Iterative Multiuser Detection and Channel Decoding for DS-CDMA Using Harmony Search
Post on 09-Mar-2017
Embed Size (px)
IEEE SIGNAL PROCESSING LETTERS, VOL. 16, NO. 10, OCTOBER 2009 917
Iterative Multiuser Detection and Channel Decodingfor DS-CDMA Using Harmony Search
Rong Zhang and Lajos Hanzo, Fellow, IEEE
AbstractA novel random-guided optimization method is pro-posed for multiuser detection (MUD) in DS-CDMA systems em-ploying the so-called Harmony Search (HS) algorithm. We specif-ically design the HS-aided MUD for the communications problemconsidered and apply it in an iterative joint MUD and channeldecoding framework. Our simulation results demonstrate that anear-single-user performance can be achieved without the employ-ment of the full-search-based optimum MUD even in extremelyhighly loaded DS-CDMA systems.
Index TermsEvolutionary Algorithm, , Harmony Search Algo-rithm, iterative receiver, Markov Chain Monte Carlo Algorithm,multiuser detection, .
I N the context of iterative detection and decoding (IDD) aided multiuser detection (MUD), stochastic global opti-mization techniques may be used in order to reduce the com-plexity, while still capturing the optimum full-search-based so-lution with a high probability using, for example, the GeneticAlgorithm (GA)  or the Metropolis-Hastings (MH) type al-gorithm .
Imitating the improvisation process of musicians, a newmeta-heuristic optimisation method referred to as the HarmonySearch (HS) algorithm was proposed in . When a musicianimprovises, the aesthetic value referred to as the fitness func-tion (FF) is determined by a set of pitches produced by themusic instruments (variables) involved. The musician seeks toproduce aesthetically pleasing harmony (the optimum solution)as determined by his/her aesthetic perception inferred fromrehearsals (iterations).
In this paper, we design a HS-aided MUD algorithm, which iscapable of approaching the optimum MUDs performance withlow complexity.
The remainder of this paper is organized as follows. InSection II, we introduce the system model and specifically de-sign the HS algorithm in the context of IDD in Section III. We
Manuscript received April 20, 2009; revised June 13, 2009. First publishedJuly 10, 2009; current version published August 12, 2009. This work was sup-ported by EPSRC as part of the Core 4 Research Programme of the VirtualCenter of Excellence in Mobile and Personal Communications, Mobile VCE,www.mobilevce.com. The associate editor coordinating the review of this man-uscript and approving it for publication was Prof. Tongtong Li.
The authors are with the School of Electrical and Computer Sci-ence, University of Southampton, Southampton SO17 1BJ, U.K. (e-mail:firstname.lastname@example.org; http://www-mobile.ecs.soton.ac.uk).
Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2009.2027159
then characterize its performance in Section IV, and concludein Section V.
II. SYSTEM MODEL AND PROBLEM FORMULATION
In the context of the IDD scheme seen in Fig. 1, we considera rate- coded BPSK modulated -user DS-CDMA systememploying user-specific -chip random spreading sequences.The assumption of a long interleaver within the IDD allowsus to focus on a particular symbol interval ,where , , and are thereceived sample vector, the transmitted symbol vector, and theDS-CDMA spreading matrix, respectively. Furthermore,
, is the complex additive white Gaussiannoise (AWGN) vector and represents theblock-invariant complex channel.
After observing the a priori information in terms of log-likeli-hood ratios (LLRs) denoted by , the extrinsic information
is delivered by the MUD to the outer channel decoder.Based on the independence of each users information, the ex-trinsic LLR of the th user is given by
, where the a posteriori LLR is
where denotes the -user vector with the th elementexcluded. We note that all possible vectors are required inthe above summation, which leads to a prohibitive complexity.This motivates the development of a range of low-complexityMUD detection algorithms, such as those in  and  and theproposed HS algorithm of the next section.
III. HARMONY SEARCH AIDED ITERATIVE RECEIVER
We approximate the summation of all legitimate -user vec-tors in terms of the Bayesian optimum of (1) by collecting a suf-ficient number of significant -user candidate vectors , wherethe collection is followed by a set of HS rules and the signif-icance is quantified by the FF to be introduced below.
A. Fitness Function of Harmony Search Aided MUD
We define the FF to be evaluated as the joint a posterioriprobability (APP) of the -user transmitted vector based onthe observation and the a priori LLRs provided by thechannel decoder, which may be expressed as
1070-9908/$26.00 2009 IEEE
918 IEEE SIGNAL PROCESSING LETTERS, VOL. 16, NO. 10, OCTOBER 2009
Fig. 1. Iterative multiuser detection and channel decoding, where ENC is short for encoding.
where denotes the a priori probabilityof each user , which may be expressed as
B. Naive Transplanting of the Harmony Search Algorithm
We first introduce a range of parameters associated with theoriginal HS algorithm . The harmony memory size spec-ifies the size of the harmony memory matrix (HMM), whichhosts the number of initial -user harmony candidates. The har-mony memory activation probability specifies the proba-bility of generating a new -user candidate from the HMM,rather than being randomly selected with a probability of
. The pitch adjustment probability specifies the prob-ability of further tuning the newly generated -user candidatefrom the HMM with a probability of . Finally, representsthe total number of improvisations executed. Explicitly, the HSalgorithm can be summarized as follows.
Initialization: We initialize the HMM hostingrandomly generated -user candidate harmony vectors
from the set of legitimate solutions.Improvisation: Three HS operations are applied during each
improvisation, namely the memory activation, pitch adjustmentand random selection. In particular, a new -user harmonyvector may be randomly selected from the alphabet
with a probability of orselected from the HMM with a probability ofaccording to , , , whichimplies inheriting the th bit of one of the candidate vectorsin the th iteration . Once the memory activa-tion was carried out, a further pitch adjustment characterised bya step of may be applied with a pitch adjustment probabilityof . In detail, the specific value of each variable ,
of the new -user harmony candidate is tuned tomatch the neighboring values in its legitimate solution alphabet
.Updating: The harmony vector having the worst FF in
is replaced by the newly generated -user harmony vector, pro-vided it is less fit than the new one, otherwise it survives for thenext iteration.
C. Pitch Adjustment in the Harmony Search Aided MUD
In the pitch adjustment step, the direct employment of the HSalgorithm would entail binary toggling of the related bit
, based on the predefined probability of andhence may lead to a dramatically different -user counterpart,which would be disadvantageous in our MUD. Instead, we thuspropose to take the soft information associated with intoaccount for determining . As a result, the specific value of
will be altered during each pitch adjustment step based onthe probability representing the related soft information.
At iteration , after randomly selecting a -user base har-mony vector from the HMM with a probability of ,the pitch adjustment is carried out by generating the th vari-able based on the marginal APP ofthe base harmony vector, where denotes the -user basevector with the th element excluded. This automatically up-dated marginal APP represents the soft information of andacts as the replacement of in the original HS proposal of. This reflects the process of musical improvisation, wherea particular instrument adjusts its pitch based on the harmonygenerated by other instruments. The LLR of this marginal APPmay be conveniently evaluated in the following way:
where the subscript indicates the sign of the th elementof the -user base vector. Hence, we have
In this case, the -based memory activation and the-based pitch adjustment merge into a single joint step.
The marginal APP-based pitch adjustment of (3) is capableof providing a sufficiently high decision reliability, hence themagnitude of the LLRs is improved during the successiveimprovisations. This implies that the selection of a new -userharmony vector for inclusion in the HMM is more appropriatethan a random choice. Hence, we may set the probability
and appropriately avoid the operation of randomselection. In other words, artificially enforcing a random se-lection with a probability of is not recommended,since the improvisations based on this random selection may bewasted. Importantly, ignoring the associated random selectiondoes not limit the exploration capability of the HS-aided MUD,because if we have a sufficiently high , randomly selectedbase vectors have first been generated, before fine-tuning thepitch adjustment. In summary, the pseudocode of our HS-aidedMUD algorithm is shown in Table I.
ZHANG AND HANZO: ITERATIVE MUD AND CHANNEL DECODING FOR DS-CDMA USING HS 919
TABLE IPSEUDOCODE OF THE HS-AIDED MUD
D. Soft Output of the Harmony Search Aided MUD
After improvisations, we may approximate (1) based on thesurviving -user candidate vectors in the final HMM ,
which can be expressed as
where and represent the FF function of a givencandidate vector having its th entry equals to 1 and 1,respectively. When considering the extrinsic LLR , thecorresponding extrinsic FF value is substituted in (4),which may be given by .
IV. PERFORMANCE EVALUATION
A. Parameters and Complexity
Consider an outer rate repetition coded DS-CDMAsystem employing user-specific random spreading sequencesof length , where the information frame length was
and the number of iterations between the MUD andthe soft decoder was set to . Furthermore, we de-fine the so-called normalised system-load as theratio of the number of users supported to the spreading sequencelength employed. An AWGN channel was assumed and a uni-formly distributed transmission block-invariant channel phasenoise was imposed, which was assumed perfectly known to thereceiver.
Let us first discuss the complexity of our HS-aided MUD interms of the required FF evaluations. Generating the a poste-riori LLRs given by (1) for users requiresevaluations of (2), while the HS algorithm requires a total of
Fig. 2. BER performance of HS-aided MUD of a -repetition codedBPSK modulated DS-CDMA system using -chip random sequencesand iterations.
evaluations of the FF of (2), In detail, it includes FF evalua-tions of the initial HMM, and a further evaluation at the end ofeach of the improvisations as well as evaluations,when generating the marginal APP based pitch adjustment ofeach of the improvisations. In addition, the soft output gen-eration requires a further evaluations of (4).
B. Simulation Results
We first compare our HS-aided MUD to the conventionalMH algorithm of a full-load uncoded DS-CDMA system. Inthis case, the soft output evaluation of (4) may be avoided bysimply considering the best harmony vector in the final HMM,i.e. , and as a result, the final additiveterm of (5) is neglected. Hence, our HS-aided MUD requires
FF evaluations. On the other hand, in the MH algo-rithm, a total of parallel seeds were generated and eachseed employed Gibbs sampling processes, resulting in atotal of FF evaluations. Fig. 2suggests that our HS-aided algorithm with less number of FFevaluations exhibit a lower error floor than that of the conven-tional MH method, which is often found at the high SNR .1
Fig. 2 shows the HS-aided MUD in the context of our iterativereceiver, which substantially mitigates the detrimental effect ofthe high correlations of random spreading sequences. It suggeststhat the performance improves upon increasing the number ofimprovisations and supporting an almost unprecedented systemload as high as is possible for our HS-aided MUD, wherethe performance is only about away from thesingle-user bound. At this system load, FF eval-uations were used instead of . Although it ispossible that a lower number of IDD iterations is needed when
1Proven convergence of the HS algorithm cannot be ensured, but experimentalevidence shows that the distribution of the bit error ratio tends to a Dirac-likefunction upon increasing the affordable complexity.
920 IEEE SIGNAL PROCESSING LETTERS, VOL. 16, NO. 10, OCTOBER 2009
Fig. 3. Extrinsic mutual information discrepancy between the HS-aided MUDand the optimum Bayesian MUD as a function of a priori mutual information,where we set , and . The results are normalisedto the maximum mutual information value observed for the perfect a priori in-formation employing the optimum MUD.
employing the optimum full-search-based MUD, when we takeinto account the number of iterations, the total number of FFevaluations for the HS-aided MUD becomes
, which is still only a fraction of even for a singleiteration, when the optimum MUD is employed. We also notethat the HS parameters were kept the same for both and
, which implies that proposed HS-aided algorithm is ca-pable of achieving a near-single-user performance without theexcessive complexity of the optimum detector and that withinlimits, the complexity of the algorithm is reasonably indepen-dent of .
Fig. 3 shows the extrinsic mutual information discrepancy be-tween the HS-aided MUD and the optimum Bayesian MUD of(1) as a function of the a priori mutual information, where we set
, and . The results are normalisedto the maximum mutual information value observed for the per-fect a priori information of employing the optimumMUD. It can be seen that by increasing the number of improvi-sations , our HS-aided MUD becomes capable of approachingthe performance of the optimum full-search-based MUD algo-rithm and the discrepancy becomes smaller upon increasing theamount of a priori mutual information, namely upon increasingthe number of iterations within the IDD.
In comparison to GA , which require the tuning of a rangeof parameters, as long as the initial HMM size of is suf-ficiently large, our designed HS-aided algorithm requires thetuning of a single parameter, namely of the number of impro-visations . This property is acquired as a benefit of the mar-ginal APP-based pitch adjustment step, which equips it with thecapability of attaining convergence from a Bayesian inferencepoint of view . At the same time, the entire search space is
visited by the HS MUD with the aid of the randomly generatedbase vector from the set of rather diverse candidate vectorsat each improvisation.2
The marginal APP-based pitch adjustment probability maybe compared to the classic Gibbs sampling of the MH method. In contrast to the successive correlated sample vectors gen-erated by Gibbs sampling, our proposed HS algorithm does notform a correlated Markov Chain...