[3] A. Plebe and A. M. Anile, “A neural-network-based approach to the double traveling salesman problem,” Neural Comput., pp. 437–471, Feb. 2001.
[4] S. Kaski, J. Sinkkonen, and J. Peltonen, “Bankruptcy analysis with self-organizing maps in learning metrics,” IEEE Trans. Neural Netw., vol. 12, no. 4, pp. 936–947, Jul. 2001.
[5] C.-H. Chang, P. Xu, R. Xiao, and T. Srikanthan, “New adaptive color quantization method based on self-organizing maps,” IEEE Trans. Neural Netw., vol. 16, no. 1, pp. 237–249, Jan. 2005.
[6] G. A. Barreto and A. F. R. Araújo, “Identification and control of dy-namical systems using the self-organizing map,” IEEE Trans. Neural Netw., vol. 15, no. 5, pp. 1244–1259, Sep. 2004.
[7] G. Dong and M. Xie, “Color clustering and learning for image segmen-tation based on neural networks,” IEEE Trans. Neural Netw., vol. 16, no. 4, pp. 925–936, Jul. 2005.
[8] H.-C. Wang, J. Badger, P. Kearney, and M. Li, “Analysis of codon usage patterns of bacterial genomes using the self-organizing map,” Mol. Biol. Evol., pp. 792–800, May 2001.
[9] M. Sultan, D. A. Wigle, C. A. Cumbaa, M. Maziarz, J. Glasgow, M. S. Tsao, and I. Jurisica, “Binary tree-structured vector quantization ap-proach to clustering visualizing microarray data,” Bioinf., vol. 18, no. 1, pp. S111–S119, 2002.
[10] R. T. Gill, E. Katsuoulakis, W. Schmitt, G. T. Oldenburg, J. Misra, and G. Stephanopoulos, “Genome-wide dynamic transcriptional profiling of the light-to-dark transition in synechocystis sp. strain PCC 6803,” J. Bacteriol., pp. 3671–3681, Jul. 2002.
[11] P. Koikkalainen and E. Oja, “Self-organizing hierarchical feature maps,” in Proc. Int. Joint Conf. Neural Netw. (IJCNN), 1990, pp. 279–284.
[12] K. K. Truong, “Multilayer Kohonen image codebooks with a loga-rithmic search complexity,” in Proc. IEEE Int. Conf. Acoustics, Speech, Signal Process., May 1991, pp. 2789–2792.
[13] P. Xu and C.-H. Chang, “Self-organizing topological tree,” in Proc. Int. Symp. Circuits Syst. (ISCAS), 2004, pp. 732–735.
[14] P. M. Sharp and W.-H. Li, “Codon usage in regulatory genes in Escherichia coli does not reflect selection for ‘rare’ codons,” Nucleic Acids Res., vol. 14, no. 19, pp. 7737–7749, 1986.
[15] M. Riley, “Functions of the gene products of Escherichia coli,” Micro-biol. Rev., vol. 57, no. 4, pp. 862–952, Dec. 1993.
A Cross-Layer Adaptation Scheme for Improving IEEE 802.11e QoS by Learning
Chiapin Wang, Po-Chiang Lin, and Tsungnan Lin
Abstract—In this letter, we propose a cross-layer adaptation scheme which improves IEEE 802.11e quality of service (QoS) by online adapting multidimensional medium access control (MAC)-layer parameters de-pending on the application-layer QoS requirements and physical layer (PHY) channel conditions. Our solution is based on an optimization approach which utilizes neural networks (NNs) to learn the cross-layer function. Simulations results demonstrate the effectiveness of our adapta-tion scheme.
Index Terms—Adaptive algorithm, IEEE 802.11e wireless local area net-works (WLAN), neural netnet-works (NNs), quality of service (QoS).
I. INTRODUCTION
With the popularity of IEEE 802.11-based wireless local area net-works (WLAN) which are capable of providing high data-rate wireless access, the demands of multimedia services are increasing for mobile users. To support quality of service (QoS) for multimedia applications in the contention-based part of 802.11 medium access control (MAC), the IEEE 802.11 standardization committee just finished a service dif-ferentiation scheme, called enhanced distributed coordination function (EDCF) [1]. It grants the higher class traffics such as voice and video traffic to access the wireless medium early in most cases by differenti-ating interframe space (IFS) and backoff parameters at MAC layer with up to eight priorities, which are also known as traffic categories (TC) [2]. Although this mechanism can improve QoS of real-time traffic, the performance obtained is not optimal since the fixed EDCF parameters cannot be adaptive to the variation of communication circumstances such as traffic characteristics and load conditions.
There have been several works about improving IEEE 802.11e QoS [2]–[4] by optimizing EDCF parameters based on traffic types or load situations. Xiao [3] developed an analytical model of 802.11 EDCF and proposed a backoff-based priority scheme for real-time services. Tinnirello et al. [4] used simulations to investigate the behavior of differentiating EDCF parameters under various conditions of traffic loads and proposed a differentiation scheme based on a joined use of minimum contention window and IFS. However, most of these works provide solutions with the assumption of ideal channel conditions or homogeneous link qualities among the participating hosts which is impractical in realistic wireless environments. The transmission qualities of hosts, e.g., bit-error rates (BER), actually are unequal at most of the times even with a link adaptation mechanism applied on 802.11 physical layer (PHY) [1] due to limited modulation and coding schemes (MCS) available. Under heterogeneous channel conditions, the EDCF parameters determined with these schemes may be no more effective to provide differentiated QoS optimally. For example, we consider a simple transmission scenario of one real-time traffic flow and one best-effort flow in the network. In case the two flows are with
Manuscript received May 29, 2006; revised July 1, 2006. This work was sup-ported in part by Yulong Corporation under Grant 95-S-C01A and by Taiwan National Science Council under Grants 95-2219-E-002-018 and 95-2221-E-002-190.
C. Wang and P.-C. Lin are with Graduate Institute of Communication Engi-neering, National Taiwan University, Taipei 106, Taiwan, R.O.C.
T. Lin is with Graduate Institute of Communication Engineering and the De-partment of Electrical Engineering, National Taiwan University, Taipei 106, Taiwan, R.O.C. (e-mail: tsungnan@ntu.edu.tw).
Digital Object Identifier 10.1109/TNN.2006.883014
identical channel qualities, it can be assured that the real-time traffic which adopts better parameters, e.g., a smaller contention window or arbitrary interframe space (AIFS) can averagely obtain more channel utilization than the best effort one does. However, if the real-time traffic in contrast experiences a much worse link quality during a certain period, it will need more retries to succeed a packet deliv-ering and thus can probably obtain less medium shares of successful transmissions despite applying better parameters. This behavior is demonstrated in Section IV. Since IEEE 802.11e QoS is affected by not only the adopted EDCF parameters but also the channel qualities of transmissions, the parameters should be online optimized with the consideration of time-varying channel conditions.
In this letter, we propose a cross-layer adaptive scheme to improve 802.11e QoS by concurrently adapting each host’s MAC-layer parame-ters based on their application-layer QoS requirements and PHY-layer channel conditions. We utilize neural networks (NNs) to real-time learn the cross-layer correlations between the adopted parameters and the performed QoS metrics. Then, we exploit the learned functions to eval-uate the gradient of the cost-reward function, which is used to quan-tify the overall QoS, with respect to each parameter by means of a backpropagation manner. The evaluated gradients can, therefore, be used to determine the multidimensional parameters jointly for opti-mizing 802.11e QoS in terms of miniopti-mizing the cost. We have ap-plied this NN-based learning technique on legacy 802.11 DCF to adjust the packet size for improving the throughput [9] and to determine the backoff parameters for providing fair multimedia QoS [10].
In this letter, we explore the average throughput as the QoS indicator, and both the minimum contention window and AIFS as adaptable argu-ments. We conduct the simulation scenarios which involve traffic flows of different access categories (AC) to demonstrate the effectiveness of our adaptive scheme. The simulation results show that under a variety of channel conditions and traffic loads, the proposed algorithm can ef-fectively determine the 802.11e MAC parameters to guarantee differen-tial QoS in accordance with service classes, and also provide absolute QoS given that the total demand is below the provision bandwidth.
II. OPTIMIZATIONPROBLEM OF802.11EQOS
To investigate the optimization problem associated with IEEE 802.11e QoS, we explore the average throughput as the QoS indicator and both the minimum contention window size and AIFS as the adaptable arguments. ConsiderK active traffic flows in the network. To flowi, the throughput requirement for satisfying QoS is given as T T HRi; the minimum contention window size and AIFS adopted
isCWi;minand AIFSi, respectively; the throughput isTi. Since the achievable throughput of a given flow is affected by the parameters adopted of its own as well as those of other flows, the overall through-puts performed of the K flows are then modeled as a correlation functionf(1) associated with the joint setting of parameters. That is (T1; T2; . . . TK) = f(CW1;min; AIFS1; CW2;min;
AIFS2. . . CWK;min; AIFSK): (1) The parameters can be chosen based on an optimization approach by minimizing a cost-reward function. The cost-reward function is designed such that 802.11e QoS will be optimized while totally the achievable throughputs are closest to the specific requirements. The cost-reward functionCQoSis
CQoS= K i=1
(Ti0 T T HRi)2
T T HRi (2)
, where the denominatorT T HRiis utilized to normalize the differ-ence between the achievable throughput and the corresponding require-ment.CQoSwill be minimized if the channel utilization is shared in proportion toT T HRi.
To minimizeCQoS, each parameter is iteratively updated based on the gradient ofCQoSwith respect to itself. To calculate the gradient, the knowledge off(1) is needed. However, f(1) strongly depends on the characteristics of communication environments such as the collision probability among flows and the link quality of each flow. For example, if some flows experience link quality degradation, the channel sharing of them as well as that of others will be influenced, leading to a skewed sharing of overall throughputs. Hence,f(1) is a nonlinear, complicated, and time-variant function which is rather difficult to be depicted with a given analytical formula [5]. We are thus motivated to exploit NNs to model the complicated functionf(1). Thus, we can utilize the learned
^
f(1) to evaluate the gradient of CQoSwith respect to each parameter for minimizingCQoSby the technique of the backpropagation manner. The function-modeling and parameter-determining procedure of our adaptive scheme will be described in Section III.
III. PROPOSEDNN-BASEDADAPTIVEALGORITHM We exploit multilayer perceptron (MLP), which is the most common representative of NNs, to model the correlation function between the parameters and corresponding throughputsf(1). The exploited MLP consists of2K0x0K sensory units at the input layer, the hidden layer, and the output layer, respectively, to model the nonlinear function from the2K parameters to the corresponding K throughputs. The output of theith neuron at the lth layer can be described as
ui(l) = N
j=1
!ij(l)aj(l 0 1) + i(l) (3) ai(l) = hl(ui(l)); 1 i Nl; l = 1; 2 (4)
whereNlis the number of neuron at thelth layer; and ui(l) and ai(l)
are the activation and output values of theith neuron at the lth layer. The input units are represented byai(0) and the output units by ai(2).
!ij(l) refers to the weight connecting the output from the jth neuron
at the(l 0 1)th layer to the input of the ith neuron at the lth layer. i(l) refers to the bias associated with theith neuron at the lth layer. The used transferring functionhl(1) is sigmoid at hidden layer (l = 1) and is linear at output layer(l = 2).
The nonlinear function can be modeled with MLP gradually by re-cursively adjusting !ij(l) and i(l) to minimize the
mean-squares-error (MSE) between the achievable throughputTiand actual outputs ai(2) E = 12 M m=1 K i=1 T(m) i 0 a(m)i (2) 2 (5) whereM is the number of teacher patterns. The universal approxima-tion theorem [6] shows that MLP can approximate the nonlinear func-tion to an arbitrary degree of the accuracy.
Now, we describe our adaptive algorithm on adaptively determining the values of parameters for minimizing the cost-reward function CQoS. Here, we denote parameteri of the nth adaptation as i(n). The gradient-learning formula to minimizeCQoSwith respect to i(n)is
(n+1)
i = (n)i + 1 (n)i (6)
1 (n)i = 0 1 @CQoS=@ i(n); 1 i 3K (7)
where is the adjusting rate. To evaluate the minus gradient of CQoS
with respect to i(n), 0@CQoS=@ i(n), we add a virtual input layer
ai(01) and a virtual weight layer vwiunderai(0). The virtual weight
TABLE I SYSTEMPARAMETERS
Fig. 1. MSE varying with the adaptation sequence.
virtual weights, we can separately modelf(1) in the original NNs and thus utilize the learned ^f(1) to evaluate the gradient of CQoSin the ex-tending networks.
All the units in the virtual input layerai(01) are set to 1, and the
transferring function at the input layerai(0) is linear. That is a(n)i (0) =
i(n). Thus, the problem of evaluating the minus gradient ofCQoSwith respect to i(n),0@CQoS=@ i(n), is equivalent to that with respect to a(n)i (0), 0@CQoS=@a(n)i (0). Based on (3) and (4), 0@CQoS=@a(n)i (l)
can be derived with0@CQoS=@a(n)i (l + 1) from upper layer with the backpropagation and chain-rule manners
0@CQoS=@a(n)i (l) = N
j=1
0@CQoS=@a(n)j (l + 1)
2hl+10(uj(l + 1)) !ji(l + 1): (8)
In particular, the minus gradient ofCQoSwith respect to output layer 0@CQoS=@a(n)i (2) can be derived from the cost-reward function as
shown in (2). Therefore,0@CQoS=@a(n)i (l) can be successively
de-rived in the order ofl =2, 1, and 0. Since the value of a(n)i (0) is equal to i(n), (6) becomes
i(n+1)= i(n)0 1 @CQoS=@a(n)i (0): (9)
From (9), the multidimensional parameters (n+1)1 . . . 2K(n+1)for min-imizingCQoScan be determined concurrently.
New parameters are, therefore, applied into the system. The training data will be updated with the recent used parameters and the corre-sponding throughputs in order to learnf(1) in accordance with current wireless environments. In a recursive manner, the proposed algorithm alternately models the nonlinear function and adjusts the parameters for minimizingCQoS. To summarize, this algorithm consists of the fol-lowing four steps.
Step 1) Collecting up-to-date training data: Collect the data set of most recently used parameters and the corresponding throughputs for training.
Step 2) Online modeling the nonlinear function: With up-to-date training patterns,f(1) is online learned by recursively ad-justing the weights and biases of MLP.
Step 3) Calculating the gradients of parameters: The learned ^f(1) is utilized to evaluate the gradient ofCQoSwith respect to each parameter by the technique of backpropagation. With these gradients, the set of multidimensional parameters for minimizingCQoScan then be computed.
Step 4) Applying new parameters and refreshing training data: New parameters are applied to the system and then the new training data are collected. The process, therefore, returns back to Step 1) to online learnf(1).
When this adaptation framework is applied in 802.11 WLAN envi-ronments, the information of throughput for the training data is col-lected by means of a measurement-based approach. Practically, the co-herence time of 802.11 wireless channels is on the order of multiple packet transmission times [7]. This fact indicates that our approach in practice has the potential to learn the nonlinear function accurately re-flecting the variety of 802.11 channels.
IV. EXPERIMENTALRESULTS
In this section, we show the simulation results to demonstrate that under varying heterogeneous channel conditions, using fixed 802.11e EDCF parameters cannot guarantee the differential QoS in accordance with service classes, and then to demonstrate the effectiveness obtained by our proposed adaptive scheme. The experimental scenarios are con-ducted as follows. We assume an IEEE 802.11b infrastructure WLAN in which two hosts transmit file-transfer-protocol (FTP) traffic flows and two hosts transmit video traffic flows through the access point (AP) to the corresponding receivers with basic carrier sense multiple access/collision avoidance (CSMA/CA) scheme [without request-to-send/clear-to-send (RTS/CTS) mechanisms]. Assume hosts delivering FTP applications are always in an ideal channel condition, whereas those delivering video applications are initially with a BER of 2E-5 and later suffer from quality degradation with a BER of 4E-5 as they move away from AP. All the hosts use the same MCS as differential binary phase shift keying (DBPSK) and then transmit at the data rate of 1 Mb/s. The values of the system parameters are shown in Table I.
We examine two simulation scenarios of diverse traffic load con-ditions as follows. In Scenario I, the video traffic is delivered under lighter load conditions with the throughput requirement of 220 Kb/s for lower video quality; in Scenario II, the video traffic is under heavier conditions with the requirement of 260 Kb/s for higher fidelity QoS. In both scenarios, the FTP traffic is under saturation conditions (i.e., always has a packet to transmit). The performance is indexed as av-erage throughput and the experimental results are obtained based on an analytical approach of extending a verified two-dimensional (2-D) Markov chain model proposed by Bianchi [8]. With this modified ana-lytical model, the average throughput with various settings of parame-ters under error-prone channel conditions can be derived numerically. We compare the proposed adaptive scheme with IEEE 802.11e EDCF protocol. With EDCF, the video traffic and FTP traffic are involved into access category of video (AC_VI) and access category of best effort (AC_BE), respectively, adopting differentiated parameters in the unit of slotted time as follows: the minimum contention window of(CWmin[VI]; CWmin[BE]) = (15, 31), the maximum contention
window of(CWmax[VI]; CWmax[BE]) = (31, 1023), and the AIFS number of (AIFSN[VI], AIFSN[BE])= (2, 3) [1]. For our adaptive algorithm, the adaptation space of minimum contention window is from 3 to 127, and that of AIFS number is from 2 to 16. The chosen values of minimum contention window and AIFS number will be rounded to the closest integer. In the function-modeling stage, we collect the data set of five most recently used parameters and the corresponding throughputs as the teacher patterns, and apply an online
Fig. 2. Adaptation trajectory of parameters of AC_BE and AC_VI flows, respectively. (a) Minimum contention window. (b) AIFS number.
Fig. 3. In Scenario I, the video traffic is under lighter load conditions with the throughput requirement of 220 Kb/s for lower quality, the throughput of an AC_BE and an AC_VI flows with 802.11e EDCF, and the proposed adaptive scheme, respectively.
learning strategy, i.e., the NN is updated when a new training data is available. The NN weights and biases are adjusted till the MSE falls below 1E-6 or the training epochs are over 1000 times. In the param-eter-determining stage, the parameters are updated with the adjusting rate of 0.1. The parameter T T HRiof the cost-reward function for AC_VI flows is set to 220 Kb/s for lower QoS or 260 Kb/s for higher fidelity QoS depending on the traffic load conditions; T T HRifor AC_BE flows is set to 60 Kb/s.
First, we show the simulation results associated with Scenario I. Fig. 1 shows the level of MSE varying with the adaptation sequences, i.e., the numbers of the parameters are adjusted. It is shown that MSE continues to decrease mostly when NN can learn the function gradually. While the communication environment changes (as the average BER of AC_VI flows deteriorates from 2E-5 to 4E-5) during the sequence of 11, MSE rises immediately due to the changes of the nonlinear function. With the learning capability of NN and the up-to-date teacher data, MSE decreases rapidly after the adaptation sequence of 14. The results demonstrate that our adaptive algorithm can real-time model the communication system under varying hetero-geneous channel conditions. Fig. 2 shows the adaptation trajectory of the minimum contention window and AIFS number of AC_BE and AC_VI flows. It is shown that these parameters converge during the sequences of 10 and 20 while the gradients of the cost-reward function with respect to each parameter can be calculated increasingly accurately. The results demonstrate that our algorithm can effectively
adjust the multidimensional parameters simultaneously for minimizing the cost-reward function under heterogeneous channel conditions.
Fig. 3 presents the average throughput of AC_VI and AC_BE flows with 802.11e EDCF and the proposed adaptation scheme, respectively. It is shown that throughput of an AC_VI flow with 802.11e EDCF (de-noted by the blue cyan triangle) is 217 Kb/s initially and comes to 158 Kb/s as communication environments change (the BER becomes from 2E-5 to 4E-5), while the requirement for lower quality (220 Kb/s) is never met all the time. Furthermore, we can observe that the achiev-able throughput of an AC_VI flow is even lower than that of an AC_BE flow (denoted by the red circle) as environments change. It is demon-strated that using fixed 802.11e EDCF parameters cannot be adaptive to the change of channel conditions for guarantying differential QoS in accordance with service classes.
With our adaptive scheme, it is shown that the QoS of AC_VI flows are significantly improved. The bandwidth utilization of various ser-vice classes is regulated by adapting the minimum contention windows and AIFS numbers, and, therefore, the throughput of AC_VI flows can meet their QoS requirement after the fourth adaptation sequence. While BER becomes worse from 2E-5 to 4E-5, the applied parameters cease to be effective in the present situation. Consequently, the throughput of AC_VI flows degrades to 195 Kb/s and can no more meet the pre-scribed requirement for lower quality. It is shown that AC_VI flows can satisfy QoS again after the sixteenth sequence while the parameters are adjusted according to the present circumstances. It is also shown
Fig. 4. In Scenario II, the video traffic is under heavier load conditions with the throughput requirement of 260 Kb/s for higher fidelity QoS, the throughput of an AC_BE and an AC_VI flows with 802.11e EDCF, and the proposed adaptive scheme, respectively.
the differential QoS is ensured with the proposed scheme such that the throughput of an AC_VI flow is higher than that of an AC_BE flow throughout.
Finally, we present the experimental results associated with Scenario II that the video traffic is in heavier load conditions with the throughput requirement of 260 Kb/s for higher fidelity QoS. The results of through-puts are presented in Fig. 4. With the proposed scheme, it is shown that AC_VI flows can gradually satisfy higher fidelity QoS when the traffic load is in this heavier situation. The results shown in Figs. 3 and 4 demonstrate that under a variety of channel conditions and traffic loads, our adaptation framework can intelligently determine 802.11e MAC parameters to guaranty differential QoS, and also provide abso-lute QoS given that the total demands are below the provision band-width. The 802.11e QoS improving in terms of average delay is also explored with the same technique, and these results are omitted due to limited space.
V. CONCLUSION
Providing QoS for multimedia services in 802.11e WLANs is chal-lenged by the time-varying channel conditions. In this letter, we pro-pose a cross-layer adaptive algorithm for provisioning 802.11e QoS by jointly determining multidimensional MAC parameters based on the physical-layer channel conditions and application-layer QoS re-quirements. The simulation results demonstrate the effectiveness of our adaptive scheme.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for their thoughtful comments and suggestions which have advanced the quality of this letter.
REFERENCES
[1] Draft supplement to part 11: Wireless medium access control (MAC) and physical layer (PHY) specifications: Medium access control (MAC) enhancements for quality of service (QoS), IEEE 802.11e/D5.0, Jun. 2003.
[2] K. Xu, Q. Wang, and H. Hassanein, “Performance analysis of differ-entiated QoS supported by IEEE 802.11e enhanced distributed coordi-nation function (EDCF) in WLAN,” in Proc. IEEE GLOBECOM ’03, Dec. 1–5, 2003, vol. 2, pp. 1048–1053.
[3] Y. Xiao, “A simple and effective priority scheme for IEEE 802.11,” IEEE Commun. Lett., vol. 7, no. 2, pp. 70–72, Feb. 2003.
[4] I. Tinnirello, G. Bianchi, and L. Scalia, “Performance evaluation of differentiated access mechanisms effectiveness in 802.11 networks,” in Proc. IEEE GLOBECOM ’04, Nov. 29–Dec. 3 2004, vol. 5, pp. 3007–3011.
[5] M. van der Schaar and D. S. Shankar, “Cross-layer wireless multimedia transmission: Challenges, principles, and new paradigms,” IEEE Wire-less Commun. Mag., vol. 12, no. 4, pp. 50–58, Aug. 2005.
[6] K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural Netw., vol. 2, no. 5, pp. 359–366, 1989.
[7] B. Sadeghi, V. Kanodia, A. Sabharwal, and E. Knightly, “OAR: an opportunistic auto-rate media access protocol for ad hoc networks,” in ACM MOBICOM’02, Sep. 2002, pp. 24–35.
[8] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed co-ordination function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 535–547, Mar. 2000.
[9] P. Lin, C. Wang, and T. Lin, “A context-aware approach for multimedia performance optimization using neural networks in wireless LAN en-vironments,” in Proc. IEEE Int. Conf. Maintenance Eng. (ICME ’06), Jul. 9–12, 2006, pp. 1177–1180.
[10] C. Wang and T. Lin, “A neural network based adaptive algorithm for multimedia quality fairness in WLAN environments,” in IEEE Int. Conf. Maintenance Eng. (ICME ’06), Jul. 9–12, 2006, pp. 1233–1236.
