Least squares support vector machines for fixed-step and fixed-set CDMA power control

Power control is a critical component in CDMA cellular systems. Power control combats the “near-fat” effect by adjusting the transmit power of each mobile. This technique reduces the multiple access interference and if the system capacity is within the set limits the, desired signalto-interference ratios (SIRS) are achieved at all base stations. Research in the power control field covers centralized, distributed, and stochastic power control [1],[2],[3]. In most of the published research however, it is assumed that the uplink channel gain is known, and no limitations are placed on the word-length and update rate of the transmitted power control command. The IS-95 and cdma2000/3G systems have an 800 bps up/down power control command rate; the single bit power control command is thus sent to the mobile station every 1.25 milliseconds [4]. This limits the options with regards to power control systems, but the design reduces the problem to that of generating a fixed-step power control command. Standard power control systems implemented in cellular systems use signa-to-interference ratio (SIR) estimates, hit emor rates (BER), or frame error rates (FER). Many of the published power control algorithms assume that these estimates are available and accurate. The algorithms presented in this paper require only the set of eigenvalues from the sample covariance matrix of the received signal; the signal subspace dimension and power estimates are not required. The machine learning algorithm classifies the set of eigenvalues into a SIR set which then determines the power control command. Both binary and multiclass machine learning algorithms are developed to