6. Conclusions and Future Scope of Research
6.1 Summary of the thesis
The work presented in this thesis starts with comparison of performances of some classical linear and nonlinear MUDs applied to SDMA–OFDM system. Further, the proposed OTs aided MER and NN based MUDs are tested for their suitability under various user loads. Comparison with optimal ML and linear MMSE detectors in terms of system performance and complexity analysis has been a major attempt to prove the efficacy of the proposed MUD schemes. Detailed achievements of the research are presented chapter wise as below.
Chapter–2 of the thesis provides details of the SDMA–OFDM system model a brief comparison between all classical MUD schemes. The performance the ML detector is found to be optimal at the cost of dramatically increased complexity, especially in context of a high number of users and higher order modulation schemes. By contrast, the linear ZF and MMSE exhibit low complexity while suffering from performance loss. Therefore, several other nonlinear MUDs are studied to make a tradeoff between performance and complexity. However, all these techniques fail to detect users in overload scenario. In such a condition, it is found that the CG aided MBER MUD is a better alternative. But the CG algorithm requires initial weights and differentiable cost functions, hence a new research direction to overcome these limitations is thought.
Chapter–3 is entirely devoted to discuss one of the major contributions of the work undertaken relating to the design of novel MUD schemes, that is the design of OTs aided MER MUD schemes. Metaheuristic OTs like AGA, ADEA and IWO algorithms are considered for MER weight optimization. The extensive simulation study shows that the proposed MUDs are superior in performance over MMSE and require less computational complexity compared to ML detector. Further it is also observed that, the performance of these OTs is greatly influenced by selection of control parameters. The performance AGA is influenced by Pc, Pm, Gg and Ng while ADEA is influenced by Cp, F, Gd and Nd. Similarly,
the IWO MER MUDs is subjective to the parameters like NI, Imax, Smax, σmax, σmin and m.
Selecting right combination of these control parameters yields a better performance. Hence, development of suitable methods for selection of control parameters has been included in this chapter. Among all the discussed OTs, the IWO algorithm is found to be better as it allows all of the individuals to participate in the reproduction process. Sometimes, it may be also possible that the individuals with the lower fitness may carry more useful information compared to the fitter individuals. Hence, this algorithm, gives a chance to the less fit plants to reproduce and if the seeds produced by them have good finesses in the colony, they can survive. Fitter individuals produce more seeds than less fit individuals, which improves the convergence and performance over the AGA and ADEA algorithms.
Chapter–4 of the thesis provides another major contribution of this work through the design of NN based MUD schemes. As the NN models are highly nonlinear classifiers, these are well suited for detection of the multiuser signals, which are corrupted by nonlinear channel distortion. In the NN family, the real valued MLP and RBF models are used for multiuser signal detection when all the users are transmitting BPSK signals. However, in several communication systems, the available signals are in complex form when the system is communicating higher order signals like M–QAM. In this case, the classical real valued NN models cannot be applied directly because it requires real valued inputs and activation functions. In order to extend the real valued NN models to complex signals, the activation function and training algorithms should be redefined. In the CMLP model, the sigmoid function is divided in to two components for responding to the real and imaginary portions of the input signals individually. For CRBF model, the ‘sech’ activation function is preferred over Gaussian function, because the Gaussian function always results real valued response. The conventional BP and GD algorithms used for updating the real valued NNs cannot be directly applied for complex valued NN models. Hence, suitable modifications have been
incorporated to these algorithms. The extensive simulation study shows that the proposed OTs aided MER and NN based MUDs are superior in performance compared to MMSE and OTs aided MER MUDs. The performance improvement of these MUDs over the MMSE in full load (L = P = 4) condition at 10–4 BER is given in Table 6.1.
Table 6.1: Performance comparisons of OTs aided MER and NN MUDs in terms of Eb/No gain (in dBs)
Channel
AGA ADEA IWO Real Valued
NNs
Complex Valued NNs
MBER MSER MBER MSER MBER MSER MLP RBF CMLP CRBF MIMO
Rayleigh 4.5 5.1 4.55 5.1 4.6 5.3 4.6 5.1 6.7 9
SUI 4.1 4.2 4.2 4.25 4.2 4.1 7.1 9.05 4.7 6.8
SWATM 6.8 7.6 6.8 7.5 6.9 7.7 9.8 13.5 8.3 11.6
In the complexity analysis, the percentage of complexity require for OTs aided MER MUDs in terms of computational operations is compared with the computationally exhaustive ML detector and presented in Table 6.2 when L = 6 and P = 4.
Table 6.2: Complexity comparisons of OTs aided MER and NN MUDs in terms of computational operations
MUD Technique Percentage of complexity
AGA MBER 15.43 MSER 19.145 ADEA MBER 13.87 MSER 17.51 IWO MBER 12.95 MSER 13.205 Real Valued NNS MLP 13.87 RBF 8.985 Complex Valued NNS MLP 17.51 RBF 16.49
Chapter–5 illustrates the ability of the proposed MUDs while reconstructing both gray scale and colored images, which are transmitted by multiple users simultaneously through wireless channels. A comparative analysis among all proposed and the classical MUD schemes based on some statistical parameters like Bias, SDD, RMSE, CC and PSNR. The SPIHT algorithm is considered for image compression and coding as it performs better than
the classical EZW technique. From simulation results, it is observed that the ML detector produces the best restored image, while the MMSE performs the worst. The reconstructed images of the OTs aided MER MUDs are usable images but still require improvement. Finally, both the visual and the statistical parameter comparisons of the NN MUDs are found to be qualitatively better than the MMSE and OTs aided MER detectors.
Finally, the general inference derived from the extensive simulation study is that the proposed NN based MUDs, especially the RBF and CRBF detectors, are efficient in terms of BER performance, faster convergence and computational complexity. Besides that, the NN base detectors has additional complexity gain over classical MMSE, ML and proposed OTs aided MER detectors, because these detectors do channel approximation and signal detection simultaneously. As the ML detector is a highly complex one, the RBF and CRBF MUDs are found to be the suitable alternative for MUD in the SDMA–OFDM system.