … internet of things, …

nicolas.bredeche@upmc.fr

!4

Hypothesis !

Distributed adaptation can be beneficial !in « rich » (spatial) environments

Case study: is this hypothesis true or false?

nicolas.bredeche@upmc.fr

Interaction between the population and the environment

• Very homogeneous environment • All can display the same behavior • Expected: centralized is best

• Very heterogeneous environment • Only specialist are allowed (e.g. limitations wrt. the metabolism) • Expected: distributed/specialist is best

• Inbetween • …?

!5

nicolas.bredeche@upmc.fr

Expected result !6

environment diversity

perfo

rman

ce

distributed (situated)

centralized

distributed (well-mixed)

nicolas.bredeche@upmc.fr

Expected result !7

environment diversity

perfo

rman

ce

distributed (situated)

centralized

distributed (well-mixed)

?

?

?

?

?

?

Methods

nicolas.bredeche@upmc.fr-nicolas.bredeche@isir.upmc.fr

Decoding Evaluation

!9

Initial Population"(random solutions)

Evaluation Selection Variations Replacement

desc

ript

ion fitness

continue stop end.

Evolutionary Computation with Robots

nicolas.bredeche@upmc.frnicolas.bredeche@isir.upmc.fr

Decoding Evaluation

!10

Initial Population"(random solutions)

Evaluation Selection Variations Replacement

desc

ript

ion fitness

continue stop end.

simulation setup!robots are situated in the environment!

no reset between generations

nicolas.bredeche@upmc.frnicolas.bredeche@isir.upmc.fr

Decoding Evaluation

!11

Initial Population"(random solutions)

Evaluation Selection Variations Replacement

desc

ript

ion fitness

continue stop end.

centralized vs. distributed!selection can be done wrt. robot location / behavior

nicolas.bredeche@upmc.fr

Roborobo (C++) !12

nicolas.bredeche@upmc.fr

Roadmap (tentative)

• Experimental setup : foraging ? • all agents in one environment, synchronized generation • mutation-only • selection schemes: ‣ global: (mu+lambda), (mu,lambda) ‣ local: (mu,1), (mu-1,1) (…?)

!

• Guidelines • homogeneous vs. heterogeneous environment • enforced specialist vs. possible generalist ‣ e.g.: genetically-coded metabolic function forces specialists

!13

nicolas.bredeche@upmc.fr

Roadmap (tentative)

• Open questions • dispersion and lifetime? ‣ longer life means more dispersion (ie. converge to well-mixed)

‣ vanilla version: simulate well-mixed by randomizing partners

• selection scheme for global approach? ‣ elitist vs. non-elitist schemes

• cooperation based on relatedness? ‣ low dispersion may favor altruistic cooperation

• decentralized as a key to complementary skills ‣ « more than the sum of its parts »

‣ What happen if cooperation « create » more energy (e.g. energy merging)

!14

Wrapping up

nicolas.bredeche@upmc.fr

Wrapping up

• Important question • decentralized: a constraint, or a feature?

!

• Possible audience for this contribution (if publication) ‣ biologists (limited dispersion as a winning strategy) ‣ robotics (on-line distributed learning can make things easier) ‣ general audience (distributed intelligence can be more creative)

!16

top related