Fig. 2 shows the architecture of the two-layered CNN processor. The equivalent two-layer structure for the CNN processor involves the first decision layer, zi,k1 !
, with state variable output Xi,k(τ), and the second output layer, zi,k2 !
, with state variable output ci(t, τ). The decision layer consists of N× K neurons; the output layer is with an (N + 1) × 1 array, where the output of the first neuron is the summation of all the others. The interconnections between the neurons of decision layer and those of output layer are defined by
• For the first decision layer to the second output layer, the connection weight between Xi,k(τ) and cj(t, τ) is 2−k, ∀k if j = i; is zero if j = i.
• For the second output layer feedback to the first decision layer, only the first neuron output is connected to the Xi,k(τ) of the decision layer with the interconnection weight η1· 2−k for ∀i.
following : We have proved
.
[Yi,k(2)(t)]
Fig. 2. The two-layer structure of CNN processor.
The recurrent interconnection weights and the external control weights for the first decision layer defined in (IV.19) are then modified to be
For the second output layer, there are no external inputs, and only recurrent interconnection weights exist. The interconnection weight between ci(t, τ) and cj(t, τ) is given by δ0,j with i = 0.
It can be shown that the two-layer structured CNN processor has the same energy function and the local minimum as the single-layer one defined by (IV.19). However, the complexity of
(19)
(20)
(19)
interconnections in the two-layer one is proportional to [3N × K + N], which is significantly lower than [N × K]N×K in the single-layer one.
V. SIMULATIONRESULTS ANDDISCUSSION
In the simulations, a scenario with five types of services in three classes is assumed. Type-1 service is a real-time class of traffic with peak rate 15kbps, activity factor 0.57, PD∗ = 0.05, D∗ = 40ms, and BER∗ = 10−3. Type-2 (type-3) service is a non-real-time interactive class of traffic with Pareto process [19] of which the mean rate is 8kbps (12kbps), R∗m,i=7.2kbps (R∗m,i=11kbps), and BER∗= 10−5 (BER∗= 10−5). And type-4 (type-5) service is a non-real-time best effort class of traffic in batch Poisson distribution with mean rate 6kbps (15kbps) and mean batch size 1.2k bits (1.2k bits), and BER∗ = 10−5. The proportion in the number of connections from type-1 to type-5 is kept at 1:1:1:1:1. Also, four modulation schemes, BPSK, QPSK, 16QAM, and 64QAM, are available for transmission as long as the BER requirement can be fulfilled and the remaining queue is enough.
We compare the proposed CNNU-based scheduler with the exponential rule (EXP) scheduling scheme. The performance measures are such as the average system throughput, the average packet dropping ratio of RT connections, PD, the average transmissio rate of NRT interactive connections, Rm, the ratio of RT connections in which their packet dropping ratio requirement is not guaranteed, φPD, the ratio of NRT interactive connections in which their minimum transmission rate requirement is not guaranteed, φRm, and the fairness variance index of NRT connections, Fv. The Fv is defined for measuring the variance of fairness to share the radio resource among all NRT connections. It is given by
Fv = 1
where NNRT is the number of NRT connections. The fairness variance index shows the variance of the normalized radio resource allocated and the normalized proportion of resource desired to share.
Fig. 3 shows the average system throughput. It can be found that the CNNU-based scheduler can always have a higher system capacity than the EXP scheduling scheme in all traffic load conditions; it achieves the improvement of the system throughput over the EXP scheduling scheme by more than 9% as the number of connections is greater than 200, and by higher
!# % # &(* , ./ (1 (
3 5 % # 8 (8 :<>? @ B ?(
Fig. 3. The average system throughput
than 15% as the number of connections increases up to 250. This is because the radio resource funciton makes CNN processor adapt to the link variation and allocate radio resource in an efficient way. Both RT and NRT connections with relatively worse link conditions have lower probability to be scheduled as long as their QoS requirements can be achieved in a long term sense. The fairness compensation function makes the NRT connections share the radio resource according to the location dependent fairness and thus a higher radio resource efficiency can be achieved. Also, the CNN processor can determine an optimal radio resource assignment vector in the sense that the allocation of downlink power by CNNU-based scheduler is the most efficient one, with given utilities and upper limits of the radio resource assignment. Additionally, beyond the point of 250 connections, the throughput of the EXP scheme is almost saturated, while the throughput of the CNNU-based scheduler continues to grow up but with a slightly lowering slope. It is because the CNNU-based scheduler can achieve utilization of multiuser diversity gain better than the EXP scheduling scheme.
Fig. 4 depicts performance measures of the average packet dropping ratio of type-1 RT con-nections PD and the average transmission rate of type-2 and type-3 NRT interaction connections Rm. It can be found that the PD of the CNNU-based scheduler is larger than that of the EXP scheme and it violates the PD∗ requirement as the number of users is at about 250; on the other hand, all the Rm of type-2 and type-3 connections of the CNNU-based scheduler are greater than that of the EXP scheme as the number of connections is greater than 125 but the EXP scheme violates the Rm∗ requirements as the number of users is at about 170. These indicate that the QoS
!# % # &(* , ./ (1 (
3 5 % # 8 (8 :<>? @ B ?(
Fig. 4. QoS performance measures ofPDandRm
guaranteed region by the CNNU-based scheduler is as the number of connections is less than 250, while that by the EXP scheme is as the number of connections is less than 175. The QoS quaranteed region achieved by the CNNU-based scheduler is larger than that given by the EXP scheme. This is beacause the QoS deviation function together with the priority bias designed in the CNNU-based scheduler can balance the extent of deviation of every performance measure from the QoS requirement. The worse the QoS performance measure is, the more the radio resource will be scheduled. Besides, since the CNNU-based scheduler has higher throughput performance, the more number of connections can be served in the QoS region. Moreover, if we define the maximum throughput achievable in QoS guaranteed region to be the average system throughput, the CNNU-based scheduler can have the average system throughput equal to 2160Mbps at 245 connections, while the EXP scheme can have the average system throughput equal to 1600Mbps at 175 connections. The former attains the average system throughput greater than the latter by an amount of 25%.
Fig. 5 shows the ratio of RT connections of which the packet dropping ratio requirement is not graranteed, φPD, and the ratio of NRT interactive connections of which the minimum transmission rate requirement is not guaranteed, φRm. It can be seen that the total ratio of connections with QoS requirements un-guaranteed for the CNNU-based scheduler is about 0.0435, while that for the EXP scheme is greater than 0.18 in heavy loaded situations as the number of connections is greater than 225. The total ratio of connections with QoS requirements un-guaranteed is here defined as 13φPD + 23φRM, which is weighted by the number of RT and
!# % # &(* , ./ (1 (
3 5 % # 8 (8 :<>? @ B ?(
Fig. 5. The ratioφPD for RT connections and the ratioφRm for NRT interactive connections
!# % # &(* , ./ (1 (
3 5 % # 8 (8 :<>? @ B ?(
Fig. 6. The fairness variation index for NRT connections
NRT interactive connections. This indicates that CNNU-based scheduler can achieve lower total ratio of connections in all traffic types of which the corresponding QoS requirements are not guaranteed than the EXP scheme does. The reason is that the CNNU-based scheduler can balance the allocation of radio resources among traffic types and avoid allocating excess radio resource to connections with bad link condition, while the EXP scheme prefers RT connections and overprotects them so that the QoS guaranteed region is reduced. Note that the ratios of φPD and φRm are greater than zero at any traffic load conditions due to the existence of connections with very bad link quality. These results imply that the CNNU-based scheduler will not guarantee all the QoS requirements all the time, and a properly designed call admission control is required to reject the connections with very bad link quality in terms of the current traffic load conditions.
Fig. 6 shows the fairness variance index of NRT connections. It can be found that the fairness
grows up slowly as the number of connections increases; the fairness variance index of the EXP scheme, on the other hand, increases with slightly higher slope compared to CNNU-based scheduler. This is because the fairness compensation function of the CNNU-based scheduler considers the location dependent information and aims to share the radio resource fairly as long as the minimum rate is guaranteed, while the design of the EXP scheme ignores the location dependent information and allocates rate fairly to all connections. The fairness compensation function, considering location dependent information, also facilitates the higher capacity for the CNNU-based scheduler shown in Fig. 3.
REFERENCES
[1] 3rd Generation Partnership Project Jun. 2001, QoS Concept and Architecture, 3GPP TS 23.107, [Online]
http://www.3gpp.org.
[2] A. Parekh and R. G. Gallager, ”A generalized processor sharing approach to flow control in integrated services networks:
the single node case,” IEEE/ACM Trans. Networking, vol. 1, no. 3, pp. 344-357, JUN. 1993.
[3] D. Ferrari and D. Verma, ”A scheme for real-time channel establishment in wide-area networks,” IEEE J. Select. Area Commun., vol. 8, no. 3, pp. 368-379, Apr. 1990.
[4] Y. M. Lu and R. W. Brodersen, ”Integrating power control, error correction coding, and scheduling for a CDMA downlink system,” IEEE J. Select. Area Commun., vol. 17, no. 5, pp. 978-989, May 1999.
[5] V. K. N. Lau, and Y. K. Kwok, ”On generalized optimal scheduling of high data-rate bursts in CDMA systems,” IEEE Trans. Commun., vol. 51, no. 2, pp. 261-266, Feb 2003.
[6] V. Bharghavan, S. W. Lu, and T. Nandagopal, ”Fair queuing in wireless networks: issues and approaches,” IEEE Personal Commun., vol. 6, no.1, pp. 44-53, Feb. 1999.
[7] A. C. Varsou and H. V. Poor, ”HOLPRO: a new rate scheduling algorithm for the downlink of CDMA networks,” Proc.
IEEE VTC’2000, pp. 948-954.
[8] A. L. Stolyar and K. Ramanan, ”Largest weighted delay first scheduling: large deviations and optimality,” Annal Appl.
Prob., vol. 11, no. 1, pp. 1-48, 2001.
[9] A. C. Kam, T. Minn, and K. Y. Siu, ”Supporting rate guarantee and fair access for bursty data traffic in W-CDMA,” IEEE J.
Select. Area Commun., vol. 19, no. 11, pp. 2121-2130, 2001.
[10] S. Shakkottai and A. L. Stolyar, ”Scheduling algorithms for a mixture of real-time and non-real-time data in HDR,” Bell Lab Reports, 2000.
[11] L. O. Chua and L. Yang, ”Cellular neural networks: Theory,” IEEE Trans. Circuits Syst., vol. 35, pp. 1257-1272, Oct. 1988.
[12] J. A. Nossek, ”Design and learning with cellular neural networks,” IEEE Proc. CNNA-1994, Rome, pp. 137-146, Dec. 1994.
[13] ”On sovling constrained optimization problems with neural networks: a penalty method approach,” IEEE Trans. Neural Net., vol. 4, no. 6, pp. 931-940, Nov. 1993.
[14] ”Energy function based dynamic programming”, IEEE Trans. Neural Net., vol. no. pp., 1991.
[15] A. J. Goldsmith and S. G. Chua, ”Variable-rate variable-power MQAM for fading channels,” IEEE Trans. Commun., vol.
45, no. 10, pp. 1218-1230, Oct. 1997.
[16] C. S. Chang and J. A. Thomas, ”Effective Bandwidth in High-Speed Digital Networks,” IEEE J. Select. Area Commun., vol.
13, no. 6, AUG. 1995.
[17] L. S. Chou and C. S. Chang, ”Experiments of the theory of effective bandwidth for markov sources and video traces,” Proc.
IEEE INFOCOM’96, pp.
[18] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, and R. Vijayakumar, ”Providing quality of service over a shared wireless link,” IEEE Commun. Mag., vol. 39, no. 2 , pp. 150-154, Feb. 2001.
[19] Universal Mobile Telecommunications System (UMTS), Selection procedures for the choice of radio transmission technologies of the UMTS, UMTS 30.03, version 3.2.0, 1998.
A Novel Dynamic Cell Configuration Scheme in Next-Generation Situation-Aware CDMA Networks
Ching-Yu Liao, Member, IEEE, Fei Yu, Member, IEEE, Victor C.M. Leung, Fellow, IEEE and Chung-Ju Chang, Senior Member, IEEE
Abstract
o balance the time-varying traffic load between cells, caused by user mobility and diverse applications, it is crucial for next-generation CDMA cellular networks to configure cell coverage and capacity dynamically. In this paper, we show that pilot power allocation is highly coupled to other facets of radio resource management.
We propose a novel dynamic cell configuration scheme for multimedia CDMA cellular networks, based on reinforcement-learning, which takes into account pilot, soft handoff, and maximum link power allocations as well as call admission control mechanisms. Simulation results demonstrate the effectiveness of the proposed scheme in situation-aware CDMA networks.o balance the time-varying traffic load between cells, caused by user mobility and diverse applications, it is crucial for next-generation CDMA cellular networks to configure cell coverage and capacity dynamically. In this paper, we show that pilot power allocation is highly coupled to other facets of radio resource management. We propose a novel dynamic cell configuration scheme for multimedia CDMA cellular networks, based on reinforcement-learning, which takes into account pilot, soft handoff, and maximum link power allocations as well as call admission control mechanisms. Simulation results demonstrate the effectiveness of the proposed scheme in situation-aware CDMA networks.T
I. INTRODUCTION
The growing popularity of multimedia Internet applications is a strong driving force for future cellular mobile systems. Due to user mobility and wide range of applications, the traffic pattern of each cell can vary dynamically. Thus, the current practice of engineering cell coverage and capacity based on pre-defined traffic patterns before a code division multiple access (CDMA) cellular network is deployed may lead to poor utilization of radio resources.
Chapter 4
A Novel Dynamic Cell Configuration Scheme in Next-Generation
Situation-Aware CDMA Networks
Due to asymmetric traffic and the interdependence of traffic capacity and coverage, this problem could be exacerbated in next-generation CDMA cellular networks, especially over the capacity-limited downlink [1]−[4].
To adapt to the variations of traffic load, tradeoffs between coverage and capacity in CDMA cellular systems have been considered [3]−[7]. For example, to guarantee the coverage of a cell, more power is used to reach mobile stations (MSs) near cell boundaries under power control. However, in interference-limited systems, the resulting higher inter-cell interference will reduce the system capacity significantly. Furthermore, under large traffic variations, power control may not be effective [3]−[5]. A uniform network layout with equal-sized cells, while optimal under uniform traffic, suffers significant capacity degradations if traffic loads are not balanced among all the cells [6]. To accommodate traffic load variations between cells, it is crucial for next-generation CDMA cellular networks to be aware of system situations and configure cell coverage and capacity dynamically [1], [7].
Several schemes for dynamic cell configuration (DCC) have recently been proposed [8]−[16].
Optimization of pilot power, and downlink capacity and coverage planning were considered in [8], [9]. In [10], a DCC scheme for circuit-switched micro-cellular CDMA systems was proposed to enhance the uplink performance. In [11], the competitive characteristics of network coverage and capacity were analyzed for a simple network. Only one class of service was considered in [8]−[11], and it may be difficult to extend these schemes to multiple classes of service. Some techniques based on heuristics have also been proposed for dynamic pilot power allocation (DPPA) to balance downlink traffic load while assuring service coverage [12], [13].
However, these schemes may cause “coverage failure regions” between cells where pilot signals are too weak to serve a MS [14], [15]. Moreover, a common shortcoming of the previous work [8]−[15] is that only pilot power is adjusted dynamically in the time-varying environment, without adjusting other parameters critical to radio resource management (RRM).
In fact, pilot power allocation and other RRM parameters are tightly coupled. In our previous work [16], we have shown that system performance can be improved significantly by a self-organized DCC scheme with coordinated call admission control (CAC), compared to fixed pilot power allocation (FPPA) and DPPA without taking CAC into account. Other work has shown that signal quality degradation can be prevented by configuring cell areas adaptively and setting power levels appropriately [4], [17], and soft handoff has significant impacts on the system
capacity and cell coverage [18], [19]. Therefore, an effective mechanism, link proportional power allocation (LPPA), was proposed for downlink soft handoff in [20], [21]. It was shown that LPPA can enhance system capacity in CDMA cellular systems with mixed-size cells, compared to conventional site-selection diversity transmissions (SSDT) scheme [22].
In this chapter, we show that DPPA without changing other related RRM parameters accordingly can result in performance degradations. To address this problem, we propose a novel DCC scheme based on reinforcement-learning called DCC-RL. The novelties are as follows.
1) DPPA is linked with soft handoff power and maximum link power allocations as well as CAC mechanisms. 2) Reinforcement-learning efficiently tackles optimization problems with large state spaces and action sets [23] in realistic CDMA multimedia cellular networks, which were previously deemed intractable [24]. 3) Our method does not require a priori knowledge of the state transition probabilities associated with the cellular network, which are very difficult to estimate in practice due to the varied propagation environment, diverse multimedia services, and random user mobility. 4) DCC-RL can be implemented in a distributed manner in each base station (BS), minimizing signaling overhead between BSs and radio network controllers, and the number of system states involved in computations.
We compare DCC-RL with fixed cell configuration (FIX) employing FPPA, and DPPA without changing other RRM parameters. Simulation results show that DCC-RL outperforms the others by increasing the total throughput, decreasing the frame error probability, blocking probability, and handoff forced termination probability with the price of slightly increasing the size of the active set.
The rest of this chapter is organized as follows. DCC issues are discussed in Section II.
Section III describes the system model. Section IV formulates the DCC problem taking into account of RRM, and presents the proposed DCC-RL scheme. Simulation results are presented and discussed in Section V.
II. ISSUES OFDYNAMICCELLCONFIGURATION