CIplus
Band 5/2017
Is Social Lea ning Mo e Than Pa ame e
Tuning?
Jacqueline Heine man, Jö g S o k, Ma ga i a Alejand a Rebolledo Coy,
Julien Hube , A.E. Eiben, Thomas Ba z-Beiels ein, E e Haasdijk
Is Social Lea ning Mo e Than Pa ame e Tuning?
Jacqueline Heine man
V ije Uni e si ei Ams e dam
Ams e dam, The Ne he lands
j. [email p o ec ed]
J¨
o g S o k
TH K¨
oln
K¨
oln, Ge many
Ma ga i a Alejand a Rebolledo
Coy
TH K¨
oln
Julien Hube
V ije Uni e si ei Ams e dam
A.E. Eiben
V ije Uni e si ei Ams e dam
Thomas Ba z-Beiels ein
TH K¨
oln
E e Haasdijk
V ije Uni e si ei Ams e dam
CCS CONCEPTS
•Compu ing me hodologies →E olu iona y obo ics;
Mul i-
agen sys ems;
•Theo y o compu a ion →E olu iona y algo-
i hms;
KEYWORDS
E olu iona y obo ics; social lea ning; neu al ne wo ks; pa ame e
uning
ACM Re e ence o ma :
Jacqueline Heine man, J
¨
o g S o k, Ma ga i a Alejand a Rebolledo Coy,
Julien Hube , A.E. Eiben, Thomas Ba z-Beiels ein, and E e Haasdijk.
2017. Is Social Lea ning Mo e Than Pa ame e Tuning?. In P oceedings o
GECCO ’17 Companion, Be lin, Ge many, July 15-19, 2017, 2 pages.
DOI: h p://dx.doi.o g/10.1145/3067695.3076059
1 INTRODUCTION
Social lea ning enables mul iple obo s o sha e lea ned expe iences
while comple ing a ask. The li e a u e o e s examples whe e obo s
ained wi h social lea ning each a highe pe o mance compa ed
o hei indi idual lea ning coun e pa s [e.g,
2
,
4
]. No explana-
ion has been ad anced o ha obse a ion. In his esea ch, we
p esen expe imen al esul s sugges ing ha a lack o uning o he
pa ame e s in social lea ning expe imen s could be he cause. In
o he wo ds: he be e he pa ame e se ings a e uned, he less
social lea ning can imp o e he sys em pe o mance.
To es ou hypo hesis, we gene a ed 50 pa ame e se ings us-
ing Design o Expe imen s (DoE) and es ed hem in an indi idual
lea ning con igu a ion wi h a single obo (i.e., wi h social lea n-
ing disabled). The de ini i e sc eening DoE was c ea ed wi h he
help o JMP so wa e (SAS Ins i u e Inc, JMP, Ve sion 11.1.0). The
expe imen s a e conduc ed in simula ion using JBo E ol e [
1
].
The expe imen equi es he obo s o lea n a o aging ask. The
en i onmen is a squa e a ena. Fi e pucks a e andomly placed in
he a ena a he s a o a un. The obo s mus collec he pucks
and b ing hem o he nes loca ed in he cen e o he a ena. Once
Pe mission o make digi al o ha d copies o pa o all o his wo k o pe sonal o
class oom use is g an ed wi hou ee p o ided ha copies a e no made o dis ibu ed
o p o i o comme cial ad an age and ha copies bea his no ice and he ull ci a ion
on he i s page. Copy igh s o hi d-pa y componen s o his wo k mus be hono ed.
Fo all o he uses, con ac he owne /au ho (s).
GECCO ’17 Companion, Be lin, Ge many
©2017 Copy igh held by he owne /au ho (s). 978-1-4503-4939-0/17/07...$15.00
DOI: h p://dx.doi.o g/10.1145/3067695.3076059
a puck is b ough o he nes , i is immedia ely eplaced a a andom
loca ion in he en i onmen . The pe o mance o each obo , i.e. i s
i ness, is equal o he numbe o pucks i collec ed du ing a ial
las ing 1000 ime s eps. The obo s use embedded ins ances o he
NEAT e olu iona y algo i hm o on-line lea ning [
3
]. The obo ’s
con olle is an a i icial neu al ne wo k. The neu al ne wo k has 11
inpu and wo ou pu nodes. The inpu nodes consis o 8 p oximi y
senso s, a nes senso , a puck senso , and a puck ca ying senso ;
he ou pu nodes p o ide he igh and le mo o speed.
F om he 50 pa ame e se ings o he DoE, we selec ed he 10
se ings wi h bes pe o mance and 10 wi h median pe o mance.
We compa ed he pe o mance o hese wo g oups o se ings,
whe e social lea ning is disabled, wi h wo social lea ning con igu-
a ions o 2 and 4 obo s. Social lea ning is implemen ed as ollows:
i s , he obo s sequen ially e alua e he con olle s in he cu en
gene a ion. Then, he obo s exchange in o ma ion. Each obo
andomly selec s ano he obo om which i ecei es i s cu en
bes con olle , i.e. he con olle wi h he highes i ness. The obo
compa es he ecei ed con olle ’s i ness o ha o i s own wo s
con olle . The new con olle eplaces he wo s con olle i i is
be e . The NEAT algo i hm uses he upda ed lis o con olle s and
i ness alues o c ea e he nex gene a ion. When social lea ning
is applied, hese obo s ha e he same pa ame e se ings as he
indi idual lea ning mechanisms excep o he popula ion size. The
popula ion size o he 2 and 4 obo se up is he popula ion size
om he 1 obo se up di ided by he numbe o obo s (e.g., when
he 1 obo se up has a popula ion size o 100, he social lea ning
expe imen s used a popula ion size o 50 and 25 o he 2 obo and
4 obo se up espec i ely).
The obo s ope a e in hei own a ena bu hey communica e
ac oss a enas. Consequen ly, he pe o mance o he obo is only
due o i s own ac ions and no in luenced by o he obo s in he
same a ena. Remo ing his in e obo collision allows o a be e
compa ison be ween he indi idual and he social lea ning expe i-
men s. Fo each expe imen , 20 eplica e uns a e pe o med wi h
di e en andom seeds.
2 RESULTS
Ou analysis shows ha om he o al o 21 in es iga ed pa ame e s
om he NEAT lea ning algo i hm, 14 pa ame e s ha e a signi ican
impac on he lea ning pe o mance. The a iable ep esen ing he
chance o andomly ese a weigh in he neu al ne wo k has a
GECCO ’17 Companion, July 15-19, 2017, Be lin, Ge many J. Heine man e al.
Figu e 1: Mean pe o mance wi h 95% con idence in e al o
he baseline expe imen s o all DoE pa ame e se ings. The
y-axis shows he pe o mance, measu ed as he numbe o
collec ed pucks. The x-axis shows he ank o he pa ame e
se ing. The esul s a e compiled o e 20 eplica e uns.
big e ec on he esponse and pe o mance is bes i i is u ned
o . Mo eo e , a la ge popula ion size, a high mu a ion p obabili y
and a small alue o he pe cen age o op indi iduals ha can be
pa en s ha e a signi ican bene icial in luence on he i ness. This
indica es ha a la ge and di e se popula ion o con olle s seem o
be ad an ageous o he lea ning a e.
Figu e 1 shows he mean pe o mance o he 1 obo pe o -
mance a he inal gene a ion (gene a ion 200). Pe o mance, i.e.,
numbe o collec ed pucks, is plo ed agains he ank o he mean
pe o mance o he 50 pa ame e se ings om he DoE. The da a
in igu e 1 con i ms ha pa ame e se ings signi ican ly in luence
he pe o mance o he con olle s (Pea son’s (50)= -0.9754128, p
<2.2e-1).
Figu e 2 shows he impac o social lea ning on pe o mance o
he bes and median pa ame e se ings. Ranks 1-10 e e o he
bes se ings and anks 21-30 o he median se ings. E e y se ing
is es ed o a se up wi h 2 ( ed) and 4 (blue) obo s in 20 eplica e
uns. The impac o social lea ning is measu ed as he a io be ween
he pe o mance wi h social lea ning and he baseline pe o mance.
A a io highe han 1 means an inc ease o pe o mance due o
social lea ning. The e is a signi ican posi i e co ela ion be ween
he ank numbe and he imp o emen a io wi h social lea ning
( o 2 obo s Pea son’s (20) = .56, p
<
0.011; o 4 obo s Pea son’s
(20) = .63, p
<
0.003). Figu e 2 shows ha be e pa ame e se ings
bene i less om social lea ning, indica ed by pe o mance a io
alues a ound 1.
3 CONCLUSIONS
Wi h his s udy we gained a be e unde s anding o he bene i s o
social lea ning. Exis ing li e a u e in social lea ning compa es indi-
idual lea ning wi h social lea ning o only one pa ame e se ing.
Figu e 2: Pe o mance a io o 2 (blue) and 4 ( ed) obo s
compa ed o he baseline expe imen s. The y-axis shows he
pe o mance a io, measu ed as he mean pe o mance o
2 espec i ely 4 obo s di ided by he baseline pe o mance.
The x-axis shows he ank o he pa ame e se ing ( ank 1
is he ank esul ing in he highes pe o mance).
Resul s show an inc eased pe o mance and inc eased lea ning
speed. This s udy ex ended his compa ison by using di e en pa-
ame e se ings. We showed ha he quali y o he pa ame e
se ings in luences how much social lea ning can imp o e he sys-
em pe o mance: he be e he pa ame e se ings, he less social
lea ning can con ibu e. The e o e, his s udy se es as a eminde
ha uning he pa ame e s can impac he conclusions d awn om
an expe imen . Ne e heless, uning can be compu a ionally ex-
pensi e o no e en possible when he op imal pa ame e s may
depend on he unknown en i onmen ha he obo s ope a e in.
The e o e, social lea ning can be a bene icial app oach o inc ease
pe o mance and se e as an al e na i e o pa ame e uning.
4 ACKNOWLEDGMENTS
This wo k is suppo ed by he Eu opean Union’s Ho izon 2020
esea ch and inno a ion p og amme unde g and ag eemen No
640891 (DREAM p ojec ).
REFERENCES
[1]
Miguel Dua e, Fe nando Sil a, Tiago Rod igues, Sancho Mou a Oli ei a, and
Ande s Lyhne Ch is ensen. 2014. JBo E ol e : A e sa ile simula ion pla o m
o e olu iona y obo ics. In P oceedings o he 14 h In e na ional Con e ence
on he Syn hesis & Simula ion o Li ing Sys ems. MIT P ess, Camb idge, MA,
210–211.
[2]
Jacqueline Heine man, Massimiliano Rango, and A.E. Eiben. 2015. E olu ion, In-
di idual Lea ning, and Social Lea ning in a Swa m o Real Robo s. In P oceedings
o he 2015 IEEE In e na ional Con e ence on E ol able Sys ems (ICES). IEEE P ess,
New Yo k, NY, USA, 1055–1062. DOI:h p://dx.doi.o g/10.1109/SSCI.2015.152
[3]
Kenne h O S anley and Ris o Miikkulainen. 2002. E ol ing neu al ne wo ks
h ough augmen ing opologies. E olu iona y compu a ion 10, 2 (2002), 99–127.
[4]
Wesley Tansey, Eliana Feasley, and Ris o Miikkulainen. 2012. Accele a ing e o-
lu ion ia egali a ian social lea ning. In P oceedings o he 14 h annual con e ence
on Gene ic and e olu iona y compu a ion, Te ence Soule (Ed.). ACM, New Yo k,
NY, USA, 919–926.
Kon ak /Imp essum
Diese Ve ö en lichungen e scheinen im Rahmen de Sch i en eihe "CIplus". Alle Ve ö -
en lichungen diese Reihe können un e
h ps://cos.bibl. h-koeln.de/home
abge u en we den.
Die Ve an wo ung ü den Inhal diese Ve ö en lichung lieg beim Au o .
Da um de Ve ö en lichung: 30.08.2017
He ausgebe / Edi o ship
P o . D . Thomas Ba z-Beiels ein,
P o . D . Wol gang Konen,
P o . D . Bo is Naujoks,
P o . D . Ho s S enzel
Ins i u e o Compu e Science,
Facul y o Compu e Science and Enginee ing Science,
TH Köln,
S einmülle allee 1,
51643 Gumme sbach
u l:
www.ciplus- esea ch.de
Sch i lei ung und Ansp echpa ne / Con ac edi o ’s office
P o . D . Thomas Ba z-Beiels ein,
Ins i u e o Compu e Science,
Facul y o Compu e Science and Enginee ing Science,
TH Köln,
S einmülle allee 1, 51643 Gumme sbach
phone: +49 2261 8196 6391
u l:
h p://www.spo se en.de
eMail: homas.ba z-beiels ein@ h-koeln.de
ISSN (online) 2194-2870
