inCoo d: In en -based Coo dina ion in he
Mul i-domain Cloud-Edge Con inuum
And ea Mo iche a 1, Juan B enes 2, Mikhail Kolobo 2, Djawida Dib 3, Thijs Me sch 3, Anna Lackinge 1,
C e a Capo a 1, Hui Song 4, Rus em Dau o 4, Ahmed Khalid 5,
Sigmund Akselsen 6, A ne Munch-Ellingsen 6, Schah am Dus da 1
1Dis ibu ed Sys ems G oup, TU Wien, Aus ia; 2Nex wo ks, Pisa, I aly; 3In el Labs, In el Co po a ion, Ge many;
4SINTEF, Oslo, No way; 5Dell Resea ch, Dell Technologies, Co k, I eland; 6Teleno ASA, No way.
Abs ac —The compu ing con inuum aims o b eak he iso-
la ion o edge and cloud compu ing, c ea ing a smoo h and
he e ogeneous in as uc u e su ace o deploying applica ions.
Howe e , he con inuum is p ac ically agmen ed, wi h in as-
uc u e managed in isola ion by a ious pa ies in a e ical
dimension, i.e., edge, og, and cloud laye s, and ho izon ally,
i.e., compu ing, ne wo k, and s o age domains. This scena io
nega i ely impac s he ul illmen o he applica ion objec i es.
We p opose inCoo d, an in en -awa e solu ion ha enables he
c ea ion o a uni ied compu ing con inuum by coo dina ing i s
ins ances o ul ill applica ion objec i es. Each ins ance ep esen s
a clus e o (compu e, s o age, o ne wo k) nodes wi h hei own
manage componen . T adi ionally, sys ems ake low-le el ac ions
on hese ins ances. In con as , inCoo d lea ns hei eme ging
beha io s and adap s hei manage s’ objec i es o ul ill he
applica ion’s in en s. He e, h ough a Rein o cemen -Lea ning-
based P oo o Concep , we show he po en ial o his sys em
o unde s and eme ging beha io and manage mul i-domain,
independen ly managed ins ances.
Index Te ms—compu ing con inuum, ein o cemen lea ning,
coo dina ion, in en -based managemen , mul i-domain
I. INTRODUCTION
The p omise o he compu ing con inuum pa adigm is o
o e a seamless in e ace o he dis ibu ed in as uc u e ha
spans om he edge o he cloud. Howe e , in ope a ional
e ms, a dis ibu ed in as uc u e is s ill s uc u ed as a com-
posi ion o clus e s o nodes, i.e., ins ances, handled by di -
e en pa ies. Fu he mo e, whe eas he ocus is ypically on
compu ing nodes, managing eal in as uc u es also equi es
managing he ne wo k and s o age componen s. The e o e, as-
pi ing o ha e comple e con ol o e i is un ealis ic. Exis ing
solu ions, like “Sky Compu ing” [1] o o he comme cial 123
ones, aim o build a uni ied in e ace o applica ion owne s.
They o e he illusion o deploying on a con inuous in as-
uc u e, hiding how i is composed. While help ul, modeling
he compu ing con inuum as an o ches a ion abs ac ion laye
does no sol e he conc e e p oblem o accu a ely cap u ing
and handling he applica ions’ eal- ime dynamics.
This wo k has been suppo ed by he Eu opean Union’s Ho izon Eu ope
esea ch and inno a ion p og am unde g an ag eemen No. 101135576
(INTEND).
1Red Ha OpenShi : h ps://www. edha .com/en/ echnologies/
cloud-compu ing/openshi
2meshS ack (meshCloud): h ps://www.meshcloud.io/en/p oduc /co e/
3IBM Cloud Pak: h ps://www.ibm.com/p oduc s/cloud-pak- o -in eg a ion
Taking he applica ion owne ’s pe spec i e, a key challenge
hey ha e in he compu ing con inuum is o iden i y he igh
in as uc u e equi emen s o gua an ee a eliable un ime
o hei se ice. In en -based managemen aims o sol e his
conce n by le ing use s de ine high-le el objec i es and ans-
la ing hem in o conc e e in as uc u e equi emen s. How-
e e , such amewo ks ypically suppo one single domain,
such as he ne wo k [2] o , ecen ly, he compu e [3], [4].
The wo p esen ed p oblems a e o hogonal, bu bo h in e sec
a he same poin : managing he compu ing con inuum. The
i s conce ns in as uc u e con ol, and he second applica-
ion managemen ; bo h need o abs ac away he inhe en
complexi y o he compu ing con inuum and c ea e a smoo h
abs ac ion o i .
We add ess hese gaps by in oducing inCoo d, an in en -
awa e solu ion o ha monizing in as uc u e ins ances in he
compu ing con inuum. Gi en an applica ion deployed in a se
o ins ances, inCoo d ecei es applica ion-le el and ins ance-
le el in en s (o objec i es). Each ins ance manage op imizes
i s s a egy o gua an ee i s in en ul illmen . F om a an age
poin , inCoo d cap u es eme ging beha io s and hei impac
on he o e all applica ion in en . Consequen ly, i uses his
in o ma ion o adjus he ins ance-le el in en s, igge ing an
upda e o he ins ance manage s’ s a egies. The no el y o
his app oach lies in no equi ing a di ec managemen o he
in as uc u e. inCoo d makes i possible o ely on indepen-
den ly managed ins ances, lea ing o he ins ance manage s he
eedom o deploy hei s a egies. Fu he mo e, in his way,
inCoo d c ea es a p ope compu ing con inuum abs ac ion.
Indeed, i does no di e en ia e be ween in-p emise o hi d-
pa y nodes, no ne wo k, compu e, o s o age domains, no
be ween edge o cloud laye s. Finally, i allows o a holis ic,
sel -adap i e s a egy o ha monize he compu ing con inuum,
equi ing minimal human in e en ion. inCoo d’s p inciples
make he selec ion o he sel -adap i e me hod lexible, open-
ing up he esea ch o he bes app oach.
In his pape , we o e a de ailed de ini ion o ou a -
chi ec u e, including how o ob ain conc e e applica ion and
ins ance objec i es om high-le el ins uc ions, and how o
adap hem a un ime. Fu he mo e, we implemen a P oo
o Concep o showcase inCoo d’s po en ial. We se up a
simula ed en i onmen wi h a ideo s eaming use case, using
Rein o cemen Lea ning (RL), in pa icula P oximal Policy979-8-3315-0376-5/25/$31.00 ©2025 IEEE
Op imiza ion (PPO) o he sel -adap i e s a egy. The esul s
show ha inCoo d has he po en ial o allow simple bu e -
ec i e ha monizing o ins ances in he compu ing con inuum,
gi en a high-le el applica ion in en .
The pape is o ganized as ollows. In Sec ion II we e iew
he s a e o he a . Sec ion III depic s an example o in en
de ini ion o he compu ing con inuum, la e e alua ed wi h
ou use case. In Sec ion IV we in oduce he inCoo d a chi-
ec u e and i s mechanisms. Sec ion V p esen s he esul s o
applying inCoo d in a simula ed use case, while Sec ion VI
concludes he pape .
II. RELATED WORK
Recen e o s in he ield ha e in es iga ed he use o
in en -d i en o ches a ion as a me hod o managing com-
plex and dynamic compu ing en i onmen s. One app oach
[5] in oduces he concep o inco po a ing ole a ions in o
in en s, allowing o g ea e lexibili y in how objec i es and
expec a ions a e me . This me hod aims o imp o e o ches-
a ion ou comes by enabling sys ems o con inue ope a ing
e ec i ely unde a ying o subop imal condi ions, which he
app oaches in his pape can le e age.
A b oade examina ion o in en -based o ches a ion is p o-
ided in [3], [6], and [7]. In [3] he au ho s ou line i s co e
p inciples and associa ed bene i s. These include abs ac ion
om low-le el con ol, imp o ed au oma ion, and enhanced
scalabili y ac oss dis ibu ed sys ems. The s udy emphasizes
he u ili y o exp essing high-le el goals o acili a e mo e
e icien and esponsi e o ches a ion p ocesses. In [6] LLMs
a e used as he cen al decision-making en i y o add ess iola-
ions o in en -d i en pe o mance c i e ia. I subsequen ly au-
oma es deploymen econ igu a ions based on he LLMs’ ec-
ommended co ec i e ac ions. [7] p oposes a gene ic app oach
o managing in en -d i en o ches a ion o dis ibu ed appli-
ca ions. A hie a chical decision-making scheme is employed,
whe e compu e and ne wo k o ches a ion componen s collab-
o a i ely manage pa s o he applica ion. Expanding on hese
concep s, [8] p esen s a amewo k in which in en s p o ide
seman ic con ex o esou ce p o ide s. This addi ional con ex
enables dynamic adap a ion o esou ce beha io in esponse
o changing condi ions, con ibu ing o mo e au onomous and
con ex -awa e o ches a ion o ackle sus ainabili y challenges.
[9] add esses he op imiza ion p oblem o se e less applica-
ions based on high-le el in en s. A Rein o cemen Lea ning-
d i en au o-scaling mechanism is used o op imize esou ce
u iliza ion. Addi ionally, a scheduling op imize is p oposed
o minimize link u iliza ion, ans o ma ion cos s, and powe
consump ion o e icien unc ion placemen . Toge he , hese
wo ks suppo he shi owa d modula and ede a ed co-
o dina ion models, whe e decla a i e speci ica ions, a he
han p ocedu al con ol, guide he ope a ion o he e ogeneous
sys ems ac oss he compu ing con inuum.
Se e al esea che s ha e o e ed con ibu ions o he co-
o dina ion o in en s h ough au oma ed agen s. Some so-
lu ions [10], [11] ocus on using Rein o cemen lea ning
me hods in he ne wo king domain, wi h ne wo k se ice
assu ance as a i s -class ci izen. Ano he esea ch s eam
uses RL o compu ing en i onmen s. While, e.g., Song e
al. [12] p opose an in en -based cogni i e con inuum o IoT-
Edge-Cloud en i onmen s, o he wo ks ocus on se e less
o ches a ion [13] o dis ibu ed applica ion managemen [14].
In con as , ou wo k ocuses on c oss-domain coo dina ion,
whe e ein o cemen lea ning is used no wi hin a single
in as uc u e domain, bu o media e and align objec i es
ac oss compu e, ne wo k, and s o age domains simul aneously.
I is wo h highligh ing ha om he s anda diza ion pe -
spec i e, se e al s anda ds and indus y speci ica ions, such
as ETSI ZSM [15], 3GPP [16] and [17] and TMFo um [18]
p opose in en -based managemen a chi ec u es wi h a ying
ocus and app oaches. Howe e , Common o all is he de -
ini ion o wo oles (in en owne and in en handle ), wi h
high-le el APIs o in en submission and s a us epo ing.
This a icle ollows his app oach and de ails how he com-
pu ing con inuum managemen can coo dina e ac oss di e en
domains ia s anda d in en -managemen APIs.
III. EXAMPLE INTENT
A ideo s eaming se ice ha deli e s con en o use s
o e he in e ne , like Ne lix, equi es e icien p ocessing
o ideo ames and handling o la ge olumes o da a o
ensu e high-quali y playback. Fo ins ance, such a se ice
can ha e a business in en o p ocessing ideo ames in
unde 100 milliseconds. A coo dina ion se ice’s ole is o
disassemble his applica ion-le el in en in o ins ance-le el
in en s, as shown in Figu e 1. In he ne wo k domain, his can
be ansla ed o an in en o ne wo k slices wi h la ency below
50 milliseconds and bandwid h g ea e han 10Mbps. In he
compu e domain, he in en can be speci ied as a equi emen
o p ocess ideo ames wi hin 30 milliseconds. In he s o age
domain, he in en can ocus on ensu ing apid access o
ideo con en wi h a ead la ency o less han 10 milliseconds
and main aining da a a ailabili y a 99.99%. To accommoda e
eal-wo ld a iabili y, occasional de ia ions om hese a ge s
can be ole a ed, such as allowing p ocessing imes o each
110 milliseconds du ing sho pe iods o peak demand, while
ensu ing ha 99% o he ime, he p ocessing emains unde
100 milliseconds. This ole a ion is c ucial o main aining
se ice quali y wi hou o e -p o isioning esou ces.
The way hese in en s a e me is p op ie a y o each domain
p o ide . The ne wo k p o ide can le e age knowledge abou
he a ailable ne wo k esou ces in each a ea o he compu ing
con inuum o es ima e he bounda ies o he la ency ha can
be p o ided in each a ea. The compu e p o ide can use
a ious scaling and esou ce alloca ion s a egies o handle
he compu a ional demands e icien ly. The s o age p o ide
can implemen caching and eplica ion echniques o en-
hance a ailabili y and educe ead la ency. Ins ance manage s
con inuously p o ide eedback o he coo dina ion se ice,
allowing o dynamic adjus men based on eal- ime da a.
Sco ing unc ions can be employed o e alua e how well each
domain is mee ing i s speci ic in en . This mechanism guides
op imiza ion e o s o ensu e ha he applica ion-le el in en
is consis en ly achie ed.
Ne wo k
In en Manage
Compu ing
In en Manage
De e mine
S o age
In en Manage
In en Coo dina ion
Ta ge o
Con ols Con ols Con ols
Ta ge o Ta ge o Ta ge o
Compu e-domain in en
ideo ame p ocessing < 30ms
Ne wo k-domain in en
ne wo k slice la ency < 50ms
bandwid h > 10Mbps
S o age-domain in en
ead la ency < 10ms
a ailabili y > 99.99%
Applica ion-le el in en
ideo QoS (low la ency, high esolu ion);
i.e., ame p ocessing < 100ms
Fig. 1: Example o in en coo dina ion.
IV. INTENT COORDINATION ACROSS DOMAINS
He e, we display he a chi ec u e o inCoo d. The goal is
o build a lexible and sel -adap i e solu ion o coo dina ing
mul i-domain in as uc u e ins ances, based on in en s. In
complex and hie a chical sys ems, such as he compu ing
con inuum [19], [20], swi ching om “how” o “wha ” is
essen ial. Ou key idea is o achie e ha by only con olling
he ins ance-le el in en s. We une hese in en s, as knobs, o
ul ill he applica ion-le el in en and handle con lic s ac oss
in as uc u e domains. This app oach makes he solu ion
lexible and independen o in en s and in as uc u es, as i
au oma ically igge s s a egy adap a ions o he in ol ed
ins ance manage s.
A. F amewo k and me hodology o coo dina ion ac oss do-
mains
inCoo d
in en Ful illmen
Top
<< equi emen >>
Ful ill business-le el
in en s o e ime
Mid-le el
<< equi emen >>
Con lic s esolu ion
Low-le el
<< equi emen >>
Con ol esou ce-
le el in en s
Low-le el
<< equi emen >>
Lea n sys em
dynamics
«Coo dina o »
decisionMake Componen
«plugin»
in en T ansla ion
In en Knowledge
«plugin»
eedbackComponen Resou ce domain manage s
«ex e nal»
Compu ing In en
Manage
«ex e nal»
Ne wo k
In en Manage
«ex e nal»
S o age
In en Manage
Repo
In en Modi ica ion
Repo
Repo
In en Con ol
In en Modi ica ion
In en Modi ica ion
Fig. 2: Componen diag am o inCoo d’s main a chi ec u e
and in e aces.
To ha e a clea o e iew o he inCoo d ole in a mul i-
domain compu ing in as uc u e, we can analyze he main
componen s depic ed in Figu e 2. inCoo d ope a es based on
high-le el ins uc ions ansla ed and decomposed in o ac ion-
able in en s. This se con ains a leas one applica ion-le el
in en ha e lec s he objec i e o he a ge applica ion, and
a leas one ins ance-le el in en o each domain ins ance. A
dedica ed componen (“in en T ansla ion”) pe o ms
his ansla ion h ough an ex e nal ool o an in e nal plugin
in inCoo d. In [21] we implemen ed a dedica ed “In en - o-
Lea ning” plugin ha , h ough a mul i-agen sys em buil on
La ge Language Models (LLMs), ansla es applica ion-le el
in en s in o execu able Rein o cemen Lea ning (RL) en i on-
men s. The “in en Ful illmen ” holds he p ocedu e,
whe he i is a ule o a mo e complex s uc u e, such as an
RL ewa d unc ion, o e alua e whe he he coo dina ion suc-
cess ully ul ills he applica ion-le el in en . This eedback can
be p oduced by an ex e nal ool du ing he in en ansla ion
phase o manually cons uc ed by a domain expe . Again,
s uc u ing i as a plugin lea es i open o adap ing o e e y use
case. In ac , a un ime, inCoo d elies on eleme y om he
in as uc u e manage ins ances and applica ion, con inuously
e alua ing each in compliance wi h he de ined ins ance-le el
and applica ion-le el in en s, espec i ely. This in o ma ion
acili a es eal- ime decision-making.
The “decisionMake Componen ” is he inCoo d co e
elemen , as i holds he s a egy’s logic. By lea ning he
sys em’s dynamics, his componen decides a un ime whe he
o adjus he in en o each ins ance. The upda e ac ions
consis o o wa ding he new ins ance-le el in en s o each
domain manage ins ance so ha , in u n, hey can make
low-le el decisions on he in as uc u e. These ecu en ad-
jus men s ha e wo main e ec s. The i s one is o ailo
he ins ance-le el in en s o he eal equi emen s o he
applica ion. E en i ca e ully de ined h ough some expe
knowledge o applica ion p o iling, i is s ill possible ha
he in en s migh be di e en . Fo example, he compu ing
demand migh be elaxed, allowing he applica ion owne o
sa e on in as uc u e cos s. The second impo an aspec is
esol ing con lic s in a mul i-domain, possibly mul i-p o ide
in as uc u e, whe e condi ions can change a un ime.
We en ision inCoo d in e ope abili y h ough he de ini ion
o APIs. The API enables equi emen s “Inges ion,” i.e.,
whe e in en - ansla ion ools can submi bo h applica ion- and
ins ance-le el in en s; u he mo e, inCoo d o e s epo ing,
whe e i sha es cu en choices and me ics. E en -d i en
no i ica ions could allow inCoo d o ecei e eal- ime upda es,
e.g., iola ions, ollowing a subsc ip ion model. Fo example,
he amewo k p oposed by TM Fo um enables he ecep ion
o epo s associa ed wi h he in en s.
O e all, he inCoo d a chi ec u e aligns wi h he TMFo um
p oposed in en -managemen a chi ec u e, which on he one
side enables he applica ion owne s o ecei e pe iodic epo s
ega ding he ul illmen o he in en objec i es, and on he
o he side p o ides mechanisms o de e mine he es ima ed
alue an in en handle can gua an ee o a ce ain objec-
i e [22]. This la e is done h ough he bes in en epo
expec a ion p oposed in [22]. In p ac ice, his allows inCoo d
o es ablish he bounda ies o he applica ion- and ins ance-
le el in en s ha impac he di e en esou ce le el domains.
Fo example, he applica ion-le el end- o-end la ency o he
ideo-s eaming se ice needs o be balanced and coo di-
na ed be ween he la ency o he compu e esou ce domain
(p oduced by he ideo s eam p ocessing), and ha o he
ne wo k connec ing he applica ion o he end-use s. I is wo h
highligh ing ha bo h he applica ion-le el in en and in e nal
coo dina ion policies help de e mine ole ance o ul illing a
speci ic objec i e. This bes -e o bounda y es ima ion aligns
wi h how inCoo d con inuously adjus s and e-aligns ins ance-
le el objec i es ac oss domains. In pa icula , he abili y o
inCoo d o dynamically p obe easibili y and adap o eal-
ime eleme y mi o s he in en o he BEST/PROPOSAL
p ac ice. The API s uc u e and un ime loop wi hin inCoo d
can inco po a e such eedback by eques ing bounda ies om
esou ce domains and e ining goals acco dingly, wi hou
o e s epping domain-speci ic au onomy.
obse e he cu en s a e o esou ce g oups
calcula e necessi y o adjus esou ce-le el in en s
need o adjus in en alue?
send new in en alues
eedback on business-in en ul illmen
yes no
no u he ac ion
Fig. 3: High-le el wo k low o he inCoo d componen .
Wo k low Fo he inCoo d un ime, we en ision a con ol
loop ha imp o es i s decision-making o e ime, allowing o
sel -adap ing cha ac e is ics. Figu e 3 summa izes he ac i i y
and highligh s he majo s eps. In pa icula , he deployed
me hod should lea n which ins ance-le el in en alues a e
mo e app op ia e. To do so, inCoo d moni o s whe he each
ins ance manage has b eached i s in en . I so, ha likely
means ha i should adjus he a ge s. Fu he mo e, and mos
impo an ly, i should check he impac on he applica ion-le el
in en , i.e., he applica ion pe o mance. Based on ha , he
mechanism should eason whe he some upda es a e necessa y
and o wha magni ude. Downs eam, i should be able o
unde s and i s impac on he o e all equilib ium. The e o e,
he inCoo d mechanism should be able o lea n om p e i-
ous expe ience and de elop an unde s anding o he sys em
dynamics.
V. EVALUATION
To demons a e he p ac ical applicabili y o ou coo dina-
ion amewo k, we p esen a simpli ied use case based on he
in en de ined in Sec ion III. The p ima y s akeholde in his
scena io is a ideo s eaming se ice p o ide whose business
objec i e is o deli e consis en high-de ini ion con en o
end use s. The a ge objec i e is o main ain he se ice’s
quali y o expe ience (QoE) wi hou educing he ideo quali y
(se a 1440p). Fo his use case, we limi he coo dina ion
o wo manage ins ances, one o he ne wo k domain and
he o he o he compu ing one. Hence, he coo dina ion
200 250 300
Time s eps
0.0
1.5
3.0
La ency (s)
Ne wo k La ency Compu ing La ency 720p
1080p
1440p
Resolu ion
Fig. 4: Exce p o he aining da ase , wi h ne wo k and
compu ing la encies and co esponding ideo quali y.
agen mus lea n o balance esou ce alloca ion ac oss hese
domains, unde s anding ha imp o emen s in one a ea migh
compensa e o limi a ions in ano he .
Da ase : We c ea e wo da ase s, he aining and es ing,
wi h 1,000 and 10,000 elemen s, espec i ely. These syn he ic
da ase s a e based on agg ega ed pe o mance da a om an
ac ual s eaming se ice o ensu e ha ou expe imen s e lec
eal-wo ld ope a ional pa e ns. Each da a en y cap u es he
mul idimensional na u e o s eaming pe o mance h ough
key me ics: o al ne wo k la ency, which ep esen s end- o-
end ne wo k delay; o al compu ing la ency, which encom-
passes all necessa y ope a ions such as ideo encoding and
p ocessing o e head. These wo can in luence he h oughpu ,
hus e lec ing a ailable bandwid h o con en deli e y and,
as a consequence, he esolu ion, i.e., he a ge me ic. In he
ollowing Figu e 4, we epo a b ie exce p o he gene a ed
ne wo k and compu ing la encies o he aining se , and
he esul ing esolu ion associa ed wi h hese alues. I is
e iden how la ge ne wo k and compu ing la encies lead o
pe o mance deg ada ion, o cing a sub-pa esolu ion o 720p.
A. inCoo d Mechanism Implemen a ion
In he con ex o sel -adap i e sys ems, esea che s and
p ac i ione s ha e de eloped se e al solu ions, whe e hey
demons a ed he use ulness o Rein o cemen -lea ning-based
app oaches [23]–[26] in mul i-objec i e coo dina ion s a e-
gies. Howe e , e en hough he e has been a ocus on de ining
he bes h esholds o au o-scaling using RL solu ions [27]–
[29], wi h p omising esul s, no one has applied his whole
app oach o mul i-domain use cases. Wi h inCoo d, we aim
a achie ing a sel -adap i e coo dina ion s a egy o mul i-
domain in as uc u e ins ances. In pa icula , o his e alua-
ion, we compa e h ee app oaches.
1) MLP-based DQN: Based on he Deep Q-Ne wo k a chi-
ec u e p oposed by DeepMind (2013), using deep neu al
ne wo ks o policy app oxima ion.
2) PPO: P oximal Policy Op imiza ion, in oduced by Ope-
nAI (2017), is a policy g adien me hod ha es ic s la ge
policy upda es o aining s abili y.
3) GRPO: G oup Rela i e Policy Op imiza ion ex ends PPO
by compa ing an ac ion o a sampled g oup o al e na-
i es, emphasizing ela i e ad an age.
The inCoo d agen goal is o main ain he a ge s eaming
QoE by con olling la ency in en s o ne wo k and compu ing
0 25 50 75 100
Pe cen age (%)
GRPO
DQN
PPO
Ou come
Succeeded
Pa ially success ul
Unsuccess ul
Fig. 5: Pe o mance o RL models in main aining 1440p ideo
esolu ion
domain manage s. The agen can decide o keep he in en s
as-is, o igh en ei he he ne wo k o compu ing objec i e,
o bo h. The new la ency a ge is he esul o an adap i e
s ep mechanism s ha dynamically adjus s he magni ude o
la ency educ ions based on cu en h oughpu Rccondi ions.
Ra he han applying ixed educ ion alues, we employ an
exponen ial scaling ac o ha becomes mo e agg essi e when
h oughpu is below a ge le els and mo e conse a i e when
adequa e h oughpu is achie ed. We con ol he adap a ion
wi h he equa ion s= educ max ×e(−α×(Rc
R )) whe e αcon-
ols he agg essi eness o he educ ion, educ max de ines he
maximum possible educ ion s ep, Rcis he cu en h oughpu
and he a ge h oughpu R is se o 15,000 kbps o suppo
1440p s eaming.
B. Resul s
We e alua e he ained models on 10,000 unseen es
en ies, measu ing hei abili y o main ain he a ge QoE.
Ou comes a e classi ied as success ( a ge me ), pa ial suc-
cess (co ec ac ion di ec ion bu insu icien magni ude), o
ailu e (no e ec i e co ec ion). As shown in Figu e 5, PPO
pe o ms bes , wi h o e 70% success ul ac ions, indica ing
i s s ong sui abili y o his ype o coo dina ion ask. GRPO
unde pe o ms, likely due o poo adap a ion o his scena io,
calling o edesign o u he uning. DQN achie es mode -
a e pe o mance (a ound 50%), bu may bene i om mo e
sophis ica ed a chi ec u es o hype pa ame e uning.
The e ec i eness o PPO is depic ed in Figu e 6a and
Figu e 6b, whe e i is immedia ely isible how he la ency
alues a e success ully educed. The iolin plo s show he
dis ibu ion o da a poin s on a iables. Each iolin is d awn
using a ke nel densi y es ima e o he unde lying dis ibu ion.
PPO is imp o ing bo h compu ing and ne wo k la ency, show-
ing p omising esul s in unde s anding he ac ions needed o
gua an ee be e ideo quali y. No ably, while PPO and GRPO
exhibi simila la ency p o iles, he DQN beha es di e en ly
be ween he wo. This di e ence is highligh ed by Figu e 7.
The e, we depic he dis ibu ion o ac ions aken by each
me hod. We can see how, in e es ingly, DQN is pe o ming
a mo e di e se se o ope a ions, by deciding o change
ei he compu ing o ne wo k la ency mo e equen ly han he
o he wo me hods. In con as , GRPO is e y conse a i e,
leading o signi ican ly mo e iola ions han he o he me hods.
O e all, PPO achie es good esul s by “simply” deciding o
dec ease bo h ne wo k and compu ing la encies. While his
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Ne wo k La ency (ms)
PPO
DQN
GRPO
O iginal
Med: 0.94
Med: 1.10
Med: 1.06
Med: 1.19
(a) Ne wo k la ency change wi h each RL me hod.
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
Ne wo k La ency (ms)
PPO
DQN
GRPO
O iginal
Med: 0.94
Med: 1.05
Med: 1.06
Med: 1.19
(b) Compu ing la ency change wi h each RL me hod.
Fig. 6: Violin plo s showing he change in oduced by he h ee
es ed RL solu ions.
0 25 50 75 100
Pe cen age (%)
GRPO
DQN
PPO
Ac ion
Keep la ency
Dec ease bo h la encies
Dec ease ne wo k la ency
Dec ease compu ing la ency
Fig. 7: S acked ba cha o ecommended ac ions o he h ee
es ed RL algo i hms.
0 5 10 15 20
Episode
-100k
0
200k
400k
Cumula i e Rewa d
Algo i hm
DQN
PPO
Fig. 8: Cumula i e ewa d e olu ion o DQN and PPO o e
20 episodes.
gua an ees he bes esul s, mo e es s a e necessa y o in es-
iga e solu ions ha o e mo e di e si ied ac ions, as his can
help be e sol e con lic s. Finally, i is in e es ing o look a
how he wo mo e compa able models, and he ones o which
we plan o pe o m mo e accu a e es s, wo k du ing aining.
To do so, we analyze he cumula i e ewa d. This inspec ion
is s anda d in RL pe o mance e alua ion and allows us o
see i and how an RL algo i hm lea ns, o e ime, o pick he
bes ac ions. We depic he esul in Figu e 8. Su p isingly,
while PPO clea ly has a p og essi e imp o emen , ou DQN
algo i hm seems o su e du ing lea ning.
Takeaways These esul s show ou app oach’s capabili y o
adap in a dynamic en i onmen , adjus ing he in as uc u e
o main ain he applica ion-le el in en ul illed. A he same
ime, he e alua ion highligh s he need o in es e o in
adjus ing he RL s a egies. Bo h DQN and PPO would bene i
om p ecise ine- uning, which we in end o include in u u e
wo k. Gi en he lexible a chi ec u es, we plan o in oduce
s a egies based on o he p inciples a a la e s age, e.g.,
he p omising ac i e in e ence [30]. Finally, we aim o es
he gene alizabili y o ou inCoo d-based app oach ac oss
di e en p oblems, inco po a ing addi ional ins ances om
di e se domains, such as s o age.
VI. CONCLUSION
In his pape , we p esen ed inCoo d, an in en -awa e so-
lu ion o ha monizing in as uc u e ins ances in he mul i-
domain compu ing con inuum. inCoo d, con e sely om o he
app oaches, does no equi e di ec in as uc u e managemen .
On he con a y, i allows ins ance manage o deploy hei
s a egies. The con ol akes place by upda ing he ins ance-
le el in en s, hus p omp ing changes in he manage s’ s a e-
gies. We o e ed a de ailed de ini ion o he inCoo d a chi-
ec u e, oge he wi h a P oo o Concep . The simula ion
esul s in a ideo s eaming use case show a low numbe o
iola ions, highligh ing he p oposed solu ion’s sel -adap i e
po en ial, gi en i s capabili y o adap quickly o he en i on-
men dynamics. In u u e wo k, we plan o ex end his solu ion
o mo e challenging and di e se use cases and include mo e
di e se domain ins ances.
REFERENCES
[1] I. S oica and S. Shenke , “F om cloud compu ing o sky compu ing,”
in P oceedings o he Wo kshop on Ho Topics in Ope a ing Sys ems,
pp. 26–32, 2021.
[2] A. Lei adeas and M. Falkne , “A su ey on in en -based ne wo king,”
IEEE Communica ions Su eys & Tu o ials, ol. 25, no. 1, pp. 625–655,
2022.
[3] T. Me sch, M. Vik o sson, A. Hoban, M. Vi ali, R. Iye , and E. Elm o h,
“In en -d i en o ches a ion: En o cing se ice le el objec i es o cloud
na i e deploymen s,” SN Compu . Sci., ol. 4, no. 3, p. 268, 2023.
[4] A. Mo iche a, N. Sp ing, P. Rai h, and S. Dus da , “In en -based
managemen o he dis ibu ed compu ing con inuum,” in 2023 IEEE In-
e na ional Con e ence on Se ice-O ien ed Sys em Enginee ing (SOSE),
pp. 239–249, IEEE, 2023.
[5] M. D. Lei e de Souza and T. Me sch, “Ad ancing o ches a ion: Le e -
aging dis ance unc ions o mee applica ion in en s,” in 2024 IEEE/ACM
17 h In e na ional Con e ence on U ili y and Cloud Compu ing (UCC),
pp. 82–87, 2024.
[6] N. Akba i, J. G undy, A. Cheema, and A. N. Toosi, “In en con inuum:
Using llms o suppo in en -based compu ing ac oss he compu e
con inuum,” a Xi p ep in a Xi :2504.04429, 2025.
[7] A. Za ei opoulos, E. Fo opoulou, C. Vassilakis, I. Tzane is, C. Lom-
ba do, A. Ca ega, and R. B uschi, “In en -d i en dis ibu ed applica ions
managemen o e compu e and ne wo k esou ces in he compu ing
con inuum,” in 2023 19 h In e na ional Con e ence on Dis ibu ed
Compu ing in Sma Sys ems and he In e ne o Things (DCOSS-IoT),
pp. 429–436, IEEE, 2023.
[8] T. Me sch and A. Hoban, “B eaking ba ie s: F om black box o in en -
d i en, use - iendly powe managemen ,” in P oceedings o he 14 h
In e na ional Con e ence on Sma Ci ies and G een ICT Sys ems,
SMARTGREENS 2025, pp. 24–35, SCITEPRESS, 2025.
[9] N. Filinis, I. Tzane is, D. Spa ha akis, E. Fo opoulou, I. Dimoli sas,
A. Za ei opoulos, C. Vassilakis, and S. Papa assiliou, “In en -d i en
o ches a ion o se e less applica ions in he compu ing con inuum,”
Fu u e Gene a ion Compu e Sys ems, ol. 154, pp. 72–86, 2024.
[10] S. K. Pe epu, J. P. Ma ins, R. S. S, and K. Dey, “In en -based mul i-
agen ein o cemen lea ning o se ice assu ance in cellula ne wo ks,”
in GLOBECOM 2022 - 2022 IEEE Global Communica ions Con e ence,
(Rio de Janei o, B azil), pp. 2879–2884, IEEE, Dec. 2022.
[11] C. V. Nahum, V. H. L. Lopes, R. M. D ei ue s , P. Ba is a, I. Co ea,
K. V. Ca doso, A. Klau au, and R. W. Hea h, “In en -Awa e Radio
Resou ce Scheduling in a RAN Slicing Scena io Using Rein o cemen
Lea ning,” IEEE T ansac ions on Wi eless Communica ions, ol. 23,
pp. 2253–2267, Ma . 2024.
[12] H. Song, A. Soylu, and D. Roman, “Towa ds Cogni i e Sel -
Managemen o IoT-Edge-Cloud Con inuum based on Use In en s,” in
2022 IEEE/ACM 15 h In e na ional Con e ence on U ili y and Cloud
Compu ing (UCC), (Vancou e , WA, USA), pp. 1–4, IEEE, Dec. 2022.
[13] N. Filinis, I. Tzane is, D. Spa ha akis, E. Fo opoulou, I. Dimoli sas,
A. Za ei opoulos, C. Vassilakis, and S. Papa assiliou, “In en -d i en
o ches a ion o se e less applica ions in he compu ing con inuum,”
Fu u e Gene a ion Compu e Sys ems, ol. 154, pp. 72–86, May 2024.
[14] A. Za ei opoulos, E. Fo opoulou, C. Vassilakis, I. Tzane is, C. Lom-
ba do, A. Ca ega, and R. B uschi, “In en -D i en Dis ibu ed Appli-
ca ions Managemen O e Compu e and Ne wo k Resou ces in he
Compu ing Con inuum,” in 2023 19 h In e na ional Con e ence on
Dis ibu ed Compu ing in Sma Sys ems and he In e ne o Things
(DCOSS-IoT), (Pa os, Cyp us), pp. 429–436, IEEE, June 2023.
[15] “Ze o- ouch ne wo k and Se ice Managemen (ZSM); Closed-Loop
Au oma ion; Pa 3: Ad anced Topics,” G oup Repo GR ZSM 009-3
V1.1.1, ETSI, Sophia An ipolis, F ance, Aug. 2023.
[16] 3GPP, “3GPP TS 28.312: Telecommunica ion managemen ; S udy on
managemen and o ches a ion o 5G ne wo ks and ne wo k slicing;
S age 1,” Technical Speci ica ion 28.312, 3 d Gene a ion Pa ne ship
P ojec (3GPP), 2023. Accessed: 2025-07-11.
[17] 3GPP, “3GPP TS 28.912: Telecommunica ion managemen ; S udy on
a chi ec u e o nex gene a ion sys em,” Technical Speci ica ion 28.912,
3 d Gene a ion Pa ne ship P ojec (3GPP), 2020. Accessed: 2025-07-
11.
[18] TM Fo um, “TMF921A In en Managemen API P o ile,” Technical
Speci ica ion TMF921A, TM Fo um, Mo is own, NJ, USA, Ma . 2022.
Published 31 Ma 2022; TM Fo um App o ed.
[19] A. Mo iche a, V. C. Pujol, and S. Dus da , “A oadmap on lea ning and
easoning o dis ibu ed compu ing con inuum ecosys ems,” in 2021
IEEE In e na ional Con e ence on Edge Compu ing (EDGE), pp. 25–
31, IEEE, 2021.
[20] A. Mo iche a, A. Lackinge , and S. Dus da , “Cohabi a ion o in elli-
gence and sys ems: Towa ds sel - e e ence in digi al ana omies,” in 2024
IEEE In e na ional Con e ence on Se ice-O ien ed Sys em Enginee ing
(SOSE), pp. 102–110, IEEE, 2024.
[21] C. Capo a, A. Mo iche a, A. Lackinge , and S. Dus da , “In en ans-
la ion o mul i-domain cloud con inuum managemen ,” in 2025 IEEE
33 d In e na ional Con e ence on Ne wo k P o ocols (ICNP), IEEE,
2025.
[22] TM Fo um, “T 291c: P oposal o bes in en model,” ech. ep., TM
Fo um, 2022.
[23] G. Minelli and M. Musolesi, “Comix: A mul i-agen ein o cemen
lea ning aining a chi ec u e o e icien decen alized coo dina ion and
independen decision-making,” a Xi p ep in a Xi :2308.10721, 2023.
[24] P.-A. D agan, A. Me zge , and K. Pohl, “Towa ds he decen alized
coo dina ion o mul iple sel -adap i e sys ems,” in 2023 IEEE In e -
na ional Con e ence on Au onomic Compu ing and Sel -O ganizing
Sys ems (ACSOS), pp. 107–116, IEEE, 2023.
[25] C. K ¨
ohe , L. Ge ling, and K. Schmid, “Con ol ac ion ypes-pa e ns
o applied con ol o sel -adap i e sys ems,” in 2023 IEEE/ACM 18 h
Symposium on So wa e Enginee ing o Adap i e and Sel -Managing
Sys ems (SEAMS), pp. 32–43, IEEE, 2023.
[26] N. Nguyen, D. Nguyen, J. Kim, G. Rizzo, and H. Nguyen, “De-
cen alized coo dina ion o mul i-agen da a collec ion in dynamic
en i onmen s,” IEEE T ansac ions on Mobile Compu ing, 2024.
[27] S. Ho o i z and Y. A ian, “E icien cloud au o-scaling wi h sla objec i e
using q-lea ning,” in 2018 IEEE 6 h In e na ional Con e ence on Fu u e
In e ne o Things and Cloud (FiCloud), pp. 85–92, IEEE, 2018.
[28] F. Rossi, V. Ca dellini, F. L. P es i, and M. Na delli, “Dynamic mul i-
me ic h esholds o scaling applica ions using ein o cemen lea ning,”
IEEE T ansac ions on Cloud Compu ing, ol. 11, no. 2, pp. 1807–1821,
2022.
[29] T. Tou nai e, H. Cas el-Taleb, and E. Hyon, “E icien compu a ion o
op imal h esholds in cloud au o-scaling sys ems,” ACM T ansac ions on
Modeling and Pe o mance E alua ion o Compu ing Sys ems, ol. 8,
no. 4, pp. 1–31, 2023.
[30] T. Pa , G. Pezzulo, and K. J. F is on, Ac i e in e ence: he ee ene gy
p inciple in mind, b ain, and beha io . MIT P ess, 2022.
