A Predictive Interference Management Algorithm for URLLC in Beyond 5G Networks
Nurul Huda Mahmood, Onel Alcaraz Lopez, Hirley Alves, Matti Latva-aho
AA Predictive Interference Management Algorithm forURLLC in Beyond 5G Networks
Nurul H. Mahmood, Onel A. L´opez, Hirley Alves and Matti Latva-aho ∗ September 22, 2020
Abstract
Interference mitigation is a major design challenge in wireless systems,especially in thecontext of ultra-reliable low-latency communication (URLLC) services. Conventional average-based interference management schemes are not suitable for URLLC as they do not accuratelycapture the tail information of the interference distribution. This letter proposes a novel in-terference prediction algorithm that considers the entire interference distribution instead ofonly the mean. The key idea is to model the interference variation as a discrete state spacediscrete-time Markov chain. The state transition probability matrix is then used to estimatethe state evolution in time, and allocate radio resources accordingly. The proposed schemeis found to meet the target reliability requirements in a low-latency single-shot transmissionsystem considering realistic system assumptions, while requiring only ∼
25% more resourcesthan the optimum case with perfect interference knowledge.
Keywords:
Beyond 5G/6G networks, intelligent resource management, interference predic-tion, URLLC.
Ultra-reliable low-latency communication (URLLC) is a novel service class introduced in thelatest fifth generation (5G) New Radio (NR) wireless network standard. URLLC targets strin-gent reliability performances, e.g., in the order of 99 . (i) uncertainty in thetraffic arrival, (ii) channel impairments such as fading, and (iii) random interference. Address-ing these challenges efficiently mandate a departure from the average utility based approach ofconventional radio resource management (RRM) practices to a framework that considers the tailbehavior of the reliability, latency and throughput performance [3].URLLC enablers in 5G NR are primarily centered around redesigning the system numerologyto meet low latency requirements [1], and ensuring high reliability through over-provisioning ofresources [4]. While such an approach is known to enable URLLC under sparse and controlled en-vironments, it is neither scalable nor resource efficient, and therefore does not adequately addressthe fundamental and novel challenges imposed by URLLC. Alongside, emerging URLLC use casesin future beyond 5G/sixth generation (6G) networks will require QoS guarantees that are muchmore diverse and stringent than those considered in 5G NR [5]. Novel solutions incorporatingintelligent and predictive RRM algorithms are therefore needed to enable URLLC in a scalableand resource efficient manner in future wireless networks.Link adaptation is a well known RRM approach in conventional interference management.It involves estimating future interference values from past samples, which is then mitigated by ∗ The authors are with 6G Flagship, Centre for Wireless Communications, University of Oulu, Finland. Cor-responding e-mail: nurulhuda.mahmood@oulu.fi . This work is supported by the Academy of Finland 6GenesisFlagship program (grant no. 318927).
Submitted to IEEE for possible publication. a r X i v : . [ c s . I T ] S e p dapting the link parameters accordingly [6]. Conventional link adaptation schemes operate bycondensing the entire interference distribution into a single mean value, and are therefore notsuitable for URLLC applications where the tail behavior of the interference distribution needs tobe considered [3]. Instead, interference prediction strategies that capture information about theentire interference distribution are better suited. Generally, such interference prediction involvesbuilding a model to reflect the interference variation across time, which is then used to predictfuture interference values.RRM algorithm design for URLLC applications based on interference prediction is still at itsinfancy (see for example [7]). We aim at contributing to this emerging field by proposing a novelinterference prediction based RRM algorithm in this work. We model the interference variationas a discrete state space discrete-time Markov chain (DTMC). The state transition probabilitymatrix, obtained by observing the history of state transitions, is used to predict future interferencestates. Finally, the predicted interference is mitigated through efficient allocation of resources.The proposed approach is validated using extensive Monte-Carlo simulations. We observe thatvery low outage probabilities can be achieved with the proposed scheme, which are otherwise notpossible by adopting a conventional average-based approach. We consider the downlink of a wireless network. The focus is on a desired URLLC linkoperating in the presence of N interferers distributed in R space. The desired channel is assumedto have a mean signal to noise ratio (SNR) of ¯ γ D , whereas the mean interference to noise ratio(INR) corresponding to each interfering link - i.e., the mean SNR of the interference signal - isconsidered to be uniformly distributed in the range [¯ γ I,min , ¯ γ I,max ] . We assume ¯ γ D ≥ ¯ γ I,max , i.e., users are served by the transmitter with the strongest mean SNR. All nodes are assumedto operate independently, i.e., there is no cooperation among them. A single-antenna Rayleighblock-fading channel model is adopted.The desired transmitter transmits a short packet of D bits with a target outage/block errorrate (BLER) (cid:15). The transmitter estimates the signal to interference plus noise ratio (SINR), andthen allocates the required resources accordingly. The goal is to estimate the SINR as accuratelyas possible to ensure efficient resource allocation.URLLC aims to realize low latency and high reliability simultaneously. This is usually done byanalyzing the achieved reliability considering a given latency budget [8]. In this work, we considera very tight latency budget that cannot accommodate any retransmission, i.e., transmissions areassumed to be single shot [9]. Such an assumption allows us to analyze the lower bound of URLLCperformance since retransmissions improve reliability, albeit at the cost of additional latency [10].URLLC transmissions usually occur over mini-slots of duration ∼ . A model-based approach to interference prediction is adopted in this work, where the in-terference distribution is modeled as a discrete state space DTMC. The sketch of the proposedinterference prediction algorithm is pictorially outlined in Fig. 1 and detailed in the rest of thissection. The proposed scheme is link direction agnostic and works equally for the uplink as well.
Page 2igure 1: Sketch of the proposed algorithm.
As the first step, the initially observed (continuous) interference space I is discretized intothe state space L (cid:44) {S , . . . , S L } such that interference values in the range [ I l − , I l ) are assignedto state S l . A straightforward option would be to discretize I into L equally spaced stateswithin the range of I . However, this treats stronger and weaker interference values equally and iscontradictory to the risk-sensitive approach recommended for URLLC [3]. We therefore discretize I into L equally spaced states and define the state boundaries of L to be the squareroots of thosevalues. Thus, larger interference values are represented by a higher number of states, as depictedin Fig. 1. Finally, we set I = 0 and I L = ∞ to capture any future interference value that may lieoutside the range of I . It is worth mentioning that, the choice of L results in a trade-off betweenalgorithm complexity and performance accuracy. Next, we obtain the transition matrix P describing the transition probabilities of the statesin L . We first define ˜ I (cid:44) {S , S , . . . } as the set of observed interference states at time index t = 0 , , . . . . The transition probability p ij denotes the probability of a transition from the states S i to S j , and is defined as p ij = (cid:80) t [ S t +1 ( S j ) | S t ( S i )] (cid:80) t [ S t ( S i )] , (1)where A ( x ) is the indicator function which equals 1 if x ∈ A and 0 otherwise. All the entries ofthe transition matrix P are thus obtained by evaluating (1) ∀ i, j ∈ { , , . . . , L } . In order to utilize the tail statistics of the interference distribution, we introduce the confidencelevel parameter η <
1, which can be viewed as a risk-sensitivity index [3]. Suppose, we wouldlike to utilize the right end of the distribution tail up to a point ζ . η is then the area underthe left side of the PDF up to point ζ . Hence, the closer η approaches one, the closer to theright end of the distribution tail is utilized for the prediction, as illustrated in Fig. 1. Morespecifically, η characterizes the likelihood that the predicted interference level is higher than theactual interference. Thus, a larger η corresponds to a more conservative interference estimation.Suppose, S t = S i , i.e., the interference state at time t is S i . Our proposed scheme predicts theinterference at time t +1 to be such that the predicted interference (cid:16) ˆ I t +1 (cid:17) is greater than the actual Page 3 nterference (cid:0) I t +1 (cid:1) with probability ≥ η. Mathematically, this is stated as
P r [ I t +1 ≤ ˆ I t +1 ] ≥ η .To ensure this, we first predict the next state ˆ S t +1 to be S j , where j is the smallest integer suchthat (cid:80) jl =1 p il ≥ η . The predicted interference level at time index t + 1 is thenˆ I t +1 = (cid:40) I j if j (cid:54) = L I L − − I L − if j = L , (2)where I j is the right-endpoint of state S j . Please note that, we have earlier set I L = ∞ and henceif ˆ S t +1 = S L , we risk obtaining ˆ I t +1 = ∞ . To avoid this, the second step in (2) sets a dummy right-endpoint of S L . Note that the dummy endpoint can be set to any value greater than I L − for the proposed scheme to work. The last step of the proposed algorithm involves updating the transition matrix P . Suppose,a transition is made from state S i to state S j . P is then updated as follows, step 1, update: p ij → p ij + ω ij step 2, normalize: normalize i th row s.t. L (cid:88) j =1 p ij = 1 . Here, 0 ≤ ω ij (cid:28) ω ij = ω i ∀ j and let its value be inversely proportional to the number of elements in state S i . The proposed interference prediction algorithm is of low-complexity with minimal additionalsignaling overhead. The only additional step during runtime involves looking up the transitionmatrix (a table of size L ) to predict the next interference state. In addition, the algorithm requiresthe receiver to feed back the experienced interference level along with the CSI, and the transmitterto update P upon receiving this information. All of these steps are rather simple to execute andincurs negligible processing delay. The performance of the proposed scheme is evaluated and compared against two baselineschemes in this section.
The conventional weighted average based interference estimator, which is adopted as an es-timator in link adaptation for conventional enhanced mobile broadband (eMBB) services [6], isconsidered as the first baseline scheme. In this scheme, the interference measurement at time t isfiltered with a low-pass first-order Infinite Impulse Response (IIR) filter, resulting in the followinginterference estimate, ˆ I t +1 = αI t − + (1 − α ) ˆ I t , (3)where 0 < α < We also consider a genie-aided estimator where the exact interference condition is known apriori at the transmitter, leading to optimum resource allocation. Even though such a scheme isnot feasible in practice, it is considered as an indicator of the optimum performance bound.Page 4 .2 Resource Allocation
Let γ = σ/ ˆ I be the predicted SINR, where σ is the SNR of the desired transmission (whichwe assume to be known using CSI estimates) and ˆ I is the predicted interference level as describedin the preceding section. Using results from finite blocklength theory, the number of informationbits D that can be transmitted with decoding error probability (cid:15) in R channel uses in an additivewhite Gaussian noise (AWGN) channel with SINR γ is given as [11] D = RC ( γ ) − Q − ( (cid:15) ) (cid:112) RV ( γ ) + O (log R ) , (4)where C ( γ ) = log (1 + γ ) is the Shannon capacity of AWGN channels under infinite blocklengthregime, V ( γ ) = (cid:16) − γ ) (cid:17) is the channel dispersion (measured in squared informationunits per channel use) and Q − ( · ) is the inverse of the Q-function. Using the above, the channelusage R can be approximated as [12] R ≈ DC ( γ ) + Q − ( (cid:15) ) V ( γ )2 C ( γ ) (cid:34) (cid:115) DC ( γ ) Q − ( (cid:15) ) V ( γ ) (cid:35) . (5)Eq. (4) is extended in [13] to derive the achievable maximum coding rate spanning over multiplecoherence intervals considering the more practical (and assumed) case of correlated Rayleighblock-fading channels. However, this only changes the absolute performance but not the relativeperformance measures of the different schemes, and hence does not alter the key findings of thispaper. The performance of the proposed interference prediction method is numerically evaluatedagainst the two baseline schemes presented in Section 4.1. Unless stated otherwise, the simulationparameters presented in Table 1 are adopted.
We first evaluate the outage probability as a function of the target outage (cid:15) as plotted in Fig. 2.Four different values of the risk-sensitivity index are considered, namely η = { . , . , . , . } . Since the genie-aided estimator is assumed to know the achieved SINR beforehand, it can allocatethe exact amount of resources required to meet the target outage. On the other hand, theconventional IIR filter based estimator performs very poorly and is only able to meet a BLERtarget of about 10% . It is worth mentioning that the conventional IIR filter based approach isknown to be quite resource efficient for eMBB services, where the BLER target is usually around10%.The performance of the proposed scheme depends on η. Outage targets as low as 5 × − can be fulfilled with a conservative value of η = 0 .
95. Obviously, this comes at the cost of higherresource usage (RU), as discussed next. It is worth highlighting that this outage is achieved with asingle shot transmission, which means that the outage performance can be significantly improvedwith retransmission or other diversity techniques, albeit at the cost of higher latency.The average RU corresponding to the above setup is shown in Fig. 3. As expected, the lowestRU is achieved when the SINR is exactly known a priori as it leads to the optimum resourceallocation. The IIR filter based baseline scheme is very close to the optimum performance interms of RU. As indicated earlier, this is the conventional scheme in the case of eMBB serviceswhere optimum RU is prioritized over high reliability and low latency.The RU of the proposed scheme depends on the risk-sensitivity index η , which can be usedto balance the trade-off between reliability and RU. A higher η reflects a more conservativeprediction. This leads to a lower SINR estimation, and subsequently higher resource allocationwhile also resulting in a better outage performance. Page 5able 1: General simulation parametersParameter ValueMean SNR, ¯ γ D
20 dBMean INR range, [¯ γ I,min , ¯ γ I,max ] [ − ,
5] dBNumber of interferers, N D
50 bitsNumber of states, L η . α . η . The performance of our proposed scheme under different interference conditions is evaluatedin this section. We consider three different scenarios, as follows
Strong SNR, strong interference:
Mean SNR and the INR range set to ¯ γ D = 20 dB, and[¯ γ I,min , ¯ γ I,max ] = [0 ,
20] dB, respectively;
Strong SNR, weak interference: ¯ γ D = 20 dB and [¯ γ I,min , ¯ γ I,max ] = [ − ,
5] dB; and
Weak SNR, weak interference: ¯ γ D = 5 dB and [¯ γ I,min , ¯ γ I,max ] = [ − ,
5] dB.Fig. 4 presents the outage probability as a function of the target outage under the consideredinterference scenarios. As a general trend, neither the baseline nor the proposed scheme is stronglysensitive to the interference scenario. The baseline scheme behaves as expected with the strongSNR, weak interference scenario demonstrating the best performance, followed respectively by weak SNR, weak interference and strong SNR and strong interference scenarios. On the otherhand, given its nature, the proposed interference prediction scheme can better track the inter-ference when it is strong. Hence, a better performance is observed for the scenarios where theinterference strength is comparable to the SNR.
In order to meet the low latency requirements, URLLC transmissions in 5G NR take place overmini-slots with rather short TTIs [1]. Hence, the channel coherence time in practice spans overPage 6 -5 -4 -3 -2 -1 Target outage20304050607080 R e s o u rce u s ag e ( c h a nn e l u s e s ) about 2 RUabout 1.3 RU = {0.95, 0.9, 0.85, 0.8} Figure 3: Resource usage in channel uses vs. target outage probability for different values of therisk-sensitivity index η .Figure 4: Outage probability vs. target outage under different interference scenarios, η = 0 . η = 0 . -5 -4 -3 -2 -1 Target outage R e s o u rce U s ag e ( c h a nn e l u s e s ) about 25% more RU about 100% more RU Correlated trafficUncorrelated traffic
Figure 6: Resource usage in channel uses vs. target outage probability with correlated traffic, η = 0 . URLLC mandates a departure from the conventional average-based resource management ap-proach. Instead, intelligent and risk sensitive designs are needed to meet the stringent reliabilityand latency requirements while guaranteeing high resource efficiency. A novel interference pre-diction based RRM algorithm is proposed in this letter. The interference distribution is modeledas a discrete state space DTMC. Future interference states are predicted by utilizing the statetransition probability matrix along with a risk sensitivity parameter. The reliability of the pro-posed scheme is found to be significantly better than that of the considered baseline RRM scheme.Alongside, the proposed scheme requires only slightly more resources than the optimum case. Asnext steps, we plan to investigate more resource efficient implementations of the proposed scheme.Page 8 eferences [1] J. Sachs et al. , “5G radio network design for ultra-reliable low-latency communication,”
IEEENetwork , vol. 32, no. 2, pp. 24–31, 2018.[2] G. Berardinelli, N. H. Mahmood, I. Rodriguez, and P. E. Mogensen, “Beyond 5G wirelessIRT for Industry 4.0: Design principles and spectrum aspects,” in
Proc. GC Workshops ,UAE, Dec. 2018.[3] M. Bennis, M. Debbah, and H. V. Poor, “Ultrareliable and low-latency wireless communi-cation: Tail, risk, and scale,”
Proceedings of the IEEE , vol. 106, no. 10, pp. 1834–1853, Oct2018.[4] N. H. Mahmood et al. , “Multi-channel access solutions for 5G new radio,” in
Proc. IEEEWireless Communications and Networking Conference (WCNC) Workshops , Marrakech, Mo-rocco, Apr. 2019, pp. 1–6.[5] ——,
White Paper on Critical and Massive Machine Type Communication towards 6G , ser.6G Research Visions, nr. 11, N. H. Mahmood et al. , Eds. Oulu, Finland: University ofOulu, Jun. 2020.[6] G. Pocovi, B. Soret, K. I. Pedersen, and P. Mogensen, “MAC layer enhancements for ultra-reliable low-latency communications in cellular networks,” in , May 2017, pp. 1005–1010.[7] M. Li, C. Chen, C. Hua, and X. Guan, “A learning-based pre-allocation scheme for low-latency access in industrial wireless networks,”
IEEE Transactions on Wireless Communica-tions , vol. 19, no. 1, pp. 650–664, Jan. 2020.[8] N. H. Mahmood et al. , “Uplink grant-free access solutions for URLLC services in 5G new ra-dio,” in
Proc. 16th International Symposium on Wireless Communication Systems (ISWCS) ,Oulu, Finland, Aug. 2019, pp. 607–612.[9] N. H. Mahmood, N. Pratas, T. H. Jacobsen, and P. E. Mogensen, “On the performance of onestage massive random access protocols in 5G systems,” in
Proc. 9th International Symposiumon Turbo Codes and Iterative Information Processing, (ISTC) , Brest, France, Sep. 2016, pp.340–344.[10] P. Popovski et al. , “Wireless access for ultra-reliable low-latency communication: Principlesand building blocks,”
IEEE Network , vol. 32, no. 2, pp. 16–23, Mar. 2018.[11] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the finite blocklengthregime,”
IEEE Transactions on Information Theory , vol. 56, no. 5, pp. 2307–2359, May2010.[12] A. Anand and G. de Veciana, “Resource allocation and HARQ optimization for URLLCtraffic in 5G wireless networks,”
IEEE Journal on Selected Areas in Communications , vol. 36,no. 11, pp. 2411–2421, Nov. 2018.[13] A. Lancho, T. Koch, and G. Durisi, “On single-antenna rayleigh block-fading channels atfinite blocklength,”