AI-Powe ed Da a Syn hesis o Ad anced
Simula ion in 5G/6G mmWa e In eg a ed Access
and Backhaul Ne wo ks
Ami Ash a i Ga ga i∗, Fa had Rezazadeh∗, Ma co Gio dani†, Sand a Lag´
en∗, Lingjia Liu‡,
And a Lu u§, Michele Zo zi†
∗Cen e Tecnol´
ogic de Telecomunicacions de Ca alunya (CTTC/CERCA), Ba celona, Spain,
email: [email p o ec ed]
‡Wi eless Resea ch G oup, Vi ginia Tech, Blacksbu g, USA, email: [email p o ec ed]
§Tele ´
onica Resea ch, Mad id, Spain, email: [email p o ec ed]
†Depa men o In o ma ion Enginee ing, Uni e si y o Pado a, I aly, email: [email p o ec ed]
Abs ac —In eg a ed Access and Backhaul (IAB) is a cos -
e ec i e and adap able solu ion o he deploymen o ul a-
dense nex -gene a ion (5G and 6G) cellula ne wo ks o inc ease
he likelihood o Line-o -Sigh (LOS) co e age. This echnology
allows wi eless backhaul connec ions o be es ablished using
he same echnology and speci ica ions as a ailable in he
access links. Howe e , he absence o a physical es bed o
a da ase ha can be used o simula ion in he millime e
wa e (mmWa e) band p e en s esea che s’ alida ion o he
p oposed algo i hms in he IAB scena io. In his pape , we
p opose a no el da a gene a o based on Gene a i e Ad e sial
Ne wo k (GAN), ained on a eal da ase om a mobile ne wo k
ha ope a es in Eu ope, and main ains a signi ican ma ke
sha e ha e u ns accu a e a ic da a o an IAB ne wo k.
Fu he mo e, we in eg a e his da a gene a o wi h he SeBaSi
simula o (an IAB simula o based on Sionna) which pe mi s
o ob ain accu a e, da a-consis en , ealis ic, and end- o-end
IAB simula ion esul s. The pe o mance esul s indica e ha
he da a gene a o success ully passes he Kolmogo o –Smi no
(KS) c i e ion, so i could ope a e as a e i ied da a gene a o .
Fu he mo e, we use he SeBaSi simula o , in eg a ed wi h he
da a gene a o , o e alua e he pe o mance o an IAB ne wo k
in he London Ci y scena io.
Index Te ms—6G, GAN, IAB, Sel -backhauling, Wi eless
Backhaul, GAN, Da a Gene a ion
I. INTRODUCTION
The signi ican inc ease in da a a e capaci y a sub-
e ahe z (THz) and millime e wa e (mmWa e) equencies
ha enable da a-hung y use cases such as Ex ended Reali y
(XR) and mobile me a e se applica ions in 5 h gene a ion
(5G) and 6 h gene a ion (6G) cellula sys ems. To mi iga e
he e ec s o he se e e p opaga ion en i onmen a highe
equencies, wi eless ne wo ks will be deployed wi h an
excep ionally high densi y o inc ease he p obabili y o Line-
o -Sigh (LOS) co e age. The 3 d Gene a ion Pa ne ship
P ojec (3GPP) has s anda dized an ex ension o 5G New
Radio (NR), known as In eg a ed Access and Backhaul
(IAB) [1], o make ul a-dense deploymen s mo e easible and
economically sus ainable. This ex ension educes he equi ed
numbe o ibe d ops by u ilizing he same wa e o m and
p o ocol s ack o p o ide wi eless backhaul o Nex Gene -
a ion Node Bases (gNBs), which a e he IAB nodes [2].
IAB is a wi eless backhauling echnique ini ially sugges ed
in 3GPP Release 16 [3]. By enabling sel -backhauling, i
p o ides a p ac ical solu ion o he challenges encoun e ed
by dense mobile ne wo ks. This echnique employs a chain
s uc u e in which nume ous IAB nodes a e ul ima ely linked
o an IAB dono , combining he esou ces o access [4]
and backhaul links in he gNB (an example o his ne -
wo k is shown in Fig. 1). The 3GPP NR Release 18 [1]
cu en ly includes a Wo k I em (WI) ha is dedica ed o
he examina ion o a chi ec u es, adio p o ocols, and he
physical laye o IAB. The objec i e is o acili a e he
sha ing o adio esou ces be ween access and backhaul
links. This WI on IAB an icipa es a mo e sophis ica ed and
adap able solu ion, ea u ing dynamic esou ce mul iplexing,
mul ihop communica ions, and a plug-and-play design o
low-complexi y deploymen s.
Nume ous esea ch pape s ha e in es iga ed he de elop-
men o IAB ne wo ks o bo h backhaul and access links,
ocusing especially on he iden i ica ion o op imal solu ions
o esou ce managemen [5], [6]. Addi ionally, IAB can
u ilize a signi ican ly g ea e bandwid h a mmWa es han in
legacy sub-6 GHz spec um, and he inhe en di ec ionali y
a his equncies mi iga es he in e e ence o concu en
access and backhaul ansmissions. Ne e heless, he design
o a high-pe o mance IAB ne wo k emains an open esea ch
challenge, despi e he consensus ega ding IAB’s capaci y o
educe deploymen and managemen cos s. Mul i-Hub (MH)
backhauling, an impo an aspec o IAB echnology, has he
po en ial o enhance ne wo k h oughpu and co e age [7],
[8]. In o de o e ec i ely add ess he blockage issue and
in e e ence managemen in an IAB ne wo k, i is c ucial
o le e age spa ial euse h ough MH backhauling [9], [10].
So, he IAB ne wo k design and in pa icula backhaul
ne wo k is an impo an and complica ed p oblem because i
in ol es managing he ne wo k’s opology, choosing ou es,
and adjus ing he esou ces ha a e sha ed be ween backhaul
and access links [11]. The MH-IAB ne wo k deploymen and
con igu a ion a e lexible, which can p esen addi ional chal-
lenges. Consequen ly, i is impo an o implemen an e icien
ne wo k design o he access and backhaul connec ions o
MH-IAB, which should conside he in e e ence gene a ed
by in-band wi eless backhaul and blockage issues.
P io o ac ual deploymen , i is essen ial o alida e all
sugges ed me hods o IAB, including IAB nodes and dono
placemen and backhaul schedule s [12]. In his sense, gi en
he absence o publicly a ailable es beds o expe imen al
se ups, ne wo k simula ion is a mo e easonable choice o
he pe o mance e alua ion o IAB ne wo ks. Among o he
solu ions, he ns3-mmwa e-IAB1[10] module is buil on op
o he ns-3 simula o , and can be used o implemen , design,
dimension and e alua e end- o-end IAB ne wo ks. Howe e ,
he module had no been upg aded o suppo he newes 5G-
NR s anda d speci ica ions, and was incapable o handling
he simula ion o la ge-scale ne wo k deploymen s. Recen ly,
we in oduced a simula o called Sel -Backhauling-Simula o
(SeBaSi) [13], which is publicly a ailable2and has he abili y
o eplica e an IAB ne wo k wi h much s onge connec ions
o he 3GPP s anda d han he cu en e sion o he ns3-
mmwa e-IAB module. Ne e heless, he design o SeBaSi has
no been ained o alida ed wi h eal da a om IAB ne wo k
deploymen s, which could lead o disc epancies be ween
simula ion esul s and eal a ic pa e ns and deploymen s.
Because he numbe o eal da ase s is ex emely limi ed, i
may be di icul o model a la ge numbe o scena ios and o
ain enough gene ali y [14]. The e o e, i is c ucial o lea n
om he eal and a ailable da ase s, and eplica e simila da a
in mo e di e en and he e ogeneous scena ios.
In his con ex , GAN has become a popula echnique o
c ea e da ase s s a ing om eal da a, including da a om
wi eless communica ion ne wo ks [15]. Gene a i e Ad e sial
Ne wo ks (GANs) is a unique ype o Deep Neu al Ne -
wo k (DNN) ha can p oduce da a by acqui ing he p ecise
s a is ical cha ac e is ics o a gi en da ase h ough indi ec
me hods. In his esea ch, we p opose a new eal da ase
gene a o based on GAN o he IAB ne wo k. This gene a o
can p oduce ealis ic IAB aces in a ious scena ios and has
been ained using a eal da ase . The s a is ical e alua ion,
based on he Kolmogo o –Smi no (KS) es , demons a es
ha he gene a o and he ac ual da ase a e compa able. We
in eg a e he p oposed gene a o in o he SeBaSi simula o
o enable ull-s ack simula ion o MH-IAB ne wo ks using
eal da a, which gua an ees he accu acy and ealism o he
esul s. Fo ins ance, we demons a e he pe o mance o an
MH-IAB deploymen in London Ci y and p esen he main
simula ion me ics and esul s.
The es o he pape is o ganized as ollows. Sec. II
in oduces he IAB sys em model and a summa y o he
SeBaSi simula o . Sec. III desc ibes he p oposed da a gen-
e a o amewo k. Sec. IV demons a es s a is icaly esul s
1h ps://gi hub.com/signe labdei/ns3-mmwa e-iab
2h ps://gi hub.com/TUDA-wise/sa ehaul in ocom2023
BSnode
BSdono
UE
Access
Link
Backhaul
Link
Fig. 1: Illus a ion o he IAB scena io (wi h backhaul and access
link), wi h NI= 3 IAB nodes.
and nume ical esul s o he London Ci y scena io. Finally,
Sec. V concludes he pape .
II. IAB SYSTEM MODEL
This sec ion p o ides an o e iew o he sys em model and
o ou assump ions in Sec. II-A, as well as a b ie in oduc ion
o SeBaSi in Sec. II-B.
A. Sys em Model
We conside a Time Di ision Mul iple Access (TDMA)
sys em, as illus a ed in Fig. 1, whe e NUUse Equipmen s
(UEs) exchange da a wi h a single IAB dono ha has ibe
connec i i y o he Co e Ne wo k (CN). In o de o p o ide
s able co e age, he dono is suppo ed by NIIAB nodes,
which may be linked ei he di ec ly o he dono o o
nea by base s a ions, po en ially c ea ing a mul i-hop wi eless
backhaul. Wi hou making any assump ions ha limi he
scope o he si ua ion, we only conside uplink a ic.
We di ide he ime esou ces in o T adio sub ames, each
wi h a du a ion o Tsub = 1 ms, whe eas all nodes a e
equipped wi h ansmission bu e s. Consequen ly, he da a
ha node i ansmi s o gNB k(ei he he IAB dono o an
IAB node) du ing sub ame is s o ed in i s bu e Bk( )i.
This da a ep esen s ei he he packe s ha he dono has
success ully ecei ed o he da a ha will be o wa ded o
he nex hop along he pa h du ing sub ame +1 in he case
o IAB nodes. We assume ha he backhaul links ope a e
in he mmWa e spec um, and each IAB node is equipped
wi h wo Radio F equency (RF) chains, so ha wo an enna
sys ems can be lexibly used o he backhaul and access
communica ions. The Sionna Ray T acing (RT) 3 ool is used
o model he channel and calcula e Powe Spec al Densi y
(PSD) in his esea ch. The Signal o In e e ence plus Noise
Ra io (SINR) o a packe om sou ce node s o des ina ion
node d,δs,d can be exp essed as
δs,d =|hs,d|2σ2
x
σ2
n+Pi∈I σ2
i
,(1)
3h ps://n labs.gi hub.io/sionna/api/ .h ml
whe e hs,d ep esen s he equi alen channel esponse
be ween he communica ion endpoin s, Ideno es he se o
in e e e s, σ2
x,σ2
iand σ2
na e he powe s o he ansmi ed
signal, he i- h ecei ed in e e ing signal, and he he mal
noise a he ecei e , espec i ely. The co esponding access
(backhaul) h oughpu βA
j,k( )(βB
s,d( )) eads
βA
s,d( ) = 1
Tsub
B
s
X
l=1
1
nˆ
bl(δs,d) = blo,(2)
whe eas gNB k= 0, . . . , NI, wi h index 0deno ing he
IAB dono , ecei es da a om node s, whe e B
sdeno es he
numbe o bi s ansmi ed om use (IAB node) j o gNB d
du ing sub ame and ˆ
bl(δj,k)is he l- h decoded bi a he
ecei e , as a unc ion o δj,k.
B. SeBaSi
SeBaSi [16] is a sys em-le el simula o , buil on op o
he open-sou ce SionnaTM [17] simula o , which is used
o modeling he physical laye o 5G and beyond 5G
ne wo ks. SeBaSi is speci ically designed o model 3GPP
Release 17 IAB cellula ne wo ks. I is w i en in Py hon
and ope a es on op o any link-le el simula o , such as
Sionna, and simula es essen ial componen s. In o de o
inco po a e sel -backhauling IAB capabili ies in o Sionna, we
ha e success ully in eg a ed se e al sys em-le el ea u es in o
SeBaSi. The ex ensions, desc ibed in de ail in [13], consis
o a schedule a he Medium Access Con ol (MAC) le el,
laye -2 bu e s, and algo i hms o selec ing he backhaul
pa h. In addi ion, we implemen ed 5G-NR p ocedu es, such
as codebook-based beam o ming and SINR compu a ion, o
imp o e he alignmen o Sionna’s physical laye wi h he
la es 5G-NR s anda ds. In addi ion, we enhanced SeBaSi o
include suppo o sub-THz links in he backhaul [18]. This
is enabled by he ex ension o SeBaSi channel modeling o
suppo sub-THz by simula ed aces in Te asim [19]. So
he links can be con igu ed o unc ion a ei he mmWa e,
sub-THz, o a combina ion o bo h equencies. The pu pose
o his enhancemen is o e alua e he pe o mance o sub-
THz equencies o IAB, which is in line wi h he la es
esea ch and s anda diza ion ac i i ies on 6G. A he physical
laye , Sionna and SeBaSi implemen he 3GPP TR 38.901
model o he mmWa e channel, e en hough he mos ecen
e sion o SeBaSi also includes a buil -in ay ace ool o
model he channel. To add ess ou ing wi hin he wi eless
backhaul ne wo k, we implemen ed he Backhaul Adap a ion
P o ocol (BAP) laye in he uppe laye s o SeBaSi [20].
This laye uses a MAC-le el schedule ha ope a es in a
TDMA manne . Addi ionally, i u ilizes hop-by-hop Radio
Link Con ol (RLC) channels o simula e laye -2 bu e ing
and da a ansmission.
SeBaSi enables use s o cus omize a ious simula ion pa-
ame e s, including he du a ion and mode o he simula ion,
he size o he packe s, and he a e o da a ansmission om
ei he indi idual use equipmen o he en i e sys em. The
simula ion modes being conside ed a e he un mode and he
debug mode. The debug mode o e s ex a con ol signals
and ela ed in o ma ion. In addi ion, use s ha e he abili y o
pe sonalize he scena io by selec ing he quan i y and loca ion
o UEs and base s a ions, as well as he IAB opology,
which e e s o he wi eless backhaul links be ween gNBs.
The backhaul schedule algo i hm, which de e mines which
backhaul links o schedule in each ime slo , allows use s o
ei he c ea e cus om policies o selec om p ede ined op ions
such as SCAROS [21], MLR [6], Sa ehaul [13], and SINR-
based [18].
The simula o gene a es a comp ehensi e collec ion o
sys em-le el Key Pe o mance Indica o s (KPIs), including
end- o-end la ency, h oughpu , and packe d op a e. Each
o hese me ics can be collec ed and shown o each IAB
node o o he en i e ne wo k. Fu he mo e, SeBaSi p o ides
in e nal and/o lowe laye me ics, such as he imes amp
o packe gene a ion and a i al, des ina ion UE, and he
backhaul pa h. In addi ion, i p o ides in o ma ion on he
load o each IAB node o each ime s ep, including bo h he
access and he backhaul in e aces.
III. DATA GENERATION FRAMEWORK
Da a D i en-SeBaSi (DD-SeBaSi) is a amewo k ha
models IAB ne wo ks using a eal da ase in eg a ed on op
o SeBaSi. DD-SeBaSi inco po a es all he exis ing ea u es
o SeBaSi and in oduces an addi ional unc ionali y ha
allows use s o selec be ween gene a ing pa ame e s (e.g.,
he numbe o connec ed UEs pe gNB, sys em Ra e, and
Re e ence Signal Recei ed Powe (RSRP)) andomly o using
da a a ic, ei he om a eal da ase o om ou da a
gene a o . Figu e 2 demons a es he p oposed da a gene a o
in collabo a ion wi h SeBaSi. In o de o accomplish his, we
ini ially collec he da ase (desc ibed in Sec. III-A). Nex ,
we employ he GAN a chi ec u e o ain he model using
his da ase and gene a e new syn he ic da a which is s ill
accu a ely ep esen a i e o he o iginal da a. Ul ima ely, we
in eg a ed he GAN model in o he SeBaSi as a new ea u e o
ha e DD-SeBaSi. I is no ewo hy ha use s ha e he abili y
o de ine any scena io, and he da a gene a o will a emp
o c ea e da a ha is ele an o ha scena io, using he eal
da ase as a basis. In he ollowing sec ions, we will begin by
p o iding an o e iew o he da a collec ion p ocess. Nex ,
we will p esen he p oposed da a gene a o . Finally, we will
demons a e how we seamlessly inco po a e i in o SeBaSi.
A. Da ase collec ion
In o de o conduc ou in es iga ion, we u ilize eal-
wo ld da ase s ha we collec ed om a mobile ne wo k ha
ope a es in Eu ope, and main ains a signi ican ma ke sha e.
The da ase comp ises indi idual measu emen samples o a
a ie y o me ics om end-use de ices ha we e collec ed
du ing 2 mon hs in 2023. The da ase con ains housands o
samples, each o which is associa ed wi h he co espond-
ing adio sec o iden i y and geog aphical coo dina es. We
concen a e ou analysis on London, which is he p ima y
inno a ion hub o he ope a o as a esul o he high
popula ion densi y and g owing se ice demand.
Simula o
nonPHY
Da aTypes
sub Class()
PHY
Physical
iab()
Da aType
Classes
iab id
cell id
loca ion
An enna & BF
Tx Bu e
Rx Bu e
Se up()
Class
GAN-based
Da a Gen
Da a
Connec ed UE
pe gNB(γ)
h oughpu (β)
Re e ence Signal
Recei ed Powe
(RSRP) (λ)
Classes
Topology()
simula ion mode
un ime
Access()
Backhaul()
packe ()
Da aType
Classes
packe id
packe size
ame id
UE id
Cell id
Delay
passed pa h lis
pa h()
Da aType
Classes
pa h id
om IAB
o IAB
SINR
pa h a e
pa h equipped
Da aFlow Con ol Plane
Bu e
RLC-like
Classes
RX DU RLC
TX MT RLC
BAP-like
Classes
3GPP S anda d
TS 38.340
Pa h Selec ion
Policy
Classes
Sa ehaul
SCAROS
MLR
Schedule
Classes
RoundRobin()
Sionna Ex ended
Modules
Sub-THz
Channel T aces
NS-3 Te asim
Classes
3GPP TR 38.901
Channel models
MU-MIMO
OFDM
5G-NR
Classes
Physical Based
THz Channel
models
140GHz
ho n An enna
Classes
Beam o ming
ICI & SINR
Fig. 2: O e all design o DD-SEBASI F amewo k ex ension. The blue block is he p oposed da a gene a o , he ed blocks a e SeBaSi,
and he g een block ep esen s he Sionna simula o [17].
B. Da a Gene a o
Wi hin he collec ed da ase , we ha e selec ed h ee pa-
ame e s ha a e likely o ha e a signi ican in luence on
he pe o mance: he numbe o connec ed UEs pe gNB
(γ), he sys em a e pe UE (β), and RSRP (λ). Gi en
he una ailabili y o eal-wo ld mmWa e deploymen s a he
ime o w i ing, all da a is limi ed o cellula deploymen s
ope a ing a sub-6 GHz equencies. As a i s s ep, we
es ablish a simula ion scena io u ilizing he Sionna RT ool
o gene a e ays a mmWa e spec um o ob ain RSRP alues
wi h co esponding deploymen posi ions. RT is an excellen
ool o modeling he en i onmen and gene a ing pa hs;
howe e , i is es ic ed o speci ic scena ios. The e o e, in
o de o ob ain he app op ia e channel and, in pa icula , he
RSRP alue, i is necessa y o model he en i onmen wi h
all de ails o each scena io. So o do simula ion wi h DD-
SeBaSi, i s we model he London Ci y scena io (same a e
eal deploymen ) in RT o ob ain p ope RSRP alues. We
u ilize exclusi ely eal da ase s o he γand βpa ame e s.
GANs a e a dis inc ype o DNNs ha ha e he capaci y o
gene a e da a by lea ning he p ecise s a is ical cha ac e is ics
o a gi en da ase h ough indi ec me hods.The in e es ing
ea u e o GANs is ha hey can be ained using a limi ed
amoun o a ailable eal da a o gene a e syn he ic da a in
di e en scena ios and condi ions. This syn he ic da a can
hen be used in da a gene a ion o wi eless communica ion
ne wo ks. GANs in ol e wo key componen s: he gene a o
G, which ans o ms a andom sample om a uni o m
dis ibu ion in o a sample ha ollows he da a dis ibu ion,
and he disc imina o D, which assesses whe he a gi en
sample is ep esen a i e o he da a dis ibu ion o no . To
lea n a gene a o dis ibu ion pgo e he da ase , i cons uc s
a mapping unc ion G(z) om a noise dis ibu ion pz(z)
whe e zis he inpu noise o he ou pu o gene a o (X ).
The disc imina o D(x) e u ns a single scala alue ha
ep esen s he p obabili y ha Xis de i ed om he eal
da ase (X ) a he han pg(X ), he e o e showing he
au hen ici y o he da a. The nega i e ela ionship be ween
he wo componen s o he GAN is e lec ed in he min-max
equa ion, which is a undamen al componen o he aining
objec i e. Concu en aining o bo h Gand Dis conduc ed,
wi h pa ame e s modi ied o G o minimize he cos unc-
ions log10(1 −D(G(z))) and log10 D(x)acco dingly, using
he min-max alue unc ion V(D, G)4which is de ined as:
min
|{z}
G
max
|{z}
D
V(D, G) = EX∼pda a (X)[log D(X|y)]
+Ez∼pz(z)[log(1 −D(G(z|y)))]
(3)
The min-max equa ion encapsula es he ad e sa ial na u e
o GAN aining, in which he gene a o and disc imina o
a e pe pe ually enhancing hemsel es in esponse o each
o he ’s ad ancemen s, esul ing in imp o ed da a gene a ion
o e ime. The unde lying p emise o min-max is ha D(X)
is a emp ing o op imize i s accu acy by inc easing he
p obabili y o dis inguishing be ween eal and ake da a.
G(X)is a emp ing o educe he disc imina o ’s capaci y
o do so by ensu ing ha he ake da a appea s as eal as
possible. D(X)should be unable o di e en ia e be ween
eal and ake da a when he GAN eaches equilib ium, which
implies ha D(G(z))) = 0.5 o all z. A his s ep, D(X)is
a i s mos con used, and he gene a o gene a es da a ha is
indis inguishable om eal da a.
Figu e 3 illus a es he a chi ec u e o he DD-SeBaSi GAN
da a gene a o model, whe eas Z, X , and X ep esen
he inpu noise, {γ, β, λ} eal,{γ, β, λ} ake5, espec i ely.
Figu es 3a and 3b ep esen he a chi ec u e o Gand D
whe eas Gand D ake as inpu Z, and {X , eal}o {X ,
ake}, espec i ely.
4Fo a mo e de ailed desc ip ion o he GAN aining p ocess, we e e
he in e es ed eade s o [22].
5Du ing aining we e e o ou pu o Gene a o (X )as ake, a e
aining, du ing he simula ion campaign in DD-SeBaSi, we e e wi h
G,{γ, β, λ} ake ={γ, β, λ}G
Finding he bes pa ame e s o aining a GAN is a chal-
lenging ask because i in ol es op imizing mul iple hype pa-
ame e s, including he lea ning a e, ba ch size, epoch coun ,
gene a o and disc imina o laye coun , ac i a ion unc ions,
egula iza ion echniques, and mo e. We pe o med aining
on he model using se e al combina ions o hype pa ame e s.
We e alua ed he pe o mance using a speci ic measu emen
and chose he hype pa ame e s ha p oduced he mos ad-
an ageous ou comes. The mos a o able hype pa ame e s
a e as ollows. We u ilize he ec i ied linea uni (ReLU)
as he ac i a ion unc ion. The aining me hod comp ises
80,000 epochs, wi h a ba ch size o 32. We employ Adam
as he op imize , u ilizing a lea ning a e o 0.001. The loss
unc ion o he gene a o Gand disc imina o Dis ob ained
by u ilizing he mean absolu e e o o Gand bina y c oss-
en opy o D.
IV. PERFORMANCE EVALUATION
In Sec. IV-A we p o ide a s a is ical analysis ha alida es
he accu acy o he p oposed da a gene a o . In Sec. IV-B we
alida e ia simula ion he implemen a ion o he DD-SeBaSi
simula o using syn he ic da a om he da a gene a o , and
e alua e he pe o mance o an IAB ne wo k conside ing
di e en backhaul schedule s.
A. S a is ical Analysis
The ini ial s ep in using he da a gene a o in he SeBaSi
simula o is o e i y ha he p oposed da a gene a o is
consis en wi h he ac ual da ase [23]. G aphically ep e-
sen ing he Cumula i e Dis ibu ion Func ions (CDFs) o wo
dis ibu ions is an e ec i e app oach o isulaize he deg ee
o simila i y o any s ong dispa i y be ween hem. In o de
o accomplish his, we plo he CDF o eal and gene a ed
alues o and G γ, β, λ in Fig. 4. While we see ha he
wo cu es almos o e lap, which indica es ha ou da a
gene a o is accu a e, we use he KS es o o mally e i y
ha he gene a ed da a is s a is ically consis en wi h he
ac ual da ase . KS is an impo an me ic used in s a is ical
analysis o compa e he dis ibu ions o wo da ase s. I s
pu pose is o de e mine whe he a da ase adhe es o a speci ic
dis ibu ion [24]. The es yields a P- alue ha signi ies he
likelihood o achie ing he obse ed dispa i y in dis ibu ions
due o andom chance. A highe P- alue sugges s ha he
wo da ase s a e p obably sampled om he same dis ibu-
ion. The D- alue, also e e ed o as he KS s a is ic, is a
quan i a i e measu e ha cap u es he la ges , also e e ed
o as e ical dis ance, be ween he CDFs o wo da ase s
unde compa ison. Consequen ly, he da ase s exhibi g ea e
simila i y when he D- alue is small.
We conduc ed he KS o he CDF o {γ, β, λ} eal and
{γ, β, λ}G. The esul s o he es a e p esen ed in Table I. I
is shown ha all pa ame e s ({γ, β, λ}) a e passing he es
e y well: An ex emely low D- alue and high P- alue o all
pa ame e s sugges ha he dis ibu ion o he wo da ase s is
compa able. The e o e, ou KS es does no ejec he null
hypo hesis, which indica es ha he e is insu icien e idence
Gene a o
X
X
Z
Sionna RT
γ
β
λ
γ
β
λ
γ
β
GAN-based Da a Gene a o
Disc imina o
X
G
eal
(a) GAN-based Da ase gene a o amewo k a chi ec u e
Gene a o
7 Con + MaxPool
Laye s
Con
(64, 1x7)
MaxPool
(1x2) (512) (256)
3 Dense Laye s
+ BachNo maliza ion+
Relu
Relu
Laye
(128) (3)
Fla en
XF
Z
(b) A chi ec u e o Gene a o o GAN
Disc imina o
X
7 Con + MaxPool
Laye s
Con
(32, 1x7)
MaxPool
(1x2) (128)
2 Dense Laye s
+ BachNo maliza ion+ Relu
Relu
Laye
(64) (1)
Fla en
Real / Fake
(c) A chi ecu e o Disc imina o o GAN
Fig. 3: S uc u e o he p oposed da a gene a o , including DNN-
based a chi ec u e
o demons a e ha he sample dis ibu ion de ia es om he
e e ence dis ibu ion.
TABLE I: A goodness o i ( wo sample KS Tes ) o .
Pa ame e P- alue D- alue
γ0.714 0.073
β0.795 0.0124
λ0.892 0.0091
B. IAB simula ion Resul s
In his sec ion we use SeBaSi o un IAB simula ions
in di e en scena ios, and used ou GAN da a gene a o ,
ha was p e iously alida ed in Sec. IV-A, o model he
channel o bo h he access and he backhaul links. Simula ion
esul s a e gi en a unc ion o he numbe o IAB nodes and
012345
Uplink Cell Th oughpu 1e6
0.0
0.2
0.4
0.6
0.8
1.0
CDF
Real Da a
Gene a ed Da a
(a)
0 50 100 150 200 250 300 350 400
A e age Connec ed Use s
0.0
0.2
0.4
0.6
0.8
1.0
CDF
Real Da a
Gene a ed Da a
(b)
160 140 120 100 80
Re e ence Signal Recei ed Powe (dBm)
0.0
0.2
0.4
0.6
0.8
1.0
CDF
Real Da a
Gene a ed Da a
(c)
Fig. 4: CDF o α, β, λ compa ison o gene a ed and eal da ase s
TABLE II: Simula ion pa ame e s.
Pa ame e Value
Ca ie equency and bandwid h 30 GHz and 400 MHz
IAB RF chains 2 (1 access + 1 backhaul)
Numbe o BS-nodes N100
IAB Backhaul and access an enna a ay 8H×8V and 4H×4V
UE an enna a ay 4H×4V
IAB and UE heigh 15 m and 1.5 m
IAB an enna gain 33 dB
Noise Figu e 10 dB
o di e en schedule implemen a ions. We plo he mean
h oughpu , la ency, and packe d op a e.
Simula ion Scena io We simula e he ac ual cellula ne -
wo k deploymen con igu a ion o London Ci y. Speci ically,
we conside N= 100 5G-NR base s a ions wi hin a 15 Km2
a ea, as shown in Fig. 5. The speci ic simula ion pa ame e s
a e ou lined in Table II.
Fig. 5: Loca ions o BS-nodes ( ed do s) in London Ci y
Nume ical Resul s We use h ee di e en IAB schedule s
a ailable in SeBaSi: (i)Sa ehaul [13], a isk-a e se lea ning
me hod o ensu ing eliabili y in mmWa e sys ems which
uses a Rein o cemen lea ning algo i hm o inc ease eliabil-
i y in he ne wo k; (ii)Scalable and Robus Sel -backhauling
Solu ion (SCAROS), an online lea ning-based echnique ha
educes he a e age backhaul scheduling la ency in he ne -
wo k [21]; and (iii)Maximum-local- a e (MLR), a g eedy
s a is ical me hod ha aims a maximizing h oughpu by
choosing links wi h he highes da a a e. This app oach
ope a es o line, g ea ly acili a ing i s applica ion in eal-
wo ld si ua ions, bu a he cos o dec eased pe o mance.
1) Scena io 1: A e age Ne wo k Pe o mance: E alua ion
o he algo i hms’ pe o mance o e ime is essen ial o
de e mining he a e a which he lea ning-based app oaches,
namely Sa ehaul and SCAROS, con e ge. The e o e, in Fig. 6
we display he mean la ency, h oughpu , and packe loss a e
o he IAB ne wo k o e ime. In Fig. 6a, we can obse e
ha Sa ehaul apidly con e ges o an a e age la ency o
app oxima ely 6.5ms which is 11% and 48.4% lowe han
he la ency o SCAROS and MLR, espec i ely. The high
pe o mance o Sa ehaul s ems om he join minimiza ion
o he a e age la ency and he expec ed alue o i s ail
loss, which esul s in a oiding isky si ua ions whe e la ency
goes beyond Tmax (maximum ime be o e packe d op in
he ne wo k, a e which a packe is conside ed as d opped,
50 ms in ou simula ion campaign). This is no he case
o SCAROS whe e we obse e a high peak in he la ency
be o e con e gence. I is exac ly he a oidance o such
ansien s in Sa ehaul ha leads o highe eliabili y in he
sys em. The eliabili y o e ed by Sa ehaul allows Mobile
Ne wo k Ope a o s (MNOs) o deploy sel -backhauling in an
online ashion and wi hou dis up ing he ne wo k ope a ion.
Figu e 6b illus a es ha he a e age ansmission o he
ne wo k is no ad e sely a ec ed by he isk-a e sion capa-
bili ies o Sa ehaul. Sa ehaul’s pe o mance is app oxima ely
79.3 Mbps, which is 11.7% highe han ha o MLR and
compa able o SCAROS. The beha io obse ed in Fig. 6a is
consis en wi h he pe o mance depic ed in Fig. 6c in e ms
o packe d op a e. Sa ehaul ob ains he lowes packe d op
a e among he e e ence schemes, which is 30.1% (84.0%)
lowe han SCAROS (MLR).
2) Scena io 2: Impac o he IAB Con igu a ion: In Fig. 7
we e alua e he pe o mance o he IAB ne wo k as a unc ion
o he numbe o IAB nodes, ha we change om 25 o 100.
Simul aneously, we augmen he ne wo k’s bu den by inc eas-
ing he numbe o UEs (2 UEs pe IAB node). We demon-
s a e ha Sa ehaul consis en ly ob ains be e pe o mance in
compa ison o he e e ence schemes. This demons a es ha
Sa ehaul accomplishes he in ended op imiza ion objec i e,
which is he join minimiza ion o he a e age end- o-end
la ency and i s an icipa ed ail loss. As he numbe o IAB-
nodes inc eases, Sa ehaul is capable o main aining a nea ly
cons an la ency, as illus a ed in Fig. 7a. In pa icula , he
a ia ion o la ency wi h Sa ehaul is 56.1% and 71.4% lowe
han ha wi h SCAROS and MLR, espec i ely. Addi ionally,
Sa ehaul ob ains an 11.1% and 43.2% lowe la ency han
SCAROS and MLR, espec i ely. The la e ’s high a iance
is a esul o he lack o adap a ion capabili ies.
The a e age h oughpu o he lea ning-based app oaches,
i.e., Sa ehaul and SCAROS, emains cons an as he numbe
0 1000 2000 3000 4000 5000
Simula ion ime [ms]
4
6
8
10
12
14
16
La ency [ms]
Sa ehaul
SCAROS
MLR
(a) A e age pe -UE end- o-end la-
ency
0 1000 2000 3000 4000 5000
Simula ion ime [ms]
50
60
70
80
90
Th oughpu [Mbps]
Sa ehaul
SCAROS
MLR
(b) A e age pe -UE h oughpu
0 1000 2000 3000 4000 5000
Simula ion ime [ms]
0
5
10
15
20
Packe d op a e [%]
Sa ehaul
SCAROS
MLR
(c) A e age pe -UE packe d op a e
Fig. 6: A e age ne wo k pe o mance o 50 UEs(Scena io 1).
25 50 75 100
Numbe o IAB-nodes
10
12
14
La ency [ms]
Sa ehaul
SCAROS
MLR
(a) Pe -UE end- o-end La ency
25 50 75 100
Numbe o IAB-nodes
36
37
38
39
40
Th oughpu [Mbps]
Sa ehaul
SCAROS
MLR
(b) Pe -UE h oughpu
25 50 75 100
Numbe o IAB-nodes
0
2
4
6
8
10
12
Packe d op a e [%]
Sa ehaul
SCAROS
MLR
(c) Pe -UE packe d op a e
Fig. 7: Ne wo k pe o mance o {25,50,75,100}BS-node, 2UEs pe BS-nodes on a e age, and 40 Mbps pe -UE sou ce a e (Scena io
2).
o IAB nodes inc eases, as illus a ed in Fig. 7b. Sa ehaul, on
he o he hand, achie es he lowes a ia ion, di e en ial o
maximum and minimum alue, in h oughpu , wi h a alue
o 0.90, as opposed o he benchma k schemes’ 1.9 and
2.8. The eliabili y capabili ies o Sa ehaul a e co obo a ed
by he esul s in Fig. 7c, whe e we plo he packe ailu e
a e s. he numbe o IAB nodes. I is wo h no ing ha
Sa ehaul consis en ly ou pe o ms he e e ence schemes and
exhibi s he leas a ia ion in esul s (a leas 47.3% lowe
han he benchma ks). When he g ea es ne wo k size and
a ic a e aken in o accoun , namely 200 BS-nodes and 400
UEs, Sa ehaul achie es a 49.3% and 81.2% lowe packe d op
a e han SCAROS and MLR, espec i ely.
Summa y We p esen an example scena io o London Ci y
end- o-end pe o mance me ics esul s, u ilizing he GAN-
based da a gene a o o demons a e how a ious schedules
can be alida ed in he SeBaSi simula o . I is e iden ha
om he ob ained esul s, among he backhaul schedule s,
Sa ehaul has he po en ial o a ain supe io pe o mance,
which is di ec ly consis en wi h he esul s ob ained in [13]
o Manha an Ci y using a andom da a gene a o .
V. CONCLUSIONS AND FUTURE WORK
The scope o his esea ch was o alida e he in eg a ion
be ween SeBaSi and he da a gene a o , and p esen ( o he
i s ime) ealis ic IAB esul s, i.e., ob ained conside ing
syn he ic ( hough alida ed wi h eal aces) da a o he
IAB channel. The nume ical esul s align wi h p io ends
and demons a e he applicabili y o app oaches o imple-
men a ion in he ac ual sys em. In u u e wo k, u he eal
da ase s will be ga he ed om a ious ope a o s o enhance
he gene ali y o he amewo k and enable he u iliza ion
o he da a gene a o in o he simula o s, such as ns-3. In
addi ion, we would ex end simula ions o do localiza ion o
ind he bes spo o ins all he IAB node in a ious scena ios.
VI. ACKNOWLEDGEMENT
This wo k was pa ially suppo ed by he EU MSCA
ITN p ojec MINTS “MIllime e -wa e Ne wo king
and Sensing o Beyond 5G” (g an no. 861222) and
unded by G an PID2021-126431OB-I00 unded by
MCIN/AEI/10.13039/501100011033 g an (ANEMONE),
Spanish MINECO g an TSI-063000-2021-54 (6G-DAWN
ELASTIC), TSI-063000-2021-55 (6G-DAWN RESILIENT),
TSI-063000-2021-56 (6G-BLUR SMART), TSI-063000-
2021-57 (6G-BLUR JOINT), and “ERDF A way o making
Eu ope” by Gene ali a de Ca alunya g an 2021 SGR
00770.” Also he wo k was pa ially suppo ed by he SNS
(Sma Ne wo ks and Se ices) JU and i s membe s, unded
by he Eu opean Union unde G an Ag eemen numbe
101139161.
REFERENCES
[1] 3GPP, “NR; In eg a ed Access and Backhaul (IAB) adio ansmis-
sion and ecep ion,” Technical Speci ica ion (TS) 38.340, July 2024,
.18.5.0.
[2] A. A. Ga ga i, A. O iz, M. Pagin, W. de Somb e, M. Zo zi, and
A. Asadi, “Risk-a e se lea ning o eliable mmwa e sel -backhauling,”
IEEE/ACM T ansac ions on Ne wo king, pp. 1–15, 2024.
[3] 3GPP, “NR; S udy on in eg a ed access and backhaul,” Technical
Speci ica ion (TS) 38.340, July 2019, .16.0.0.
[4] F. Rezazadeh, S. Ba achina-Mu˜
noz, H. Che gui, J. Mangues, M. Ben-
nis, D. Niya o, H. Song, and L. Liu, “Towa d explainable easoning
in 6g: A p oo o concep s udy on adio esou ce alloca ion,” IEEE
Open Jou nal o he Communica ions Socie y, ol. 5, pp. 6239–6260,
2024.
[5] O. P. Ada e, H. Babbili, C. Madapa ha, B. Makki, and T. S ensson,
“Uplink powe con ol in in eg a ed access and backhaul ne wo ks,” in
IEEE In e na ional Symposium on Dynamic Spec um Access Ne wo ks
(DySPAN), 2021, pp. 163–168.
[6] M. Pagin, T. Zugno, M. Polese, and M. Zo zi, “Resou ce manage-
men o 5G NR in eg a ed access and backhaul: A semi-cen alized
app oach,” IEEE T ansac ions on Wi eless Communica ions, ol. 21,
no. 2, pp. 753–767, Jul. 2022.
[7] C. Fang, C. Madapa ha, B. Makki, and T. S ensson, “Join schedul-
ing and h oughpu maximiza ion in sel -backhauled millime e wa e
cellula ne wo ks,” in 17 h In e na ional Symposium on Wi eless Com-
munica ion Sys ems (ISWCS), 2021.
[8] C. Madapa ha, B. Makki, C. Fang, O. Teyeb, E. Dahlman, M.-
S. Alouini, and T. S ensson, “On in eg a ed access and backhaul
ne wo ks: Cu en s a us and po en ials,” IEEE Open Jou nal o he
Communica ions Socie y, ol. 1, pp. 1374–1389, Sep. 2020.
[9] M. Polese, M. Gio dani, T. Zugno, A. Roy, S. Goyal, D. Cas o , and
M. Zo zi, “In eg a ed access and backhaul in 5g mmwa e ne wo ks:
Po en ial and challenges,” IEEE Communica ions Magazine, ol. 58,
no. 3, pp. 62–68, Ma 2020.
[10] M. Polese, M. Gio dani, A. Roy, S. Goyal, D. Cas o , and
M. Zo zi, “End- o-end simula ion o in eg a ed access and backhaul
a mmwa es,” in IEEE 23 d In e na ional Wo kshop on Compu e
Aided Modeling and Design o Communica ion Links and Ne wo ks
(CAMAD), 2018.
[11] C. Madapa ha, B. Makki, H. Guo, and T. S ensson, “Cons ained de-
ploymen op imiza ion in in eg a ed access and backhaul ne wo ks,” in
IEEE Wi eless Communica ions and Ne wo king Con e ence (WCNC),
2023.
[12] F. Rezazadeh, A. A. Ga ga i, S. Lag´
en, J. Mangues, D. Niya o, and
L. Liu, “Genone : Gene a i e open xg ne wo k simula ion wi h mul i-
agen llm and ns-3,” a Xi p ep in a Xi :2408.13781, 2024.
[13] A. Ash a i Ga ga i, A. O iz, M. Pagin, A. Klein, M. Hollick, M. Zo zi,
and A. Asadi, “Sa ehaul: Risk-A e se Lea ning o Reliable mmWa e
Sel -Backhauling in 6G Ne wo ks,” in IEEE Con e ence on Compu e
Communica ions (INFOCOM), New Yo k, USA, 2023.
[14] J. Kim, Y. Ahn, and B. Shim, “Massi e da a gene a ion o deep
lea ning-aided wi eless sys ems using me a lea ning and gene a i e
ad e sa ial ne wo k,” IEEE T ansac ions on Vehicula Technology,
ol. 72, no. 1, pp. 1302–1306, Sep 2022.
[15] E. Ayanoglu, K. Da aslioglu, and Y. E. Sagduyu, “Machine lea ning
in nex g ne wo ks ia gene a i e ad e sa ial ne wo ks,” IEEE T ans-
ac ions on Cogni i e Communica ions and Ne wo king, ol. 8, no. 2,
pp. 480–501, Feb 2022.
[16] A. A. Ga ga i, M. Pagin, A. O iz, N. M. Gholian, M. Polese, and
M. Zo zi, “Demo:[sebasi] sys em-le el in eg a ed access and backhaul
simula o o sel -backhauling,” in 2023 IEEE 24 h In e na ional
Symposium on a Wo ld o Wi eless, Mobile and Mul imedia Ne wo ks
(WoWMoM), 2023, pp. 355–357.
[17] J. Hoydis, S. Camme e , F. A. Aoudia, A. Vem, N. Binde , G. Ma cus,
and A. Kelle , “Sionna: An open-sou ce lib a y o nex -gene a ion
physical laye esea ch,” a Xi p ep in a Xi :2203.11854, 2022.
[18] A. A. Ga ga i, M. Pagin, M. Polese, and M. Zo zi, “6G In eg a ed Ac-
cess and Backhaul ne wo ks wi h sub- e ahe z links,” in 18 h Wi eless
On-Demand Ne wo k Sys ems and Se ices Con e ence (WONS), 2023,
pp. 13–19.
[19] Z. Hossain, Q. Xia, and J. M. Jo ne , “Te asim: An ns-3 ex ension o
simula e Te ahe z-band communica ion ne wo ks,” Nano Communica-
ion Ne wo ks, ol. 17, pp. 36–44, Sep embe 2018.
[20] 3GPP, “NR; Backhaul Adap a ion P o ocol (BAP) speci ica ion,” Tech-
nical Speci ica ion (TS) 38.340, May 2022, .17.0.0.
[21] A. O iz, A. Asadi, G. H. Sim, D. S einme ze , and M. Hollick,
“SCAROS: A scalable and obus sel -backhauling solu ion o highly
dynamic millime e -wa e ne wo ks,” IEEE Jou nal on Selec ed A eas
in Communica ions, ol. 37, no. 12, pp. 2685–2698, Oc . 2019.
[22] I. Good ellow, J. Pouge -Abadie, M. Mi za, B. Xu, D. Wa de-Fa ley,
S. Ozai , A. Cou ille, and Y. Bengio, “Gene a i e ad e sa ial ne -
wo ks,” Communica ions o he ACM, ol. 63, no. 11, pp. 139–144,
Oc 2020.
[23] X. Liao, J. Xie, and J. Zhou, “A da a-d i en op imiza ion me hod o
simula ing a bi a ily dis ibu ed and spa ial– empo al co ela ed ada
sea clu e ,” IEEE T ansac ions on Geoscience and Remo e Sensing,
ol. 61, pp. 1–15, Oc 2023.
[24] A. Zeimbekakis, E. D. Schi ano, and J. Yan, “On misuses o he
kolmogo o –smi no es o one-sample goodness-o - i ,” The Ame i-
can S a is ician, pp. 1–18, May 2024.
