Adaptive Antenna Arrays Using a Genetic Algorithm Randy L. Haupt The Pennsylvania State University Applied Research Laboratory P. O. Box 30 State College, PA 16804-0030 E-mail: haupt@ieee.org Phone: 814-865-7299 x210, Fax: 814-863-6239 Abstract—This paper presents the application of a genetic algorithm (GA) to adapting an antenna array's response in order to reject interference. Since the GA reduces the total power output of the array, constraints are used to prevent nulling of the desired signal received by the main beam. Constraints take the form of using only the least significant bits or a subset of the elements to perform the nulling. Examples demonstrate adaptive nulling using amplitude and phase, phase only, and amplitude only weights. I. INTRODUCTION Adaptive antennas enhance desired signal reception by placing nulls in the antenna pattern in the directions of the interference while minimally perturbing the main beam. The original adaptive antenna consisted of a highly directional reflector antenna with a small omni-directional antenna. This sidelobe canceller [1] adjusted the output of the omni-directional antenna until it cancelled the interference entering a sidelobe. Since the omni-directional antenna gain is small compared to the main beam of the high gain antenna, there was not much impact on the main beam. Adaptive arrays improved on the sidelobe canceller by adjusting the weights of an array to both receive the desired signal as well as cancel the sidelobe interference [2]. The problem with most adaptive algorithms is that they need a receiver at each element in order to get the amplitude and phase of the signals to form the covariance matrix. The receivers are expensive and require regular calibration. Another approach to adaptive nulling is to minimize the total output power of the array [3]. Since the total output power consists of both the desired signal and interference signals, some constraints are needed to insure that only the sidelobes are nulled and not the main beam. This paper shows how the constraints are implemented through using the least significant bits of the amplitude and phase weights or only a fraction of the elements in the array. As with a sidelobe canceller, the signal strength associated with the 1-4244-0166-6/06/$20.00 ©2006 IEEE. least significant bits is too small to create a null in the main beam. A genetic algorithm (GA) performs the adaptation by manipulating the weights until the total output power is minimized. Results have been previously reported for a computer model of phase-only adaptive nulling with linear arrays [4] and experimental results for amplitude and phase adaptive nulling with an eight element conformal array [5]. In both cases, GAs with small population sizes were used to create deep nulls in the antenna patterns. II. COST FUNCTION A GA massages the variables of a cost function until the output or cost is minimized. In this case, the cost function is a linear array with variable amplitude and/or phase weights, and the cost is the total output power. The linear array model of isotropic point sources is shown in Fig. 1. The elements in the array are often approximated by isotropic point sources which have a corresponding array factor given by N AF = ∑ wn e jk ( n −1) d cos φ (1) n =1 where wn = an e jδn k = 2π / wavelength d = spacing between elements φ = angle relative to x-axis N = number of array elements Controlling the weights modifies the main beam peak and nulls. Usually, the array weights are digital and have a finite number of settings. The problem with this cost function formulation is that the desired signal and the interfering signals are mixed together. Minimizing the output power will decrease the desired signal in addition to the interfering signals unless the desired signal is assumed to enter the main beam and the adaptive weights are constrained to small values that cannot place a null in the main beam. Fig. 2. Array factor at the main beam (φ=90ο) and a sidelobe (φ=42ο) when all combinations of 1 through 4 least significant bits out of a total of 6 bits in the phase shifters are tried. III. THE ADAPTIVE ALGORITHM Fig. 1. Diagram of an adaptive array controlled by a GA. If a six element array has uniform amplitude weights and six bit phase shifters, then the six bit correspond to the following six phase settings: bit 1 2 3 4 5 phase π π π π π 32 16 8 4 2 6 π (2) The least significant bits have small phase values that have correspondingly small effects on the main beam. Assume that the array is symmetric about its center and the center two elements have zero phase. The two phase shifters on the right side of the array have the negative phase of the two phase shifters on the left side of the array. Fig. 2 shows the array output in dB at the main beam ( φ = 90o ) and at a sidelobe ( φ = 45o ) as a function of phase shifter settings using up to four out of six bits. Note that at least four bits are needed to place a null in the sidelobe. If this were a low sidelobe array, then fewer bits may also place a null in the sidelobe. Also note that using up to 4 least significant bits will not significantly perturb the main beam. If five or six bits are used, then a null can be placed in the main beam. The cost function must be selected so that nulls can be placed in the sidelobes while nulls cannot be placed in the main beam. In this way, the desire signal will be minimally perturbed while the interfering signals will be nulled. As a result, this adaptive nulling configuration is simply minimizing the total array output power. Each chromosome in the GA population represents a few of the least significant bits of the phase and/or amplitude settings at each element in the array. Adjusting these settings has a small effect on the main beam but can place nulls in the sidelobes. The goal of the GA is to minimize the total output power of the antenna by adjusting these array settings. Since the algorithm must be fast and a global minimum is not necessary, the GA uses a small population size. Fig. 3 is a flowchart of the GA used for adaptive nulling. Each chromosome represents the least significant bits of the phase shifters in the antenna array. The least significant bits are sent to the antenna array and the output power is measured. In this way, every population member has an associated cost. Members of the population with high costs are discarded. The surviving members form a mating pool. The parents are combined in some manner (such as single point crossover as shown in Fig. 4) to produce offspring. The offspring replace the discarded chromosomes. Fig. 4 demonstrates the process of parent selection and the use of single point crossover to create two new offspring. The next step randomly mutates a certain percentage of the population by changing bits from one to zero or from zero to one (Fig. 5). Normally, the best chromosome is not mutated. After mutation, the process repeats by measuring the output power associated with the new population. they require many function calls to find an acceptable solution. These slow algorithms will not work for real time applications like adaptive nulling. There has been strong evidence that GAs with small population size and high mutation rates find good solutions fast [6], [7].. IV. AMPLITUDE AND PHASE ADAPTIVE NULLING Fig. 3. Flow chart of the GA. Fig. 4. The process of natural selection and mating. Fig. 5. After random mutations occur in the population, the output power is measured for each chromosome and the process repeats. Most GAs have a large population size and low mutation rate. Although these implementations have been successful, The GA functions as an adaptive antenna algorithm by minimizing the total output power of the array. This approach only works if the desired signal is not present or if the adaptive algorithm is constrained to making small amplitude and phase perturbations at each element. Although adapting while the desired signal is absent may work for a stationary antenna with relatively stationary interference sources, it would fail to provide reasonable protection for most communications and radar systems. Limiting the amplitude and phase perturbations at each element is actually quite easy to do, especially with a GA, since the GA inherently constrains its variables. Most phase shifters and attenuators are digital, so a binary GA naturally works with binary control signals. The first example has a 20 element array of isotropic point sources spaced 0.5λ apart. Each element has six bit amplitude and phase weights. There is a 20 dB, n = 3 low sidelobe Taylor amplitude taper with two least significant bits of the amplitude weights and three of the phase weights dedicated to adaptive nulling. The desired signal has an amplitude of 0 dB and is incident on the peak of the main o beam at φ=90 . Two 30 dB interference sources are incident at 111o and 117 o . GA parameters include a population size of 8, a 50% selection rate, roulette wheel selection, uniform crossover, and a 10% mutation rate. The quiescent and resulting adapted patterns appear in Fig. 6. Deep nulls are created in the pattern with little perturbation to the main beam. The nulls come at a cost of increased average sidelobe level. Convergence of the algorithm is shown in Fig. 7. The GA uses elitism, so the maximum number of output power measurements is 8 + 17 × 7 = 127 . The power output decreases monotonically, but the individual interference source received power varies. At times, the power received by the array from one interference source will go up or down relative to the power received from the second interference source. The output power of the desired signal shows little variation, because the main beam remains virtually unperturbed. Fig. 8 is a graph of the signal to interference ratio for the best chromosome in each generation. A second example has two 30 dB interference signals at o o 50 and 130 . The desired signal is at 0 dB and is incident at the peak of the main beam. More bits were needed to perform the nulling in this case. Consequently, four amplitude and three phase least significant bits were used. Fig. 9 shows the resulting adapted pattern. Pattern distortion is more noticeable than in the previous non-symmetric example. Fig. 6. The solid line is the adapted array factor and the dashed line is the quiescent array factor. Fig. 9. Adapted pattern when there are two symmetric interference sources. V. PHASE ONLY ADAPTIVE NULLING Fig. 7. GA convergence for amplitude and phase adaptive nulling. Fig. 8. Signal to interference ratio as a function of generation for amplitude and phase adaptive nulling. Phased arrays have variable phase weights to steer the main beam, but rarely have variable amplitude weights. Thus, it is very desirable to do the adaptive nulling using only digital phase shifters. The same idea applies where only a few of the least significant bits of the phase shifters are used for nulling. The non-symmetric example is repeated for the phase only case. The three least significant bits from the six bit phase shifters were used by the same GA to minimize the output power from the array. Fig. 10 shows the adapted array factor. The resulting pattern has little main beam perturbation. Sidelobes are generally higher, though. Convergence of the phase only GA is shown in Fig. 11. Both interference signals are minimized. Fig. 12 is a plot of the signal to interference ratio. The GA improves the ratio by about 20 dB. Nulling interference sources that are at symmetric angles about the main beam is more difficult for phase only nulling than for amplitude and phase nulling. Using three least significant phase bits only place one nulling in the pattern. Adding a fourth bit produced the adapted pattern shown in Fig. 13. Although the desired nulls appear in the pattern, the main beam gain dropped slightly, and the sidelobes increased significantly. Fig. 10. The solid line is the adapted array factor and the dashed line is the quiescent array factor. Fig. 13. Adapted pattern when there are two symmetric interference sources. VI. AMPLITUDE ONLY ADAPTIVE NULLING Fig. 11. GA convergence for phase only adaptive nulling. Amplitude only adaptive nulling is not normally considered a viable option, because arrays have adjustable phase shifters and not adjustable amplitude weights. A new concept where smart materials are used to adjust the amplitude of the array elements provides adequate justification for trying this approach. The approach to the adaptive nulling is slightly different here. Instead of using only the least significant bits of the variable amplitude weights, only a few of the edge element are given variable amplitude weights. The GA can use the full range of the amplitude weights without limiting to small amplitude values. Because only some of the elements are adaptive, the main beam receives limited perturbations. As an example, consider an array of 16 elements that are spaced λ/2 apart. Three edge elements on both ends of the array have continuous variable amplitude settings. A continuous variable GA is used in place of the binary GA to do the adaptation. The array is assumed to start with a uniform amplitude distribution. There are two interference signals: one 20 dB signal incident at 72o and one 24 dB signal incident at 133o. The resulting adapted amplitude weights are given by w = [.064 .68 .69 1 1 1 1 1 1 1 1 1 1 .69 .68 .064] Fig. 12. Signal to interference ratio as a function of generation for phase only adaptive nulling. Fig. 14 is the adapted pattern corresponding to these weights. On the one hand, sidelobes go down due to the effective amplitude taper that places the nulls. On the other hand, the main beam gain goes down and the beam width expands. Limiting the number of elements with variable amplitude weights is an alternative to using only the least significant bits of the amplitude weights. Both approaches place nulls without undue main beam variations. constraining the size of the weight perturbations allows the GA to minimize the output power to reject the interference while at the same time not placing a null in the direction of the desired signal. References [1] [2] [3] [4] Fig. 14. Array factor due to amplitude only nulling with six out of sixteen elements. VII. CONCLUSIONS The GA has proven quite successful as an adaptive antenna algorithm for arrays. Using a small population size and high mutation rate helps the GA to quickly place nulls. In this paper several examples were presented to show how [5] [6] [7] S.P. Applebaum, "Adaptive arrays," Syracuse University Research Corporation Report SPL TR 66-1, Aug 1966. B. Widrow, et.al., "Adaptive antenna systems," IEEE Proc., Vol. 55, No. 12, Dec 1967, pp. 2143–2159. C. A. Baird and G. G. Rassweiler, "Adaptive sidelobe nulling using digitally controlled phase-shifters,'' IEEE AP Trans., Vol 24, No. 5, pp. 638-649, Sep 76. R.L. Haupt, "Phase-only adaptive nulling with genetic algorithms," IEEE AP-S Trans., Vol. 45, No. 5, Jun 97, 1009-1015. R.L. Haupt and H.L. Southall, "Experimental adaptive nulling with a genetic algorithm," Microwave Journal, vol. 42, no. 1, Jan 99, pp. 7889. R.L. Haupt and S.E. Haupt, "Optimum population size and mutation rate for a simple real genetic algorithm that optimizes array factors," Applied Computational Electromagnetics Society Journal, Vol. 15, No. 2., July 2000, pp. 94-102. nd R.L. Haupt and S.E. Haupt, Practical Genetic Algorithms, 2 ed. New York: Wiley, 2004.