1616 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 16, NO. 6, NOVEMBER 2005 Using Adaline Neural Network for Performance Improvement of Smart Antennas in TDD Wireless Communications Adnan Kavak, Halil Yigit, and H. Metin Ertunc Abstract—In time-division-duplex (TDD) mode wireless com- munications, downlink beamforming performance of a smart antenna system at the base station can be degraded due to vari- ation of spatial signature vectors corresponding to mobile users especially in fast fading scenarios. To mitigate this, downlink beams must be controlled by properly adjusting their weight vectors in response to changing propagation dynamics. This can be achieved by modeling the spatial signature vectors in the uplink period and then predicting them to be used as beamforming weight vectors for the new mobile position in the downlink transmission period. We show that ADAptive LInear NEuron (ADALINE) network modeling based prediction of spatial signatures pro- vides certain level of performance improvement compared to conventional beamforming method that employs spatial signature obtained in previous uplink interval. We compare the perfor- mance of ADALINE with autoregressive (AR) modeling based predictions under varying channel propagation (mobile speed, multipath angle spread, and number of multipaths), and filter order/delay conditions. ADALINE modeling outperforms AR modeling in terms of downlink SNR improvement and relative error improvement especially under high mobile speeds, i.e., . Index Terms—Adaline network, autoregressive model, beam- forming, linear prediction, smart antennas, vector channel. I. INTRODUCTION S MART antenna systems are proven to provide significant capacity increase and performance enhancement at the base station of wireless communications [1]–[3]. Spatial signature vector or channel vector describes the propagation character- istics of the signals present at an antenna array of a smart an- tenna system (SAS). For downlink beamforming in time-divi- sion-duplex (TDD) systems where uplink and downlink share the same carrier frequency, the SAS conventionally uses the last known spatial signature estimated during the uplink interval as the weight vector. This conventional approach, known as spatial signature based beamforming [4], [5], performs well as long as channel characteristics are almost the same between consecu- tive time intervals. When the mobile terminal is stationary or moving at low speed, it has been demonstrated that spatial sig- nature variations are not significant, and that direction of arrivals Manuscript received March 29, 2004; revised January 6, 2005. A. Kavak is with the Department of Computer Engineering, Kocaeli Univer- sity, 41040, Izmit, Turkey (e-mail: akavak@kou.edu.tr). H. Yigit is with the Department of Electronics and Computer Education, Ko- caeli University, 41380, Izmit, Turkey (e-mail: halilyigit@kou.edu.tr). H. M. Ertunc is with the Department of Mechatronics Engineering, Kocaeli University, 41040, Izmit, Turkey (e-mail: hmertunc@kou.edu.tr). Digital Object Identifier 10.1109/TNN.2005.857947 (DOAs) are almost unchanged [6]. However, if the mobile user moves at relatively high speed, spatial signature vectors change rapidly due to fast fading effects induced by Doppler shift at each multipath. Under such circumstances, using the spatial sig- nature of the previous uplink time slot as the downlink weight vector for the new mobile position results in performance degra- dation. This can be avoided by accurately updating transmission weight vectors to control downlink beams as depicted in Fig. 1. Within small movement of the mobile, spatial signatures are as- sumed to vary due only to Doppler shifts induced on multipaths so that each element of the spatial signature vector can be mod- eled as the sum of sinusoids [7]–[10]. One way of overcoming the fast fading effect and improving downlink performance of the SAS is to predict spatial signatures based on their autoregressive (AR) modeling as demonstrated in [10] where Arredondo et al. have used the fact that model coeffi- cients are set during the uplink period and assumed unchanged for a number of downlink periods. Recently, neural networks have been considered for the applications in wireless communi- cations [11], [12], and proposed to perform computational tasks required in antenna array processing [13] such as array pat- tern synthesis [14], [15], direction finding [16], [17], multiple source tracking [18], and beamforming [19], [20]. Our motiva- tion in this work is to explore a neural network based prediction model that continuously updates prediction filter coefficients as new spatial signature samples collected so that we can obtain improved performance over AR modeling. In other words, we want to study whether we can improve conventional AR predic- tion performance by applying a linear class of neural network that responds to changes in its environment as it is operating. An ADAptive LInear NEuron (ADALINE) network, much like the perceptron, can only solve linearly separable problems [21]. Nevertheless, the ADALINE is quite suitable for practical ap- plications such as prediction [22] or noise cancellation owing to its simple structure based on linear activation function and the least mean squares (LMS) learning rule which is much more powerful than the perceptron learning rule. The problem studied in this paper has autoregressive structure under certain assump- tions as mentioned above, and thus we focus on comparing AR with only ADALINE network rather than with other nonlinear class of neural network architectures for the prediction of spa- tial signatures. We evaluate the performance of the ADALINE and AR prediction models under varying mobile speed , multipath angle spread , number of multipaths , and prediction filter order/delay conditions. The performance measures 1045-9227/$20.00 © 2005 IEEE