Wu et al.

: Planning System for Indoor Wireless Network

73

PLANNING SYSTEM FOR INDOOR WIRELESS NETWORK

Rong-Hou Wu, Yang-Han Lee and Shih-An Chen Department of Electrical Engineering Tamkang University Tamsui, Taipei Hsien, Taiwan 251, Republic of China E-mail: yhlee@ee.tku.edu.tw

ABSTRACT A novel Wireless LAN prediction tool using genetic algorithm and neural network has been proposed in this paper. We establish a site survey tool system to predict the Received Signal Strength Index (RSSI) in indoor environment. The system includes six items. (1) The fading function: It corrects the functional characteristics of the RSSI for different kinds of Wireless LAN card in free space. (2) Setting the attributes of obstacles in indoor environment: The idea of single attribute of local area is proposed in this paper. If there are same obstacles in one area, we set the area as one attribute. (3) Genetic Algorithm: We use Reproduction, Crossover and Mutation to obtain the propagation loss through the different obstacles (Li). (4) Neural Network: We use Neural Network concept to correct the prediction error arisen from the multipath effect in indoor environment. ( 5 ) The auxiliary judgment for the sampling points: The method is helpful to users in establishing the best sampling points. (6) The calibration of prediction results: We use calibration to correct the prediction error arisen from Li.

been the two methods: the ray-tracing technique [3]-[5] and the path loss model [ 11, [6]-[SI. It has been found that no matter what the method is employed, the parameters of computation equation have to be exactly specified. To overcome several disadvantages of conventional methods, the integrated technology of the prediction system for wireless network is proposed in this paper. This prediction system consists of two parts: genetic calculator and neural network calculator. The genetic calculator is employed to computerize the environment parameters without experiencing the trial-and-error process. Moreover, the prediction accuracy can be increased by the neural network calculator which is used to modify the prediction error resulted from the multipath effect. In addition, the communication quality in terms of received signal strength index (RSSI) can be read out real time from this prediction system. Furthermore, during the experimental process, it is not necessary to prepare any special experiment equipment except the normal notebook computers and wireless LAN cards.

1. INTRODUCTION Because there are many advantages in wireless communication system, such as roaming, easy maintenance, elimination of wiring around the indoor environment, and the flexibility of relocating equipment, it becomes the most attractive alternative to traditional wire-based and optical-fiber communication system [ 11-[2]. In order to develop a unique and compatible interface for users, there have been two architectures defined by the IEEE stand in the media access control (MAC) protocol, which are named Infrastructure and Ad Hoc network, respectively. Of these two, the Infrastructure network is a more popular technique because of its convenience for the connection to the Internet. In designing a wireless system based on Infrastructure system, it is difficult to decide the numbers of access points (AP) and the place where AP should be placed in a building. To obtain the suitable locations of AP, there have
Conti-ibuted Paper Original manuscript received October 3 1, 2000

2. DESIGN OF THE NOVEL PREDICTION SYSTEM 2.1 Fading Function If we input AutoCAD file to the prediction tool, considering only the interference of the environment and disregarding the functional characteristics of different Wireless LAN cards (Fig. I), the prediction tool will yield wrong result. To avoid this, we use the RSSI-distance from a different Wireless LAN card for the correction process.

0098 3063/00 510.00 2001 IEEE

74

IEEE Transactions on Consumer Electronics, Vol. 47, No. 1>FEBRUARY 2001

ex, zx,2 ... cx;+l

C.Y,2

Ex,

...

E,Y; -rb,,b,

cPL(X),

(4)

CX,'

"'

p:"' c x ; P L ( x ) , B, =
c x , * " b,, c-r"PL(X),

obtained from the genetic algorithm. In the calculation process, the following equations arc chosen as the fitness function

EX; ZX:"'...

Bi

=cis, -GA,/
k=m

(5)

ID

30

35

40

45

50

55

60

65

TX.&

Fig. 1 Characteristic among the different Wireless LAN cards

2.2 Genetic Algorithm Genetic algorithm, as first proposed by John Holland [9], has become a very important tool applied in machine leamtng, multiprocessor scheduling, and function optimization Unlike many optimization methods, it can simultaneously search several paths and its convergence rate is still faster than other techniques. Fig. 2 shows the operation procedure by using the concept of genetic algorithm in the proposed prediction system. At first, the population of N, i.e. N chromosomes will be generated according to the algorithm. In this research, N=20 has been chosen. In general, the specific chromosome can be represented by a 20-bit sequence which is composed of 0 and 1, and each sequence represents a particular solution. After randomly processing the population, the fitness value of each chromosome will be calculated. According to this value, the candidate for suitable reproduction will be generated. Following this procedure, the random selection of one-to-one among all the candidates will continue proceed the crossover and mutation process in order to generate the next generation. All of these procedures will be repeated, including the calculation of the fitness value, reproduction, crossover, and mutation, until the predetermined value is presented. In this research, the value of environmental parameter will be represented by chromosome, while the suitable solution is

Fig. 2 The operation procedure of the genetic algorithm We will only consider the received signal strength index (RSSI) in direct path without considering the multipath effect. The following equation will be employed to determine the RSSI of each neuron.

(7)
wherc S is the received signal strength, pT(,y) is the signal value which only path loss is taken into account, 0, is the number of the ith obstacle, L, is the propagation loss through the ith obstacle, and n is the index of a specific obstacle. If the value of total error is smaller, the larger fitness value would be given. After calculating through a maximum of 4,000 times; the prediction system would list out the most suitable environmental parameters by referring to total estimation error in the sampling points.
2.3 Neural Network A computation system constructed from the neural network demonstrates the capability of simulating the neural network of living creature

Wu et al.: Planning System for Indoor Wireless Network

by using a large amount of the artificial neuron. The propagation path of signal between several neurons is called connection. There exists a weight factor IVY belonging to each connection in order to represent the degree of influence which the Lth neuron imposes to the jth one. According to the concept of neural network, the following equations are employed to correct the prediction error:

Where Srrepresents the signal value at neuron t after correction, lfiu is the weighting factor between the neuron t and the neuron U, SOis the signal value at neuron t before correction, Su is the largest signal value among SL, and Stv is signal values among the other neurons which are located at the surrounding of the neuron t. Fig. 3 illustrates the neural network technique which is employed to modify the prediction error arisen from the multipath effect in indoor environment. From this figure, it is obvious that it will modify the prediction value at the neuron t by that at the neuron U.

2. The program chooses the first one among the 12 points (a) Among the sampling points, we find the sampling point between which and AP there are the most categories. The one with the most categories will be selected first. (b) When there are more than one sampling points with the same amount of categories, it will find the sampling points with the largest amount of obstacles between it and AP. (c) When there are more than one sampling points meeting with the above requirements, it find the sampling point with the longest distance to AP. (d) When there are still more than one sampling points that meet with the above requirements, it will pick any one up randomly.

3. The program chooses the second one among the 12 points (a) There should be no less than 18 degrees between any two points. (b) The process way is as the same as described above from 2(a) to 2(d). D. For the rest 8 points following the steps in C
\

TX

1
Fig. 4 The auxiliary judgment for the sampling points

Fig. 3 Illustration of the neural network technique which is employed to modify the prediction error arisen from the multipath effect

2.4 Sampling Points When Genetic Algorithm is used to obtain the propagation loss through different obstacles, we cannot know in advance where the best sampling points are. We thus design a program to show up the sampling points automatically. A. As shown in Fig. 4, sampling points are located within a radius of 30m for the center is AP. B. In this circumstance it established 10 sampling points in which every 72 degrees has 2 sampling points. C. The program chooses 2 sampling points for each 72 degrees 1. Choose one sampling point at every 6 degrees to have a total of 12 sampling points.

2.5 Calibration Procedure The calibration of large range When the error rate of RSSI is between 20% and 30%, and as long as any one of the obstacles modified its RSSI loss, others obstacles will have the same attributes, too.

(10)
n=I

,=I

where B, is the loss of RSSI caused by the ilh obstacle after correction. L, is the loss of RSSI

16

IEEE Transactions on Consumer Electronics, Vol. 47, No. 1, FEBRUARY 2001

caused by the ithobstacle. B,, is to correct the prediction error of the ittl obstacle for the n t h sampling point. L is the modified coefficient, ,7 O<p < I . P , and M, represent the predicted and measured values of RSSI arising for the M,I, sampling point. Or,,is the number of i,h obstacles between AP and the ntllsampling point.
The calibration of small range When the error rate of RSSI is over 20% after the calibration of large range, it only modifies the RSSI loss between which obstacle and AP, it does not care whether other obstacles have the same attributes.

3. EXPERIMENTAL RESULTS (a) We choose I O sampling points for the first experiment, as shown in Fig. 6. We use four sampling points (4, 5 , 6 and 7 ) for GA process. All the values of RSSI are shown in Table 1 We use the calibration of large range to modify point 1 for the reason of its error rate of RSSI is 3 0 2 4 % We use the calibration of small range for point 1 again, because its error rate of RSSI is still over 20%. The error rate of RSSI for point 1 is 18 21% finally.

where T, is the loss of RSSI caused by the i,~, obstacle after correction, L, is the loss of RSSl caused by the
ith

Fig 6 Positions of sampling points (for the first experiment)

(b) We choose 27 sampling points for the second experiment, as shown in Fig. 7 . We use five sampling points (3, 5, 8, 20 and 27) for CA process. All the values of RSSI are shown in Table 2. We use the calibration of large range to modify points 1, 4, 6, 9, 13 and 24 for the reason of their error rates of RSSI are between 20% and 30%. According to Table 2, the error rate of these six sampling points are decreased from 23.9%, 21.65%, 22.62%, 28.97%, 29.63% and 29.38% to 18.21%, 6.4%, 12.39%, .14.71%, 24.15% and 20.82%, respectively. We use the calibration of small range for points 8, 10, 11, 12, 13, 14, 22, 24 and 25 again, because their error rates of RSSI are still over 20%. The error rate of these nine sampling points are decreased from 27.26%, 34.84%, 37%, 38.51%, 24.1WO, 3@.4%, 27.85%, 20.82% and 35.81% to 12.43%, 11.62%, 23.26%, 23.06%, 6.07'36, 14.8%, 6.8lY0, 18.44% and 30.23%, respectively. The error rate of RSSl of sampling points 11, 12 and 22 are still over 20%, we call it " the revise calibration effect" of this phenomenon. Table 3 shows the RSSI loss of the different obstacles from experimental.

obstacle.

is the modified

coefficient, O<p<1. P,, and M, represent the predicted and measured values of RSSI arising for the nth sampling point. On,is the number of it], obstacles between AP and the ntllsampling point. The flow chart of the proposed prediction system is shown in Fig. 5.

Fig. 5 The flow chart of the proposed prediction system

Fig. 7 Positions of sampling points (for the second experiment)

Wu et al.:

Planning System for Indoor Wireless Network

(c) Finally, we use the prediction results to predict the optimal locations of access points (AP) in indoor environment and illustrate the method of site survey tool application. Fig. 8 shown the prediction result using one AP which denote by black dot, where the gray area, white area and dark gray area are representing RSSI of 63-51, 50-31 and 30-0, respectively. In Fig. 9 the gray area, which begins with a smaller area. is becoming larger when using two APs. Fig 10 shown the best prediction result, where the three black dots represent the optimal AP positions.

and prediction errors also render more accurate prediction values. Finally we use the prediction results to predict the optimal locations of access points (AP) in indoor environment and illustrate the method of site survey tool application.
ACKNOWLEDGMENT This work was supported by thc National Science Council, Taipei, Taiwan, R. 0. C. under Contract NSC 89-221 5-E-032-001 REFERENCES D.C.Cox, Universal portable radio communications, IEEE Trans. Veh. Technol., vol. VT-34, no. 3, pp.117-121, Aug. 1985. M. A. Panjwani, et. al., Interactive computation of coverage regions for wireless communication in multifloored indoor environments, IEEE Journal on Selected Areas in Commun., vol. 14, no. 3, pp. 420-429, April 1996. J . W. McKown, et. al., Ray tracing as a design tool for radio networks, T E E Network Mag., vol. 5 . pp.27-31, Nov. 1991. W. Honcharenko, et. al., Mechanisms governing propagation between floors in buildings, IEEE Trans. Antennas Propaga., vol. 41, no. 6, pp.787-790, 1993. T. S. Rappaport, et. al., Site-specific propagation prediction for PCS system in Wireless Personal design, Communications, M. J. Feuerstein and T.S. Rappaport, Eds. Norwell MA: Kluwer, 1993, pp. 281-315. K. W. Cheung, et. al., A new empirical model for indoor propagation prediction, IEEE Trans. Veh. Technol., vol. 47, no. 8, pp. 996-1001, Aug. 1998. C. C. Chiu, et. al., Coverage prediction 171 in indoor wireless communication, IEICE Trans. Commun., vol. E79-B, no. 9, pp. 1346-1350, Sep. 1996. H. Lee, ct. AI., Site sury tool for wireless network based on auto-calibrat ion. J . H. Holland, Adaptation i n Natural and Artificial Systems. University of Michigan Press 1975. (Second edition: MIT press, 1992)

Fig. 8 The prediction result using one AP

Fig. 9 The prediction result using two APs

Fig. 10 The best prediction result of the optimal AP positions

4. CONCLUSION The advantages of the site survey tool are as follows: By classifying different obstructers by color, the users can obtain a clear view on the location distribution of obstructers in the indoor environment. The auxiliary judgment for the sampling points is helpful to users in establishing the best sampling points. The applications of Genetic Algorithm and method of Neural Network result in more precise prediction values and an increased calculation speed. The calibration of Wireless LAN cards

78

IEEE Transactions on Consumer Electronics, Vol. 47, No. 1, FEBRUARY 2001

Table 1. Implementation result of the first experiment

Table 2. Implementation result of the sccond experiment

Table 3. The RSSI loss ofthe different obstacles

Obstacle Wood wall Cement wall Lobby wall

RSSI loss 1.5-2.5 4-6 6-8

Wu et al.: Planning System for Indoor Wireless Network

BIOGRAPHIES
Rong-Hou Wu was born in Taipei, Taiwan, Republic of China, on October 12, 1956 He received his B.S. degrec in Electronic Engineering in 1981 from the Tamkang University, Tamsui, Taiwan. He received his M.S. degree in Electrical and Computer Engineering in 1987 from the University of Massachusetts Dartmouth. He is currently a Ph. D. candidate of the Department of Electrical Engineering, Tamkang University, Tamsui, Taiwan. His research interests include Fiber-optical communication and computer communication networks.

Yang-Han Lee was born in Taipei, Taiwan, Republic of China, in 1964. He received the B.S., M.S., and Ph.D degrees in electrical engineering from National Taiwan University, Taipei, in 1987, 1989, and 1991, respectively. From 1992 to 1994, he was on duty in the Air Force He joined the faculty of the Department of Electrical Engineering, Tamkang University, Taipei, as an Associate Professor. His main research interests include optical fiber communication systems and communication electronics.

Shih-An Chen received thc B.S. degree from Feng Chia University, Taiwan, R.O.C. in 1998 and the M.S. degree from Tamkang University, Taiwan, in 2000, all in electrical engineering. His research interests include wireless communication, optical communication, and communication theory.