i want summary please

Survey paper When machine learning meets congestion control: A survey and comparison Huiling Jiang s", Qing Lit ker, Yong Jiang ", , GengBiao Shen ", Richard Sinnott ", Chen Tian', ATICLE INNO A \& S A A CT 1. Introduction 1.1. Tradidonal congrinion cosurol As a fursalanental component of computer netweeks, congestion The loternet transmission protocol is based on packet switching over control () plays a signiflicant role in impreving the network resource besteffort network forwanting [1], Find 40-end transmivion contrel is utilization to achieve betier performance. With the development of a required to pronide a reliable service for applications. To aveid netwark large number of widely waed technologie, es. dath conten (DC), degradation caused by congestion, oC alyorithms are typically emWif, 50, and satellite comerankations, the complesify and tive. ployed to improve reliable tranamistion over the setwork. Cangestion sity of network tranamission wesarios and peotocols have tereased in a state of the netweek in which it is not capable to deliver the dramatically. This has brought significast challenges to tranamission serice it was designed for. Netwoek congestion occun when eacrasive protocol design. A rich variety of CC algorithus have brow designed aumben of data padiess are sem over the netwark by hosis [2]. The for rpecific wenarion. However, the variety of netwerk wemarion and objective of OC alporithm in to achieve higher netwock twoughpoit bemely difficult to design efficient generic C Algonithms. Therefore, Wleally also guaraniee faimess berween end-to-end sesions. CC algorithass broed on makhune leamieg (ML.) have been perpoied The traditianal CE alpocithms can be caiegoriznd into two type: different network senarios, In thin paper, we provide a backgound proshes cely require the eallaboration of arnden and receivers, and analysis of traditional CC. Alaied on thik, we iavetigate correst works bence they do not rely an any explicit signals from the setwork. that research challenges in the application of ML. in the field of OC. Network-atisted appeoaches require the suppoit of network devices, - Cirrerponsing auther. Availite alline 31 March 2021 af. Alay at al fairness and eesponaiveseat in complex networking acrnariat. technigon. Superviked and unaupervised Jearning techeigues bave For end-to-end OC, une of the main challenges is to identify net-. beea adrly emplayed to improve the performance of network CC IIti, packet loss. Lasses are generaied when the buffer in a given nerwor . hgher delline leaming capability [24,25]. At peesent, inuch research device is overloaded, then lose-based approaches are supposed te attain foctors se Mt-based cr, whemet. high theoughput by making ase of the link hasdwidth. Hirwever, for Howrver, leaming based C, is still in its infancy. Most learningacote drlay-ienuitive applikations, lower tranamiaion time fannot be bund CC alporithms adjust the congention windos (CWND) to control suaranieed. Besides, the packet loss nay not be trigered by network. the sending not inatead of adpuating the sending rate directiy. Therecongrstice (e.g., rasiom packet droppingt. whikh may mislead or ferr, burwinen is atill a genblem in highspeed networks because the decisions. CWSD can itcrease sharply when multiple MCi's arrive [20]. Current Therelore, delay-based appruache such as Tirely [12) have bern lrarning haved CE alyorahms sach as [2%,251 generally focus on endlays caused by the network. Compared with Jotb-baned approaches, 0 , algurihan are not saitable for realistic netwerks due to Eime random packet lines, However, falculatiag the esatt tranmibsinn delmy remains a significant challenge. For esample, if there exids a alyefr change nhen the pachet is procrased in the Linar stack, the measered the the meased delsy cinnot be controlled precincly. appeoaches such as Veso [13] asd an adaptive and fair rapid increave stocarth, there is no comprebehsive survey en this aapect. There are and tranimissoe drlay, To solve this peoblem, network-asised OC, approaches such as the mork insurs include wheduling. localizatbos, data aggregarion, and so proposed. The network devices provide explicit sighals related to the hanod algorithess in wiseless wennor netwerks. Some echer surveys on network seatus for houts to make serding rate dechiont. When the gyencife setwork ewsinunments related to ML. techniques are groposed signal. The receiver will send hack the PCN signal in the ACK and the that here been employed for oognitive radine. In [11], the sevey cender will adjust the seruling rate accordingly. The ICN signal for gave an everview of the moet Mel. fechuitues encsentered in cellular provides finer'grained CC With the dramatically ifcreased complesity and divenity of net-. Aurvey ovven alousdant aspects nelated to netwoeking. it only gives traffic patiems in eee network scenario may also affect the perlormunot . Inarning-hased oct ander dynamic setworks and campare them with 1.2. croming hased cangeution contrel porithmb pries readere eahausive insight and lays the foendatica for The dynamic nature, divenity, amel complecity of network somario In trallatic networls, the inteliementation of learning-based CC algohave browth sipnificant challenges for CC,As ach, it is diffieult as design a generic scheme for all network scenarion. Furthenacin, the ridhen has abewn that they are not efficient as supposcat to be because dynamic natare of even the same network can make the performunoe foulge the pros and eons of deciaion models, we conduct comprratensive of CC unatable, Current network mvieonments and aloo incloble beeh wired networks and wireless oetworks, making the detectios of pacief loss mare dafficult [14-20]. roperimenth of varicus schrmes by taing the NS3 emulatur [3.]. lis the nimulation, we compare the Hl. hased CC algorishums of Deep control decisions instead of using predetermined rules. This allos threr s firrrat soenarins with different configurations of bandwidth them to have betier adaptability to dynamic and complek netwod. Ant Acligy. The network with high bandwideh and low delay simulates Scenarios. 11. Nave at al These three netwiwk Esilroumests represent the diverse mvinenments Slow start. In the dewic dear start provesh, rash time a good we generate the DC iraffe which includes bull data trander which egponently aver rime. phase, and larrane the valae of CWND each time baind en the aine nows for evperinents as well. alower than the oponorntal groth rate durity the aliow atart. The experimental results show that Ieaming based 0 , algunichims Retrannmialion. Rretratomiuise incfules timeout retranamiution width is low or the link delay in low, learning.bascel CC alpurithms are wart wable a certain time, the desu is netransmitied until the transmistoo aggreaive so learn with great atability, Morewer, the perion ante sion is nacenala. Jart metratimbion requares the receiver to vend a of thene thrre learning hased : algorithme shuws no defference in avor duplicane ACX immetlatrly afrre ruveiving an iot-of sequence seghent simulabed environments because the ecomplenity of the mviroements in as that the mater lam a mien an ponaible that a mejnent has limited. Therefore, all of theis ran handle these network scenarion. In rraliseic scenaxios, 8l-hased OC algorithms are influensed to the compotation time needed for R1. This imparts the frasibiay af confirmution. The setrantimien mechanim in CC enatires that data these schemes. Therefore, we propose three poiemial soluticns to dral with this problent. Firuly. ineign ligheseighy models hacel on map Fant recovery. Fart rnewwry arane that when the mender seceives ping sables of sates and actions to decruase the time connumpeion three depleate NCXI it noecmive, it executen a multiplikation redocof karning decitiont, Secondly, decreme the ferquency. of decinium tien alporing and halso the alor start throhold to pervent netwouk. to prowide better feasibility under low-dynamic tetwerk scenaries. congeation. The CWKD increses in an accumulative manter, cwating. vergence, incomparillility and fairnem. Based oe the wadersanding and analyait of the curnat learning tased OC aclutions, we blentify trendh 2.2. Clent ongatin ountral alourthis in learning-tased C. Firstly, because of thrir capability for draling algorithms to give end hoss better insights inte network status. Third. given the excesive time and cout of harning decisions, lightwright learnirg thased CC will to a key research direction. Finally, an open mechanismi, is muppoted to be designet. conduct simulations and compare performances between BL based CC algoritum and traditinnal OC algorithm, in Secvise An, we ostline the been morivel, ster fast mover mechanism will conider the pachet conclude the paper. 2. Harkground and rubuetien aftrr packin lask but the drawbick in that in enly , parkert an los, it is comilend enngation. Therffore, Newilime [D] is In this section, we will introduce relerant donain kewcoledge in proponal as deal with thin problrin, which mainly inpruve the faut algorithme, and the performance metrics. parkes are mhraksoninted and rereived confirmation will the sender 2.1. Copgetion catrol methaniurn exit. Althosph Nrefins can wolve ithe problem of a large manber of parket lowsa, Nirntlemo cas ally have ene perket bese erior per aTT uime. In oeder to deal with the lose of a large number of dafa peckets procedure of CC, we adept the windembased CC to Hens. Therefore, the nender as know which data has been received developenent of kew-based C e algorithms. fairness and Cubic [as] is proponed to opeimize Bic, Cubic uses a cabac However, in some specific network scenarios, opeimized CC: algo- functinn to replace the growth function of Mic. Also, the most aritical are destgned to deal with this iswae, Bic LII] in an example, Bic uses growth rate of the CWND is completely independent of the netwoul the idea of binary search. When a packet lins wccurs, indicating that Mtr. Thar, fairness is ensured. The eptimal window value sheuld be smaller than this value, then Ifighspoed |l9] and Hobla [tid are suitable for high-speed nerworks Bie wets the CWND at this time to mrs, Msr, and the value after the as well. Mighipoed modifies the reaction functicon of the standard multiplication is reduced to minjus, The optimal value of the CWND Tor protocol, which is affected by the conbincd effect of the growth and improve throughput. and congetion losh. While the measurement of delay can also te However, in DCa, the Mif is relatively small, and sitwaer timewhed to inder the netvurk congrstion, which can coarpenate for the impact of the sandom packet loss. Thns, delay-based CE alporithms are atarfuing at promidnt by modem Network lateface Cards (NiCo) 50 jiretented. tive queue management scalable TCP (DAST) [47], How latrneice TCP in Timely [11]. (Eala) [4f, and Timely [12] The main topec of delay based TCP as to eimiar the mesturnmeth (LDAst) [43) adopen a ene-way delay bo meanure the congevtion of of delay such as RTT, the one-way delay, and equeve delay to obtain netwerks. The ahraetege of the cee-way delay ower init is that it does that is based on the growth function of Cubic by aning a fair fliw halancing mechanism. Though loila has betier inter-flow ETI fairnesi friendy to Cuble fleran. and better fiTT eatimation, if cannat caenint fairly with loss toanod CC; 1. Jarr n all Capan Nuratu ha caury naws saime ratical problem. Once in detcets abourmal traffic, it will make feodback frem the control channel, thua data channel is possible to used with QUIC, Copa provides higher qualify and lower latency for family is alko a anethod of dynamically adjusting the rate by detecting mabile live breadcas than Cubic and BRi, Moreover, Copa has lower available network resourcea including FCC. [5.], FOC Vivace [lie], and traniminsoe fTt add lower low overhead. IceC Preteus [59]. They show geat performance tusal an carefully Hybrid C E algorithms use loss and delay t0 jointly evaluate coe- deslgnod utility functions that cover hasic performance megies such getion. In this way, IHigh throughput and low latency may be achieved as round-trip time (BTT). Compared to IBER, INCE converges slower at the same time. There are some fypieal hybrid CC algorithms ach becane of PCCis eceacroative increasing mechanian. as Veno [13]. Africa [14]. Conipound [t50], Libre [5]L, and Coogle Reno. Network-atsisted CC, algorithms require she coordination of the Veno combines Vegas and leno. Vegas cart mrause the number of sender anal reociver, and require routers te perforth apecific procestdata packets belonging to this connection in the sotwork botileneck ling on packets. Typical algorithms consist of ECN-tased cC algorouicr. Veno uses this variable to distinguhh random parket lois from eichms [Se] and Qaantiked Congeation Notificution-bused (oCE.based) congecioe packet kns and takes differrnt actinns. Veno also improve algorithms [5V]. The nummary of network-anaisted CCE algarthni is the growth function of CWED that is the batic part of Meso. When peraented in Tahie 4 , the number of packet belonging to this connection in the network ICV is an estension to the Internet Frotocel and Tranumission bottleneck router escecds a cenain value, the growth rate of CWND will Control Protocol. ECN allows ced-to-nad notification of CR: to avoid be slowed down and thas packet losses will be reduccd. Experiments packet losi and requires specific support freen the laternet layer and thow that Veno can improve throdgheut aignificantly without advenely the transport layer. Generally spcaking. TC.P/lntereet Frotocol (ti) affecting other concurrent TCP connections comparci wish Vegas and ectworks indicate channel cungestion by dropping data packets. In the Reno (1]. case af successful BCN negutiation, the DCN-aware router can set a Airica [t.4] is based an Rese as well, which can switch modes mark in the IlP header lintead of dibcarding the packet to indicate that cient, Arlica uses an atgressive, scalable window increase mechanism. Freponds to the sending end in indication, reducing ita tranamision rate ETT is wed to meature congecion as well. If the congotice level thows as if if had detected pucket-bost in the unual way, There are ithan high thronghput but also enaures falmes. Compound (50] is also a bybrid algerithm thed to solve the pred. aveld sethansmissian and reduce wasting time, njecially eetwark jirter. lem of poor MTT faimess among lob-hased CC algorithets. Compound QCX is a set of end-io-end eongestinn notification mechanisms maintains two CWNDS, con is a atandard windose aimilar to the Heno applied to 2, Throgh an active revene notification, the padket liss mechanium and the scoond is a scalable delay window hased on Vegas. zate and delay in the network are reduced, thereby inproviag netThe CWWND is calculated based es the symmation of these two windews. work perfommance. QCY inchudes twu parz: Cosgestion Foint (CP) and standard losshased component. When the network is congested, the equipment samples the data frames that ace bring sent in the sending delay.based component will reduce the sending rate significantly. But befier. If congestion ocrurs, it will genwerate a Congestion Notifirarion the throughput will be lower benunded by Reno. Therefore, there is Mesage (CNM) to the fil af the sampled dara frame. When the Bd limited EIT unfairness due to delay-based achemes, and the throughput excrives the CNM informatice, it will limit the sending rate of the ATT unfaimess, Labra [SI] vises noelinear optimization to garantee its icsoling rate to detect the available bandwidth and recerer the rate fairnes amting TCR flows regardies of ftrt. Different from the above algorithus, GOC: [52] is applied in We- DCTCR. The proposes algorithm dramatically inproves the throughyet beKC which ases UDP.baced BTP be tranamif media data instead of and fairness of the traffic. Thowgh QCN is efficient in controlling the TCP, In GOC;, the rase ceeerel based on the parket bost nate funs on quese length and assists the OC, QCX cantion be deployed and ased the irTce na messege from the receiver at the sender and dynamically setworks. adjusts the oode rate at the sender acconfing to the packet lows rate indormation carried in its Meport Block. Delay-based rate coetrol runs on the receiving end. Hesides, there are scme innevative hybrid algorthus that have CK; algorihhms are expected tio achieve various goals and objectives Control Protocol (scr) (54). Perfoemanceociented Congestion Control Throughpot represents the amount of data that pases through a net(POC) [S5). JCC Vivace (SE), and FCC Nroteut [5\%], in Bemy [53], the utilly function condists of thoughput and delay. itheam high link utilization. Masimicing throphput is erucial. Given To maximize the evpocted value of the utility fanction, Reary finds the link handwidth, high throughput inaticates high efficiency in transthe mapping based on a pre-compubed looknp takle. Thence, the coere- ferring data. For instance, Bic is aimied bo achieve hightr throughgnat sponding arnding rate is evtimatrd. To converge to the egtimal sending . hased on a more aggressive mechanim compared with Reno (It). rale and fully utiline the network, BBA estimaies the wvailatle hand. HTT meavures the time including the transimission time, the gropat of adjustieg the sending fate by detecting the network stanis. Hearning time (FC;T) indicates the time required io tranafer the flowi. Itr and Cantrol Protiool (SCP) in (54). To achieve fast convergence, SCR panti- that they may have to tolerate. However, it may he the case that the actual sending mate in the data chansel, SCP serulu stand-in packets High throughapul means making use of the link bandwidth as ituch is the control channel to probe the netwerk condition. Based on the as possible, which fan give riae to an increased geeue length that if. Aary at at. Cinwane Nureets EA2 cat'1] Javet may cause drlays, In Vegan [4h, the cruclal insight in to peedict the becass optimal tratfic clasaification policies can promete the reasoncongestion level bowed en measured MtT. Moreover, the lower ilt and ahle allocacion of neriark msources and redaer the peobability of Ny deaigning two channels, aere packet loss is ponsibie. agplication of apervied learning alporithms in the field of CE. Fairness is a measure of equality of the resousce allocatinn of the There in a begr amowet of supervied learning technioues used POC Vivace [50] is the improsed speed ad responve. application, which laprowes the acouracy aignificantly. In this paper, These objectives are important for all CC algorithus. ter they are the acraracy in about 965 . The inaight proviled in this mesearch is that also have different priarities, and hence trade-offs are nocemary. Based fording the Gasalan ansoptins and ingruving the quality af dis: drtail. fartaicit dhcriminant poeve among shitinctive elawses. To reduce the the presented nodel, nulbigle binary SVM dawafien are organinal into a toormaneme wenchare, wheh can ifrimatically dectease the number of an opeimal madel, and then use this model to map all inputs to corre. prove accurary and relure oumputational cent. The aimulation shown learaling techniques have the atelay an perforent dase dassification. times. forests, Aayes, regrescion, and necural networks. In the network field, supervised learning methode are used is im: inces nach an rlaniflocien acoeracy and computarion cost. Eacept length for metwork wasisted networks which affect the performance of CE directly. Congestien signal predictioe conalsts of los clawificarion and deloy prediction, As mentioned before, congestian is defecard implicity hased in packet loss of delay when coegrnzin occurs in traditinnal CC algotithms. In supentied learning thased O C. alporithm, congewion in estimated in adeance haved on curcent and previcen network stato woch as the packet arrival interval and the network delay. The key hans for this approsch is that netwurk atates form a cuntinuous tinse series, where the future atate can be pendicted by past states. An for qarue lengih managenent, supervised leamlig methodi play an impoetant role in the aceurate and efficient gerediction of queue length. In the neat two parts, a more detailed drscription will be shoren. 2.2. Cocpoution detection in end to-and neneurly 3. 1.1. Lai dounficanion the lack al hullens is Cesh, contestion loss is generatal when there is evential to undentand CC to fons clasiticatises in wirrles networks hased on staditicnal ce Roondrring losk canase be ignored in networls with inulti-channed arrival time in elasify wireira lons and congestion loss if the packet inter-arrival time is confined to a rarye, the missing perkrts are lod S Fervistal lraraitg buwell ce algorithms art able to seal with the due to whrekess lons. Otherwile, the kns is conaldered a cengestion line. In (enf). the number of loves and BOTI were used to distanguish the types of losies. The presented algurithm called Frghes in more efficient le medation, wirrlens lons, euntentien bom, and ecordering loss inguct the dechection of evepestion loss. Sopervised learning technilues while Spike [69] ahowi letter performance in wireiras hackhone tepolegy with multiple flows. Digzag 6j is relatively more general, and Irartang thand or algorhhats hence ia able to satisfy different topology acenarion but it is Mrnative Miscluaification at ane iswev, in winelese newwerks, jredefined pafameters detirmine the enurs in elataifying congevtion losa and wireas the clasisier. The minimum BTT value and the current HitT valae ane regarded as input atates. While in 7FO/, the arrival time internal of ACX; and miniman AIT are choom at fratures and the Nalioos algorithan in uned to dlasify the low into wireless loss and cengratica loss for satelIite networka in (18)], the one-way delay wat inter-packit times were used as states to predict boss caicgories. in (19), the quruing delmy. neod ts be onawered earefally to balance performance in diflenent the interarivel time, and lists of porkets were used as ingrits. Berides, diverse supenised learning techniques were applied, In [711, deciuin treck decision tree macemblet, bagking, random fremts, ratra trees, boosting, and mult - layer perceptreen were used to ctaidif the types of loss Simulations show that these iniclligent kon chasifiers actirne heh 312. Exlig gredinan accuracy in differest netwock scenarios. Is conchation, when classifying. As a cengerion signal, the delay of tranamininis reflects the the wireless boss and the cungention lons, delay information is the kernel anvost of is flight data, which show the overall lowd on the netuork. aves the sourees due io wavelength reacration. Honecer, tercause of shoers and Table 7 cenchals, nupeniod letrning lechatiques have high N. Nens n at Tupersioed harulas algeriline reacting equickly to avold congestion. the ervirunment. Some trward har ibown that the furure queue length exponentially weighted moving average technique demonurates a moee time series of proviout tratfic an ingot without considering diverve accurate al gorithm. parameters in the actave. A a tenily, these algoriahans leave space other parameters in the network, In [ bol, lintar regesaion was used to the quicue length. eatakinh the relatiombilp between 8rT and the seadiog rate. In fith. a Bayesian sechnique wa sned is simulate the distribution betwern 4. Unsupervised linarsing besed congeation coetrol algorithms delay and the wending rate asd then to predirt drlay based on the sending rate. This is necoled in teal eime video applications and wirches In this arctian, another eatryary of leaming banal Cr algorithms netwoeks. hight reaponsiveness have been proposed. Further rerarch is needed to mation carnot te folly provided, the unsopervised learning techniques techniques to improve delay predictions. unauporvised learning notbodn can be imployed as traffic clanifiers 3.2. Qurue longth managowent in nefwork asueted tweworis at well. Thus, hood an chasen gronated by unsepentiod leara: Qurue Ingib management is a hyy focun tor netwoek-aniared CC queve management (AQM) familly of ISV toshniquen. Homere, the (BME) 104]. is usel as a basic clustering method. The features selected in this work As for the detectoon ol coegarios, Engerviaed learning al gorithms the number of the puibed data packnts, efr. Blased on X-means, the authors adopt the feature selection to find an optimal featwee ant and los transformation to impeove the accuracy. The experimerus stwow that 4.1. Congonan derinion in end to-nd cargration hantral alporither the proposed method can obtain up to 80 overall accuracy. Compared with K-thenans, 1MM is generally used for Cavusian Mtixture 4. t.t. Leas chaturies Model. In K-meank, an unknown data point mast belone to a single cluater, wadle a data point can be mapped into meltiple clusters lused on EM. In [94l. the authoes use an Autodlas appraach which is an B6.55. However, the performance of clustering depends on the paraku: congeited pariet lom, of congention aigal will be trigernd, and the K-mean in coetplex data sets, but this advantage is limited. ing thravars to cantral the amdiet ratr. A detalled aummary ia ahown Nefeork exten mainly osealat of lossdelay pairs anat the namber of bunes berrwen felluees. Thas, moee research it eequired to broaden thin fiek. 4.1.2. frily prodican There are oely a lemeted number of uhuspervised leaming-based CC and the amociuted mechunine are presented in hy. 5 and Talle 12 . Ms shows in Fig. 3 , the sember obtains the states in wetwoeks ach an the menare siac. Ther. delly chatering it conducted hused un unsupervined learning meethode. The prisratied clusters can be used to adjus the meswar alm, the valifity of inesugr, the distance betiveen vehicles and Baris, and the tyje of mencege is diviled inbo different grosph and the bownt delwy it each group in selecked an the certumication parameter for cach elentre. Bricd on the comanumicatico parameter, a specife aralint ratir will be anigned to each cluster. Therefoee, hased on the meaksinarnt of deloy, CC, can be actirinvel. Similar to low chuntrring hated an antepervised leamiag technisock, the revedreh of the obervation, in wireless netwarks, losses can he chawifind inte congestion lins and wireiess hins, In thas rescard, die diatribution of loss-delay pairs ean be eaphared by a Hiddrat Marhov Madel (ECMO). 5. NH-baned coegentica ceetrol algorithme The IBtSt is trained to associate the fype of losues with a state haned cn model is more eftirient than Vegas predictors. rgooific arrions given the network utate, to determine if a given action research, the losses clustering problem is formulated as a statistical valae fanction then calcelates the value of the action and update it loe Righer. losses an well. In OAS, the aunhnen in [7J] find that the number af bunts between failutrs can be wed to differentiate congention and valae of acticns dirvitey. twe chantering methods to improve the performance of Cr. In the the large handmath of matWave and the high cust af infrastracture if dimer at at Cepane Nowntu ra2 cansis sonus Tatele 13 atrategy to impreve the probehelities of moccuful tranamisube ever topologies and divenifird fown are a major challenge fir]. Traditional in edge computing and redge cache also uses the Det. model such rithe, the m. atchigat in wad to opdate CWND based on different important role in the network feld including Cr Amongst the dafferent Mahehaied CC algarithms, th. hat gatsod the moit attention. Diflerent frem nupenioed learning methods, A. Alporithms monitoe the aatue of the rmvirunment cuntibueckly and rract to the envirenment to eptimise a stilify function. Therofoer, 31 , al gorithma are more suitable for variable and unutable netieurk envi: 5.1. Wind upieriat it end o-end ivetwarla renments. Two main trends are telated to this kind of ketwork. First, Intral of predicting congotion loss and delay as widh wivervised and a1 Irarn the OC nulen directly hawed an different emilentment infuenation. Since fil. algorithm can incorporate scal-time netwerk conditions and define arions sccontiegly, real-time contrul is powible in BL. algoeithins: use 7it. to update CWND for ppecific scetarius. The mechanise af It. based CC: algonibuas is shawn in fig. 6. the summary is shows in Table 11 and 12 , which show more ideialled information including atates, rewards, actions, pron, cons, etc. 5. 7.1. Ayschnincia tranufer mode netwerls Aynchronous transfer mode (ATM) is a typical network mitable for Bl-baied CC algorithms. ATM nerworla are clank netwerk that mpport mulb-media applications. For differras multimedia traffic, A7M offers ditferent QoS such as crll loss rate (CaM) and delay. Nowever, in ATM, highly time-varying traftic pettene can increase the ahcertainty of network traffic. Soccover, the small cell transminion time and low buffer siars in ATM networks require mare adaptive and high responsive OC algorithnis in foilL an MC algonthin is applied te doal with these peoblems, In the proposed CC algorithen, AC focuses 0de the the algocithm measures the action accoeding to the performance. In this way, diflerent traffic patierne ase connected with correpponding actions. Simulation rewults show that the Cur in low and volice quality in maintained, Comparod with clasical optimal control aleorithme which if dimer at at Cepane Nowntu ra2 cansis sonus Tatele 13 atrategy to impreve the probehelities of moccuful tranamisube ever topologies and divenifird fown are a major challenge fir]. Traditional in edge computing and redge cache also uses the Det. model such rithe, the m. atchigat in wad to opdate CWND based on different important role in the network feld including Cr Amongst the dafferent Mahehaied CC algarithms, th. hat gatsod the moit attention. Diflerent frem nupenioed learning methods, A. Alporithms monitoe the aatue of the rmvirunment cuntibueckly and rract to the envirenment to eptimise a stilify function. Therofoer, 31 , al gorithma are more suitable for variable and unutable netieurk envi: 5.1. Wind upieriat it end o-end ivetwarla renments. Two main trends are telated to this kind of ketwork. First, Intral of predicting congotion loss and delay as widh wivervised and a1 Irarn the OC nulen directly hawed an different emilentment infuenation. Since fil. algorithm can incorporate scal-time netwerk conditions and define arions sccontiegly, real-time contrul is powible in BL. algoeithins: use 7it. to update CWND for ppecific scetarius. The mechanise af It. based CC: algonibuas is shawn in fig. 6. the summary is shows in Table 11 and 12 , which show more ideialled information including atates, rewards, actions, pron, cons, etc. 5. 7.1. Ayschnincia tranufer mode netwerls Aynchronous transfer mode (ATM) is a typical network mitable for Bl-baied CC algorithms. ATM nerworla are clank netwerk that mpport mulb-media applications. For differras multimedia traffic, A7M offers ditferent QoS such as crll loss rate (CaM) and delay. Nowever, in ATM, highly time-varying traftic pettene can increase the ahcertainty of network traffic. Soccover, the small cell transminion time and low buffer siars in ATM networks require mare adaptive and high responsive OC algorithnis in foilL an MC algonthin is applied te doal with these peoblems, In the proposed CC algorithen, AC focuses 0de the the algocithm measures the action accoeding to the performance. In this way, diflerent traffic patierne ase connected with correpponding actions. Simulation rewults show that the Cur in low and volice quality in maintained, Comparod with clasical optimal control aleorithme which Survey paper When machine learning meets congestion control: A survey and comparison Huiling Jiang s", Qing Lit ker, Yong Jiang ", , GengBiao Shen ", Richard Sinnott ", Chen Tian', ATICLE INNO A \& S A A CT 1. Introduction 1.1. Tradidonal congrinion cosurol As a fursalanental component of computer netweeks, congestion The loternet transmission protocol is based on packet switching over control () plays a signiflicant role in impreving the network resource besteffort network forwanting [1], Find 40-end transmivion contrel is utilization to achieve betier performance. With the development of a required to pronide a reliable service for applications. To aveid netwark large number of widely waed technologie, es. dath conten (DC), degradation caused by congestion, oC alyorithms are typically emWif, 50, and satellite comerankations, the complesify and tive. ployed to improve reliable tranamistion over the setwork. Cangestion sity of network tranamission wesarios and peotocols have tereased in a state of the netweek in which it is not capable to deliver the dramatically. This has brought significast challenges to tranamission serice it was designed for. Netwoek congestion occun when eacrasive protocol design. A rich variety of CC algorithus have brow designed aumben of data padiess are sem over the netwark by hosis [2]. The for rpecific wenarion. However, the variety of netwerk wemarion and objective of OC alporithm in to achieve higher netwock twoughpoit bemely difficult to design efficient generic C Algonithms. Therefore, Wleally also guaraniee faimess berween end-to-end sesions. CC algorithass broed on makhune leamieg (ML.) have been perpoied The traditianal CE alpocithms can be caiegoriznd into two type: different network senarios, In thin paper, we provide a backgound proshes cely require the eallaboration of arnden and receivers, and analysis of traditional CC. Alaied on thik, we iavetigate correst works bence they do not rely an any explicit signals from the setwork. that research challenges in the application of ML. in the field of OC. Network-atisted appeoaches require the suppoit of network devices, - Cirrerponsing auther. Availite alline 31 March 2021 af. Alay at al fairness and eesponaiveseat in complex networking acrnariat. technigon. Superviked and unaupervised Jearning techeigues bave For end-to-end OC, une of the main challenges is to identify net-. beea adrly emplayed to improve the performance of network CC IIti, packet loss. Lasses are generaied when the buffer in a given nerwor . hgher delline leaming capability [24,25]. At peesent, inuch research device is overloaded, then lose-based approaches are supposed te attain foctors se Mt-based cr, whemet. high theoughput by making ase of the link hasdwidth. Hirwever, for Howrver, leaming based C, is still in its infancy. Most learningacote drlay-ienuitive applikations, lower tranamiaion time fannot be bund CC alporithms adjust the congention windos (CWND) to control suaranieed. Besides, the packet loss nay not be trigered by network. the sending not inatead of adpuating the sending rate directiy. Therecongrstice (e.g., rasiom packet droppingt. whikh may mislead or ferr, burwinen is atill a genblem in highspeed networks because the decisions. CWSD can itcrease sharply when multiple MCi's arrive [20]. Current Therelore, delay-based appruache such as Tirely [12) have bern lrarning haved CE alyorahms sach as [2%,251 generally focus on endlays caused by the network. Compared with Jotb-baned approaches, 0 , algurihan are not saitable for realistic netwerks due to Eime random packet lines, However, falculatiag the esatt tranmibsinn delmy remains a significant challenge. For esample, if there exids a alyefr change nhen the pachet is procrased in the Linar stack, the measered the the meased delsy cinnot be controlled precincly. appeoaches such as Veso [13] asd an adaptive and fair rapid increave stocarth, there is no comprebehsive survey en this aapect. There are and tranimissoe drlay, To solve this peoblem, network-asised OC, approaches such as the mork insurs include wheduling. localizatbos, data aggregarion, and so proposed. The network devices provide explicit sighals related to the hanod algorithess in wiseless wennor netwerks. Some echer surveys on network seatus for houts to make serding rate dechiont. When the gyencife setwork ewsinunments related to ML. techniques are groposed signal. The receiver will send hack the PCN signal in the ACK and the that here been employed for oognitive radine. In [11], the sevey cender will adjust the seruling rate accordingly. The ICN signal for gave an everview of the moet Mel. fechuitues encsentered in cellular provides finer'grained CC With the dramatically ifcreased complesity and divenity of net-. Aurvey ovven alousdant aspects nelated to netwoeking. it only gives traffic patiems in eee network scenario may also affect the perlormunot . Inarning-hased oct ander dynamic setworks and campare them with 1.2. croming hased cangeution contrel porithmb pries readere eahausive insight and lays the foendatica for The dynamic nature, divenity, amel complecity of network somario In trallatic networls, the inteliementation of learning-based CC algohave browth sipnificant challenges for CC,As ach, it is diffieult as design a generic scheme for all network scenarion. Furthenacin, the ridhen has abewn that they are not efficient as supposcat to be because dynamic natare of even the same network can make the performunoe foulge the pros and eons of deciaion models, we conduct comprratensive of CC unatable, Current network mvieonments and aloo incloble beeh wired networks and wireless oetworks, making the detectios of pacief loss mare dafficult [14-20]. roperimenth of varicus schrmes by taing the NS3 emulatur [3.]. lis the nimulation, we compare the Hl. hased CC algorishums of Deep control decisions instead of using predetermined rules. This allos threr s firrrat soenarins with different configurations of bandwidth them to have betier adaptability to dynamic and complek netwod. Ant Acligy. The network with high bandwideh and low delay simulates Scenarios. 11. Nave at al These three netwiwk Esilroumests represent the diverse mvinenments Slow start. In the dewic dear start provesh, rash time a good we generate the DC iraffe which includes bull data trander which egponently aver rime. phase, and larrane the valae of CWND each time baind en the aine nows for evperinents as well. alower than the oponorntal groth rate durity the aliow atart. The experimental results show that Ieaming based 0 , algunichims Retrannmialion. Rretratomiuise incfules timeout retranamiution width is low or the link delay in low, learning.bascel CC alpurithms are wart wable a certain time, the desu is netransmitied until the transmistoo aggreaive so learn with great atability, Morewer, the perion ante sion is nacenala. Jart metratimbion requares the receiver to vend a of thene thrre learning hased : algorithme shuws no defference in avor duplicane ACX immetlatrly afrre ruveiving an iot-of sequence seghent simulabed environments because the ecomplenity of the mviroements in as that the mater lam a mien an ponaible that a mejnent has limited. Therefore, all of theis ran handle these network scenarion. In rraliseic scenaxios, 8l-hased OC algorithms are influensed to the compotation time needed for R1. This imparts the frasibiay af confirmution. The setrantimien mechanim in CC enatires that data these schemes. Therefore, we propose three poiemial soluticns to dral with this problent. Firuly. ineign ligheseighy models hacel on map Fant recovery. Fart rnewwry arane that when the mender seceives ping sables of sates and actions to decruase the time connumpeion three depleate NCXI it noecmive, it executen a multiplikation redocof karning decitiont, Secondly, decreme the ferquency. of decinium tien alporing and halso the alor start throhold to pervent netwouk. to prowide better feasibility under low-dynamic tetwerk scenaries. congeation. The CWKD increses in an accumulative manter, cwating. vergence, incomparillility and fairnem. Based oe the wadersanding and analyait of the curnat learning tased OC aclutions, we blentify trendh 2.2. Clent ongatin ountral alourthis in learning-tased C. Firstly, because of thrir capability for draling algorithms to give end hoss better insights inte network status. Third. given the excesive time and cout of harning decisions, lightwright learnirg thased CC will to a key research direction. Finally, an open mechanismi, is muppoted to be designet. conduct simulations and compare performances between BL based CC algoritum and traditinnal OC algorithm, in Secvise An, we ostline the been morivel, ster fast mover mechanism will conider the pachet conclude the paper. 2. Harkground and rubuetien aftrr packin lask but the drawbick in that in enly , parkert an los, it is comilend enngation. Therffore, Newilime [D] is In this section, we will introduce relerant donain kewcoledge in proponal as deal with thin problrin, which mainly inpruve the faut algorithme, and the performance metrics. parkes are mhraksoninted and rereived confirmation will the sender 2.1. Copgetion catrol methaniurn exit. Althosph Nrefins can wolve ithe problem of a large manber of parket lowsa, Nirntlemo cas ally have ene perket bese erior per aTT uime. In oeder to deal with the lose of a large number of dafa peckets procedure of CC, we adept the windembased CC to Hens. Therefore, the nender as know which data has been received developenent of kew-based C e algorithms. fairness and Cubic [as] is proponed to opeimize Bic, Cubic uses a cabac However, in some specific network scenarios, opeimized CC: algo- functinn to replace the growth function of Mic. Also, the most aritical are destgned to deal with this iswae, Bic LII] in an example, Bic uses growth rate of the CWND is completely independent of the netwoul the idea of binary search. When a packet lins wccurs, indicating that Mtr. Thar, fairness is ensured. The eptimal window value sheuld be smaller than this value, then Ifighspoed |l9] and Hobla [tid are suitable for high-speed nerworks Bie wets the CWND at this time to mrs, Msr, and the value after the as well. Mighipoed modifies the reaction functicon of the standard multiplication is reduced to minjus, The optimal value of the CWND Tor protocol, which is affected by the conbincd effect of the growth and improve throughput. and congetion losh. While the measurement of delay can also te However, in DCa, the Mif is relatively small, and sitwaer timewhed to inder the netvurk congrstion, which can coarpenate for the impact of the sandom packet loss. Thns, delay-based CE alporithms are atarfuing at promidnt by modem Network lateface Cards (NiCo) 50 jiretented. tive queue management scalable TCP (DAST) [47], How latrneice TCP in Timely [11]. (Eala) [4f, and Timely [12] The main topec of delay based TCP as to eimiar the mesturnmeth (LDAst) [43) adopen a ene-way delay bo meanure the congevtion of of delay such as RTT, the one-way delay, and equeve delay to obtain netwerks. The ahraetege of the cee-way delay ower init is that it does that is based on the growth function of Cubic by aning a fair fliw halancing mechanism. Though loila has betier inter-flow ETI fairnesi friendy to Cuble fleran. and better fiTT eatimation, if cannat caenint fairly with loss toanod CC; 1. Jarr n all Capan Nuratu ha caury naws saime ratical problem. Once in detcets abourmal traffic, it will make feodback frem the control channel, thua data channel is possible to used with QUIC, Copa provides higher qualify and lower latency for family is alko a anethod of dynamically adjusting the rate by detecting mabile live breadcas than Cubic and BRi, Moreover, Copa has lower available network resourcea including FCC. [5.], FOC Vivace [lie], and traniminsoe fTt add lower low overhead. IceC Preteus [59]. They show geat performance tusal an carefully Hybrid C E algorithms use loss and delay t0 jointly evaluate coe- deslgnod utility functions that cover hasic performance megies such getion. In this way, IHigh throughput and low latency may be achieved as round-trip time (BTT). Compared to IBER, INCE converges slower at the same time. There are some fypieal hybrid CC algorithms ach becane of PCCis eceacroative increasing mechanian. as Veno [13]. Africa [14]. Conipound [t50], Libre [5]L, and Coogle Reno. Network-atsisted CC, algorithms require she coordination of the Veno combines Vegas and leno. Vegas cart mrause the number of sender anal reociver, and require routers te perforth apecific procestdata packets belonging to this connection in the sotwork botileneck ling on packets. Typical algorithms consist of ECN-tased cC algorouicr. Veno uses this variable to distinguhh random parket lois from eichms [Se] and Qaantiked Congeation Notificution-bused (oCE.based) congecioe packet kns and takes differrnt actinns. Veno also improve algorithms [5V]. The nummary of network-anaisted CCE algarthni is the growth function of CWED that is the batic part of Meso. When peraented in Tahie 4 , the number of packet belonging to this connection in the network ICV is an estension to the Internet Frotocel and Tranumission bottleneck router escecds a cenain value, the growth rate of CWND will Control Protocol. ECN allows ced-to-nad notification of CR: to avoid be slowed down and thas packet losses will be reduccd. Experiments packet losi and requires specific support freen the laternet layer and thow that Veno can improve throdgheut aignificantly without advenely the transport layer. Generally spcaking. TC.P/lntereet Frotocol (ti) affecting other concurrent TCP connections comparci wish Vegas and ectworks indicate channel cungestion by dropping data packets. In the Reno (1]. case af successful BCN negutiation, the DCN-aware router can set a Airica [t.4] is based an Rese as well, which can switch modes mark in the IlP header lintead of dibcarding the packet to indicate that cient, Arlica uses an atgressive, scalable window increase mechanism. Freponds to the sending end in indication, reducing ita tranamision rate ETT is wed to meature congecion as well. If the congotice level thows as if if had detected pucket-bost in the unual way, There are ithan high thronghput but also enaures falmes. Compound (50] is also a bybrid algerithm thed to solve the pred. aveld sethansmissian and reduce wasting time, njecially eetwark jirter. lem of poor MTT faimess among lob-hased CC algorithets. Compound QCX is a set of end-io-end eongestinn notification mechanisms maintains two CWNDS, con is a atandard windose aimilar to the Heno applied to 2, Throgh an active revene notification, the padket liss mechanium and the scoond is a scalable delay window hased on Vegas. zate and delay in the network are reduced, thereby inproviag netThe CWWND is calculated based es the symmation of these two windews. work perfommance. QCY inchudes twu parz: Cosgestion Foint (CP) and standard losshased component. When the network is congested, the equipment samples the data frames that ace bring sent in the sending delay.based component will reduce the sending rate significantly. But befier. If congestion ocrurs, it will genwerate a Congestion Notifirarion the throughput will be lower benunded by Reno. Therefore, there is Mesage (CNM) to the fil af the sampled dara frame. When the Bd limited EIT unfairness due to delay-based achemes, and the throughput excrives the CNM informatice, it will limit the sending rate of the ATT unfaimess, Labra [SI] vises noelinear optimization to garantee its icsoling rate to detect the available bandwidth and recerer the rate fairnes amting TCR flows regardies of ftrt. Different from the above algorithus, GOC: [52] is applied in We- DCTCR. The proposes algorithm dramatically inproves the throughyet beKC which ases UDP.baced BTP be tranamif media data instead of and fairness of the traffic. Thowgh QCN is efficient in controlling the TCP, In GOC;, the rase ceeerel based on the parket bost nate funs on quese length and assists the OC, QCX cantion be deployed and ased the irTce na messege from the receiver at the sender and dynamically setworks. adjusts the oode rate at the sender acconfing to the packet lows rate indormation carried in its Meport Block. Delay-based rate coetrol runs on the receiving end. Hesides, there are scme innevative hybrid algorthus that have CK; algorihhms are expected tio achieve various goals and objectives Control Protocol (scr) (54). Perfoemanceociented Congestion Control Throughpot represents the amount of data that pases through a net(POC) [S5). JCC Vivace (SE), and FCC Nroteut [5\%], in Bemy [53], the utilly function condists of thoughput and delay. itheam high link utilization. Masimicing throphput is erucial. Given To maximize the evpocted value of the utility fanction, Reary finds the link handwidth, high throughput inaticates high efficiency in transthe mapping based on a pre-compubed looknp takle. Thence, the coere- ferring data. For instance, Bic is aimied bo achieve hightr throughgnat sponding arnding rate is evtimatrd. To converge to the egtimal sending . hased on a more aggressive mechanim compared with Reno (It). rale and fully utiline the network, BBA estimaies the wvailatle hand. HTT meavures the time including the transimission time, the gropat of adjustieg the sending fate by detecting the network stanis. Hearning time (FC;T) indicates the time required io tranafer the flowi. Itr and Cantrol Protiool (SCP) in (54). To achieve fast convergence, SCR panti- that they may have to tolerate. However, it may he the case that the actual sending mate in the data chansel, SCP serulu stand-in packets High throughapul means making use of the link bandwidth as ituch is the control channel to probe the netwerk condition. Based on the as possible, which fan give riae to an increased geeue length that if. Aary at at. Cinwane Nureets EA2 cat'1] Javet may cause drlays, In Vegan [4h, the cruclal insight in to peedict the becass optimal tratfic clasaification policies can promete the reasoncongestion level bowed en measured MtT. Moreover, the lower ilt and ahle allocacion of neriark msources and redaer the peobability of Ny deaigning two channels, aere packet loss is ponsibie. agplication of apervied learning alporithms in the field of CE. Fairness is a measure of equality of the resousce allocatinn of the There in a begr amowet of supervied learning technioues used POC Vivace [50] is the improsed speed ad responve. application, which laprowes the acouracy aignificantly. In this paper, These objectives are important for all CC algorithus. ter they are the acraracy in about 965 . The inaight proviled in this mesearch is that also have different priarities, and hence trade-offs are nocemary. Based fording the Gasalan ansoptins and ingruving the quality af dis: drtail. fartaicit dhcriminant poeve among shitinctive elawses. To reduce the the presented nodel, nulbigle binary SVM dawafien are organinal into a toormaneme wenchare, wheh can ifrimatically dectease the number of an opeimal madel, and then use this model to map all inputs to corre. prove accurary and relure oumputational cent. The aimulation shown learaling techniques have the atelay an perforent dase dassification. times. forests, Aayes, regrescion, and necural networks. In the network field, supervised learning methode are used is im: inces nach an rlaniflocien acoeracy and computarion cost. Eacept length for metwork wasisted networks which affect the performance of CE directly. Congestien signal predictioe conalsts of los clawificarion and deloy prediction, As mentioned before, congestian is defecard implicity hased in packet loss of delay when coegrnzin occurs in traditinnal CC algotithms. In supentied learning thased O C. alporithm, congewion in estimated in adeance haved on curcent and previcen network stato woch as the packet arrival interval and the network delay. The key hans for this approsch is that netwurk atates form a cuntinuous tinse series, where the future atate can be pendicted by past states. An for qarue lengih managenent, supervised leamlig methodi play an impoetant role in the aceurate and efficient gerediction of queue length. In the neat two parts, a more detailed drscription will be shoren. 2.2. Cocpoution detection in end to-and neneurly 3. 1.1. Lai dounficanion the lack al hullens is Cesh, contestion loss is generatal when there is evential to undentand CC to fons clasiticatises in wirrles networks hased on staditicnal ce Roondrring losk canase be ignored in networls with inulti-channed arrival time in elasify wireira lons and congestion loss if the packet inter-arrival time is confined to a rarye, the missing perkrts are lod S Fervistal lraraitg buwell ce algorithms art able to seal with the due to whrekess lons. Otherwile, the kns is conaldered a cengestion line. In (enf). the number of loves and BOTI were used to distanguish the types of losies. The presented algurithm called Frghes in more efficient le medation, wirrlens lons, euntentien bom, and ecordering loss inguct the dechection of evepestion loss. Sopervised learning technilues while Spike [69] ahowi letter performance in wireiras hackhone tepolegy with multiple flows. Digzag 6j is relatively more general, and Irartang thand or algorhhats hence ia able to satisfy different topology acenarion but it is Mrnative Miscluaification at ane iswev, in winelese newwerks, jredefined pafameters detirmine the enurs in elataifying congevtion losa and wireas the clasisier. The minimum BTT value and the current HitT valae ane regarded as input atates. While in 7FO/, the arrival time internal of ACX; and miniman AIT are choom at fratures and the Nalioos algorithan in uned to dlasify the low into wireless loss and cengratica loss for satelIite networka in (18)], the one-way delay wat inter-packit times were used as states to predict boss caicgories. in (19), the quruing delmy. neod ts be onawered earefally to balance performance in diflenent the interarivel time, and lists of porkets were used as ingrits. Berides, diverse supenised learning techniques were applied, In [711, deciuin treck decision tree macemblet, bagking, random fremts, ratra trees, boosting, and mult - layer perceptreen were used to ctaidif the types of loss Simulations show that these iniclligent kon chasifiers actirne heh 312. Exlig gredinan accuracy in differest netwock scenarios. Is conchation, when classifying. As a cengerion signal, the delay of tranamininis reflects the the wireless boss and the cungention lons, delay information is the kernel anvost of is flight data, which show the overall lowd on the netuork. aves the sourees due io wavelength reacration. Honecer, tercause of shoers and Table 7 cenchals, nupeniod letrning lechatiques have high N. Nens n at Tupersioed harulas algeriline reacting equickly to avold congestion. the ervirunment. Some trward har ibown that the furure queue length exponentially weighted moving average technique demonurates a moee time series of proviout tratfic an ingot without considering diverve accurate al gorithm. parameters in the actave. A a tenily, these algoriahans leave space other parameters in the network, In [ bol, lintar regesaion was used to the quicue length. eatakinh the relatiombilp between 8rT and the seadiog rate. In fith. a Bayesian sechnique wa sned is simulate the distribution betwern 4. Unsupervised linarsing besed congeation coetrol algorithms delay and the wending rate asd then to predirt drlay based on the sending rate. This is necoled in teal eime video applications and wirches In this arctian, another eatryary of leaming banal Cr algorithms netwoeks. hight reaponsiveness have been proposed. Further rerarch is needed to mation carnot te folly provided, the unsopervised learning techniques techniques to improve delay predictions. unauporvised learning notbodn can be imployed as traffic clanifiers 3.2. Qurue longth managowent in nefwork asueted tweworis at well. Thus, hood an chasen gronated by unsepentiod leara: Qurue Ingib management is a hyy focun tor netwoek-aniared CC queve management (AQM) familly of ISV toshniquen. Homere, the (BME) 104]. is usel as a basic clustering method. The features selected in this work As for the detectoon ol coegarios, Engerviaed learning al gorithms the number of the puibed data packnts, efr. Blased on X-means, the authors adopt the feature selection to find an optimal featwee ant and los transformation to impeove the accuracy. The experimerus stwow that 4.1. Congonan derinion in end to-nd cargration hantral alporither the proposed method can obtain up to 80 overall accuracy. Compared with K-thenans, 1MM is generally used for Cavusian Mtixture 4. t.t. Leas chaturies Model. In K-meank, an unknown data point mast belone to a single cluater, wadle a data point can be mapped into meltiple clusters lused on EM. In [94l. the authoes use an Autodlas appraach which is an B6.55. However, the performance of clustering depends on the paraku: congeited pariet lom, of congention aigal will be trigernd, and the K-mean in coetplex data sets, but this advantage is limited. ing thravars to cantral the amdiet ratr. A detalled aummary ia ahown Nefeork exten mainly osealat of lossdelay pairs anat the namber of bunes berrwen felluees. Thas, moee research it eequired to broaden thin fiek. 4.1.2. frily prodican There are oely a lemeted number of uhuspervised leaming-based CC and the amociuted mechunine are presented in hy. 5 and Talle 12 . Ms shows in Fig. 3 , the sember obtains the states in wetwoeks ach an the menare siac. Ther. delly chatering it conducted hused un unsupervined learning meethode. The prisratied clusters can be used to adjus the meswar alm, the valifity of inesugr, the distance betiveen vehicles and Baris, and the tyje of mencege is diviled inbo different grosph and the bownt delwy it each group in selecked an the certumication parameter for cach elentre. Bricd on the comanumicatico parameter, a specife aralint ratir will be anigned to each cluster. Therefoee, hased on the meaksinarnt of deloy, CC, can be actirinvel. Similar to low chuntrring hated an antepervised leamiag technisock, the revedreh of the obervation, in wireless netwarks, losses can he chawifind inte congestion lins and wireiess hins, In thas rescard, die diatribution of loss-delay pairs ean be eaphared by a Hiddrat Marhov Madel (ECMO). 5. NH-baned coegentica ceetrol algorithme The IBtSt is trained to associate the fype of losues with a state haned cn model is more eftirient than Vegas predictors. rgooific arrions given the network utate, to determine if a given action research, the losses clustering problem is formulated as a statistical valae fanction then calcelates the value of the action and update it loe Righer. losses an well. In OAS, the aunhnen in [7J] find that the number af bunts between failutrs can be wed to differentiate congention and valae of acticns dirvitey. twe chantering methods to improve the performance of Cr. In the the large handmath of matWave and the high cust af infrastracture if dimer at at Cepane Nowntu ra2 cansis sonus Tatele 13 atrategy to impreve the probehelities of moccuful tranamisube ever topologies and divenifird fown are a major challenge fir]. Traditional in edge computing and redge cache also uses the Det. model such rithe, the m. atchigat in wad to opdate CWND based on different important role in the network feld including Cr Amongst the dafferent Mahehaied CC algarithms, th. hat gatsod the moit attention. Diflerent frem nupenioed learning methods, A. Alporithms monitoe the aatue of the rmvirunment cuntibueckly and rract to the envirenment to eptimise a stilify function. Therofoer, 31 , al gorithma are more suitable for variable and unutable netieurk envi: 5.1. Wind upieriat it end o-end ivetwarla renments. Two main trends are telated to this kind of ketwork. First, Intral of predicting congotion loss and delay as widh wivervised and a1 Irarn the OC nulen directly hawed an different emilentment infuenation. Since fil. algorithm can incorporate scal-time netwerk conditions and define arions sccontiegly, real-time contrul is powible in BL. algoeithins: use 7it. to update CWND for ppecific scetarius. The mechanise af It. based CC: algonibuas is shawn in fig. 6. the summary is shows in Table 11 and 12 , which show more ideialled information including atates, rewards, actions, pron, cons, etc. 5. 7.1. Ayschnincia tranufer mode netwerls Aynchronous transfer mode (ATM) is a typical network mitable for Bl-baied CC algorithms. ATM nerworla are clank netwerk that mpport mulb-media applications. For differras multimedia traffic, A7M offers ditferent QoS such as crll loss rate (CaM) and delay. Nowever, in ATM, highly time-varying traftic pettene can increase the ahcertainty of network traffic. Soccover, the small cell transminion time and low buffer siars in ATM networks require mare adaptive and high responsive OC algorithnis in foilL an MC algonthin is applied te doal with these peoblems, In the proposed CC algorithen, AC focuses 0de the the algocithm measures the action accoeding to the performance. In this way, diflerent traffic patierne ase connected with correpponding actions. Simulation rewults show that the Cur in low and volice quality in maintained, Comparod with clasical optimal control aleorithme which if dimer at at Cepane Nowntu ra2 cansis sonus Tatele 13 atrategy to impreve the probehelities of moccuful tranamisube ever topologies and divenifird fown are a major challenge fir]. Traditional in edge computing and redge cache also uses the Det. model such rithe, the m. atchigat in wad to opdate CWND based on different important role in the network feld including Cr Amongst the dafferent Mahehaied CC algarithms, th. hat gatsod the moit attention. Diflerent frem nupenioed learning methods, A. Alporithms monitoe the aatue of the rmvirunment cuntibueckly and rract to the envirenment to eptimise a stilify function. Therofoer, 31 , al gorithma are more suitable for variable and unutable netieurk envi: 5.1. Wind upieriat it end o-end ivetwarla renments. Two main trends are telated to this kind of ketwork. First, Intral of predicting congotion loss and delay as widh wivervised and a1 Irarn the OC nulen directly hawed an different emilentment infuenation. Since fil. algorithm can incorporate scal-time netwerk conditions and define arions sccontiegly, real-time contrul is powible in BL. algoeithins: use 7it. to update CWND for ppecific scetarius. The mechanise af It. based CC: algonibuas is shawn in fig. 6. the summary is shows in Table 11 and 12 , which show more ideialled information including atates, rewards, actions, pron, cons, etc. 5. 7.1. Ayschnincia tranufer mode netwerls Aynchronous transfer mode (ATM) is a typical network mitable for Bl-baied CC algorithms. ATM nerworla are clank netwerk that mpport mulb-media applications. For differras multimedia traffic, A7M offers ditferent QoS such as crll loss rate (CaM) and delay. Nowever, in ATM, highly time-varying traftic pettene can increase the ahcer