International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.

5, 2011

Fuzzy-based hybrid MIMO Channel Estimator with variable Membership and Inference Rule Set
Kandarpa Kumar Sarma, Abhijit Mitra
Department of Electronics and Electrical Engineering, Indian Institute of Technology Guwahati Guwahati-781039, Assam, India. Email: s.kandarpa@iitg.ernet.in.

Abstract
Performance of a fuzzy-neural (FN) estimator of Multi-Input-Multi-Output (MIMO) wireless channels depends on the membership and inference rule sets adopted. This pertains to the fact that these two factors decide the ability of the fuzzy-estimator to capture the subtle variations in the input signal patterns and provide corresponding responses with optimum performance. Also, the precision and processing speed of the system is related to these two factors. Larger is the input norm set and inference states, better shall be the precision at the cost of corresponding increase in timecomplexity and computational requirements. An optimal choice, therefore, is critical in determining the ability of the fuzzy system while applying it for MIMO channel estimation. Two different sets of input norms and inference rules are formulated to ascertain the suitability of these choices. Experimental results show that a set formulated with seven linguistic hedges and six inference states resolves the trade-off between processing speed and precision generated while tackling a 4-path MIMO set-up with time-varying signals.

Keywords: MIMO, Estimation, Fuzzy, Fuzzy-Neural. 1. Introduction

Channel modeling is an area full of uncertainty. It is more so with Multi-Input-Multi-Output (MIMO) wireless channels which show time-dependent behaviour and interferences between signal components. A host of methods have been proposed which have successfully dealt with such a stochastic medium of wave propagation. Artificial Neural Network (ANN) and Fuzzy system (FS), as important components of soft-computation, have received attention in diverse fields that have constituent uncertainty. This is primarily due to the fact that such tools can learn from the surrounding environment and use the knowledge thus acquired for subsequent processing including prediction and estimation. Applications of ANN for estimation of MIMO wireless channels have been highlighted in [1] - [3]. The work in [3] is related to a class of Recurrent Neural Network (RNN) s applied for MIMO channel estimation which provides satisfactory performance. Yet the opportunity always exists to explore other options including combinations of ANN and FS to formulate hybrid systems which can model the uncertainty related with MIMO channels with further improvement. The objective is to reduce system and time complexity and enhance performance in terms of lower bit error rate (BER) values and processing speeds. This work is related to the formulation of membership function and the inference rule sets for application with a fuzzy-based MIMO channel estimator implemented to deal with the inherent timevarying nature. The success of fuzzy-based systems depends on the membership function and the inference rule set adopted. This work shows how variable performance can be achieved with different membership and inference rule sets while designing a fuzzy-based MIMO channel estimator.

November Issue

Page 9 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Fig. 1: Components of an FNS

Fig.2. System model of fuzzy system based MIMO channel estimation

2. Fuzzy Neural System (FNS) MIMO Channel Estimator- Design Consideration

Fuzzy systems are applicable where sufficient expert knowledge about a process in available [5][6] while ANN is comfortable with cases that have sufficient process data [4]. Therefore, while ANNs possess numeric-quantitative capability, fuzzy systems exhibit symbolicqualitative capacity [4] [5] [6]. Thus, hybrid systems formed by combinations of fuzzy and ANN methods have adaptability, parallelism, non-linear processing, robustness and learning in data rich environment and modeling uncertainty. These attributes enable integration of fuzzy and neural systems as either neuro-fuzzy sys-tem (NFS) or fuzzy-neural system (FNS) depending upon the requirement. The composite and hybrid characteristics demonstrated by NFS or FNS enables them to acquire numeric-qualitative expert level decision making capacity and demonstrate greater adaptability and robustness while handling unknown process or situations. These attributes of the fuzzy based systems make them fit for applications in wireless communication. A detailed account of some applications of fuzzy systems in communication is provided in [6]. The work cited in [7] proposes an adaptive NF inference system (ANFIS) for channel estimation in orthogonal frequency division multiplexing (OFDM) systems. Fuzzy systems (Fig 1) either in NFS or FNS forms have their advantages and disadvantages while designing and configuring them for specific applications. FNS is more oriented towards real-world applications hence are considered for MIMO channel estimation. This work provides some experimental results showing the superiority of a FNS MIMO estimator over its NFS counterpart. The FNS provides a fuzzy input to output mapping unlike the NFS which relates a crisp input to a corresponding output. As a result, the FNS has greater ability to keep track of finer contextual variations unlike the NFS. Also, FNS allows automation of generation of fuzzy rules and has the ability to perform combined learning of numerical data as well as expert-knowledge expressed

Table 1: Linguistic steps used to condition the inputs

Case 1 2 3 4 5 6 7 State Negative Large Negative Medium Negative Small Close to Zero Positive Large Positive Medium Positive Small Notation NL NM NS CZ PL PM PS

November Issue

Page 10 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Table 2: Set of norms used to condition the inputs Case Set of Norms
f ( x) NL, 0.66 x 0.99; NM , 0.33 x 0.66; NS , 0 x 0.33; CS , x 0; (2)

PL, 0.66 x 0.99; PM , 0.33 x 0.66; PS , 0 x 0.33;

f ( x) Minus 9, 0.88 x 0.99 Minus8, 0.77 x 0.88 Minus 7, 0.66 x 0.77 Minus 6, 0.55 x 0.66 Minus 5, 0.44 x 0.55 Minus 4, 0.33 x 0.44 Minus 3, 0.22 x 0.33 Minus 2, 0.11 x 0.22 Minus1, 0.01 x 0.11 Z, x 0 (3) Plus1, 0.1 x 0.22 Plus 2, 0.22 x 0.33 Plus 3, 0.33 x 0.44 Plus 4, 0.44 x 0.55 Plus 5, 0.55 x 0.66 Plus 6, 0.66 x 0.77 Plus 7, 0.77 x 0.88 Plus 8, 0.88 x 0.99

as fuzzy if-thenelse rules. Moreover, FNS have smaller networks and faster process times compared to ANNs and NFS [8], hence are more suitable for applications like adaptive receiver design high data rate mobile communication. Figure 2 shows the system model of a FN MIMO channel estimator. The system is formed by a 4 x 4 MIMO set-up transmitting signal symbols which are used to train the FNS system to perform channel estimation. It can also perform symbol recovery as the system is made familiar with the transmission content of the signals.

3. Membership function selection and implementation

The working of the FNS system is related to the fuzzification and membership function generation process.
3.1. Conditioning the Inputs The MIMO channel input-output relationship is expressed as
xn H ns(n) v(n) (1)

where x is an M x 1 vector with xi (n), i =1, 2..., M as the elements, s(n) representing the signal symbols, v(n) denoting additive background noise with H n being the M x N channel matrix which can include Rayleigh, Rician and Nakagami multipath fading. The respective channel sets are generated and the signals are mixed following time and frequency domain considerations. These samples have SNR variation between -10 to 10 dB, are normalized and are confined to a few linguistic hedges [5] as shown in Table 1. The inputs are also constrained by the set of norms shown in Table 2.

November Issue

Page 11 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Fig. 4: Fuzzy recurrent neuron

Fig. 3: Different layers of the FNS MIMO estimator

Two separate sets of input conditioning norms are taken to ascertain the effect of variable size of rules and segments during the fuzzification process. The inputs are divided into in -phase and quadrature components to allow the FNS to learn the indiv idual signal segments separately. Fig 3 shows the different layers of the FNS MIMO estimator. The first is the input layer which passes on the samples to the fuzzification layer. Each node in this layer is formed by a fuzzy recurrent neuron (FRN) (Fig 4). The inputs (X kj), weights (W j) and outputs are all fuzzified [5] and the training process is carried out by a fuzzified Decoupled Extended Kalman Filter (DEKF) algorithm. These neurons have the capacity to deal with time -varying inputs and track the subtle variation in magnitudes. Each neuron in the fuzzification layer represents a fuzzy term of the input linguistic variables of which two different sets have been taken as shown by (2), (3) in Table 2. All the decision making regarding the classification and judging the difference between actual and desired response is carried out by the inference layer where multiple sets are implemented to check the effectiveness and performance derived. The response of the system is de-fuzzified by the next layer which is evaluated as per the decision risk functional norm. The learning process makes bi-directional flow of the samples from the output to the preceding layers which carry out an adaptive modification of the connectionist links between the neurons. 3.2. Fuzzification and Membership function generation The fuzzy sets of the respective inputs are generated following two ways: In the first method, a Bell-shaped membership function is used to generate the member-grades of each input. A one-to-one correspondence is established between input and fuzzy sets. This association is taken to train a Multi Layer Perceptron (MLP) to generate the membership function. At the end of the training, the MLP becomes an automatic membership generator (Fig. 5). The second method is to use a Self Organizing Map (SOM) to create clusters for each of the samples forming the input matrix. This clustering is used to train a MLP to act as an automated membership generator (Fig. 6).

4. Inference Rule Generation

The inference rule for the FNS system to perform MIMO channel estimation is fixed by considering transmit-receive conditions. The set-up is constituted by a FNS estimator which is designed to handle the inputs and rely on an inference mechanism of fast decision making. Let

November Issue

Page 12 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Fig. 5: Membership grade generation using Bell-shape function and trained MLP

Fig. 6: Membership grade generation using SOM and trained MLP

Table 3: Inference rule for training FNS based MIMO channel estimator
Set yi
Max. Min.

yT Inference

h Inference
0.99 hii 0.1 hij

1 0.9 1 0.7 0.5 0.3 0.3 0.99 0.87 0.76 0.65 2 0.54 0.43 0.22 0.1 0.08

0.1 0.1 0.3 0.5 0.5 0.7 0.88 0.77 0.66 0.55 0.44 0.33 0.11 0.09

yT 0.9 y i hii 0.1 yi hij

yT 0.7 yi hii 0.3 yi hij
yT 0.5 y i hii 0.5 yi hij

yT 0.99 y i hii 0.01 yi hij

0.9 hii 0.1 hij

0.7 hii 0.3 hij

0.5 hii 0.5 hij

yT 0.3 y i hii 0.7 yi hij

yT 0.99 y i hii 0.01 yi hij

0.3 hii 0.7 hij

0.99 hii 0.1 hij

0.87 hii 0.13 hij
0.65 hii 0.35 hij
0.43 hii 0.57 hij

yT 0.87 y i hii 0.13 yi hij

yT 0.76 y i hii 0.24 yi hij
yT 0.54 y i hii 0.44 yi hij

yT 0.65 y i hii 0.35 yi hij

0.76 hii 0.24 hij 0.54 hii 0.44 hij

0.22 hii 0.78 hij
0.1 hii 0.9 hij

yT 0.22 y i hii 0.78 yi hij yT 0.08 y i hii 0.92 yi hij

yT 0.1 y i hii 0.9 yi hij

yT 0.43 y i hii 0.57 yi hij

0.08 hii 0.92 hij

[h 11, ., h 21,, h 31,, h 41,, h 44] be known estimates of the channel coefficients. Similarly, let after a few training sessions the estimate of the channel coe fficients generated by the FNS be [h11,..., h21,..., h31,..., h41,..., h44 ]. So the inference engine should minimize a risk functional given as
d (hi , j , hi , j ) where d (.) represents a Euclidean distance between expected and actual estimates of

the channel coefficients. For a given estimation iteration, h ii represents a significant coefficient while h ij denotes a co-channel interference (CCI). Hence, the estimation process minimizes the effects of CCI and prioritizes the significant components to mak e the procedure effective. The training is made resilient enough to enable the system derive discrimination boundaries between significant and CCI coefficients. The training set, therefore, contains significant and CCI coefficients with unique correspondence with decision logic such that the mean square error (MSE) is low and receiver end and h be the channel coefficient inference. By following these notations the inference rule formulated for the FNS, shown as two separate sets, is su mmarized in Table 3. Appropriate encoding schemes are generated for implementing the inference rules. The inference engine is designed using Time Delay Fully Recurrent Neural Network (TDFRNN)

November Issue

Page 13 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Fig. 7: Comparative % performance of membership grade generation methods blocks with split activation for in-phase and quadrature components. In this form these blocks are called Complex TDFRNN (CTDFRNN) which is fuzzified by the use of a FRN. The fuzzified CTDFRNN (FCTDFRNN) blocks have excellent capacity to dea l with time variation observed in MIMO transmissions. The FRN receives a fuzzy input and is related to an inference which needs to be defuzzified.

5. Experimental Results and Discussion

A comparative depiction of the performance of the above two membership grade generation methods is shown in Figure 7. While the SOM-MLP method is less accurate on an average by about 4% but it is faster by atleast 21%, hence it is more acceptable for speed critical applications but for situations where precision is of greater importance the Bell-MLP method is preferable. This method of extracting membership grades of the inputs works well with static and slowly varying signal samples. For timevarying cases including fast fading, a better option is the use of TDFRNN blocks in place of the MLP. The primary motivation is that the RNN has the capacity to deal with time varying inputs due to the presence of atleast one feed-backward loop for which learning accumulates over an extended period of time. Moreover, the RNN is faster compared to the MLP, hence takes less process time, in this case it shows atleast 23% improvement. Therefore the functional membership grade block is constituted by a SOM and a TDFRNN block. The TDFRNN blocks uses split activation and are trained with DEKF algorithm. Segregated real and imaginary inputs are applied to different TRFRNN blocks with the output going to respective sections.

The samples are accumulated for a 4 x 4 MIMO OFDM set-up for three different AWGN values viz.-3dB, 1dB and 3dB for the generic, Rayleigh and Rician faded channels. The testing includes a range of signal conditions with SNR values ranging from -10 to 10 dB. The testing carried out with inputs from the receiver side calculates channel coefficients and compares them to the theoretically generated values for a frequency range of 0.2 Ghz to 8 GHz. Some of the values taken for training the system to deal with Rayleigh and Rician channels are generated using Clarke-Gans model. In case
of the FTDFRNN structures, split-activation technique is adopted to train real and imaginary sections separately with specific signal segments. The FTDFRNN structures are trained with fuzzy DEKF algorithm. A training window of n seconds is given to the two FTDFRNN structures during which several estimated of the signal samples are generated. The performance of the system is dependent on the inference rule set adopted for carrying out the estimation process. A standard set of six rules gives optimal performance but experiments are carried out to see the effect of variation of the inference stage. Table 4 summarizes the results obtained. The implementation of the inference engine using six set of rule shows a dependence on the network structure adopted for the purpose. Table 5 summarizes the variation in p erformance of the FNS with change in the network structure adopted for the implementation of the inference engine. The inference network structure is considered to be formed by an input, two hidden and one output layer. Four different FNS structures are used to implement the inference rules. While all the four networks using six inference rules generate a MSE convergence between 0.24 x 10 -5 and 0.33 x 10 -5 , the value comes down significantly to a range of 0.03 x 10 -5 to 0.05 x 10-5

November Issue

Page 14 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Table 4: Effect of variation of number of inference rules No. of MSE Precision Epochs inference attained in % rules x 10-5 4 0.5 93 34 6 0.28 97 31 8 0.15 97 43 9 0.09 97 58 10 0.08 97 75 12 0.03 98 91

Table 5: Effect in performance due to variation in network structure adopted for implementation of inference engine. Network No. MSE Epochs Precisio size of x 10-5 n (%) rules 6 0.3 34 96.5 20-24-12-4 (N1) 12 0.05 105 97.3 6 0.26 33 96.8 20-30-15-4 (N2) 12 0.04 99 97 6 0.33 39 96 20-40-20-4 (N3) 12 0.04 104 97 6 0.24 44 96 20-50-25-4 (N4) 12 0.03 102 97

Fig. 8: MSE convergence plot of four FNS networks upto 50 epochs with twelve inference rules. It indicates that the MSE convergence rate falls with more inference stages. With more inference rules, the networks learn better and approach the level of optimality with greater closeness. It amounts to an improvement between 84 to 87% in MSE convergence rates. But this happens at the cost of greater processing time. Number of epochs increases by about 1.32 to 1.7 times when twelve inference rules are used compared to case when the system is designed with a set of six such sets. But this increase in processing time and lowering of the MSE values result in an improvement of precision marginally between 1.04 to 1.5%. This marginal improvement is further compounded by increase in design complexity. Hence, the six inference rule format is adopted to carry out the FNS based MIMO channel estimation. The four networks are trained separately and the MSE convergence plots are shown in Figure 8. The network named N4 with a structure of 20-50-25-4 gives the best performance. This happens at the cost of greater processing time. Number of epochs increases by about 1.32 to 1.7 times when twelve inference rules are used compared to case when the system is designed with a set of six such sets. But this increase in processing time and lowering of the MSE values result in an improvement of precision marginally between 1.04 to 1.5%. This marginal improvement is further compounded by increase in design complexity. Hence, the six inference rule format is adopted to carry out the FNS based MIMO channel estimation. The four networks are trained separately and the MSE convergence plots are shown in Fig. 8.

November Issue

Page 15 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

Table 6: Suitability of NFS or FNS for MIMO estimator design Parameter NFS Case FNS
VOIP Data Size 128bit blocks, 50 no.s Digital Video, 1000 no.s of 6400bit blocks
Network Size Precision (%) MSE Inference rules Epochs % time Network Size Precision (%) MSE Inference rules Epochs % time

20-50-59-61-25-4 95.2 0.8 x 10-5 6 to10 42 68 20-50-59-61-25-4 95.1 0.88 x 10-5 6 to10 584 73

20-50-25-4 96.1 0.24 x 10-5 6 31 61 20-50-25-4 96 0.26 x 10-5 6 412 65

The network named N4 with a structure of 20-50-25-4 gives the best performance. MSE value converges, on an average to about 0.24 x 10-5 during the stipulated training slots below 50 epochs with the data set taken. This network has an input layer of 20 FRNs in the input layer. Similarly, there are 50 and 25 FRNs in the two hidden layers. The output layer has four FRNs as it needs to retain only four sets of data for a 4 x 4 MIMO set-up designed for the purpose. The significant part of the data set are retained, interpretation derived after de-fuzzification and BER values calculated. A set of experiments are also perform to see the effectiveness of the fuzzy MIMO estimator in NFS and FNS forms. The two configurations are tested for a VOIP based voice and digital video broadcast using a 4 x 4 MIMO wireless set-up infested with Rayleigh fading. The VOIP voice grade transmissions involve 128-bit size OFDM data blocks of around fifty while the digital video broadcast involves over a thousand trails. Similarly there is enhancement of BER performance as well over the entire SNR spectrum for all the multipath fading channels considered. The fuzzy-based methods clearly provide advantages of faster processing time, lower BERs and better precision while carrying out symbol recovery. In this performance enhancement, the membership generation and the inference mechanism plays a decisive role. The choice of the specific membership norms and inference rule set is critical in resolving the speed-precision stand-off.

6. Conclusion
In this work, we have shown the membership generation and the inference rule formulation of a FNS configured for MIMO channel estimation. The membership generation process is dependent on the method adopted for implementation. We here found that a SOM-MLP based membership generation method is a compromise between processing speed and precision generated. Similarly a six rule inference block formulated using a 20-30-15-4 network structure of FRNs give the best performance in terms of precision and processing speed. Compared to statistical and ANN approaches fuzzy methods of MIMO channel estimation provide appreciable improvement in performance.

November Issue

Page 16 of 84

ISSN 2249 8923

International Transactions on Electrical, Electronics and Communication Engineering, Vol. 1, No.5, 2011

7. Reference:
[1] K. K. Sarma and A. Mitra: Estimation of MIMO Wireless Channels using Artificial Neural Networks chapter of Cross-Disciplinary Applications of Artificial Intelligence and Pattern Recognition: Advancing Technologies, IGI Global, 701 E. Chocolate Ave. Hershey, PA 17033, USA, 2011. K. K. Sarma and A. Mitra: A Class of Recurrent Neural Network (RNN) Architectures with SOM for Estimating MIMO Channels, Springer-Verlag ACC 2011 Proceedings, vol. III, pp. 512-521, 2011. K. K. Sarma and A. Mitra: Estimation of Multipath Fading Channel of MIMO-OFDM System using ANN, Chapter of Advanced Computing Applications Databases and Networks, Narosa Publishing House, New Delhi, June, 2011. S. Haykin: Neural Networks A Comprehensive Foundation, 2nd ed., Pearson Education, New Delhi, 2003. T. J. Ross: Fuzzy Logic with Engineering Applications, 2nd ed., Wiley India, New Delhi, 2008. S. Mitra and Y. Hayashi: Neuro-Fuzzy Rule Generation: Survey in Soft Computing Framework, IEEE Transactions on Neural Networks, vol. II, no. 3, pp. 748-768, 2000. M. Nuri, Seyman, N. Taspinar: Channel Estimation Based on Adaptive Neuro-Fuzzy Inference System in OFDM, IEICE Transactions on Communications, vol. E91.B, no. 7, pp. 2426-2430, 2010. R. A. Aliev, B. G. Guirimov, B. Fazlollahi and R. R. Aliev: Evolutionary algorithm-based learning of fuzzy neural networks. Part 2: Recurrent Fuzzy Neural Networks, Elsavier Journal of Fuzzy Sets and Systems, Vol. 160, pp. 2553-2566, 2009.

[2] [3]

[4] [5] [6] [7] [8]

Authors Profile

Kandarpa Kumar Sarma obtained his MTech degree in Signal Processing from Indian Institute of Technology Guwahati, Assam, India in 2005. Presently he is with the department of Electronics and Communication Technology, Gauhati University as an Assistant Professor, Guwahati, Assam, India. His areas of interest include Applications of Soft-computational Tools, Mobile Communication and Pattern Recognition. He is a member of IEEE (USA) and IETE (India) and has authored two books and several research papers.

Dr. Abhijit Mitra his PhD from IIT Kharagpur, India, in Electronics and Communication Engineering. He is with the department of Electronics and Electrical Engineering (erstwhile, department of ECE) at IIT Guwahati, India, since 2004, where currently he is an Associate Professor. His broad area of research and teaching is in signal processing and wireless communication with a special emphasis on low power implementation of the said systems. Dr. Mitra received several research related awards. He is a Fellow of IETE (India) and a Senior Member of IEEE (USA). He serves as an Associate Editor of Recent Patents on Electrical Engineering (Bentham Science, USA) since 2007 and an Associate Editor of Current Trends in Signal Processing (STM Journals, India) since 2011. He serves as an Associate Editor of Recent Patents on Electrical Engineering (Bentham Science, USA) since 2007. Previously, he has served as an Associate Editor of WASET journals, a reviewer of several IEEE, IEE and Elsevier journals and a program committee member of many national/international conferences.

November Issue

Page 17 of 84

ISSN 2249 8923