IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 20, NO. 2, FEBRUARY 2002 463 Genetically Modified Multiuser Detection for Code Division Multiple Access Systems Saied Abedi, Associate Member, IEEE, and Rahim Tafazolli Abstract—The problem of multiple access interference (MAI) and intersymbol interference (ISI) suppression in code division multiple access (CDMA) systems is considered. By combining the theory of multiuser detection (MUD) and evolutionary computation, a hybrid genetic engine is proposed, suitable for detection of CDMA signals in presence of MAI and ISI. The proposed hybrid detector structure can be extended to most of multiuser detectors, used as the base detector within the structure. Using random selection, mutation, and crossover operators and unique chromosome structure, the genetic algorithm evolves the base detector to a group of more efficient detectors in terms of bit-error rate performance. First a new packet-level genetic MUD technique, using conventional single user detector as the base detector, for asynchronous CDMA (ACDMA) with negligible ISI is proposed. Then the signal-subspace-based minimum mean square error detector is chosen as a base detector and wrapped inside the hybrid genetic engine to evolve to a better structure to nearly eliminate both ISI and MAI. The novelty of the proposed structure is the way the deterministic closed-form solution of the base detector is mapped to a genetic engine resulting to a group of more efficient and adaptive detectors. Index Terms—Code division multiple access, genetic algorithms, intersymbol interference, multiple access interference, multiuser detection. I. INTRODUCTION F OR A nonorthogonal code division multiple access (CDMA) system the multiple access interference (MAI) is the major factor that limits the system performance and capacity. MAI between data symbols of different users arises due to the poor cross-correlation properties of signatures of different users. In asynchronous CDMA (ACDMA), the users transmit the signal independently. Therefore, the arrival time of the transmitted signals of different users are different at the receiver. Because of their relative time delays the cross corre- lation between signals of different users is nonzero effectively increasing the level of MAI in ACDMA. The conventional detector (CD) [14] consists of a bank of matched filters (MFs), which detects the signal of a user and treats other users’ signal as additive white Gaussian noise (AWGN). Verdu [12], [18], [19] has shown that the computational complexity of optimum multiuser detector (OMD) increases exponentially when the number of users is Manuscript received November 27, 2000; revised April 18, 2001. S. Abedi is with the Fujitsu Laboratories of Europe Ltd. (FLE), Hayes Park Central, Hayes, Middlesex, UB4 8FE, U.K. (e-mail: S.Abedi@fujitsu.co.uk). R. Tafazolli is with the Center for Communication Systems Research, University of Surrey, Guildford, Surrey GU2 7XH, U.K. (e-mail: R.Tafa- zolli@eim.surrey.ac.uk). Publisher Item Identifier S 0733-8716(02)00991-5. increased. Therefore, in the synchronous and the ACDMA, optimum multiuser detection (MUD) is a nondeterministic polynomial time (NP)-complete problem and implementation of OMD becomes impractical. Other suboptimum techniques need to be found, for a better suppression of MAI compared with the CD whilst requiring lower computational complexity compared with that required for the OMD. Several of such suboptimum techniques have been proposed in the literature such as interference suppression techniques in [13], [14] and multistage detector (MSD) for ACDMA proposed by Varanasi and Aazhang [17]. Kechriotis and Manolakos [10], [11] pro- posed Hopefield neural network implementation of the optimal CDMA multiuser detector. Honig et al. [8] also introduced a blind adaptive interference suppression technique. Finite memory-length linear MUD is another technique proposed by Juntti and Aazhang [9]. For high rate CDMA system, where the data bit duration is shorter than the multipath delays, intersymbol interference (ISI) becomes another performance limiting factor [20]. The signal-subspace-based methods for estimation of parameters and blind MUD in the presence of both ISI and MAI have been reported in [3], [16], and [20]. The MUD problem in CDMA can be realized as a search problem in a noisy space of possible solutions. Genetic algo- rithm (GA) is directed random search techniques, inspired by the mechanics of natural selection. GA and hybrid techniques based on GA have been used as an optimization tool in many fields of engineering. This paper shows the feasibility of GA as a powerful solution for MUD problem. We are not just looking for the best combination of data bits at the output of the re- ceiver. The aim is to improve the structure of detector (base de- tector), which results to better estimation of the data bits. For example, in the linear multiuser detector setup, a filter is de- fined by a number of taps, which help to estimate the trans- mitted information symbols. In the proposed structure, GA is used to optimize these tap coefficients through a number of it- erations (generations). As GA requires the process of initializa- tion, the first estimation of the tap coefficients performed by the base detector, is used as an initial configuration that forms the first generation of GA. In a way, the multiuser detector struc- ture called here base detector is wrapped in the GA structure. During the subsequent generations, these initial set of coeffi- cients are improved based on a predefined objective function. In this paper, two different detectors, i.e., CD and signal-subspace- based minimum mean square error (MMSE) were chosen as the base detectors. Their equivalent structures with GA are called packet-level asynchronous genetic multiuser detector (AGMD) and hybrid genetic multiuser detector (HGMD), respectively. The proposed AGMD and HGMD were investigated in different 0733–8716/02$17.00 © 2002 IEEE