2
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 30, NO. 1, FEBRUARY 2000
A Neural-Fuzzy System for Congestion Control in ATM Networks Shie-Jue Lee, Member, IEEE, and Chun-Liang Hou
Abstract—We propose the use of a neural-fuzzy scheme for ratebased feedback congestion control in asynchronous transfer mode (ATM) networks. Available bit rate (ABR) traffic is not guaranteed quality of service (QoS) in the setup connection, and it can dynamically share the available bandwidth. Therefore, congestion can be controlled by regulating the source rate, to a certain degree, according to the current traffic flow. Traditional methods perform congestion control by monitoring the queue length. The source rate is decreased by a fixed rate when the queue length is greater than a prespecified threshold. However, it is difficult to get a suitable rate according to the degree of traffic congestion. We employ a neural-fuzzy mechanism to control the source rate. Through learning, membership values can be generated and cell loss can be predicted from the status of the queue length. Then, an explicit rate is calculated and the source rate is controlled appropriately. Simulation results have shown that our method is effective compared with traditional methods. Index Terms—ATM, cell loss, cell rate, congestion control, fuzzy logic, neural-fuzzy networks.
I. INTRODUCTION
A
SYNCHRONOUS transfer mode (ATM) is a modern technology enabling the integration of different traffic types within a single communication network [2], [21]. ATM networks are being widely developed to carry voice, video, and data traffics. Various service classes have been defined in ATM for the support of traffic with different quality-of-service (QoS) requirements. These classes consist of constant bit rate (CBR), real-time variable rate (rt-VBR), nonreal time variable bit rate (nrt-VBR), available bit rate (ABR), and unspecified bit rate (UBR). ABR allows applications to fully utilize the available bandwidth in the network by adjusting their instantaneous transmission rates to the available capacity. The quality (QoS) of transmission cannot be determined in the connection setup time for ABR. Therefore, there is a need for a congestion management scheme to adaptively control the rate of transmission by closed-loop. A congestion control scheme is essential for the support of ABR traffic to utilize the available network bandwidth without causing congestion. Among various control schemes, the rate-based congestion control framework is widely used. A rate-based flow-control scheme is an end-to-end feedback mechanism. The scheme consists
Manuscript received January 24, 1999; revised October 31, 1999. This work was supported in part by the National Science Council under Grant NSC-872213-E-110-012. This paper was recommended by Associate Editor A. Kandel. The authors are with the Department of Electrical Engineering, National Sun Yat-Sen University, Kaohsiung, Taiwan, R.O.C. (e-mail:
[email protected]). Publisher Item Identifier S 1083-4419(00)01395-9.
of one source end system and one destination end system for each feedback loop. The source starts to send data cells with the rate of initial cell rate (ICR), and subsequent traffic flow is regulated by resource management (RM) cells. Each RM cell is allowed to flow all the way to the destination end system, and travels in the reverse direction. Intermediate switches mark down the rate in the reverse RM cells. The smallest allocation is therefore the value in the cell when it reaches the source. The source may then use this rate for subsequent transmission until a new resource management cell is received. Many rate-based control schemes [1], [3], [12], [14], [22] have been proposed. In the EFCI marking scheme [10], when a data cell from the source reaches a switch which is under congestion, the EFCI field of the cell header is set to one. If the destination checks that the EFCI field is one, it will send an RM cell back to the source. When the source receives an RM cell, it reduces the cell rate; otherwise the cell rate is increased. However, difficulties may arise if RM cells are lost or delayed. To overcome this disadvantage, Branhart [4] proposed the proportional rate control algorithm (PRCA) scheme. The source sends a data cell, in a constant rate, in which EFCI is set to zero. When the destination checks that EFCI is zero, it sends an RM cell back to the source and the source is informed that the cell rate can be increased. The source will decrease the cell rate when RM cells are not received in the expected period. Therefore, the source continues to decrease the cell rate if RM cells are lost or delayed. However, a long path VC has a large probability of EFCI bit being set; a switch under congestion sets EFCI on all VC’s. Roberts [18] proposed the enhanced proportional rate control algorithm (EPRCA). Congestion is detected at switches based on the queue length. When the queue length is greater than a prespecified threshold, the source is asked to reduce the cell rate by a fixed rate. However, it is difficult to get a suitable rate according to the degree of traffic congestion. Furthermore, this scheme may cause a high cell loss rate. Recently, fuzzy logic and neural networks have been applied to the traffic control in ATM networks [6]–[8], [16], [17], [20]. In this paper, we propose a neural-fuzzy rate-based feedback controller. We employ a neural-fuzzy mechanism to control the source rate. By monitoring the current queue state and the rate of change of the queue length, the future queue behavior is predicted. This closed-loop control scheme can maximize the traffic flow and avoid congestion. Simulation results have shown that our method is effective compared with EPRCA. The rest of the paper is organized as follows. Section II provides a brief introduction to neural-fuzzy networks. An overview of our method is given in Section III. Section IV describes the generation of fuzzy rules and the prediction
1083–4419/00$10.00 © 2000 IEEE
LEE AND HOU: NEURAL-FUZZY SYSTEM FOR CONGESTION CONTROL IN ATM NETWORKS
3
output-linguistic variable. FALCON does not allow the slopes of a trapezoidal membership function to be adjustable. This may affect the speed of learning. Also, the max operator of layer four is not suitable to rule extraction. III. OVERVIEW OF THE CONTROLLER
Fig. 1.
Congestion control in ATM networks.
Fig. 2.
Block diagram of our neural-fuzzy controller.
of cell loss by a neural-fuzzy network. Section V presents a self-tuning fuzzy inference engine. Simulation results are given in Section VI. Finally, conclusions are summarized in Section VII. II. NEURAL-FUZZY NETWORKS Neural-fuzzy networks [11], [19], [23] borrow the ideas from fuzzy logic and neural networks. They have many advantages, such as learning abilities, optimization abilities, human-like IF-THEN rule thinking, and ease of incorporating expert knowledge. There are basically two approaches for building neural-fuzzy networks. One can construct the structure from a set of input-output training data pairs through learning, or one can use a predefined architecture determined previously by some other means. Of course, the former is more desirable. Therefore, we adopt this kind of approach in this paper. Lin and Lin [15] proposed a fuzzy adaptive learning control network (FALCON) which can automatically perform fuzzy partitioning by learning from training examples. FALCON consists of five layers. Layer 1 is the input layer, containing input nodes (input linguistic variables). The output of layer one is linked directly to layer two which is called the term layer (e.g., “small”, “medium”, and “large”) and acts as membership functions representing the terms of respective linguistic variables. Layer 3 contains rule nodes which represent fuzzy rules. Links outgoing from layer-three denote preconditions of the rule nodes. Nodes at layer four are called output-term nodes, representing membership functions of the terms in output-linguistic variables. Layer 5 is the output layer. Each node in this layer is called an output-linguistic node and corresponds to one
The goal of congestion control is to smooth down the burstiness of the input traffic to avoid congestion [2], [21]. Consider an end-to-end path shown in Fig. 1. Data cells are sent from the source to the destination through a series of switches. Congestion controllers monitor the status of switches and notify the source to increase or decrease the rate of transmission, in order to maximize the throughput and avoid congestion. Our congestion controller consists of two parts, a neural-fuzzy network and a fuzzy inference engine, as shown in Fig. 2. The neural-fuzzy network is responsible for the generation of fuzzy rules and the prediction of the future cell loss, while the fuzzy inference engine calculates an explicit rate to regulate the cell rate of the source. The neural-fuzzy network predicts cell loss by monitoring the current normalized queue , current normalized queue change rate , length . and previous normalized queue change rate Fuzzy rules and membership values are obtained by learning. Information concerning the weights of obtained fuzzy rules is passed to the fuzzy inference engine, shown by the symbol . , is predicted and also Another parameter, cell loss passed to the fuzzy inference engine. A performance function is used to maintain the queue length. If the future cell loss is greater than one, the switch sets CI (congestion indicator) and notifies the previous switch in an RM cell that passes through the switch. Also, the switch calculates an explicit rate to adjust the local sources attached to it. Therefore, the system is a closed-loop scheme between sources and switches. IV. PREDICTION OF CELL LOSS As mentioned in Section III, a neural-fuzzy network is used , based on the current normalto predict cell loss, , current normalized queue change rate ized queue length , and previous normalized queue change rate . Learning of the neural-fuzzy network is divided into two phases. In the first phase, the architecture of the network is built by applying the fuzzy ART algorithm [5]. In the second phase, parameters are adjusted by a back-propagation algorithm. The architecture of our neural-fuzzy network contains four layers, as shown in Fig. 3, in contrast to five layers of FALCON represents described in Section II. In this figure, represents , and represents . The function of each layer is described below. 1) Layer 1: Like fuzzy ART, nodes in this layer just transmit input signals and their complements to the next layer didenote the output of node in layer . We rectly. Let have (1) Note that each node of layer 1 outputs two values, , to layer 2.
and
4
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 30, NO. 1, FEBRUARY 2000
where is the desired output. Output nodes provide output values to the outside world. They act as defuzzifiers. The activation function of a such node is defined, as in [15], to be the center of area (5) where and are the end points of a range within which values are allowed to exist. Construction of the network structure is performed as follows. Training input patterns are fed into the input layer. Let be the vector containing the input training pattern and its complement. We compute the following value: Fig. 3. Architecture of our neural-fuzzy network.
(6) where
is defined as (7)
for term node in layer 2. Note that . Let between the input pattern and
denotes the similarity
(8) Then, Fig. 4. Membership function of node i in layer 2.
is the winner and we do vigilance test to check if (9)
2) Layer 2: Each node of this layer represents a term of an input-linguistic variable. The membership function of each term node is trapezoidal, as shown in Fig. 4, and can be expressed as follows:
(2) otherwise. and are normalized. Also, note that +Note that in contrast to FALCON, we allow the slopes of the two and , of a trapezoidal function to be sides, i.e., trainable. 3) Layer 3: This layer is called the rule layer and each node in this layer represents one fuzzy logic rule. The output of a rule node in this layer is calculated by the product operation as follows:
holds. If the test fails, then mismatch reset occurs. Node of layer 2 is disabled and the process restarts to search for the next term node until a match is obtained. If no such node is found, a new rule node is created in layer 3 and a corresponding term node is created in layer 2. When a term node goes through vigilance test, we have to perform vigilance test for the desired output. If the test fails, a new rule node and a corresponding term node are created at the same time. When a node is selected and passes the associated vigilance test, we modify the weights by taking the AND operation between old weights and training patterns. When the structure of the network is built, we proceed to do the second phase of learning. In this phase, we intend to minimize errors with the gradient descent method, by adjusting the parameters associated with membership functions. The error function is defined to be (10)
(3) 4) Layer 4: This layer is called the output-linguistic layer and contains two kinds of nodes, training nodes and output nodes. Training nodes feed desired output data downward to the network. Their function is similar to that of layer 1, namely, (4)
where is the desired output value and is the actual output value. Obviously, only layer 2 and layer 4 have parameters associated with them. These parameters control the shape of a trapezoidal membership function. We use the back-propagation algorithm to do these adjustments. To tune the parameters of layer 2, we have to transmit error signals from layer 4 to layer 2 through layer 3. We describe the learning process for layers 4, 3, and 2 as follows.
LEE AND HOU: NEURAL-FUZZY SYSTEM FOR CONGESTION CONTROL IN ATM NETWORKS
5
and
• Layer 4: In this layer, as done in [15], values are only and . Therefore, we only allowed to exist between adjust these two parameters. Clearly, we have
(23)
otherwise (11)
(24)
otherwise using (5). Then,
is updated by
(25)
otherwise. (12) V. FUZZY INFERENCE ENGINE where
is the learning rate. Similarly,
is updated by (13)
• Layer 3: The error signal is computed by (14) and can be expressed as
(15)
by using (5). , and • Layer 2: In this layer, four parameters of each membership function have to be trained. Note that (16)
The purpose of control is to influence the behavior of a system by changing inputs to that system according to a set of rules that model how the system operates. Classic control theory uses a mathematical model to define a relationship that transforms the desired state (requested) and observed state (measured) of the system into inputs that will alter the future state of that system. As the complexity of the system increases it becomes highly difficult to formulate that mathematical model. Fuzzy control replaces the classic mathematical model with one that is built from a number of fuzzy IF-THEN rules. The process of inference binds the rules together to produce desired outputs. We use a self-tuning fuzzy inference engine, one that is similar to that proposed in [9], for the congestion controller to control the traffic flow and make the network”s available bandwidth maximally utilized. Our method intends to minimize the difference between the future queue length and the desired queue length. Therefore, the desired queue length can be maintained and queue overflow can be avoided. Our fuzzy inference is based on the fuzzy IF-THEN rules that are automatically obtained from the neural-fuzzy network is expressed as described in Section IV. Each rule IF
where
is
and
is
and
is
THEN otherwise
(17)
(18) from (2) and (3), respectively. Therefore,
is updated by (19)
Similarly, we have (20)
(21)
where
is the action function and is defined to be
(26) , and are where is the desired queue length, and coefficients to be described later. Equation (26) is adopted from [9], except for the addition of the last term. Note that is the predicted cell loss from the neural-fuzzy network of the to be reduced if cell loss previous section. We would like increases. The system can calculate the rate of cell transmission, i.e., cell rate, by the following equation [9]: (27) which is the explicit rate that will be sent back to the source to is the weight of the adjust its transmission rate. Note that th rule, passed from the neural-fuzzy network described in the previous section, and is defined as
(22) (28)
6
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 30, NO. 1, FEBRUARY 2000
where
denotes the membership degree of on , or ( can be obtained easily from the term nodes in layer 2 of the neural-fuzzy network). In order to avoid congestion, the system is designed to minimize a performance function based on the gradient descent method. The performance function is defined as (29) The future queue length,
, can be expressed as [9] (30)
is the output rate, is the interval time, and MCR where is the mean cell rate of other VC’s. , and , Note that there are four coefficients, in (26). These coefficients can be obtained by performance . For example, can be obtained as learning on follows:
Fig. 5.
Cell loss rate versus queue size.
From (27), (26), and substitution, we have (35), shown at the bottom of the page. Therefore, (see (36), shown at the bottom of the page). Initial conditions are defined as follows [9]: (37) Similarly,
, and
can also be obtained.
(31) VI. EXPERIMENTAL RESULTS and (32) From. (30) and substitution, we have
(33) where
is the backward shift operator. Then
(34)
In this section, we show some experimental results. A comparison with EPRCA is also made. First of all, we run an experiment to show the power of our controller in reducing the cell loss rate. In this experiment, one switch and two attached sources are simulated. The result is shown in Fig. 5. We can see that without the controller, the cell loss rate is high. However, cell loss is nearly zero using the controller. In the second experiment, we use a model shown in Fig. 6 in which two ATM switches, Sw1 and Sw2, are cascaded. Two sources, S1 and S2, are connected to Sw1, and two destinations, D1 and D2 are connected to Sw2. Values for related parameters about sources and switches are listed in Table I in which ICR stands for “initial cell rate,” PCR for “peak cell rate,” MCR for “mean cell rate,” AIR for “additive increase rate,” QT stands for “queue threshold,” DQT for “queue very congested threshold,” MRF for “major reduction factor,” and ERF for “explicit reduction factor.” We use the interrupted Bernoulli Process (IBP) as the traffic type for sources. The IBP state diagram is shown in Fig. 7, in which 0 and 1 represent the idle state and the burst
(35)
(36)
LEE AND HOU: NEURAL-FUZZY SYSTEM FOR CONGESTION CONTROL IN ATM NETWORKS
7
Fig. 6. Model with two switches. TABLE I PARAMETER VALUES FOR SOURCES SWITCHES
Fig. 7.
AND
Fig. 10.
Throughput of each switch.
Fig. 11.
Queue occupation rate with EPRCA.
Fig. 12.
Queue occupation rate with our method.
Fig. 13.
Throughput with different P values.
IBP state diagram.
Fig. 8. Allowed cell rate with EPRCA.
Fig. 9. Allowed cell rate with our method.
state, respectively, and and are transfer probabilities between states with the following properties:
In our experiments, we set . Also, is 0.9 for is a fixed value. Figs. 8 and 9 the experiments in which if show the allowed cell rate (ACR) obtained from EPRCA and our method, respectively. Apparently, the ACR of EPRCA varies much more rapidly than that of our method. Furthermore, our
method allows a larger throughput at each switch, as indicated in Fig. 10. Figs. 11 and 12 show queue occupation rate at each ), switch obtained by EPRCA and our method (with respectively. Obviously, our method offers a much smaller varivalues ation in queue occupation rate. Also, we use different to compare the throughput of our method and that of EPRCA at each switch. Fig. 13 shows the simulation results, demonstrating a higher throughput obtained by our method, especially is high. when traffic is high, i.e., In the third experiment, we use another model shown in Fig. 14 in which four ATM switches, Sw1, Sw2, Sw3, and Sw4, are cascaded [13]. Two sources, S1 and S2, are connected to Sw1; One source, S3, and one destination, D1, are connected to
8
Fig. 14.
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 30, NO. 1, FEBRUARY 2000
Another network model.
Fig. 16.
Allowed cell rate with our method.
Fig. 17.
Throughput of each switch.
Fig. 18.
Throughput with different P values.
TABLE II PARAMETER VALUES FOR SOURCES
TABLE III PARAMETER VALUES FOR SWITCHES
Fig. 15.
Allowed cell rate with EPRCA.
Sw2; One source, S4, is connected to Sw3; Three destinations, D2, D3, and D4, are connected to Sw4. Values for related parameters about sources and switches are listed in Tables II and III, respectively. Fig. 15 shows the ACR obtained from EPRCA. Again, this figure shows that ACR oscillates, meaning that EPRCA fails to have an adaptive capability in traffic control. Fig. 16 shows the ACR obtained from our method. This figure shows a very much smoother curve, meaning that our method has a strong adaptive capability in traffic control.
Fig. 17 compares the throughput of each switch. Obviously, our method results in a higher throughput at each switch. Finally, we demonstrate the effectiveness of our method with different . The result is shown in Fig. 18 which shows values of that our method provides a higher throughput than EPRCA, especially when traffic is high. VII. CONCLUSION An ATM congestion controller based on a neural-fuzzy network and a fuzzy inference engine has been presented. The
LEE AND HOU: NEURAL-FUZZY SYSTEM FOR CONGESTION CONTROL IN ATM NETWORKS
neural-fuzzy network predicts cell loss and derives fuzzy rules for the fuzzy inference engine. The fuzzy inference engine then calculates an explicit rate which can be used to control the cell rate of the sources. This closed-loop control scheme can maximize the traffic flow and avoid congestion. Simulation results have shown that our method can increase throughput and reduce cell loss rate. ACKNOWLEDGMENT Comments from the anonymous referees are appreciated.
9
[18] L. Roberts, “Enchanced PRCA (Proportional Rate-Control Algorithm),”, ATM Forum Contribution 94-0735R1, Aug. 1994. [19] P. K. Simpson, “Fuzzy min-max neural networks—Part 2: Clustering,” IEEE Trans. Fuzzy Syst., vol. 1, pp. 32–45, Feb. 1993. [20] A. Tar and I. Habib, “Congestion control mechanism for ATM networks using neural network,” in Proc. IEEE Int. Conf. Commun., 1995, pp. 206–210. [21] A. Tarraf and I. Habib, “A neurocomputing approach to congestion control in an ATM multiplexer,” J. High Speed Networks, vol. 5, pp. 329–446, Aug. 1998. [22] “Traffic Management Specification Version 4.0,”, The ATM Forum Technical Committee, 1996. [23] Y. Q. Zhang and A. Kandel, “Compensatory neurofuzzy systems with fast learning algorithms,” IEEE Trans. Neural Networks, vol. 9, pp. 83–105, Jan. 1998.
REFERENCES [1] A. Atai, “A rate-based feedback traffic control in B-ISDN,” in Proc. IEEE Int. Conf. Commun., 1994, pp. 1605–1615. [2] J. Aweya, “Neurocontroller for buffer overload in a packet switch,” Proc. Inst. Elect. Eng., Commun., vol. 145, pp. 227–223, Aug. 1998. [3] L. Benmohamed and S. M. Meerkov, “Feedback control of congestion in packet switching networks,” IEEE/ACM Trans. Networking, vol. 6, pp. 693–708, Dec. 1993. [4] A. W. Branhart, “Baseline performance using PRCA rate-control,” presented at the ATM Forum, 1994. [5] G. A. Carpenter and S. Grossberg, “Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system,” Neural Networks, vol. 4, pp. 759–771, 1991. [6] R. G. Cheng and C. J. Chang, “Design of a fuzzy traffic controller for ATM networks,” IEEE/ACM Trans. Networking, vol. 4, pp. 460–469, Nov. 1996. [7] R. Durand and Y. C. Liu, “Neural network based congestion control for single and multiple sources in ATM networks,” in Proc. IEEE/MACS Computer Engineering for Systems Applications, July 1996, pp. 1100–1105. [8] I. Habib, A. Tarraf, and T. Saadawi, “A neural network controller for congestion control in ATM multiplexers,” Computer Networks and ISDN Systems, pp. 325–334, 1997. [9] Q. Hu and D. W. Petr, “Self-tuning fuzzy traffic rate control network,” in Proc. IEEE Int. Conf. Commun., Dallas, Tx, 1996, pp. 424–428. [10] Traffic control and congestion control in B-ISDN, in Draft Recommendations, ITU-T, 1995. [11] J. S. Jang, “ANFIS: Adaptive-network-based fuzzy inference systems,” IEEE Trans. Syst., Man, Cybern., vol. 23, pp. 665–685, May 1993. [12] R. Krishnan, “Rate-based control schemes for ABR traffic-design principles and performance comparison,” Comput. Networks ISDN Syst., pp. 583–593, 1997. [13] W. G. Lai, “A Flow Control Scheme on ATM Networks with Max-Min Fairness,” Nat. Sci. Council, Taiwan, R.O.C., Tech. Rep., 1997. [14] Y. C. Lai and Y. D. Lin, “Performance analysis of rate-based flow control under a variable number of sources,” Comput. Networks ISDN Syst., pp. 2139–2152, 1998. [15] C. J. Lin and C. T. Lin, “An ART-based fuzzy adaptive learning control network,” IEEE Trans. Fuzzy Syst., vol. 5, pp. 447–496, Nov. 1997. [16] Y. C. Liu and C. Douligeris, “Rate regulation with feedback controller in ATM network—A neural network approach,” IEEE J. Select. Areas Commun., vol. 15, pp. 200–208, Feb. 1997. [17] A. Pitsillides, Y. A. Sekercioglu, and G. Ramamurthy, “Effective control of traffic flow in ATM networks using fuzzy explicit rate marking (FERM),” IEEE J. Select. Areas Commun., pp. 209–225, 1997.
Shie-Jue Lee (M’90) was born at Kin-Men, Taiwan, R.O.C., on August 15, 1955. He received the B.S.E.E. and M.S.E.E. degrees from National Taiwan University, Taipei, in 1977 and 1979, respectively, and the Ph.D. degree from the Department of Computer Science, University of North Carolina, Chapel Hill, in 1990. He joined the faculty of the Department of Electrical Engineering, National Sun Yat-Sen University, Kaohsiung, Taiwan, in 1983, where he has been a Professor since 1994. He is also serving as the Director of the Center for Telecommunications Research and Development, National Sun Yat-Sen University. His research interests include artificial intelligence, telecommunication networks, and VLSI signal processing. He served as the Director of the Southern Telecommunications Research Center, National Science Council, Taiwan, from 1998 to 1999. Dr. Lee received the Distinguished Teachers Award of the Ministry of Education, Taiwan, in 1993. He was awarded by the Chinese Institute of Electrical Engineering for Outstanding M.S. Thesis Supervision in 1997. He also received the Distinguished Paper Award of the Computer Society of the Republic of China in 1998, and the Distinguished Research Award of National Sun Yat-Sen University in 1998. He served as the program chair for the International Conference on Artificial Intelligence (TAAI-96), Kaohsiung, Taiwan, December 1996, and the International Computer Symposium—Workshop on Artificial Intelligence, Tainan, Taiwan, December 1998. He is a member of the IEEE Systems, Man, and Cybernetics Society, the International Society of Applied Intelligence, the Association for Automated Reasoning, the Institute of Information and Computing Machinery, and the Taiwanese Association of Artificial Intelligence (TAAI). He is also an executive committee member of TAAI.
Chun-Liang Hou was born in Kaohsiung, Taiwan, R.O.C. on April 10, 1962. He received the B.S.E.E. and M.S.E.E. degrees from National Cheng Kung University, Tainan, Taiwan, in 1992 and 1994, respectively. He is currently pursuing the Ph.D. degree at the Department of Electrical Engineering, National Sun Yat-Sen University, Kaohsiung. His main research interests include artificial intelligence, ATM networks, and pattern recognition.
