9ROXPH1XPEHU % 2FWREHU,661;
I C I C International ,&,&([SUHVV/HWWHUV
$Q,QWHUQDWLRQDO-RXUQDORI5HVHDUFKDQG6XUYH\V
(GLWRUVLQ&KLHI -XQ]R:DWDGD:DVHGD8QLYHUVLW\-DSDQ ,QGH[HGE\ (L&RPSHQGH[ (OVHYLHU 6FRSXV (OVHYLHU ,163(& ,(7 3XEOLVKHGE\,&,&,QWHUQDWLRQDO KWWSZZZLMLFLFRUJLFLFHOKWP ICIC Express Letters Editors-in-Chief Junzo Watada, Graduate School of Information, Production & Systems, Waseda University 2-7, Hibikino, Wakamatsu, Kitakyushu 808-0135, Japan Yan Shi, Graduate School of Science and Technology, Tokai University 9-1-1, Toroku, Kumamoto 862-8652, Japan Advisory Board Steve P. Banks, UK Tianyou Chai, China Tom Heskes, Netherlands Lakhmi C. Jain, Australia Jerry M. Mendel, USA Masaharu Mizumoto, Japan Witold Pedrycz, Canada Jeng-Shyang Pan, Taiwan Peng Shi, UK Jianrong Tan, China Takeshi Yamakawa, Japan Lotfi A. Zadeh, USA Kaiqi Zou, China Associate Editors Ramesh Agarwal, USA Jamal Ameen, UK Rawshan Basha, UAE Michael V. Basin, Mexico Yasar Becerikli, Turkey Genci Capi, Japan Ozer Ciftcioglu, Netherlands Joshua Dayan, Israel Vasile Dragan, Romania Kei Eguchi, Japan Keiichi Horio, Japan Chao-Hsing Hsu, Taiwan Xiangpei Hu, China Gerardo Iovane, Italy Karim Kemih, Algeria Dongsoo Kim, Korea Hyoungseop Kim, Japan Huey-Ming Lee, Taiwan Zuoliang Lv, China Magdi Mahmoud, Saudi Arabia Anatolii Martynyuk, Ukraine Tshilidzi Marwala, South Africa Subhas Misra, India Takuo Nakashima, Japan Nikos Nikolaidis, Greece Sing Kiong Nguang, New Zealand Pavel Pakshin, Russia Jiqing Qiu, China Chin-Shiuh Shieh, Taiwan Mika Sato-Ilic, Japan Guifa Teng, China Gancho Vachkov, Japan Edwin Engin Yaz, USA Jianqiang Yi, China Zhong Zhang, Japan Lindu Zhao, China Sponsored by QinHuangDao YanDa Intelligent Informatics Inc. 2010 ICIC International ISSN 1881-803X Printed in Japan ICIC EXPRESS LETTERS Volume 4, Number 5(B), October 2010 CONTENTS Heuristics for Joint Decisions in Domain Implementation and Verification 1735 Zhiqiao Wu and Jiafu Tang A First Approach to Artificial Cognitive Control System Implementation Based on the Shared Circuits Model of Sociocognitive Capacities 1741 Alfonso Sanchez Boza and Rodolfo Haber Guerra Study of Shore-Based AIS Network Link Capacity 1747 Chang Liu and Xinyu Wang Ontology and Rule Combined Reasoning Framework Design 1753 Guanyu Li, Yi Liu and Buwei Chen Vulnerability Assessment of Distribution Networks Based on Multi-Agent System and Quantum Computing 1761 Xiangping Meng, Kaige Yan, Jingweijia Tan and Zhaoyu Pian A New Multiple Kernel Learning Based Least Square Support Vector Regression and Its Application in On-Line Gas Holder Level Prediction of Steel Industry 1767 Xiaoping Zhang, Jun Zhao and Wei Wang The Parameterization of all Stabilizing Simple Repetitive Controllers with the Specified Input-Output Characteristic 1773 Iwanori Murakami, Tatsuya Sakanushi, Kou Yamada, Yoshinori Ando Takaaki Hagiwara and Shun Matsuura Image Noise Removal Based on Image Detail Preserving 1779 Yunfeng Zhang, Caiming Zhang and Hui Liu A Novel Epidemic Model with Transmission Medium on Complex Networks 1785 Chengyi Xia, Junhai Ma and Zengqiang Chen Output Feedback Switching Controller Design for State-Delayed Linear Systems with Input Quantization and Disturbances 1791 Zhaobing Liu, Huaguang Zhang, Qiuye Sun and Dongsheng Yang Fuzzy Variable of Econometrics Based on Fuzzy Membership Function 1799 Xiaoyue Zhou, Kaiqi Zou and Yanfang Wang Applying Image Processing Technique to Radar Target Tracking Problems 1805 Kuen-Cheng Wang, Yi-Nung Chung, Chao-Hsing Hsu and Tsair-Rong Chen Fault Detection and Diagnosis for Process Control Rig Using Artificial Intelligent 1811 Rubiyah Yusof, Ribhan Zafira Abdul Rahman and Marzuki Khalid Syntactic Feature Based Word Sense Disambiguation of English Modal Verbs by Naive Bayesian Model 1817 Jianping Yu, Jilin Fu and Jianli Duan A Robust Adaptive Beamforming Optimization Control Method via Second-Order Cone Programming for Bistatic MIMO Radar Systems 1823 Fulai Liu, Changyin Sun, Jinkuan Wang and Ruiyan Du (Continued) 2010 ICIC International ISSN 1881-803X Printed in Japan ICIC EXPRESS LETTERS Volume 4, Number 5(B), October 2010 CONTENTS (Continued) Study of Parameters Selection in Finite Element Model Updating Based on Parameter Correction 1831 Linren Zhou and Jinping Ou Dynamic Modeling of 3D Facial Expression 1839 Jing Chi Chaos in Small-World Cellular Neural Network 1845 Qiaolun Gu and Tiegang Gao Evaluating the Quality of Education via Linguistic Aggregation Operator 1851 Ying Qiao, Xin Liu and Li Zou Collaborative Filtering Algorithm Based on Feedback Control 1857 Baoming Zhao and Guishi Deng MVC Algorithm Using Depth Map through an Efficient Side Information Generation 1863 Ji-Hwan Yoo, Young-Ho Seo, Dong-Wook Kim, Manbae Kim and Ji-Sang Yoo Design of Node with SOPC in the Wireless Sensor Network 1869 Jigang Tong, Zhenxin Zhang, Qinglin Sun and Zengqiang Chen Research on the Sensorless Control of SPMSM Based on a Reduced-Order Variable Structure MRAS Observer 1875 Lipeng Wang, Huaguang Zhang, Zhaobing Liu, Limin Hou and Xiuchong Liu Production Planning Based on Evolutionary Mixed-Integer Nonlinear Programming 1881 Yung-Chien Lin, Yung-Chin Lin and Kuo-Lan Su Hiding Secret Information in Modified Locally Adaptive Data Compression Code 1887 Chin-Chen Chang, Kuo-Nan Chen and Zhi-Hui Wang Genetic Programming Based Perceptual Shaping of a Digital Watermark in the Wavelet Domain 1893 Asma Ahmad and Anwar M. Mirza A Competitive Particle Swarm Optimization for Finding Plural Acceptable Solutions 1899 Yu Taguchi, Hidehiro Nakano, Akihide Utani, Arata Miyauchi and Hisao Yamamoto Design and Implementation of a Novel Monitoring System for Container Logistics Based on Wireless Sensor Networks and GPRS 1905 Kezhi Wang, Shan Liang, Xiaodong Xian and Qinyu Xiong Spatially Adaptive BayesShrink Thresholding with Elliptic Directional Windows in the Nonsubsampled Contourlet Domain for Image Denoising 1913 Xiaohong Shen, Yulin Zhang and Caiming Zhang Application of Type-2 Fuzzy Logic System in Indoor Temperature Control 1919 Tao Wang, Long Li and Shaocheng Tong Traffic Flow Forecasting and Signal Timing Research Based on Ant Colony Algorithm 1925 Wenge Ma, Yan Yan and Dayong Geng Statistical Layout of Improved Image Descriptor for Pedestrian Detection 1931 Ming Bai, Yan Zhuang and Wei Wang (Continued) 2010 ICIC International ISSN 1881-803X Printed in Japan ICIC EXPRESS LETTERS Volume 4, Number 5(B), October 2010 CONTENTS (Continued) Adaptive Control for Missile Systems with Parameter Uncertainty 1937 Zhiwei Lin, Zheng Zhu, Yuanqing Xia and Shuo Wang Application of Plant Growth Simulation Algorithm on SMT Problem 1945 Tong Li, Weiling Su and Jiangong Liu A Controller Design for T-S Fuzzy Model with Reconstruction Error 1951 Hugang Han and Yanchuan Liu An Analysis for Parameter Configuration to Find a Trigger of Change 1959 Rika Ito and Kenichi Kikuchi Complexity Reduction Algorithm for Enhancement Layer of H.264/SVC 1965 Kentaro Takei, Takafumi Katayama, Tian Song and Takashi Shimamoto Modeling of Enterprises Risk Management and Its Robust Solution Method 1973 Min Huang, Yanli Huo, Chunhui Xu and Xingwei Wang Fundamental Study of Clustering Images Generated from Customer Trajectory by Using Self-Organizing Maps 1979 Asako Ohno, Tsutomu Inamoto and Hajime Murao Interface Circuit for Single Active Element Resistive Sensors 1985 Amphawan Julsereewong, Prasit Julsereewong, Tipparat Rungkhum Hirofumi Sasaki and Hiroshi Isoguchi Analytic Solution of Shock Waves Equation with Higher Order Approximation 1991 Valentin A. Soloiu, Marvin H.-M. Cheng and Cheng-Yi Chen Fuzzy Opinion Survey Based on Interval Value 1997 Lily Lin, Huey-Ming Lee and Jin-Shieh Su Certificate of Authorization with Watermark Processing in Computer System 2003 Nai-Wen Kuo, Huey-Ming Lee and Tsang-Yean Lee Weighted Similarity Retrieval of Video Database 2009 Ping Yu Job Scheduling of Retrieving Dynamic Pages from Online Auction Websites on Grid Architecture 2015 Chong-Yen Lee, Hau-Dong Tsui and Ya-Chu Tai An Improvement on Li and Hwang's Biometrics-Based Remote User Authentication Scheme 2021 Wen-Gong Shieh and Mei-Tzu Wang Multiple Robot System Applying in Chinese Chess Game 2027 Song-Hiang Chia, Kuo-Lan Su, Sheng-Ven Shiau and Chia-Ju Wu Towards a Dynamic and Vigorous SOA ESB for C4I Architecture Framework 2033 Abdullah S Alghamdi, Iftikhar Ahmad and Muhammad Nasir Joint Multiple Parameters Estimation for Vector-Sensor Array Using Biquaternions 2039 Fei Wang, Hailin Li and Jianjiang Zhou 2010 ICIC International ISSN 1881-803X Printed in Japan ICIC EXPRESS LETTERS Aims and Scope of ICIC Express Letters The Innovative Computing, Information and Control Express Letters (abbreviated as ICIC Express Letters, or ICIC-EL) is a peer-reviewed English language journal of research and surveys on Innovative Computing, Information and Control, and is published by ICIC International bimonthly. The primary aim of the ICIC-EL is to publish quality short papers (no more than 6 printing pages, basically) of new developments and trends, novel techniques and approaches, innovative methodologies and technologies on the theory and applications of intelligent systems, information and control. The scope of the ICIC-EL is in the broad areas of intelligent systems, information and control, includes but not limited to: 1. Intelligent Systems and Innovative Computing Artificial intelligence and expert systems, approximate reasoning, fuzzy systems and rough sets, fuzzy optimization and decision making, evolutionary computation, intelligent information systems, image and speech signal processing, human-machine learning, neural networks, knowledge discovery, data mining, clustering and data analysis, decision support systems, intelligent fault detection and identification, knowledge-based systems, hybrid intelligent systems, soft computing, management science 2. Information and Control Information theory and applications, optimal, adaptive, predictive, robust, fuzzy and reliable control and filtering on linear and nonlinear systems, stochastic systems, hybrid systems, supervisory systems, guidance and navigation systems, real-time systems, discrete event systems, signal and image processing, system identification, optimization techniques, robustness analysis, learning systems, fault detection, robotics and motion control, manufacturing systems, power systems, networked control systems, environmental and bio-engineering, complex and large scale systems, bioinformatics, control education, transportation systems Subscription Library/Institutional/ Individual Rates: Hard copy (Printing): JPY33,000 CD: JPY16,500 Hard copy with CD: JPY45,000 The charge includes postage, packing, handling and surface delivery of all issues. ICIC-EL Editorial Office Tokai University, Kumamoto Campus 9-1-1, Toroku, Kumamoto 862-8652, Japan Phone: +81-96-386-2666 Fax: +81-96-386-2666 E-mail: [email protected] 2010 ICIC International ISSN 1881-803X Printed in Japan ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1735–1740 HEURISTICS FOR JOINT DECISIONS IN DOMAIN IMPLEMENTATION AND VERIFICATION Zhiqiao Wu and Jiafu Tang Department of Systems Engineering Key Lab of Integrated Automation of Process Industry of MOE Northeastern University Shenyang 110004, P. R. China [email protected]; [email protected] Received February 2010; accepted April 2010 Abstract. One of the critical objectives of software product line (SPL) development is minimizing the development cost of software systems. Thus, costs of domain implementa- tion and verification, which may involve substantial expenses in SPL development, should be reduced. In addition, both should not be considered individually but in an integrated manner to further reduce development cost. In this paper, an integrated decision model for SPL development is proposed to assist decision-makers in selecting reuse scenarios for domain implementation and in simultaneously determining the optimal number of test cases for domain verification. An objective of the model is to minimize SPL develop- ment cost while satisfying the required system and reliability requirements. The Lagrange relaxation decomposition (LRD) method with heuristics was developed to solve integrated decision problems. Based on LRD, the non-linear model boils down to a 0-1 knapsack problem for the sub-problem of reuse scenario selection and an integer knapsack problem of the number of tests assignment sub-problem. Keywords: Software product line, Integrated decisions, Lagrange relaxation 1. Introduction. In the past, reuse-based software development has been implementati- on-centric. On account of increasing scale and complexity as well as decreasing develop- ment cost and time of software products for maintaining market competitiveness, it is difficult to develop reliable software products using conventional techniques [1]. There- fore, a tradeoff between cost and quality of reuse-based software products has become a major concern in software product development. To address this issue, various software engineering techniques, such as reuse economics and analysis/design methods, have been developed. Software product line (SPL) engineering ranks among the most advanced software practices based on these results [2]. One of the important and specific activities of SPL is domain engineering. Here, products in a product line are assembled from the organization’s core assets. Each core asset has an attached process for analysis and com- parison of reuse scenarios or approaches that include without reuse, with white-box reuse, and with black-box reuse. A design of domain implementation involves the inclusion of descriptions of how the attached processes cooperate to yield products, and how the or- ganization builds products for SPL. The next step after implementation is verification, which validates and verifies reusability for each core asset’s specifications. After successful testing of core assets in different use cases and scenarios, the domain engineering phase is complete and core assets are ready for reuse. One of the major benefits of software product line development (SPL) is its ability to curb software-intensive systems development cost. To achieve this goal, previous studies have examined cost optimization of domain engineering in the implementation [3] and verification processes [4,5]. However, verification at all times accounts for between 30 to 50% of a project’s effort [6]. Separate consideration of implementation and verification 1735 1736 Z. WU AND J. TANG cannot yield optimum development. Hence, more effort has been placed on increasing efficiency of implementation cost estimation [7], as well as testability technology for im- proving the efficiency of verification cost estimation [8,9]. Particularly, in recent studies with the advanced methodology of test cost estimate, models for cost optimization have tended to focus on coordinating joint decisions on the whole life cycle extending the system implementation decision to the quality verification management [10]. Afterwards, domain engineering decisions should be made in an integrated way so as to lead substantial savings in the main expenditures of SPL development. However, due to the complex and nonlinear relationship between the implementation and verification processes, research on models or algorithms optimizing development cost while simultaneously considering implementation and verification have received little at- tention. This paper attempts to tackle these still unsolved problems where domain imple- mentation and verification built by multiple reuse scenarios and multiple optional numbers of testing considered jointly in order to achieve cost savings. In this paper, an integrated decision model (IDM) is provided by introducing 0-1 variables that represented alternative industrial reuse scenarios. Integer variables are employed to represent the number of tests on alternative scenarios regarding cost objective and reliability constraint simultaneously. IDM not only selects the most cost-effective reuse scenarios but also indicates the proper optimal testing to detect faults of each scenario, with consideration of the entire SPL reliability and system requirements. An examination of non-linear model indicates that integrated decision procedures are a two-layer decision. Hence, a Lagrange relaxation decomposition method – based on a relaxation technique and decomposition method – is developed and proposed in this work to solve the problem effectively. Following the introduction, Section 2 shows the objective function formulation and constraints presentation. Then, in Section 3 gives integrated decision process. A Lagrange relaxation decomposition method to solve the problem is provided in Section 4. And finally, concluding remarks are given in Section 5. 2. Cost Analysis and Objective Function. To formulate the problem, the following notations are used throughout the paper: n Number of modules within a target SPL system; mj Number of reuse scenarios available for module j; xij Decision variable, which is equal to 1 when module j is implemented using scenario i, and 0 if otherwise; zij An integer decision variable model that represents total number of tests performed on module j when implemented using scenario i. Implementation cost formulation: Wu and Tang [7] summarized the alternative reuse scenario in a product line. They also discussed that the implementation cost analysis method which enabled product line developers to automatically evaluate and compare all possible alternative reuse scenarios implementation cost Cij. The objective function of implementation cost can be formulation as Cij × Xij. Verification cost formulation: The cost of testing is largely a function of the number of tests performed z on each component’s reuse scenario; assuming a linear function this would be τ × z where τ is the cost per test [11]. Objective function F is formulated to represent the combined cost of implementation and verification to be minimized. ∑n ∑mj ∑n ∑mj F (xij, zij) = Implementation cost + Verification cost = cijxij + τijzij (1) j=1 i=1 j=1 i=1 Objective functions to be considered in the reuse scenario selection problem are designed to minimize cost of implementation; they are likewise geared towards minimizing the number of test assignment problems to minimize cost of verification phase. ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1737 Formulation of reliability constraints. Based on the Testability and Bayes’ the- orem, Wu and Tang [7] presents a testability function, ρ(z) (equal to (1 − η) − η(1 − η)2 ln(1 − η)z), which is the probability that an instance is failure-free during a single run given that z test cases have been performed. The following constraint can be obtained for each module: ρij(zij) ≥ (1 − εj)xij ∀i, j (2) Parameter εj is accepted level of failure rate for module j in the SPL system and ηij is testability of reuse scenarios i for module j. However, estimated reliability may not improve in proportion to the number of success- ful tests. Growing reliability will end once the test has covered all necessary operations, i.e. obtained test cover test Γij. The following formulation can be obtained: ∑n ∑mj ∑n ∑mj ∗ zij ≤ Γijxij (3 ) j i j i Following Oded’s work [12], the effect of module failures can be expressed as an expo- nential function with Sojourn time (sj) of each module. Constraining the reliability of a SPL is to constrain the probability of at least one failure occurrence in a system during the execution of a run. It can be expressed as follows: ∑n ∑mj ∑n ∑mj ∑n ∑mj sij(1 − ρij(zij))xij ≤ εJ or sijρij(zij)xij ≥ sijxij − εJ (4) j=1 i=1 j=1 i=1 j=1 i=1 Parameter εJ is accepted level of failure rate for the whole SPL system. Formulation of goal satisfaction constraints. In view of the user satisfaction evaluation [13] and goal-oriented requirement engineering approach [14], the performance of module j with scenario i; it can be evaluated in terms of a weighted satisfaction of∑ each goals Satij(k) based on operational distribution χijk formulated as Satij = χijk × Satij(k). k Thus, regardless of which scenario in a module is reused, the targeted software system should fulfill an accepted overall satisfaction level of at least R. It is given as follows: ∑n ∑mj ωjSatijxij ≥R (5) j=1 i=1 Parameter ωj is the weight of module j which expresses access reuse frequencies. Figure 1. Hierarchical scheme of software system 1738 Z. WU AND J. TANG 3. Integrated Decisions Model for Domain Implementation and Verification. Without loss of generality, it is assumed that an SPL will be developed containing n modules. Each can be built using an optional alternative reuse scenario. Each module must satisfy one or several functional requirements with required levels. The goal-oriented requirement engineering approach [14] matches the goals of a system and the features of reuse scenarios through modules. Afterwards, selected components for each module are tested several times to verify their reliability (Figure 1). With this hierarchical scheme, the decision problem involves the manner of selecting reuse scenario for building each module. It involves assigning the number of times tests must be done to verify each module of a targeted system based on minimum implementation and verification cost under the constraints of functional goal satisfaction and reliability requirements. Based on the above, it can be concluded that an integrated decision model (IDM) for minimizing both implementation and verification costs is as follows: ∑n ∑mj ∑n ∑mj min F (xij, zij) = cijxij + τijzij j=1 i=1 j=1 i=1 (6) s.t. (2), (3∗), (4) and (5) ∑mj xij = 1 ∀j (7) i However, the TCP or WTCP are NP-hard in general. Therefore, majority of developed algorithms will merely provide the low bound of the least testing when SPL has a signif- icant number of operations K. It is thus necessary to introduce a relaxation technique involving the conversion of hard constraint (3∗) into an objective to add a penalty to the latter if unsatisfied. On the other hand, the IDM is characterized by nonlinear program- ming. It can be solved using traditional nonlinear programming techniques. However, only the local optimal solution may be found. Therefore, a more effective decomposition method is required to solve the IDM model. 4. Lagrange Relaxation Decomposition (LRD) Method. The IDM may be viewed as a two-layer decision. The first layer involves the decision on reuse scenario selection at the implementation stage, and the second involves the decision on the number of test assignment at the verification stage. By combining Lagrange multipliers into a two-layer decomposition, an LRD method with heuristics is developed to solve the mixed integrated decision problem. Formulation (3∗) is a hard fixed constraint, as discussed in the previous section. It provides a bridge between decisions in implementation and verification. By ∗ introducing Lagrange multipliers δij to Constraint (3 ), IDM can be decomposed into two integrated decision sub-problems corresponding to the two layers. The first pertains to reuse scenario selection sub-problem (RSS), and the second refers to the number of tests assignment sub-problem (NTA). These are presented below. For the sub-problem RSS: ∑n ∑mj ∑n ∑mj cos t min F1(xij) = cijxij + δijΓij xij j=1 i=1 j i (8) s.t. (5) and (7) For the sub-problem NTA: ∑n ∑n min F2(zij) = τjzj − δjzj j=1 j=1 (9) s.t. (2), (3∗) and (4) ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1739 The last term in the objective function (8) of sub-problem RSS represents verification- related costs, while Lagrange multiplier δij is interpreted as the verification cost for de- livering unit cost required to perform a test case on module j with reuse scenarios i. For example, it can be determined in terms of a coefficient of verification cost as δij = τijλ. Coefficient λ of Lagrange multiplier δij satisfies 0 ≤ λ ≤ 1. It can be interpreted as the degree of the role of verification costs during implementation when formulating integrated decisions. For example, if the decision of the domain engineer depends entirely on imple- mentation costs, then λ = 0. For a specified instance of the IDM, optimal λ∗ reflects the actual impact of verification planning on decisions of implementation assignment. A key point in LRD is the manner of traversal of all values of λ to obtain the global optimal solution. Conventionally, an initial value is set (e.g., λ = 0), which increases by a fixed step length (e.g., ∆λ = 0) until the stop value is obtained (e.g., λ = 1). However, it can be observed that both RSS and NTA are similar to the integer knapsack problem. Dantzig [15] demonstrated that an optimal solution for the continuous 0-1 knapsack problem may be obtained by sorting items according to their non-increasing profit-to-weight ratios (called efficiencies), and adding them to the knapsack until capacity is reached. Thus, the results indicate that LRD is more effective than previous methods in determining global optimal solutions for integrated decision problems. 5. Conclusions and Further Work. One of the major problems of domain engineer- ing involves the effective coordination of domain implementation and verification. More specifically, it is about the integration of selecting reuse scenarios at the implementa- tion phase and determining the number of tests to be performed for each scenario at the verification phase. Research described in this paper considers the joint decisions in selecting reuse scenarios and determining optimal number of tests when developing an SPL towards minimizing cost, while satisfying system and reliability requirements to a certain degree. An LRD method with heuristics is proposed to solve the integrated deci- sion problem. Based on the LRD, integrated decision problem is solved by decomposing it into two subproblems in two different layers. The contributions of this paper lie in two aspects. First, an integrated decision problem was formulated by combining decisions on domain implementation and verification. A stop-test rule mandating an extension of the testability definition was proposed for effectively developing integrated decision problem. Second, a framework and an overall procedure of LRD combined with LMDH approach were developed to solve the integrated decision problem. Acknowledgment. This work was financially supported by the National Natural Sci- ence Foundation of China (NSFC) (70721001, 70625001), the Fundamental Research Funds for the Central Universities (N090604004), the National 973 Basic Research Project (2009CB320601). REFERENCES [1] T. Kishi and N. Noda, Formal verification and software product lines – Using formal verification techniques to verify designs within a product line, Communications of the ACM, vol.49, no.12, pp.73-77, 2006. [2] P. Clements and L. Northrop, Software Product Lines: Practices and Patterns, Addison-Wesley, Boston, 2002. [3] Y. J. Li and H. Gao, A domain optimization problem with boundary penalty cost functional, Journal of Mathematical Analysis and Applications, vol.301, no.1, pp.170-186, 2005. [4] R. M. Hierons, Verdict functions in testing with a fault domain or test hypotheses, ACM Transactions on Software Engineering and Methodology, vol.18, no.4, pp.1-19, 2009. [5] Y. Liu, C. Wei and S. Gao, Research on software reliability modeling based on stochastic petri nets and module level software rejuvenation, ICIC Express Letters, vol.3, no.3(B), pp.607-613, 2009. [6] M. Ellims, J. Bridges and D. C. Ince, The economics of unit testing, Empirical Software Engineering, vol.11, no.1, pp.5-31, 2006. 1740 Z. WU AND J. TANG [7] Z. Q. Wu, J. F. Tang and L. Y. Wang, An optimization framework for reuse scenarios selection in software product line, Proc. of the Chinese Control and Decision Conference, Guilin, China, pp.1880-1884, 2009. [8] J. M. Voas and K. W. Miller, Software testability – The new verification, IEEE Computer Society, vol.12, no.3, pp.17-28, 1995. [9] V. Cortellessa, F. Marinelli and P. Potena, Automated selection of software components based on cost/reliability tradeoff, Lecture Notes in Computer Science, pp.66-81, 2006. [10] A. Kleyner and P. Sandborn, Minimizing life cycle cost by managing product reliability via validation plan and warranty return cost, International Journal of Production Economics, vol.112, no.2, pp.796- 807, 2008. [11] D. B. Brown, S. Maghsoodloo and W. H. Deason, A cost model for determining the optimal number of software test cases, IEEE Transactions on Software Engineering, vol.15, no.2, pp.218-221, 1989. [12] O. Berman and M. Cutler, Optimal software implementation considering reliability and cost, Com- puters and Operations Research, vol.25, no.10, pp.857-868, 1997. [13] L. Lin and H.-M. Lee, A new algorithm of software quality evaluation for user satisfaction, Inter- national Journal of Innovative Computing, Information and Control, vol.4, no.10, pp.2639-2648, 2008. [14] A. V. Lamsweerde and E. Letier, Handling obstacles in goal-oriented requirements engineering, IEEE Transactions on Software Engineering, vol.26, no.10, pp.978-1005, 2000. [15] G. B. Dantzig, Discrete-variable extremum problems, Operations Research, vol.5, no.2, pp.226-288, 1957. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1741-1746 A FIRST APPROACH TO ARTIFICIAL COGNITIVE CONTROL SYSTEM IMPLEMENTATION BASED ON THE SHARED CIRCUITS MODEL OF SOCIOCOGNITIVE CAPACITIES Alfonso Sanchez´ Boza1 and Rodolfo Haber Guerra1,2 1Centro de Autom´aticay Rob´otica Consejo Superior de Investigaciones Cient´ıficas- Universidad Polit´ecnicade Madrid CP 28500, Arganda del Rey, Madrid, Spain [email protected] 2Escuela Polit´ecnicaSuperior Universidad Aut´onomade Madrid CP 28049, Madrid, Spain [email protected] Received February 2010; accepted April 2010 Abstract. A first approach for designing and implementing an artificial cognitive con- trol system based on the shared circuits models is presented in this work. The shared circuits model approach of sociocognitive capacities recently proposed by Hurley [1] is en- riched and improved in this work. A five-layer computational architecture for designing artificial cognitive control systems is proposed on the basis of a modified shared cir- cuits model for emulating sociocognitive experiences such as imitation, deliberation, and mindreading. An artificial cognitive control system is applied for controlling force in a manufacturing process that demonstrates the suitability of the suggested approach. Keywords: Artificial cognitive control, Embodied cognition, Imitation, Internal model control, Mirroring, Shared circuits model 1. Introduction. There are many definitions of intelligence, one of them is the ability of human beings to perform new, highly complex, unknown or arbitrary cognitive tasks efficiently and then explain those tasks with brief instructions. It has spurred many researchers in areas of knowledge such as control theory, computer science, and artificial intelligence (AI) to explore new paradigms to achieve a qualitative change and then to move from intelligent control systems to artificial cognitive control strategies [2]. A natural cognitive system displays effective behaviour through perception, action, deliberation, communication, and both individual interaction and interaction with the environment. During cognitive or executive control, the human brain and some animal brains process a wide variety of stimuli in parallel and choose an appropriate action (task context), even in the presence of a conflict of objectives and goals [3, 4, 5]. There is a wide variety of strategies and initiatives related with the partial or full emulation of cognitive capacities in computational architectures. Each one is based on a different conception of the nature of cognitive capacity, what makes a cognitive system, and how to analyze and synthesize such a system [6, 7]. The novelty of this work is that it is based on an neuroscientific and psychological approach, the shared circuits model (SCM) [1]. SCM approach serves as the foundation for designing an artificial cognitive control system where imitation, deliberation, and mindreading processes are emulated through computational efficient algorithms in a computational architecture. Hurleys approach suggests that these capacities can be achieved just by having control mechanisms, other- action mirroring and simulation. An artificial cognitive control system should incorporate these capacities and therefore it would be capable of responding efficiently and robustly 1741 1742 A. SANCHEZ´ BOZA AND R. HABER GUERRA to nonlinearities, disturbances and uncertainties. The modifications introduced to the SCM approach make that this preliminary version can be applied to design a control architecture for a case study: a high-performance drilling process. In order to improve efficiency of a high-performance drilling process, the current study focuses on the design and implementation of a control system for drilling force. This article is organized into four sections. The modified shared circuits model (MSCM) incorporated to an architecture in which is implemented an artificial cognitive control system is explained in Section 2. Section 3 shows the experimental results of applying an implementation of the MSCM to a case study represented by a high-performance drilling process. Finally, the conclusions are presented in Section 4. 2. An Architecture for Artificial Cognitive Control. Modified Shared Circuits Model. SCM approach is supported on a layered structure to describe how certain hu- man capacities (i.e., imitation, deliberation, and mindreading) can be deployed thanks to subpersonal mechanisms of control, mirroring, and simulation (Figure 1). Basically, SCM is based on the observation of the human brain. It may be envisaged as making use of not only inverse models that estimate the necessary motor plan for accomplishing an objective in a given context, but also a forward model that enables the brain to anticipate the perceivable effects of its motor plan, with the object of improving response efficiency. Figure 1. Depiction of SCM. Layer 5 monitors simulation of input acts or evoking objects. Using SCM layers 2 and 3, SCM can perform simulation at both ends and, with layer 5, enables strategic deliberation. A computational architecture for an artificial cognitive control system is proposed for high-performance manufacturing processes, underpinned by the modified shared circuits model (MSCM, Figure 2). Therefore, it is necessary to enrich SCM approach from a computational science viewpoint. To develop a complex cognitive agent, it is necessary to make a global structure that would be a collection of information processing elements, linked by information forwarding elements layered a top physical/information interfaces [8]. Modifications to SCM are introduced to enrich and improve its capacities, taking into account the suggestions reported in the state-of-the-art and the main constraints of the SCM approach. Since a layer-based model is incorporated in a computational architecture, five modules are constructed made up of one or more processes performed by the above- described layers. At MSCM, module 1 is represented by a controller and an optimization/adjustment pro- cess for this controller to adapt it to new environmental conditions. Module 3 supports a set of inverse models that obtain an imitative output action and module 2 generates future output prediction to deliberate about module 1 actions or this module 3 imita- tive actions. Module 4 is in charge of inhibiting either module to distinguish between information from the agent himself and information from others, and MSCM module 5 gathers associations between dictated action and the possible, or counterfactual, inputs (i.e. the reaction of others). In this way, a functional parallelism between layers in SCM ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1743 and modules are established. The rest of modules 0 and a are introduced to take part in the decision of when modules act observing the control process success. Module 1 and 2 manage set of inverse and forward models respectively, and they optimize them in function of environmental changes. Several solutions could be used, like [9] or [10] if the model’s parameters are extensive. This is a modular approach that manages models, so each module can be developed separately. From the best of authors knowledge, the main novelties of this work are twofold. Firstly, the SCM is enriched and improved on the basis of the state of the art. For example, SCM does not cover the description of an executive level to manage functioning of layers, so modules in charge of this task are also introduced. Secondly, the implementation of an artificial cognitive control system has been developed using the enriched SCM, and the application to a high-performance manufacturing process corroborates on the basis of experimental results the suitability of the suggested approach. Figure 2. Simplified block diagram of the MSCM proposed system. In MSCM as in SCM, the relationships and interactions among modules make possible to artificially emulate the cognitive capacities of deliberation, imitation, and mindreading. So, in order to develop a computational framework aiming at control system design, it is necessary to address module interactions. Layers interactions is one of the main weaknesses of the SCM approach. This temporal pattern is essential to clarify how and when the system acts, and therefore it is necessary to set a method or strategy to establish a sequence of actions for each module. MSCM explores possible operating sequences of the different modules (Figure 3) attending the results already reported in the literature [1, p.8,14,15][11, 12, 13]. 1. Deliberate the possible actions to carry out, by means of interaction of modules 1, 2, and 5. 2. If any of the actions leads to success, execute it. (a) If there is noise in excess, module 2 learns the new effects that have been produced. 3. If not, (a) If noise surpasses a threshold, go to (3.iii.A). (b) If not, (i) Deliberate about the possible actions of others (imitative actions), through the interaction of modules 2, 3, and 5. (ii) If any of the imitative actions leads to success, execute it. (A) Learn this action by incorporating the corresponding instrumental association into module 1’s private set of forward models. (iii) If not, (A) Through an optimization process, acquire a new action using the process model handled by module 2, whose results are handled by said optimization process in module 1. (B) Execute a new action by means of the operation of this module. Figure 3. Algorithm of the systems action, imitation, and learning cycle. 1744 A. SANCHEZ´ BOZA AND R. HABER GUERRA 3. MSCM as the Basement to Control a Complex Process. An Application to the Case Study of High-Performance Drilling Process. In order to demonstrate the viability of an artificial cognitive control system based on MSCM foundations, the authors have selected the drilling process. Drilling is one of the most intensely used processes in the manufacturing of aircraft parts, automobile parts, and molds and dies in general. And the modified SCM (MSCM) proposed in this work offers the necessary theoretical framework to design an artificial cognitive control system. Here a neurofuzzy system like ANFIS [15] is exploited to combine the semantic transparency and intrinsic robustness of fuzzy systems [16] with the learning ability of artificial neural networks. The most adequate paradigms can be selected from among the extensive choice provided by Control Theory, Artificial Intelligence and its techniques, Computer Science, System Theory and Information Theory. The study and selection of the most viable method from among the many offered by the myriad topics mentioned above is not straightforward. One way is to exploit the advantages of neurofuzzy systems, as mentioned before. Their ca- pacities for incorporating experience, for learning, and for adaptation and self-adjustment [17] are aligned with the capacities available in the MSCM for emulating sociocognitive experiences such as imitation, deliberation, and mindreading. The role of the internal models in the MSCM approach opens the possibility of using the design method provided by Internal Model Control paradigm [18], that theoretically guarantees control system robustness and stability in the presence of external disturbances. The role of modules 2 and 3 and their interaction can be effectively represented by neurofuzzy systems (e.g., ANFIS) and the IMC scheme. The IMC paradigm and the ANFIS neurofuzzy system are used in this work to provide functioning capabilities to the modules 2 and 3. The main rationale of using both methods was previously explained. ANFIS is used to generate direct models (the model used in module 2) and inverse models (the model used in module 3) which shape the knowledge on which the system is initially based. Figure 4. A preliminary MSCM control architecture implemented in a real-world process. A networked control is performed through a distributed implementation enriching the solution designed and implemented by Gajate et al. [19]. In our case two machines are used: one receives the user’s goals and the values measured by the sensors and it does executive functions, while the other emulates cognitive processes (Figure 4). And ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1745 MSCM uses several modules to perform an artificial cognitive control instead of one unique controller, as the Gajate’s solution. Communication between machines is achieved through the intermediary communications system Ice. So that communications are transparent to the network features and the machines that perform these processes (Figure 5). interface Cognetcon { bool handshake(); idempotent bool sendNumVars(short numVars); idempotent double sendValuesRecvAction(FloatList values); }; interface Mod1Server { idempotent void init(string controllerParamFile, double a,double b,double c) throws Error; idempotent double control (double error, double reference, int responseIndex); }; interface Mod2Server { idempotent double response(double controlOutput, int responseIndex); }; Figure 5. Passage from the definition of the operations that support the executive and cognitive servers. The main goal is to obtain a good transient response without overshoot using the cutting parameters given by the tool manufacturer for this tool and workpiece material combination. Experimental trials are conducted using a machining center equipped with an open computer numerical controller (Figure 6). Figure 6. Resultant force y of the control exercised by the action of the modules 2 and 3 of MSCM. 4. Conclusion. This paper presents a first approach to design an artificial cognitive control system from the conceptual framework that is given by the shared circuit models approach to emulate sociocognitive capacities. The shared circuit model (SCM) approach is enriched and improved using the state-of-the-art on this field. Moreover, relevant reports 1746 A. SANCHEZ´ BOZA AND R. HABER GUERRA on this issue as well as the contributions of the authors are also outlined. The modified SCM (MSCM) is postulated from the viewpoint of System Theory and Computer Science. Therefore, a sequence of operating MSCM is proposed in this paper. The suggested procedure enables the activity of combined modules or a module alone. In general, MSCM goes beyond the SCM approach with the aim of a deriving a computational solution to a complex control problem. From a theoretical point of view, MSCM provides an alternative conceptual framework to perform control tasks in an efficient, robust fashion that characterizes human cognitive processes. Acknowledgment. This work was supported by DPI2008-01978 Networked cognitive control system for high-performance machining processes (COGNETCON) of the Spanish Ministry of Science and Innovation. REFERENCES [1] S. Hurley, The shared circuits model (SCM): How control, mirroring, and simulation can enable imitation, deliberation, and mindreading, Behavioural and Brain Science, vol.31, no.1, pp.1-22+52- 58, 2008. [2] J. Albus, Toward a computational theory of mind, Journal of Mind Theory, vol.0, no.1, pp.1-38, 2008. [3] R. Huerta and T. Nowotny, Fast and robust learning by reinforcement signals: Explorations in the insect brain, Neural Computation, vol.21, no.8, pp.2123-2151, 2009. [4] M. Ito, Control of mental activities by internal models in the cerebellum, Nature Reviews Neuro- science, vol.9, no.4, pp.304-313, 2008. [5] M. I. Rabinovich, P. Varona, A. I. Selverston and H. D. I. Abarbanel, Dynamical principles in neuroscience, Reviews of Modern Physics, vol.78, no.4, 2006. [6] Z. Pylyshyn, Computation and Cognition: Toward a Foundation for Cognitive Science, MIT Press, 1984. [7] L. B. Thelen and E. Smith, A Dynamic System Approach to the Development of Cognition and Action, MIT Press, Cambridge, 1994. [8] R. Sanz, C. Hern´andez,A. Hernando, J. G´omezand J. Bermejo, Grounding robot autonomy in emotion and self-awareness, Lecture Notes in Computer Science, vol.5744, pp.23-43, 2009. [9] D. Martin, R. M. del Toro, R. E. Haber and J. Dorronsoro, Optimal tuning of a networked linear controller using a multi-objective genetic algorithm and its application to one complex electrome- chanical process, International Journal of Innovative Computing, Information and Control, vol.5, no.10(B), pp.3405-3414, 2009. [10] J. Li, X. Hu, Z. Pang and K. Qian, A parallel ant colony optimization algorithm based on fine- grained model with GPU-acceleration, International Journal of Innovative Computing, Information and Control, vol.5, no.11(A), pp.3707-3716, 2009. [11] M. Nielsen, The social motivation for social learning, Behavioral and Brain Sciences, vol.31, no.1, pp.33, 2008. [12] T. Makino, Failure, instead of inhibition, should be monitored for the distinction of self/other and actual/possible actions, Behavioral and Brain Sciences, vol.31, no.1, pp.32-33, 2008. [13] A. Meltzoff, Imitation and other minds: The “like me” hypothesis, in Perspectives on Imitation: From Neuroscience to Social Science – Two Volume Set, S. Hurley and N. Chater (eds.), The MIT Press, 2005. [14] M. Tomasello, The Cultural Origins of Human Cognition, Harvard University Press, 1999. [15] J.-S. R. Jang, Anfis: Adaptive-network-based fuzzy inference system, IEEE Transactions on Systems, Man and Cybernetics, vol.23, no.3, pp.665-685, 1993. [16] R. E. Precup, S. Preitl, J. K. Tar, M. L. Tomescu, M. Takacs, P. Korondi and P. Baranyi, Fuzzy con- trol system performance enhancement by iterative learning control, IEEE Transactions on Industrial Electronics, vol.55, no.9, pp.3461-3475, 2008. [17] T. Fukuda and N. Kubota, An intelligent robotic system based on a fuzzy approach, Proc. of the IEEE, vol.87, no.9, pp.1448-1470, 1999. [18] M. Morari and E. Zafiriou, Robust Process Control, Prentice Hall, 1989. [19] A. Gajate, R. E. Haber and R. M. del Toro, Neurofuzzy drilling force-based control in an ethernet- based application, International Journal of Innovative Computing, Information and Control, vol.6, no.4, pp.373-386, 2010. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1747-1752 STUDY OF SHORE-BASED AIS NETWORK LINK CAPACITY Chang Liu1 and Xinyu Wang2 1School of Information and Science Technology Dalian Maritime University No. 1 LingShui Road, High-Tech Zone, Dalian, P. R. China [email protected] 2Industry Technology Supervision and Research Center for Building Materials Beijing, P. R. China [email protected] Received February 2010; accepted April 2010 Abstract. This paper put forward a method of the capacity estimation of the shore- based AIS (Automatic Identification System) network, which is available to ensure the vessel data link more stable and reliable. Based on the analysis of the technical character- istics and transmission parameters of the shore-based AIS, we discussed the problem of communication conflict caused by the slot reuse which affects the capacity of shore-based AIS network. By experiment we proposed the demarcation method of shore-based AIS coverage and the design of the data link capacity which will help the vessel traffic service (VTS) management more efficient. Keywords: Automatic identification system (AIS), Vessel data link (VDL), Capacity, Net throughput, Slot reuse 1. Introduction. Automatic Identification System (AIS) is a VHF digital mobile com- munication system based on self-organized TDMA (SOTDMA) technique. Some research about the capacity of ship borne AIS in a certain water area was carried out. But with the construction of the AIS network and shore-based facilities, the number of AIS land users is constantly increasing, which will have more information that transmitted through the AIS vessel data link (VDL) [1-3]. As the AIS services are security-related informa- tion transfer services, to ensure the VDL stable and reliable is essential for the vessel traffic management. How many users can the accommodate to shore-based network and whether can its capacity meet the growth requirements of the communication and also the surveillance and security information transmission in the future is an important issue. Based on the analysis and experiment of AIS data link characteristics, we discussed the capacity of the shore-based AIS network. 2. AIS Data Link Characteristics. As a mobile data communication system, the outstanding characteristic of AIS is that the mobile station can communicate with each other using self-organization protocol without the control of the base station [2]. Each AIS station has its own communication cell. The cell is a circle with the station as the center, and the radius D usually depends on the transmitting and receiving antenna height estimated by (1). (√ √ ) D = 3.6 h1 + h2 km (1) Inside (1) h1 and h2 are the height of receiving and transmitting antennas. As a cellular mobile communication system, AIS uses SOTDMA protocol to communicate and realize the share of frequency and slots. The slot reuse will occur in these two conditions [4]. (1) Intentional slot reuse: a ship closer to the shore station may occupy the slot schedule of a distant ship when the link capacity is overload. 1747 1748 C. LIU AND X. WANG (2) Automatic slot reuse: two or more ships that are non-visible to each other, that is, they are not in the same communication cell, may send data packets using the same slot. Slot reuse may affect the shore station correctly receiving the reports from the ship AIS, which will result in two conditions below [5]. (1) Confusion: AIS shore station can not correctly receive data packets from all AIS ships at the same slot. (2) Discrimination: AIS shore station can correctly receive data packets only from one of all ships in the same slot. Confusion and discrimination contribute to reduce the rate of AIS report from ship AIS to shore station. The report rate of ship AIS stations is usually evaluated by the two parameters as follows. (1) Nominal report rate (Rr): the desired report rate received by AIS stations according to the standard [2]. The maximum rate is 30 per minute for the dynamic information. (2) Net report rate (Rn): the actual report rate received by AIS stations which may be less than the nominal report rate because of the slot reuse or interference. As we have known AIS is not a connection-oriented system but a broadcast system. And also the AIS information is real time, so the miss data will not be retransmitted. In order to obtain the processing capacity of AIS and the transmission quality of the data link we proposed two parameters as follows. (1) Required capacity (C): the maximum number of AIS ship stations that the shore-base station can accommodate within its communication cell. (2) Net throughput (Nt): the ratio of actual report rate and the nominal report rate from the same sender [3]. Net throughput of link is defined as Nt = Rn/Rr (2) Because of slot reusing and interference, Nt will less than 1. These two parameters are very important for the design of AIS shore station. In the density traffic area the load of AIS data link is tend to be full, the receive data confusion and discrimination will occur which will result in the reduced net throughput. By experiment, the relationship curve of slot reuse and the required capacity is shown as in Figure 1. When the required capacity is 150%, the time slot reuse is nearly 80%, the net throughput of the AIS system may be less than 20%, and the system’s normal data transmission will be difficult to be assured. Figure 1. Curve between slot reuse and the required capacity ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1749 3. Calculation of Link Capacity and Net Throughput. From the analysis above, the capacity increases, the net throughput will decrease. In order to analyze the relation- ship about the capacity and the net throughput we suppose that the transmit power, the antenna gain and the receiver level of each AIS ship station are all equal. Suppose the distances of Ship 1 and Ship 2 away from the shore are respectively d1 and d2, shown in Figure 2. Generally, when 0.5 < d2/d1 < 2 or 0.5 < d1/d2 < 2 (3) Confusion occurs. The discrimination occurs when d2/d1 > 2 or d1/d2 > 2 (4) Figure 2. Relationship of d1 and d2 The shore station AIS can correctly receive data from Ship 1 or Ship 2, in other words, the shore station can correctly receive data from ship AIS which is closer. Figure 3. Region partition of shore-base AIS station Assume the communication cell radius of the shore station AIS is R, as in Figure 3, the distance from the shore station to the ship AIS is d. When R > d > R/2 and another ship in the region is not visible, automatic slot reuse may occur in the case of full link. Because the distance from ship to shore station does not meet (4), conflict ship data packet will be confusion. For the shore station, the AIS ships in the region selected time slot randomly, 1750 C. LIU AND X. WANG similar to the slot ALOHA system, so the region is defined as the ALOHA region. In time slot ALOHA system, the maximum transmission success probability [6,7] is: 1 ρ = (5) e When the required capacity reaches 100%, in ALOHA region the net throughput is Nt ≈ 37% (6) When R/2 > d > R/3 and another ship in the region is visible, ship communication maybe conflict with that in the ALOHA region. Suppose the distance from the ship T to the shore station is d = 0.4R, then the ship may be discriminated to the ship that the distance is bigger than 0.8R. But it does not affect the net reporting rate of ship T. But for the ships that the distance is less than 0.8R in ALOHA region, slot confusion maybe happen when communication conflict, resulting in the reduced net report rate of ship T. So the region R/2 > d > R/3 is called discrimination region. In the region R/3 > d, the ship will only conflict with ships of the distance d > 2R/3 (in the ALOHA region). At this point, if two ships conflict according to (4) the discrimination will happen, but does not affect the net reporting rate of ships at close range. So the net throughput in the region is 100%, and the region is called protected region. Table 1. Simulation result of net throughput and required capacity in different regions Required capacity (C) d/R Net throughput (Nt) 10% 50% 100% 150% 200% 0 .00 1 .00 1 .00 1 .00 1 .00 1 .00 0 .05 1 .00 1 .00 1 .00 1 .00 1 .00 0 .10 1 .00 1 .00 1 .00 1 .00 1 .00 0 .15 1 .00 1 .00 1 .00 1 .00 1 .00 0 .20 1 .00 1 .00 1 .00 1 .00 1 .00 0 .25 1 .00 1 .00 1 .00 1 .00 1 .00 0 .30 1 .00 1 .00 1 .00 1 .00 1 .00 0 .35 0 .94 0 .75 0 .60 0 .50 0 .45 0 .40 0 .92 0 .68 0 .48 0 .36 0 .29 0 .45 0 .90 0 .63 0 .40 0 .26 0 .18 0 .50 0 .90 0 .61 0 .37 0 .22 0 .14 0 .55 0 .90 0 .61 0 .37 0 .22 0 .14 0 .60 0 .90 0 .61 0 .37 0 .22 0 .14 0 .65 0 .90 0 .61 0 .37 0 .22 0 .14 0 .70 0 .90 0 .61 0 .37 0 .22 0 .14 0 .75 0 .90 0 .61 0 .37 0 .22 0 .14 0 .80 0 .90 0 .61 0 .37 0 .22 0 .14 0 .85 0 .90 0 .61 0 .37 0 .22 0 .14 0 .90 0 .90 0 .61 0 .37 0 .22 0 .14 0 .95 0 .90 0 .61 0 .37 0 .22 0 .14 1 .00 0 .90 0 .61 0 .37 0 .22 0 .14 From the analysis above, when the required capacity is 100%, the net throughput of AIS shore stations for the ship in different regions is different. The region can be divided into three parts, respectively named as ALOHA region, discrimination region and protected region. Through the theoretical calculations and simulation analysis, we got the results of the relationship of the net throughput (Nt) and the required link capacity (C) in different ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1751 regions (d/R), as shown in Table 1. Within the coverage of the AIS shore station, the net throughput in ALOHA region is lowest, and reduced with the required capacity increased. 4. Design of AIS Shore-based Station Link Capacity. The purpose of designing link capacity of AIS shore station is to obtain the lowest net throughput and the maximum capacity in the monitoring coverage of AIS shore stations, according to the required reporting rate of ships. The key problem about the link capacity analysis of AIS shore station is to make statistics and forecast of the traffic distribution in its coverage and determine the dense traffic areas [8]. Table 2 shows the vessel traffic distribution of a certain water area, in which the shore station coverage is 72 km (about 40 nautical miles). From Table 2, within the coverage of the shore station, communication of AIS ship dynamic information is 3415 slots per minute. Table 2. Distribution of traffic statistics in a certain water area Ship Type Ship Capacity Report Period (s) Report Times /min Anchor, <3kn 270 180 90 Anchor, >3kn 30 10 180 0-14kn 100 10 600 0-14kn and change course 15 3 1 /3 270 14-23kn 110 6 1100 14-23kn and change course 23 2 690 > 23kn 10 2 300 > 23kn and change course 2 2 60 Others 50 24 125 Total 610 3415 Because there are 4500 slots in AIS link [2], the dynamic information listed in Table 2 occupies 76% of the entire link capacity of AIS shore station. Consider that there is some other information to occupy the time slots, AIS shore station for the dynamic information of the link capacity will be less than 4500 slots. If the AIS dynamic information occupies 80% of the available capacity, there will be 232 slots for some other information. In this case, the net throughput is 45% in the ALOHA region (36km to 72km). For the ship whose nominal report rate is 30/min, the net reporting rate is calculated as 13.5/min according to (2), less than the required report rate for VTS surveillance [1]. In order to achieve the required rate of shore-based monitoring, according to Table 1, the required capacity of AIS shore-based station should be less than 50% in order to ensure the effective dynamic information of the vessels to be received. 5. Conclusions. In the SOTDMA communication mode, the coverage of shore-based AIS network can be divided into three regions according to the link characteristic, named as ALOHA, discrimination and protected region. In ALOHA region when the link capac- ity overloads, the net throughput will decrease evidently. By analyzing the relationship of the required capacity, the net throughput and the distance from the vessel to shore station, we design the capacity of shore-based AIS network. As a monitoring system, the capacity of the AIS stations is infinite [9]. But for the shore-based AIS, the network capacity is limited in the case of the required net throughput. By theoretical analysis and computer simulation, this paper discussed a single link capacity and its related charac- teristics parameters. For the network constructed by a number of AIS shore stations, the capacity analysis remains to be further studied. 1752 C. LIU AND X. WANG Acknowledgment. This work is partially supported by the Application Basic Research Project of Ministry of Traffic, China, under Grant No. 2005329225060. We also gratefully acknowledge the helpful comments and suggestions of the reviewers, which have improved the presentation. REFERENCES [1] IALA, Recommendation on Automatic Identification System (AIS) Shore Station and Networking Aspects Relating to the AIS Service, 2002. [2] ITU-R M.1371, Technical Characteristics for a Universal Shipborne Automatic Identification System (AIS) Using Time Division Multiple Access in the VHF Maritime Mobile Band, 2001. [3] IEC 61193-2, Maritime Navigation and Radio-communication Equipment and Systems – Automatic Identification System (AIS), 2002. [4] Y. Hu, Data link capacitance and congestion resolution for AIS, Computer Measurement & Control, pp.1631-1634, 2007. [5] R. Kjellberg, Capacity and Throughput Using a Self-Organized Time Division Multiple Access VHF Data Link in Surveillance Applications, http://www.gpcsweden.com, 2008. [6] X. Xie, Computer Network, Dalian University of Technology Publishing House, 2001. [7] G. Sun and M. Liaog, Upper bound for transport capacity of sensor networks, ICIC Express Letters, vol.3, no.1, pp.93-98, 2009. [8] S. J. Chang, Development and analysis of AIS applications as an efficient tool for vessel traffic service, MTS/IEEE OCEANS 2004, vol.4, pp.2249-2253, 2004. [9] Z. Yang and H. Imazu, Algorithm for cooperative collision avoidance through communication with automatic identification system, IEEE Computation Intelligence for Modeling, Control and Automa- tion, pp.1084-1088, 2005. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1753-1759 ONTOLOGY AND RULE COMBINED REASONING FRAMEWORK DESIGN Guanyu Li, Yi Liu and Buwei Chen Information Science and Technology College Dalian Maritime University Dalian 116026, P. R. China { rabitlee; sepc }@163.com; [email protected] Received February 2010; accepted April 2010 Abstract. It is known that ontology description language OWL DL has defects in re- gard to the expression and reasoning ability as its being restricted in description logic. Experience has proved that the combination of ontology and rule is an effective way to solve this problem. And by using SWRL, a novel framework of ontology and rule combined reasoning is proposed, which is named as SWRL Based Reasoning Mechanism Framework. In this framework, rule representation is introduced into OWL ontology, so the reasoning ability of OWL DL is significantly improved. Generally, SWRL based reasoning framework uses the descriptive logic based reasoning machine and Jess Tab to realize the transformation from OWL ontology to Jess fact base, and uses XSLT to realize the transformation from SWRL rule to Jess rule base, which results in the in- convenient operation. The proposed framework uses SWRL Bridge to combine Jess Rule Engine with Prot´eg´e-OWL, which can perform the two transformations mentioned above conveniently. The new OWL ontology obtained in such a manner includes the semantic relationships between the concepts which show the implicit knowledge in the former on- tology, so the former ontology is substantially improved semantically. The experimental result shows that the availability factor of ontology knowledge in semantic web domain can be improved to a large extent by use of this framework. Keywords: Domain ontology, Semantic web, OWL, Reasoning, SWRL, Jess 1. Introduction. Being increasingly popular, the Internet greatly changes our lives [1]. The development of ontology has been one of the motivations of semantic web since it was envisioned [2]. However, some important knowledge is given implicitly in the ontology, which causes low availability factor of ontology knowledge. Firstly, the limitation of the ontology description language OWL DL is analyzed. Sec- ondly, the SWRL-based Reasoning Framework is proposed, which is considered as a kind of ontology and rule combined reasoning framework. Lastly, based on ontology and rule, the implicit information in the ontology can be deduced by using the reasoning machine Jess. 2. Related Work. The sublanguage OWL DL of Web Ontology Language OWL defined by the W3C has the limitation to Description Logic. Being compared with OWL DL, the rule has more powerful ability to describe the logic. For example, the concept Uncle can be defined by the first-order logic as: P arent(?x, ?y) ∧ Brother(?y, ?z) → Uncle(?x, ?z). But OWL DL can not define such relationship. Therefore, the rule description should play an important role in Semantic Web, so the Semantic Web Rule Language was proposed. The organization DARPA proposed the Semantic Web Rule Language-SWRL on No- vember 2003, which mainly consists of four parts – Imp, Atom, Variable and Built-in [3]. Imp contains SWRL rules. In Imp, head describes the result of reasoning, and body de- scribes the basic form of the reasoning premise. In Imp, the basic ingredient of head and body is Atom. For example, all the clauses of Horn are used in its framework. Variable is 1753 1754 G. LI, Y. LIU AND B. CHEN to record the variable part in Atom. Built-in is the ingredient of the SWRL modulization which is utilized to record logic comparison relationship that the SWRL offers. Some specifications should be followed when editing SWRL rules. As an example, the specification of SWRL rules is as antecedent→consequent. antecedent and consequent are the extractions of the Atoms in the ontology, such as a1 ∧ ... ∧ an. Since SWRL rules consist of the head and body, it is supposed that S is SWRL knowledge base, OC is the set of the classes in OWL, OP is the attribute set of OWL, ST is the set of the constants of OWL and variables of SWRL, the specification of antecedent→consequent can be written as h1 ∧ ... ∧ hn ← b1 ∧ ... ∧ bm. Where hi, bj (1 ≤ i ≤ n, 1 ≤ j ≤ m) is Atom within C(i) (C ∈ OC , i ∈ ST ) and P (i, j)(P ∈ OP , i, j ∈ ST ). When the language SWRL is used to edit rules, the two points below should be followed. (1) All elements used in rules come from the classes or the attributes of the ontology; (2) The variables in the head of the rule must appear in the body of the rules, and it is impossible to introduce the new variables in the head of the rules, which is the safe restricted condition. In addition, when SWRL is used to edit rules, the rationality of rule descriptions is not checked, which may result in the following situation: the man-made rules and the OWL restriction contradict with each other. In brief, the prominent advantage of SWRL is the usage of ontology based rule language. The ontology language OWL supported by SWRL is the specification of W3C, and has more powerful ability to describe relationships. Meanwhile, SWRL also separates the ontology from the rule [4], and the rules described by the SWRL can be conveniently transformed to the rules in the present rule system. 3. The Rule Based Reasoning. Here, the reasoning machine of Jess is used to deal with reasoning part. Firstly, it transforms OWL ontology into the fact base, and trans- forms SWRL rules into the rule base. Then, it imports the facts and the rules into the working area of Jess reasoning engine to match the real facts and the rules. The rules are activated when the part antecedent of rule is satisfied. After matching, the activated rules are added to the pending area in order. Finally, the new facts will be gotten after the rules are run item by item. The new facts deduced are added into the fact base to serve the reasoning of the deeper level until the completion of the reasoning work. 4. SWRL-Based Reasoning Framework Design. Because the files in the format of OWL and in the format of SWRL can not be directly parsed by Jess reasoning engine, the ontology in the format of OWL and the rules in the format of SWRL should be transformed to the compatible format with Jess. To implement this mechanism, a sort of ontology and rule combined reasoning framework named as SWRL-based reasoning framework is given in Figure 1. The framework as shown in Figure 1 involves two transformations. The first is from OWL ontology to Jess fact base, and the other is from SWRL rule to Jess rule base. Prot´eg´e-OWL supplies a SWRL Bridge API. SWRL Bridge can be used to combine Jess Rule Engine with Prot´eg´e-OWL. Therefore, the framework can be used to realize the two transformations mentioned above. As to the two transformations, new reasoning fact base can be gotten by Jess reasoning, and then be transformed to the new OWL ontology. A. Transformation from OWL Ontology to Jess Factbase Because the reasoning engine Jess can not directly parse the ontology in the form of OWL, it must be transformed to the compatible form. During the transforming, two im- portant attributes must be showed: (1) the attribute of instance’s class; (2) the attribute of the instance. Besides, the information same-as and different-form in the instance must be changed. ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1755 Figure 1. SWRL based reasoning framewor Proverbially, ontology includes five basic elements, so the transformed object from OWL ontology to the Jess fact base is naturally the very five basic elements. (1) The OWL formula for class is: howl:Class rdf:ID =“Ship”/i The corresponding formula for Jess is: (deftemplate Ship (slot name)) (2) The OWL formula for relation (take the subclassof for example) is: howl:Class rdf:ID=“Ship”i hrdfs:subClassOf rdf:resource=“#WaterTransportationTool”/i h/owl:Classi The corresponding formula for Jess is: (deftemplate Ship extends WaterTransportationTool) (3) The OWL formula for instance is: hShip rdf:ID=“Swallow”/i The corresponding formula for Jess is: (assert (Ship (name Swallow))) (assert (WaterTransportationTool (name Swallow))) Here, Swallow is the instance of class Ship. During the instance transformation, the information about class Ship to which instance Swallow belongs has to be gotten firstly, and the relevant information of father class Ship has to be gotten. In addition, if class Ship also has parent class, then the information about its parent class should be gotten. Hence, the instance transformation is an iterative process. Because the semantemes which the instance implies in OWL contain both the instance of the subclass and the one of its parent class, the instance transformation is not only the problem of the format transformation to be dealt with but also the axiom it implies needs to be transformed. (4) The OWL formula for function is: hShipMan rdf:ID=“M 01”i hhas Manager rdf:resource=“#M 02”/i h/ShipMani The corresponding formula for Jess is: (assert (hasManager M01 M02)) 1756 G. LI, Y. LIU AND B. CHEN (5) The axiom in OWL ontology is the most important, so the axiom transformation is obviously vital. A simple introduction to the transformation of the axiom subclass is also the parent class is given in the front OWL Formula (3). Relatively, the axiom transformation is more complicated. Firstly it must be decided which type the axiom belongs to and secondly it is transformed correspondingly. In detail, the related algorithm of axiom conversion is given in Table 1. Table 1. Algorithm of axiom conversion Algorithm 1 AxiomConvertor() 18: hT 0i=OWLObjectPropertyAssertionAxiom 1: begin 19: else if a ∈OWLClassPropertyAssertionAxiom 2: for each element e∈Ontology do 20: hT 0i=OWLClassPropertyAssertionAxiom 3: if e ∈OWLClass 21: else if a ∈OWLPropertyPropertyAssertionAxiom 4: hT i=Class 22: hT 0i=OWLPropertyPropertyAssertionAxiom 5: else if e ∈OWLIndividual 23: else if a ∈OWLSameIndividualsAxiom 6: hT i=Individual 24: hT 0i=OWLSameIndividualsAxiom 7: else if e ∈OWLDatatypeValue 25: else if a ∈OWLDifferentIndividualsAxiom 8: hT i=DatatypeValue 26: hT 0i=OWLDifferentIndividualsAxiom 9: representation=getOWLhT iRepresentation(e) 27: else if a ∈OWLClassAssertionAxiom 10: add(ArryList, representation) 28: hT 0i=OWLClassAssertionAxiom 11: else end 29: A = (hT 0i)a 12: end if 30: representation=gethT 0iRepresentation(A) 13: end for 31: add(ArryList, representation) 14: for each axiom a∈Ontology do 32: else end 15: if a ∈OWLDatatypePropertyAssertionAxiom 33: end if 16: hT 0i=OWLDatatypePropertyAssertionAxiom 34: end for 17: else if a ∈OWLObjectPropertyAssertionAxiom 35: end B. Transformation from SWRL Rule to Jess Rulebase Firstly the edited SWRL rule as the source of Jess rule base should be transformed. The rule conversion algorithm is detailedly given in Table 2. Table 2. Algorithm of rule conversion Algorithm 2 RuleConvertor() 13: else if atom∈SameIndividualAtom atom=(atom b, atom h) 14: hT i=SameIndividualAtom 1: begin 15: else if atom∈DifferentIndividualsAtom 2: rule=Rule(i) 16: hT i=DifferentIndividualsAtom 3: if rule∈SWRLRule then 17: else if atom∈BuiltInAtom 4: atom b= getBodyAtoms() 18: hT i=BuiltInAtom 5: atom h= getHeadAtoms() 19: representation=gethT iRepresentation() 6: for all atom do 20: add(ArryList, representation) 7: if atom∈ClassAtom 21: else end 8: hT i=ClassAtom 22: end if 9: else if atom∈DatavaluedPropertyAtom 23: end for 10: hT i=DatavaluedPropertyAtom 24: end if 11: else if atom∈IndividualPropertyAtom 25: end 12: hT i=IndividualPropertyAtom The SWRL rule consists of Imp whose components are both head and body. Since the basic ingredient allowed to appear in the head and body is Atom, the transformation of SWRL rule is the one of Atom. (1) When Atom serves as ClassAtom, its formula in the SWRL rule is: ClassA(?x). The corresponding formula of Jess is: (ClassA(name ?x) or (assert(ClassA(name ?x)). (2) When Atom serves as DatavaluedPropertyAtom, its formula in the SWRL rule is: DatavaluedPropertyB(?x, 20). ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1757 The corresponding formula of Jess is: (DatavaluedPropertyB(?x, 20) or (assert(DatavaluedPropertyB ?x, 20)). (3) When Atom serves as IndividualPropertyAtom, its formula in the SWRL rule is: IndividualPropertyC (?x,?y). The corresponding formula of Jess is: (IndividualPropertyC ?x?y) or (assert(IndividualPropertyC ?x ?y)). (4) When Atom serves as SameIndividualAtom, its formula in the SWRL rule is: sameAs(?x,?y). The corresponding formula of Jess is: (sameAs ?x ?y) or (assert(sameAs?x ?y)). (5) When Atom serves as DifferentIndividualsAtom, its formula in the SWRL rule is: differentFrom(?x,?y). The corresponding formula of Jess is: (differentFrom ?x ?y) or (assert(differentFrom?x ?y)). (6) When Atom serves as ClassAtom, taking greater Than for example, its formula in the SWRL rule is: greaterThan(age1,age2). The corresponding formula of Jess is: test(> age1 age2). After the transformation from OWL ontology and SWRL rules to the fact base and rule base of Jess, the method RunRuleEngine can be employed to start the Jess engine to reason. During the reasoning, the converted facts will be matched with the condition of the rules to get the activated rules, and the rules are added into the working agenda of Jess. These rules are run item-by-item to get the result of the reasoning. Finally, the new fact deduced is added into the Jess fact base which can be used for deeper level reasoning. When reasoning is accomplished, the method of writeAssertedIndividualsAnd- Properties2OWL() supplied by the class SWRLRule-EngineBridge can be used to write the deduced Jess facts into the original OWL ontology, so that the original ontology is updated, and a new ontology that is more sophisticated is gotten. 5. Demonstration and Verifying. In order to verify the proposed framework in this paper, a detailed instance is given in Table 3. Table 3. The segment of the ‘ShipMan’ ontology code ...... rdf :ID=“has colleague”i howl:Class rdf :ID=“Captain”i howl:inverseOf hrdfs:subClassOf i rdf :resource=“#has colleague”/i howl:Class rdf :ID=“Person”/i hrdfs:domain rdf :resource=“#Sailor”/i h/rdfs:subClassOf i hrdfs:range rdf :resource=“#Sailor”/i howl:disjointWithi h/owl:SymmetricPropertyi howl:Class rdf :ID=“Sailor”/i ...... h/owl:disjointWithi hCaptain rdf :ID=“Tom”i h/owl:Classi hmanager sailor rdf :resource=“#Gin”/i howl:Class rdf:about=“#Sailor”i hmanager sailor rdf :resource=“#Nat”/i hrdfs:subClassOf hmanager sailor rdf :resource=“#Jer”/i rdf :resource=“#Person”/i h/Captaini ...... howl:ObjectProperty hCaptain rdf :ID=“Jerry”i rdf :ID=“has manager”i hmanager sailor rdf :resource=“#Ell”/i hrdfs:range rdf:resource=“#Captian”/i hmanager sailor rdf :resource=“#Cas”/i hrdfs:domain rdf :resource=“#Sailor”/i h/Captaini howl:inverseOf hCaptain rdf :ID=“Jim”i rdf :resource=“#manager sailor”/i hmanager sailor rdf :resource=“#Zach”/i h/owl:ObjectPropertyi h/Captaini howl:SymmetricProperty 1758 G. LI, Y. LIU AND B. CHEN In Table 3, there are not any direct relations between Gin, Nat and Jer as well as between Cas and Ell. If someone directly inquires the ontology for the instance of Sailor that has the colleague relation to Gin, then the result will be empty. However, Gin, Nat and Jer have the common ruler Tom, and both Cas and Ell have the common ruler Jerry, so there is the colleague relation between Gin, Nat and Jer as well as between Cas and Ell. The knowledge is implied in the ontology. In order to reason the ontology by making use of the proposed framework, a SWRL rule needs to be defined as below: Sailor(?x) ∧ Sailor(?z) ∧ Captain(?y) ∧ hasManager(?x, ?y)∧ hasManager(?z, ?y) ∧ differentFrom(?x, ?y) → hasColleague(?x, ?y) After the reasoning, the instance Sailor that has the colleague relation will be deduced, and the detailed process is shown as follows. The transformed Jess fact: (assert (has manager Ell Jerry)); (assert (has manager Cas Jerry)); (assert (has manager Gin To m)); (assert (has manager Zach Jim)); (assert (has manager Jer Tom)); (assert (has manager Nat Tom)); differentFrom(A, B). The differentFrom(A, B) mentioned above indicates that there is a relation different- From between all the instances of Sailor and all the instances of Captain. Due to the space limit of this paper, we do not show all the relations. The transformed Jess rule is: (defrule def-has colleague (has manager ?x ?z)(has manager ?y ?z) (differentFrom ?x ?y)(Sailor (name ?x)) (Sailor (name ?y)) ⇒(assert(has colleague?x ?y))(assert(OWLProperty has colleague ?x?y)) The new fact after Jess reasoning is: (assert (has colleague Jer Gin)); (assert (has colleague Gin Jer)); (assert (has colleague Nat Jer)); (assert (has colleague Jer Nat)); (assert (has colleague Nat Gin)); (assert (has colleague Gin Nat)); (assert (has colleague Cas Ell)); (assert (has colleague Ell Cas)). The deduced new fact is written into the original ontology, and the original ontology is updated to get a new ontology. Therefore, if someone inquires the ontology for the instance of the Sailor that has the colleague relation to Gin, he will get the results including both Jer and Nat. 6. Conclusions. Since OWL DL cannot adequately describe some complicated rules in reasoning level, a ontology and rule combined reasoning framework (SWRL based reason- ing framework) is proposed to get new ontology which has richer semantic relationships between concepts, so the knowledge of ontology is improved. Simply building the relevant SWRL rules in accordance with requirements, the reasoning service can be accomplished. So, not only the ontology knowledge availability factor is raised, but also the recall ratio and the precision ratio are raised in the information retrieve application. The framework can be improved to serve the uncertain ontology reasoning by referring to the advanced methods with respect to the reasoning algorithms, such as the extended neuro-fuzzy learn- ing algorithm for tuning fuzzy rules [5]. Acknowledgment. This work is sponsored by National Natural Science Fund of China (Grant No. 60972090) and by Liaoning Province Natural Fund (Grant No. 20072142). REFERENCES [1] L. Yu, X. Liu and F. Ren, Learnig to classify semantic orientation on-line document, International Journal of Innovative Computing, Information and Control, vol.5, no.12(A), pp.4637-4645, 2009. [2] X.-Q. Yang, N. Sun and T.-L. Sun, The application of latent semantic indexing and ontology in text classification, International Journal of Innovative Computing, Information and Control, vol.5, no.12(A), pp.4491-4499, 2009. ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1759 [3] L. Qian, A Inquiry System Design and Realization on the Base of the Reasoning of the Ontology and the Rule, Ph.D. Thesis, Southeast China University, 2007. [4] R. E. McGraph and J. Futrelle, Reasoning about provenance with OWL and SWRL rules, Lecture Notes in Computer Science, 2008. [5] Y. Shi, Fuzzy inference modeling based on fuzzy singleton-type reasoning, International Journal of Innovative Computing, Information and Control, vol.3, no.1, pp.13-20, 2007. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1761-1766 VULNERABILITY ASSESSMENT OF DISTRIBUTION NETWORKS BASED ON MULTI-AGENT SYSTEM AND QUANTUM COMPUTING Xiangping Meng1, Kaige Yan2, Jingweijia Tan2 and Zhaoyu Pian1 1School of Electrical Engineering and Information Technology Changchun Institute of Technology Changchun 130012, P. R. China mxp [email protected] 2College of Computer Science and Technology Jilin University Changchun 130012, P. R. China Received February 2010; accepted April 2010 Abstract. To overcome the defects of information architecture of distribution networks, a coordination of distributed control structure is presented based on multi-agent system for vulnerability assessment. The new architecture makes use of multi-agent system’s advantages of handling of complex, dynamic, and distributed problems. Meanwhile, since the classical reinforcement learning of multi-agent system would be failed, when the num- ber of agents is too large, a quantum reinforcement learning algorithm is proposed by combining quantum computing to perfect the assessment. Finally, the numerical simu- lations for both the classical Nash Q-learning and quantum reinforcement learning, and vulnerability assessment experiments based on an actual distribution network, have been investigated respectively. The results of simulation demonstrate that the improved model is more precise and highly efficient for the vulnerability assessment. Keywords: Vulnerability assessment, Multi-agent system, Reinforcement learning, Quan- tum computing 1. Introduction. For the past years, many countries have suffered from serious black- outs, and the aftermath of the events poignantly illustrates the vast social disruption and economic loss [1,2]. Although the distribution network has been proven to be surprisingly robust and reliable, there does exist the potential disturbances, even small ones, to trig- ger an unfortunate cascade of events that cause large portions of the grid to blackout. Accordingly, the research of power system vulnerability assessment should be focused on. Chen et al. proposed a complex power system vulnerability assessment based on the risk theory [3]. In [4], the artificial neural network was used in network vulnerability assess- ment. And Wu et al. studied the onset and spreading of cascading failure on weighted heterogeneous networks by adopting a local weighted flow redistribution rule [5]. How- ever, those researchers’ works neglected the concrete engineering features. The existing information architecture of distribution networks is to use only one central controller to monitor the activities of all other devices, and all the decisions are made by the central controller. But with the development of society, the operations of power devices may not be centralized control anymore. Therefore, an intelligent, distributed control system is in more urgent need to achieve accurate control in a large-scale and complex distribu- tion network. Recently, multi-agent system (MAS) has been more focused in distributed artificial intelligence, which is a coordinate operation to achieve the global goal [6]. In this paper, we implement the MAS to propose an improved model representing a coor- dination of distributed and centralized control structure for overcoming the defects of classical architecture. Meanwhile, a novel reinforcement learning algorithm is presented 1761 1762 X. MENG, K. YAN, J. TAN AND Z. PIAN by combining quantum computing (QC). Moreover, the comparison between the tradi- tional Nash Q-learning and quantum reinforcement learning, and vulnerability assessment experiments with regard to actual oilfield’s distribution network, has been investigated through numerical simulations respectively, and the results also verify the superiority of the proposed model both in precision and efficiency. Figure 1. Framework of distribution networks control 2. Classical Information Architecture of Distribution Networks. Figure 1 shows the traditional information framework of distribution networks control. As illustrations of the literature, there are two categories of control methods, decentralized control and centralized control. The signals and information (i.e., control and vulnerability informa- tion) of all load devices and substations are sent to control center (the red ellipse part) in a one-way passing mode, and all the information is dealt in the top controller. This hierarchy involves long-distance transmission of vulnerability information and hence is difficult to apply to fast control even though it can be technically implemented, moreover, large amounts of information can also cause blockage. In order to enhance the supervised effectiveness of distribution networks, more de- centralized controllers should be advocated. However, all the schemes of coordination proposed so far are based on the framework of decentralized control shown in Figure 1. There is no exchange of related information among local controllers. With the rapid devel- opment of technology, the evolution of distribution networks operation is more complex. Therefore, a more intelligent coordination of decentralized local controllers needs to be developed, which enables local controllers to exchange information together. Multi-agent theory has been widely applied in power systems recently. In this work, the MAS will be implemented to propose a new assessment architecture, which could guarantee fast and accurate vulnerability assessment of distribution networks. 3. Improved Architecture for Vulnerability Assessment Based on MAS. 3.1. Improved architecture. The most prominent property of classical assessment ar- chitecture is to only one controller directing all the actions of others. Once the central controller is at fault, it will affect all other agents, and the entire system will be collapse. To avoid these problems, we use the hybrid and hierarchical architecture of MAS for the theme concerned in this paper, which consists in using several supervisors in multiple hierarchical levels (see Figure 2). ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1763 Figure 2. Improved framework of distribution networks control According to the importance and hierarchy of the devices in distribution networks, three control layers can be defined, which constitute a vertical supervised mechanism. They are the reactive layer, coordination layer, and deliberative layer [6]. The reactive layer, attached to the lowest layer, is located in every load devices and is used to collect information of the environment. Load devices could be buses, large-scale power facilities, etc; The higher level is coordination layer. It takes the role of exchanging the information. It can trigger alarms received from the reactive layer, and can also pass the instructions of deliberative layer. Otherwise, the coordination layer is also a local controller. If a trig- gering event is under control (In the pre-set range), the control agents of the coordination layer issue instructions for the agents in reactive layer; The most top layer is the delib- erative layer, which is the global supervisor and responsible for coordinating the whole system. All the information is sent to deliberative layer, in which the plans, decisions, and commands are made according to the principle of coordination and real states. To make full use of the architecture, the similar multi-agent modules are set up on each control nodes. Every multi-agent module consists of several functional agents, which play different roles in a vulnerability assessing process. As the module has completed technical functions, every control node becomes an autonomous local controller. Otherwise, with the communication agent, different modules could exchange their information to keep coordination and accuracy of assessment. Accordingly, a new vulnerability assessment mechanism is established based on the architecture and the multi-agent modules. 3.2. The learning rate drawbacks of MAS for vulnerability assessment. As we all know, there are a large number of load devices in a distribution network, and every load device is assigned a multi-agent module. In the proposed assessment mechanism, multi-agent learning deals with the problem domains involving multiple agents, and the agent not only learns what effects its actions have, but also how to coordinate or align its action choices with those of other agents. Accordingly, the search space involved can be unusually large, and classical searching algorithm turns to be inefficient. As in reinforcement learning algorithm, when the number of agents or/and agent’s action is too large, all of the action selection methods will be failed: the speed of learning is decreased sharply. Consequently, to implement assessment mechanism described above, the problem of low learning rate must be solved firstly. 4. A New Quantum Reinforcement Learning for Correcting Learning Rate Drawbacks of MAS. To solve the problem of learning rate deficiency, we try to intro- duce quantum mechanism to cooperate multi-agent reinforcement learning. 1764 X. MENG, K. YAN, J. TAN AND Z. PIAN 4.1. Multi-agent Q-learning (QLA). Learning behaviors in a multi-agent environ- ment are crucial for developing and adapting multi-agent systems. The Q-learning is a standard reinforcement learning technique [7]. In multi-agent Q-learning, the Q-function of agent i is defined over states and joint-action vectors ~a = (a1, a2, ··· , an), rather than state-action pairs. The agents start with arbitrary Q-values, and the updating of Q value proceeds as following: [ ] i − i i · i Qt+1(s,~a) = (1 α)Qt(s,~a) + α rt + β V (st+1) (1) i where, V (st+1) is state value functions, and i i i V (st+1) = max f (Qt(st+1,~a)). (2) ai∈Ai 4.2. Quantum algorithm. 1. Quantum Bits: The unit of quantum information is the quantum bit (qubit). The qubit is a vector in a two-dimensional complex vector space with inner product, repre- sented with quantum state. It is an arbitrary superposition state of two-state quantum system: |ψi = α |0i + β |1i , |α|2 + |β|2 = 1, (3) where α and β are complex coefficients. |0i and |1i correspond to classical bit 0 and 1. |α|2 and |β|2 represent the occurrence probabilities of |0i and |1i respectively when the project |ψi is measured, the outcome of the measurement is not deterministic. 2. Grover’s operator: Grover’s idea was to represent all numbers in {0, 1} via a superposition φ of n qubits, and shift their amplitudes so that the probability to find an element out of L in a measurement is near 1. To achieve this, a superposition of n qubits in which all elements have the same amplitude is created. Then the following operations are applied alternately: • Amplitudes of all elements ∈ L are reflected at zero. • Amplitudes of all elements are reflected at the average of all amplitudes. The two operators together are called the Grover-operator. The other fundamental operation required is the conditional phase shift operation which is an important element to carry out Grover iteration [8]. 4.3. Action probing based on quantum computing. At first, we initialize the action z }|n { E 11∑··· 1 E (n) 1 (n) f(s) = as = √ |ai, then f(s) can be re-expressed as f(s) = as = n √ 2 a=00···0 √ 1 2n − 1 1 ∑ √ |aki + |φi, where |φi = |ai. Define the angle θ satisfying n 2n 2n − 1 2 E a=ak 1 (n) sin θ = √ , thus f(s) = as = sin θ |aki + cos θ |φi, then from, we know that applying 2n E (n) Grover operator UGroverL times on as can be represented as: L (n) | i | i UGrover as = sin((2L + 1)θ) ak + cos((2L + 1)θ) φ (4) Through repeating Grover operator, we can reinforce the probability amplitude of cor- responding action according to the reward value. Classically, the classical search scale O(N) is dependent√ upon structuring within the database [8]. Grover operator offers an improvement to O( N) and works just as well in a structured database as a disordered one. As Grover operator, the complexity of reinforcement learning reduces sharply. Con- sequently, we proposed a new quantum reinforcement learning. 5. Simulated Experiments and Results. ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1765 5.1. The simulation of quantum reinforcement learning. We used the pursuit games in the grid world to verify the validity of learning algorithm presented above. There are five agents: four hunt agents and one target agent. At each episode, when the target agent is rounded by hunt agents, the game is over. When the hunt agents round the target agent at the same time, the hunt agents gain the maximal reward simultane- ity. We played this pursuit game both on the Nash Q-learning algorithm and quantum reinforcement learning, the experiment results are shown in Figure 3. Figure 3. The performances of Nash Q-learning and quantum reinforce- ment learning As the Figure 3 shown, in multi-agent quantum reinforcement learning, the action se- lection method is obviously different from traditional multi-agent reinforcement learning algorithms, since Grover operator is used to explore multi-agent’s learning policies, multi- agent’s action selection method makes a good tradeoff between exploration and exploita- tion, and the updating process is carried out through parallel computation. Hence, the new quantum reinforcement learning lays a basic foundation for vulnerability assessment based on MAS. 5.2. The simulation of vulnerability assessment. According to improved assessment architecture and quantum reinforcement learning algorithm, we developed the software for vulnerability assessment of distribution networks. To verify the validity of the proposed algorithms and software, we implemented the software to assess vulnerability of one civil oilfield’s distribution network. In the experiment, the 110 KV, 35KV, and 10KV lines were selected to compute. Figure 4 shows the vulnerability ranks of 35KV lines, and Figure 5 illustrates the vulnerability value that is got by quantitative computes. Simulations show that the new assessment architecture and proposed quantum reinforcement learning make the vulnerability assessment more accurate and rapid. And the software developed with the proposed algorithms, will also be applied widely in distribution networks. 1766 X. MENG, K. YAN, J. TAN AND Z. PIAN Figure 4. Vulnerability ranks of Figure 5. Assessment results 35KV lines 6. Conclusion. In certain time and places, the distribution network is very vulnerable, and even small ones could trigger an unfortunate cascade of events. Consequently, the vulnerability of distribution network should be researched. In this paper, a vulnerability assessment model was proposed for vulnerability analysis of distribution network, in which the MAS’s superiorities were fully utilized. Otherwise, to keep the rapid running of MAS in the new assessing process, a quantum reinforcement learning algorithm was proposed based on quantum theory, which uses Grover searching algorithm to reduce computing complexity that leads the learning operation of MAS to be failed. The final tests on actual distribution network display that the proposed model is more precise and efficient. And the proposed model and software can be used to make an accurate vulnerability assessment for the distribution networks. Acknowledgment. This work was supported by the National Natural Science Founda- tion of China under Grant No. 60974055 and the Changchun Science and Technology Plan Projects under Grant No. 08k218. REFERENCES [1] H.-Y. Wu, C.-Y. Hsu, T.-F. Lee and F.-M. Fang, Improved SVM and ANN in incipient fault diag- nosis of power transformers using clonal selection algorithms, International Journal of Innovative Computing, Information and Control, vol.5, no.7, pp.1959-1974, 2009. [2] H. Yoshikawa, K. Akama and H. Mabuchi, ET-based distributed cooperative system, International Journal of Innovative Computing, Information and Control, vol.5, no.12(A), pp.4655-4666, 2009. [3] W. Chen, Q. Jiang, Y. Cao and Z. Han, Risk-based vulnerability assessmnet in complex power system, Power System Techonolygy, vol.2, pp.12-19, 2005. [4] Q. Zhou, J. Davidson and A. A. Fouad, Application of artificial neural networks in power system security and vulnerability assessment, IEEE Trans. on Power Systems, vol.9, pp.525-532, 2002. [5] Z. X. Wu, G. Peng, W. X. Wang, S. Chan and E. E. M. Wong, Cascading failure spreading on weighted heterogeneous networks, Journal of Statistical Mechanics: Theory and Experiment, vol.2008, 2008. [6] C.-C. Liu, J. Jung, G. T. Heydt, V. Vittal and A. G. Phadke, The strategic power infrastructure defense (SPID) system: A conceptual design, IEEE Trans. on Control Systems Magazine, vol.20, no.4, pp.40-52, 2000. [7] J. Hu and M. P. Wellman, Nash Q-learning for general-sum stochastic games, Journal of Machine Learning Research, vol.4, pp.1039-1069, 2003. [8] L. Grover, A fast quantum mechanical algorithm for database search, Proc. of the 28th Annual ACM Symposium on Theory of Computation, pp.212-219, 1996. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1767-1772 A NEW MULTIPLE KERNEL LEARNING BASED LEAST SQUARE SUPPORT VECTOR REGRESSION AND ITS APPLICATION IN ON-LINE GAS HOLDER LEVEL PREDICTION OF STEEL INDUSTRY Xiaoping Zhang, Jun Zhao∗ and Wei Wang Research Center of Information and Control Dalian University of Technology Dalian 116024, P. R. China zhang [email protected]; [email protected] ∗Corresponding author: [email protected] Received February 2010; accepted April 2010 Abstract. The key of gas scheduling in modern steel enterprises is to accurately pre- dict gasholder level on-line. Because of the frequent and great fluctuation of level, it is difficult to be precisely predicted by the manual experience or the mechanism model based method. In this study, a class of MKL based on reduced gradient algorithm is ex- tended using LSSVR, and the MKLLSSVR is developed. The MKLLSSVR overcomes the poor generalization of single kernel based LSSVR and the long training process of tradi- tional MKL based SVR, and can rapidly give the optimal linear combination of kernels to base the resulting regressor for better interpretation. The simulation results with the real gasholder level data show the MKLLSSVR can achieve better prediction performance compared to the traditional MKL and single kernel based LSSVR. Keywords: Gasholder level prediction, Multiple kernel learning, Reduced gradient method, Least square support vector regression 1. Introduction. With the shortage of fossil energy (such as coal and oil) and the av- ocation of green manufacturing, making full use of by-product coal gas generated during steel production can greatly improve the energy saving level [1]. In byproduct energy scheduling for gas utilization and the safety operation of gasholder simultaneously, the real-time prediction for holder level is one of the most significant tasks. Heretofore, some scholars have paid attention on the online prediction for the holder level in steel industry. Iwao et al. in Kawasaki steel plant argued that the gasholder level exhibited a high self-correlation [1]. In their study, the previous level was employed using a time series model to predict the subsequent trend. However, the effect was limited because of the neglect of some influence factors, such as the generation and consumption of gas users. Then Kim et al. established a mechanism model that regarded the current level as the difference of all the generation and consumption of previous time [2]. Likewise, such method ignores the different impact of gas unit on the holder level. As the description by Iwao and Kim, the gasholder level is strongly affected by the previous level and the generation-consumption in the gas system. Therefore, the relationship among them is complex and with high nonlinearity. Although some effective applications had been carried out, such as finance [3], electrical industry [4], image coding [5], the LSSVR is generally based on a single kernel, and de- pends on the choices of a good kernel and the corresponding feature space [6]. However, the optimized single kernel cannot fully describe the complicated regression problem. Recent studies have shown that multiple kernel learning (MKL) is able to identify the appropri- ate combination of kernel functions to enrich the interpretation of regressor. Lanchriet et al. proposed the MKL framework as a semi-definite programming, and encoded it for a yeast protein functional classification [7]. For applying to the larger-scale problems, Bach 1767 1768 X. ZHANG, J. ZHAO AND W. WANG et al. introduced a dual formulation of the semi-definite programming as a second-order cone programming [8]. Subsequently, Sonnenburg et al. simplified the learning process by expressing the second-order cone programming as a semi-infinite linear programming, and put forward a wrapper algorithm with cutting planes method [9]. Rakotomamonjy et al. solved the MKL by a simple reduced gradient method (RGM) [10]. Note that the described fast approaches divided the MKL into two cycles [9,10], in which the inner cycle employed a standard SVR with linear combination of base kernels for the objective function update, and the outer cycle optimized the objective function to obtain the kernel weight vector. So the overall efficiency of MKL was tied to the single kernel based SVR. When dealing with the practical problem with the requirements of real-time and high precision, the effect will be restrained in spite of the contribution of RGM on saving the computation time. This paper combines the LSSVR with MKL to construct a MKLLSSVR for predicting the frequent non-flat variation of gasholder level. Firstly, instead of the single kernel, a series of kernels including various forms or scales are employed to enrich the features space, which make the regression function well interpret the complex problem. Secondly, the replacement of SVR by LSSVR in inner cycle greatly reduces the computation time of MKL, and simultaneously obtains the better objective function to further improve the RGM’s global optimization. The real data experiments demonstrate the high efficiency of MKLLSSVR in predicting the gasholder level over compared to the traditional MKL, single kernel LSSVR or the common used mechanism method. 2. Problem Statement. In Baosteel, blast furnace gas (BFG) is generated from the four blast furnaces. By the gas pipeline, it is supplied to many gas users as the secondary fuel for maintaining the production, for example, coke furnace (CF), hot blast furnace (HBF), continuous casting, heat treatment in rolling process, power plants, low pressure boilers (LPB), chemical plants, etc. And the rest are transported into the gasholder that works as a buffer to temporarily store the BFG. In general, the patterns of the generation and consumption for most of BFG units are largely different [2]. The various fluctuation of generation and consumption of gas units will influence the gasholder level change via gas pipeline. To sustain a stable running for the BFG supply and avoid the severe fluctuation of the holder level, its level has to be predicted in advance for guiding the gas scheduling. 3. Multiple Kernel Learning Least Square Support Vector Regression. Given a { }N { }m ∈ p set of training data xi, yi i=1 and positive definite function kernel kk k=1, where xi R is the input data; yi ∈ R is the target value; m is the initial amount of kernels; and N is the size of training data set. Under the functional framework [10], we propose the MKLLSSVR for regression problem is to solve the following convex optimization: 1 ∑m 1 ∑N min J{f , b, e, d} = kf k2 + C e2 k k Hk i 2 dk k=1 i=1 (1) ∑m ∑m s.t. yi = fk(xi) + b + ei ∀i; dk = 1, dk ≥ 0 ∀k k=1 k=1 ∑m Here f(x) = fk + b is the regression function learned by the MKLLSSVR. Each fk k=1 is trained in a different reproduce hilbert kernel space Hk; Hk is associated with a basis kernel kk; b is the coefficient of the regressor; dk with l1-norm sparse constraint is the kernel weight which controls the squared norm of fk in the objective function; ei is the training error on the ith training data point; C is the penalty factor which balances the empirical risk and the complexity of the regression function. ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1769 Introduce the Lagrange multipliers αi, β and ηk, the Lagrangian of problem (1) is: ( ) 1 ∑m 1 ∑N ∑N ∑m L({f }, b, e, d, α, β, η) = kf k2 + C e2 − α f (x ) + b + e − y k k Hk i i k i i i 2 dk k=1( ) i=1 i=1 k=1 ∑m ∑m −β 1 − dk − ηkdk (2) k=1 k=1 We have the dual problem: ∑N max J(α, λ) = αiyi−λ i=1 ( ) ∑N 1 1 s.t. α = 0; λ ≥ αT k + I α ∀k, (3) i 2 k 2C i=1 1 ∑n where β + α2 = λ 4C i i=1 According to the minimization of the primal problem and its differentiability, we prefer to consider problem shown in Equation (1) as the following constrained optimization: ∑m min J(d), s.t. dk = 1, dk ≥ 0 k=1 1 ∑m 1 ∑N J(d) = min J{f , b, e, d} = kf k2 + C e2 (4) k k Hk i 2 dk k=1 i=1 ∑m s.t. yi = fk(xi) + b + ei k=1 By the optimality conditions, we derive the associated dual problem of (4): ( ) 1 ∑m 1 ∑N ∑N max J(a) = − α d k + I αT + α y s.t. α = 0 (5) 2 k k 2C i i i k=1 i=1 i=1 Then the gradient of J(d) with respect to dk is easily computed: ∂J 1 ∗ ∗T = − α kkα ∀k (6) ∂dk 2 According to the convex property and differentiability with Lipschitz gradient for J(·) [12], we have the reduced gradient r of J(d) in Equation (4): { ( ) ∂J ∂J ∑ ∂J ∂J rk = − ∀k =6 u; ru = − , u = arg max(dk) (7) ∂dk ∂du ∂dk ∂du { | 6 } k=6 u k dk=0 The feasible descent direction and update scheme for d with the step size s are given as: Dk = {for ∀k =6 u, if rk > 0, −dkrk; for ∀k =6 u, if rk ≤ 0, −rk; for k = u, ru (8) d ← d + sD (9) The duality gap to monitor the convergence of the MKLLSSVR is formulated as: ( ) ( ( ) ) 1 ∑m 1 1 1 − ∗ ∗ ∗T ∗ ∗T DualGap = α dkkk + I α + max α kk + I α (10) 2 2C 2 k 2C k=1 1770 X. ZHANG, J. ZHAO AND W. WANG On the basis of the above reduced gradient, kernel weight updating and duality gap expression, the solving steps of the MKLLSSVR solved by RGM (MKLLSSVR(RGM)) is given as follows. Step 1: Initialize the number of kernels and set the original weight dk = 1/m for k = 1, . . ., m; Step 2: Obtain α∗ by using LSSVR to solve the problem (4) with hybrid kernel K; Step 3: Renew the objective function J(d) of the equivalent problem (4); Step 4: Compute ∂J/∂dk, r and Dk; { } dv dk Step 5: Set the maximal step size smax = − v = arg min − ; Dv {k|Dk<0} Dk Step 6: Update the weight d and the hybrid kernel K; Step 7: With K, repeat Step2∼Step3 and obtain the new objective function Jnew (dnew); Step 8: Set d = dnew, Du = Du − Dv, Dv =0 and repeat Step 6∼Step 8 until Jnew (dnew) stops decrease; Step 9: Determine the optimal step size s∗ using the Armijo line search (s∗ = arg min J(d+ sD){s ∈ [0, smax]}), and update the weight d at the same time; Step 10: If the duality gap formulated by Equation (10) is less than the setting value, the MKLLSSVR finishes, otherwise goes back to Step 2. During the whole learning process of MKLLSSVR (RGM), the line search process in- volving the computation of the objective function is greatly speeded up by using LSSVR in inner cycle. Consequently, a substantial amount of computation time is further saved. Besides the advantage of time saving, LSSVR with hybrid kernel is easier to be precisely solved than the QP problems of SVR. This replacement ensures MKLLSSVR obtain a better objective function. Benefiting from the better objective value and feasible descent direction, the weight d can rapidly converge to KKT point, in which only several dk have the large values, and others are very small or even zero. Then, some base kernels with larger weight constitute the optimal linear combination of kernel to well extract the relevant knowledge about the problem for basing the regressor with good interpretation. 4. Experiments. In this section we evaluate the performance of MKLLSSVR (RGM) on the gasholder level prediction through comparing it with MKL based SVR (MKLSVR (RGM)), which is also solved by RGM, single kernel LSSVR optimized by 10 fold cross- validation (SKLLSSVR (10FCV)) and the holder level mechanism method (HLM) in [2]. We randomly choose 30 example sets with size of N = 2000 (98.5% as training data and 1.5% for testing) from Baosteel. The gas production-consumption flow of each unit and gas holder level is served as input and target value. For the MKL algorithms in comparison, 25 RBF-kernel sampling from the domain [0.01, 1000] and 5 Poly-kennel from [1, 5] are selected as the initial kernels. And RBF-kernel is adopted in SKLLSSVR (10FCV). The two MKL are trained for different values of hyperparameter C, where ε is set to 0.01 in MKLSVR (RGM). We choose C = 100 for MKLSVR (RGM), C = 55 for MKLLSSVR (RGM). MKLLSSVR (RGM) is implemented combined an LSSVR solver and SimpleMKL toolbox written in Matlab [9]. For a fair comparison, the same terminating criterion for all MKL algorithms are adopted, and the algorithm stops when the duality gap is lower than 0.01, or the number of iterations exceeds 1000. And the experiments are conducted on a PC with 3.6GHZ CPU and 1GB memory, all the algorithms are run 20 times. Here, we give some prediction results of 2# BFG holder level in Baosteel. Table 1 lists the prediction measures of the four algorithms in comparison. We also randomly select three groups of the results to picture on Figure 1. Regarding the pre- diction performance, it can be seen that HLM results in the largest mean absolute error (MAE) although it consumes the least time. As for the other three methods, the results are greatly improved. Whereas, the MAE obtained by SKLLSSVR (10FCV) is larger than ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1771 that by MKLSVR (RGM) and our MKLLSSVR (RGM), as the regressor through only one kernel is difficult to recover different trend of level information. Simultaneously, the training time is rather long due to the time-consuming validation. In contrast, MKLSVR (RGM) and the developed MKLLSSVR (RGM) select the most relevant combination of different kernel to construct the rich kernel space, in which the regressor with better in- terpretation is based. Moreover, by LSSVR instead of SVR, our approach can finish the 30 minutes prediction in less than one minute and generate far less MAE that meets the real time requirement of the practical prediction. Table 1. Average performance measures of the four prediction methods Method Trend MAE Time(s) Method Trend MAE Time(s) ascending 8.0965 ascending 2.2234 MKLSVR HLM stable 2.7112 0.005 stable 1.4818 756.90 (RGM) descending 6.0125 descending 2.0506 ascending 3.7617 ascending 1.0243 SKLLSSVR MKLLSSVR stable 1.7956 410.12 stable 0.9215 53.55 (10FCV) (RGM) descending 2.5947 descending 1.1938 (a) ascending trend (b) stable trend (c) descending trend Figure 1. The trend prediction result of 2#BFG holder level From Figure 1, the four algorithms in comparison can all predict the ascending, stable and descending trend of level in the future half an hour, but the deviation for HLM begins to become increasingly great as the prediction time goes on. Although, the great deviation is reduced obviously in SKLLSSVR and MKLSVR (RGM), it can also exist when the prediction time exceeds ten minutes. In contrast, the prediction result of the MKLLSSVR (RGM) proposed in this paper is the best and satisfies the precision requirement well. 5. Conclusions. The contribution of this paper is that a new MKLLSSVR has been proposed and applied to the gasholder level on-line prediction. In particular, the MK- LLSSVR overcomes the drawbacks of single kernel based LSSVR and MKLSVR. Our experimental results have shown that the developed MKLLSSVR is able to greatly im- prove the prediction precision over SKLLSSVR and reduces the prediction time over MKLSVR. Furthermore, the prediction performance prefers to the mechanism method 1772 X. ZHANG, J. ZHAO AND W. WANG used at present. So it can be generalized to other similar steel plant’s gasholder level prediction for offering effective guidance of gas real time balance scheduling. Acknowledgment. This work has been supported by the National High-Tech R&D Pro- gram of China (2007AA04Z1A9). REFERENCES [1] I. Higashi, Energy balance of steel mills and utilization of byproduct gases, Transactions of the Iron and Steel Institute of Japan, vol.22, no.1, pp.57-65, 1982. [2] J. H. Kim, H. S. Yi and C. Han, A novel MILP model for plantwide multiperiod optimization of byproduct gas supply system in the iron and steel making process, Chemical Engineering Research Design, vol.81, no.8, pp.1015-1025, 2003. [3] J. A. K. Suykens, T. V. Gestel, J. D. Brabanter et al., Least Squares Support Vector Machines, World Scientific, Singapore, 2002. [4] G. F. Duan, Y. W. Chen and T. Sukekawa, Automatic optical inspection of micro drill bit in printed circuit board manufacturing using support vector machines, ICIC Express Letters, vol.5, no.11(B), pp.4347-4355, 2009. [5] Q. S. She, H. Y. Su, L. D. Dong and J. Chu, Support vector machine with adaptive parameters in image coding, ICIC Express Letters, vol.4, no.2, pp.359-367, 2008. [6] O. Chapelle, V. Vapnik, O. Bousquet and S. Mukherjee, Choosing multiple parameters for support vector machines, Machine Learning, vol.46, no.1, pp.131-159, 2002. [7] G. R. G. Lanckriet, N. Cristianini, P. Bartlett, L. E. Ghaoui and M. I. Jordan, Learning the kernel matrix with semidefinite programming, Journal of Machine Learning Research, vol.1, no.5, pp.27-72, 2004. [8] F. Bach, G. R. G. Lanckriet and M. I. Jordan, Multiple kernel learning, conic duality, and the SMO algorithm, Proc. of the 21st International Conference on Machine Learning, Banff, Alberta, Canada, 2004. [9] S. Sonnenburg, G. Ratsch, C. Schafer and B. Scholkopf, Large scale multiple kernel learning, Journal of Machine Learning Research, vol.7, no.7, pp.1531-1565, 2006. [10] A. Rakotomamonjy, F. R. Bach, S. Canu and Y. Grandvalet, SimpleMKL, Journal of Machine Learning Research, vol.9, no.11, pp.2491-2521, 2008. [11] J. F. Bonnans and A. Shapiro, Optimization problems with pertubation: A guided tour, SIAM Review, vol.40, no.2, pp.202-227, 1998. [12] D. Luenberger, Linear and Nonlinear Programming, Addison-Wesley, 1984. ICIC Express Letters ICIC International c 2010 ISSN 1881-803X Volume 4, Number 5(B), October 2010 pp. 1773-1778 THE PARAMETERIZATION OF ALL STABILIZING SIMPLE REPETITIVE CONTROLLERS WITH THE SPECIFIED INPUT-OUTPUT CHARACTERISTIC Iwanori Murakami, Tatsuya Sakanushi, Kou Yamada, Yoshinori Ando Takaaki Hagiwara and Shun Matsuura Department of Mechanical System Engineering Gunma University 1-5-1 Tenjincho, Kiryu, Japan { murakami; t09801226; yamada; ando; t08801218; t10801252 }@gunma-u.ac.jp Received February 2010; accepted April 2010 Abstract. The simple repetitive control system proposed by Yamada et al. is a type of servomechanism for the periodic reference input. That is, the simple repetitive con- trol system follows the periodic reference input with small steady state error, even if a periodic disturbance or uncertainty exists in the plant. In addition, simple repetitive con- trol systems make transfer functions from the periodic reference input to the output and from the disturbance to the output have finite numbers of poles. Yamada et al. clarified the parameterization of all stabilizing simple repetitive controllers. However, using the method by Yamada et al., it is complex to specify the low-pass filter in the internal model for the periodic reference input of which the role is to specify the input-output charac- teristic. The purpose of this paper is to propose the parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic such that the input-output characteristic can be specified beforehand. Keywords: Repetitive control, Finite numbers of poles, Parameterization 1. Introduction. A repetitive control system is a type of servomechanism for the peri- odic reference input. That is, the repetitive control system follows the periodic reference input without steady state error, even if a periodic disturbance or uncertainty exists in the plant [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]. It is difficult to design stabilizing con- trollers for the strictly proper plant, because the repetitive control system that follows any periodic reference input without steady state error is a neutral type of time-delay control system [11]. To design a repetitive control system that follows any periodic reference input without steady state error, the plant needs to be biproper [3, 4, 5, 6, 7, 8, 9, 10, 11]. In practice, the plant is strictly proper. Many design methods for repetitive control systems for strictly proper plants have been given [3, 4, 5, 6, 7, 8, 9, 10, 11]. These studies are divided into two types. One uses a low-pass filter [3, 4, 5, 6, 7, 8, 9, 10] and the other uses an attenuator [11]. The latter is difficult to design because it uses a state variable time-delay in the repetitive controller [11]. The former has a simple structure and is easily designed. Therefore, the former type of repetitive control system is called the modified repetitive control system [3, 4, 5, 6, 7, 8, 9, 10]. Using modified repetitive controllers in [3, 4, 5, 6, 7, 8, 9, 10], even if the plant does not include time-delays, transfer functions from the periodic reference input to the output and from the disturbance to the output have infinite numbers of poles. This makes it difficult to specify the input-output characteristic and the disturbance attenuation characteristic. From the practical point of view, it is desirable that these characteristics should be easy to specify. Therefore, these transfer functions should have finite numbers of poles. To overcome this problem, Yamada et al. proposed simple repetitive control systems such that the controller works as a modified repetitive controller and transfer functions from the 1773 1774 I. MURAKAMI, T. SAKANUSHI, K. YAMADA ET AL. periodic reference input to the output and from the disturbance to the output have finite numbers of poles [14]. In addition, Yamada et al. clarified the parameterization of all stabilizing simple repetitive controllers. According to Yamada et al., the parameterization of all stabilizing simple repetitive controllers includes two free-parameters. One has the role to specify the disturbance attenuation characteristic. The other has the role to specify the low-pass filter in the internal model for the periodic reference input of which the role is to specify the input-output characteristic. However, using the method by Yamada et al., it is complex to specify the low-pass filter in the internal model for the periodic reference input. When we design a simple repetitive controller, if the low-pass filter in the internal model for the periodic reference input is settled beforehand, we can specify the input-output characteristic more easily than the method in [14]. This problem is solved by obtaining the parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic, which is the parameterization when the low-pass filter is settled beforehand. However, no paper has considered the problem to obtain the parameterization of all stabilizing simple repetitive controllers with the specified input- output characteristic. In addition, the parameterization is useful to design stabilizing controllers [15, 16, 17, 18]. In this paper, we propose the parameterization of all stabilizing simple repetitive con- trollers with the specified input-output characteristic such that the low-pass filter in the internal model for the periodic reference input is settled beforehand. 2. Simple Repetitive Controller with the Specified Input-output Character- istic and Problem Formulation. Consider the unity feedback control system given by { y(s) = G(s)u(s) + d(s) , (1) u(s) = C(s)(r(s) − y(s)) where G(s) ∈ R(s) is the strictly proper plant, C(s) is the controller, u(s) ∈ R is the control input, y(s) ∈ R is the output, d(s) ∈ R is the disturbance and r(s) ∈ R is the periodic reference input with period T > 0 satisfying r(t + T ) = r(t)(∀t ≥ 0). (2) According to [3, 4, 5, 6, 7, 8, 9, 10], the modified repetitive controller C(s) is written by the form in C(s) = C1(s) + C2(s)Cr(s), (3) where C1(s) ∈ R(s) and C2(s) =6 0 ∈ R(s). Cr(s) is an internal model for the periodic reference input r(s) with period T and written by e−sT Cr(s) = , (4) 1 − q(s)e−sT where q(s) ∈ R(s) is a proper low-pass filter satisfying q(0) = 1. Using the modified repetitive controller C(s) in (3), transfer functions from the periodic reference input r(s) to the output y(s) and from the disturbance d(s) to the output y(s) in (1) are written as y(s) C(s)G(s) = r(s) 1 + C(s){G(s) } −sT C1(s) − (C1(s)q(s) − C2(s))e G(s) = −sT (5) 1 + C1(s)G(s) − {(1 + C1(s)G(s))q(s) − C2(s)G(s)} e ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1775 and y(s) 1 = d(s) 1 + C(s)G(s) 1 − q(s)e−sT = −sT , (6) 1 + C1(s)G(s) − {(1 + C1(s)G(s))q(s) − C2(s)G(s)} e respectively. Generally, transfer functions from the periodic reference input r(s) to the output y(s) in (5) and from the disturbance d(s) to the output y(s) in (6) have infinite numbers of poles. In order to specify the input-output characteristic and the disturbance attenuation characteristic easily, Yamada et al. proposed simple repetitive control systems such that the controller works as a modified repetitive controller and transfer functions from the periodic reference input to the output and from the disturbance to the output have finite numbers of poles [14]. In addition, Yamada et al. clarified the parameterization of all stabilizing simple repetitive controllers. On the other hand, according to [3, 4, 5, 6, 7, 8, 9, 10], it is note that if the low-pass filter q(s) satisfy 1 − q(jωi) ' 0 (∀i = 0,...,Nmax) , (7) where ωi(i = 1,...,Nmax) are frequency components of the periodic reference input r(s) written by 2π ω = i (i = 0,...,N ) (8) i T max and ωNmax is the maximum frequency component of the periodic reference input r(s), then the output y(s) in (1) follows the periodic reference input r(s) with small steady state error. Using the result in [14], in order for q(s) to satisfy (7) in wide frequency range, we must design q(s) to be stable and of minimum phase. If we obtain the parameterization of all stabilizing simple repetitive controllers such that q(s) in (4) is settled beforehand, we can design the simple repetitive controller satisfying (7) more easily than the method in [14]. From above practical requirement, we propose the concept of the simple repetitive controller with the specified input-output characteristic as follows: Definition 2.1. (simple repetitive controller with the specified input-output characteristic) We call the controller C(s) a “simple repetitive controller with the specified input-output characteristic”, if following expressions hold true: 1. The low-pass filter q(s) ∈ RH∞ in (4) is settled beforehand. That is, the input-output characteristic is settled beforehand. 2. The controller C(s) works as a modified repetitive controller. That is, the controller C(s) is written by (3), where C1(s) ∈ R(s), C2(s) =6 0 ∈ R(s) and Cr(s) is written by (4). 3. The controller C(s) makes transfer functions from the periodic reference input r(s) to the output y(s) in (1) and from the disturbance d(s) to the output y(s) in (1) have finite numbers of poles. The problem considered in this paper is to propose the parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic. 3. The Parameterization of all Stabilizing Simple Repetitive Controllers with the Specified Input-output Characteristic. In this section, we clarify the parame- terization of all stabilizing simple repetitive controllers with the specified input-output characteristic defined in Definition 2.1. In order to obtain the parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic, q(s) ∈ RH∞ is assumed to be settled 1776 I. MURAKAMI, T. SAKANUSHI, K. YAMADA ET AL. beforehand. The parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic is summarized in the following theorem. Theorem 3.1. There exists a stabilizing simple repetitive controller with the specified input-output characteristic if and only if the low-pass filter q(s) ∈ RH∞ in (4) takes the form: q(s) = N(s)¯q(s). (9) Here, N(s) ∈ RH∞ and D(s) ∈ RH∞ are coprime factors of G(s) on RH∞ satisfying N(s) G(s) = (10) D(s) and q¯(s) =6 0 ∈ RH∞ is any function. When the low-pass filter q(s) ∈ RH∞ in (4) satisfies (9), the parameterization of all stabilizing simple repetitive controllers with the specified input-output characteristic is given by X(s) + D(s)Q(s) + D(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT C(s) = . (11) Y (s) − N(s)Q(s) − N(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT Here, X(s) ∈ RH∞ and Y (s) ∈ RH∞ are functions satisfying X(s)N(s) + Y (s)D(s) = 1 (12) and Q(s) ∈ RH∞ is any function. Proof of this theorem requires following lemma. Lemma 3.1. Unity feedback control system in (1) is internally stable if and only if C(s) is written by X(s) + D(s)Q(s) C(s) = , (13) Y (s) − N(s)Q(s) where N(s) ∈ RH∞ and D(s) ∈ RH∞ are coprime factors of G(s) on RH∞ satisfying (10), X(s) ∈ RH∞ and Y (s) ∈ RH∞ are functions satisfying (12) and Q(s) ∈ RH∞ is any function [18]. Using Lemma 3.1, we shall show the proof of Theorem 3.1. Proof: First, the necessity is shown. That is, we show that if the controller C(s) in (3) makes the control system in (1) stable and makes the transfer function from the periodic reference input r(s) to the output y(s) of the control system in (1) have finite numbers of poles, then the low-pass filter q(s) must take the form (9). From the assumption that the controller C(s) in (3) makes the transfer function from the periodic reference input r(s) to the output y(s) of the control system in (1) have finite numbers of poles, { } −sT G(s)C(s) C1(s) − (C1(s)q(s) − C2(s))e G(s) = −sT (14) 1 + G(s)C(s) 1 + G(s)C1(s) − {(1 + G(s)C1(s))q(s) − C2(s)G(s)} e has finite numbers of poles. This implies that (1 + G(s)C (s))q(s) C (s) = 1 (15) 2 G(s) is satisfied, that is, C(s) is necessarily G(s)C (s) + q(s)e−sT C(s) = (1 ) . (16) G(s) 1 − q(s)e−sT ICIC EXPRESS LETTERS, VOL.4, NO.5, 2010 1777 From the assumption that C(s) in (3) makes the control system in (1) stable, G(s)C(s) /(1 + G(s)C(s)), C(s)/(1 + G(s)C(s)), G(s)/(1 + G(s)C(s)) and 1/(1 + G(s)C(s)) are stable. From simple manipulation and (16), we have G(s)C(s) G(s)C (s) + q(s)e−sT = 1 , (17) 1 + G(s)C(s) 1 + G(s)C1(s) C(s) G(s)C (s) + q(s)e−sT = 1 , (18) 1 + G(s)C(s) (1 + G(s)C1(s))G(s) G(s) (1 − q(s)e−sT )G(s) = (19) 1 + G(s)C(s) 1 + G(s)C1(s) and 1 1 − q(s)e−sT = . (20) 1 + G(s)C(s) 1 + G(s)C1(s) From the assumption that all transfer functions in (17), (18), (19) and (20) are stable, G(s)C1(s)/(1 + G(s)C1(s)), C1(s)/(1 + G(s)C1(s)), G(s)/(1 + G(s)C1(s)) and 1/(1 + G(s)C1(s)) are stable. This means that C1(s) is an internally stabilizing controller for G(s). From Lemma 3.1, C1(s) must take the form: X(s) + D(s)Q(s) C (s) = , (21) 1 Y (s) − N(s)Q(s) where Q(s) ∈ RH∞. From the assumption that the transfer function in (18) is stable, q(s) (Y (s) − N(s)Q(s)) D2(s)q(s) = (22) G(s) (1 + G(s)C1(s)) N(s) is stable. This implies that q(s) must take the form q(s) = N(s)¯q(s), (23) whereq ¯(s) =6 0 ∈ RH∞ is any function. In this way, it is shown that if there exists a stabilizing simple repetitive controller with the specified input-output characteristic, then the low-pass filter q(s) must take the form (9). Next, we show that if (9) holds true, then C(s) is written by (11). Substituting (15), (21) and (23) into (3), we have (11). Thus, the necessity has been shown. Next, the sufficiency is shown. That is, it is shown that if q(s) and C(s) take the form (9) and (11), respectively, then the controller C(s) makes the control system in (1) stable, makes the transfer functions from r(s) and d(s) to y(s) of the control system in (1) have finite numbers of poles and works as a stabilizing modified repetitive controller. After simple manipulation, we have G(s)C(s) { } = X(s) + D(s)Q(s) + D(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT N(s), (24) 1 + G(s)C(s) C(s) { } = X(s) + D(s)Q(s) + D(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT D(s), (25) 1 + G(s)C(s) G(s) { ) = Y (s) − N(s)Q(s) − N(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT N(s) (26) 1 + G(s)C(s) and 1 { } = Y (s) − N(s)Q(s) − N(s)(Y (s) − N(s)Q(s))q ¯(s)e−sT D(s). (27) 1 + G(s)C(s) Since X(s) ∈ RH∞, Y (s) ∈ RH∞, N(s) ∈ RH∞, D(s) ∈ RH∞, Q(s) ∈ RH∞ and q¯(s) ∈ RH∞, the transfer functions in (24), (25), (26) and (27) are stable. In addition, 1778 I. MURAKAMI, T. SAKANUSHI, K. YAMADA ET AL. from the same reason, transfer functions from r(s) and d(s) to y(s) of the control system in (1) have finite numbers of poles. Next, we show that the controller in (11) works as a modified repetitive controller. The controller in (11) is rewritten by the form in (3), where X(s) + D(s)Q(s) C (s) = (28) 1 Y (s) − N(s)Q(s) and q¯(s) C (s) = . (29) 2 (Y (s) − N(s)Q(s)) From the assumption ofq ¯(s) =6 0, C2(s) =6 0 holds true. These expressions imply that the controller C(s) in (11) works as a modified repetitive controller. Thus, the sufficiency has been shown. We have thus proved Theorem 3.1. 4. Conclusions. In this paper, we proposed the parameterization of all stabilizing sim- ple repetitive controllers with the specified input-output characteristic. This work was supported by JSPS Grant-in-Aid for Scientific Research (A20560210). REFERENCES [1] T. Inoue, et al., High accuracy control magnet power supply of proton synchrotron in recurrent operation, Trans. Institute of Electrical Engineers of Japan, vol.C100, no.7, pp.234-240, 1980. [2] T. Inoue, S. Iwai and M. Nakano, High accuracy control of play-back servo system, Trans. Institute of Electrical Engineers of Japan, vol.C101, no.4, pp.89-96, 1981. [3] S. Hara, T. Omata and M. Nakano, Stability condition and synthesis methods for repetitive control system, Trans. Society of Instrument and Control Engineers, vol.22, no.1, pp.36-42, 1986. [4] S. Hara and Y. Yamamoto, Stability of multivariable repetitive control systems – Stability condition and class of stabilizing controllers, Trans. Society of Instrument and Control Engineers, vol.22, no.12, pp.1256-1261, 1986. [5] Y. Yamamoto and S. Hara, The internal model principle and stabilizability of repetitive control system, Trans. Society of Instrument and Control Engineers, vol.22, no.8, pp.830-834, 1987. [6] S. Hara, Y. Yamamoto, T. Omata and M. Nakano, Repetitive control system: A new type of servo system for periodic exogenous signals, IEEE Trans. Automatic Control, vol.AC-33, no.7, pp.659-668, 1988. [7] T. Nakano, T. Inoue, Y. Yamamoto and S. Hara, Repetitive Control, SICE Publications, 1989. [8] S. Hara, P. Trannitad and Y. Chen, Robust stabilization for repetitive control systems, Proc. of the 1st Asian Control Conference, pp.541-544, 1994. [9] G. Weiss, Repetitive control systems: Old and new ideas, Systems and Control in the Twenty-First Century, pp.389-404, 1997. [10] T. Omata, S. Hara and M. Nakano, Nonlinear repetitive control with application to trajectory control of manipulators, J. Robotic Systems, vol.4, no.5, pp.631-652, 1987. [11] K. Watanabe and M. Yamatari, Stabilization of repetitive control system – Spectral decomposition approach, Trans. Society of Instrument and Control Engineers, vol.22, no.5, pp.535-541, 1986. [12] M. Ikeda and M. Takano, Repetitive control for systems with nonzero relative degree, Proc. of the 29th CDC, pp.1667-1672, 1990. [13] H. Katoh and Y. Funahashi, A design method for repetitive controllers, Trans. Society of Instrument and Control Engineers, vol.32, no.12, pp.1601-1605, 1996. [14] K. Yamada, H. Takenaga, Y. Saitou and K. Satoh, Proposal for simple repetitive controllers, ECTI Transactions on Electrical Eng., Electronics, and Communications, vol.6, no.1, pp.64-72, 2008. [15] D. C. Youla, H. Jabr and J. J. Bongiorno, Modern Wiener-Hopf design of optimal controllers. Part I, IEEE Trans. on Automatic Control, vol.AC-21, pp.3-13, 1976. [16] V. Kucera, Discrete Linear System, The Polynomial Eqnarray Approach, Wiley, 1979. [17] J. J. Glaria and G. C. Goodwin, A parameterization for the class of all stabilizing controllers for linear minimum phase system, IEEE Trans. on Automatic Control, vol.AC-39, no.2, pp.433-434, 1994. [18] M. Vidyasagar, Control System Synthesis: A Factorization Approach, MIT Press, 1985. ÁÁ ÜÔÖ×× ÄØØÖ× ÁÁ ÁÒØÖÒØÓÒÐ ¾¼½¼ ÁËËÆ ½½¹¼¿ ÎÓÐÙÑ ¸ÆÙÑÖ ´µ¸ Ç ØÓ Ö ¾¼½¼ ÔÔº ½¹½ ÁÅ ÆÇÁË ÊÅÇÎÄ Ë ÇÆ ÁÅ ÌÁÄ ÈÊËÊÎÁÆ ½ ½ ¾ ½