ࡱ> |}~9 bjbjr^l2222T ;;;;?< `AAAAAVVV*`,`,`,`,`,`,`$aa cTP`VQVVVP`[X22AAe`[X[X[XV2AA*`[XV*`[X[X\]K:hv]AA `/5T 2;Vf]v]{`0`n]cVVcv][X 2222INdAM Istituto Nazionale di Alta Matematica "Francesco Severi" Numerical Methods for Local and Global Optimization: Sequential and Parallel Algorithms "Il Palazzone", Cortona (Italy) July 14-20, 2003 scientific committee Roger Fletcher (University of Dundee, U.K.) Valeria Ruggiero (University of Ferrara, Italy) Yaroslav D. Sergeyev (University of Calabria, Cosenza, Italy, and University of Nizhni Novgorod, Russia) Roman G. Strongin (University of Nizhni Novgorod, Russia) OPT2003 Contents Introduction 1 Program 2 Main lectures 3 Seminars 5 Contributed abstracts 10 List of participants 18 The school is partially supported by the Italian FIRB Project RBAU01JYPN. organizing committee Valeria Ruggiero (University of Ferrara, Italy) Emanuele Galligani (University of Modena and Reggio Emilia, Italy) Gaetano Zanghirati (University of Ferrara, Italy) addresses Il Palazzone - Scuola Normale Superiore di Pisa via Case sparse, 193 - 52044 Cortona (Arezzo) - Italy Tel: +39-050-509398/9 Fax: +39-0575-630164 OPT2003: http://dm.unife.it/opt2003 July 14-20, 2003 OPT2003_______________________________________________________________introduction Introduction The current methodology of designing highly efficient technological systems needs to choose the best combination of the parameters affecting the performance. The solution of very difficult constrained nonlinear optimization problems is typically the means to identify these parameters. However, the existence of multiple local minima and the lack of sufficient regularity of the models often requires us to consider these problems as "multiextremal" or global optimization problems, instead of as local optimization problems. Another important issue is that both the local and the global problems are large-scale optimization problems, mainly for real world applications. For these applications the development of effective numerical methods, well suited for the modern computing systems like distributed memory parallel computers, is necessary. The goal of the proposed scientific meeting is to discuss the state-of-the-art of the research of local and global large-scale optimization methods, by encouraging researchers on different topics to share their ideas and experiences and by favoring the comparison of different solution techniques. In particular, the following topics have a great relevance in the research: convergence analysis of inexact Newton methods for large-scale unconstrained nonlinear local optimization problems, with special attention to the inner iterative solver, possibly using parallel computing technique; the comparison of the Lagrangian approach (augmented Lagrangian methods, perturbed damped Newton or interior point methods) with the successive quadratic programming techniques (SQP) for local constrained nonlinear optimization problems; analysis and practice of inner iterative solvers and of the recent filter techniques that are used to globalize SQP and related methods; analysis of the main global optimization issues (computational complexity, formulation of efficient stopping criteria, identification of suitable test functions, techniques for algorithm convergence properties analysis); comparison of stochastic and deterministic techniques in global optimization; methods based on the domain covering by means of Peano-Hilbert curves for Lipschitzian or Holderian functions and their generalization to multidimensional problems, also via adap tive curves; these methods require particular techniques (e.g. "local tuning") in order to adaptively estimate the involved parameters (e.g. the Lipschitz constant or the Holder con stant) ; design, analysis and evaluation of parallel schemes based on the "nonredundant parallelism" concept. The cited topics are mostly covered by the following books: [1] R. Fletcher, Practical Methods of Optimization, 2nd ed., John Wiley and Sons, 1987. [2] J. Nocedal, S.J. Wright, Numerical Optimization, Springer Verlag, 1999. [3] R.G. Strongin, Y.D. Sergeyev, Global Optimization with non-convex constraints: sequential and parallel algorithms, Kluwer Academic Publ., 2000. Cortona, Italy 1 program OPT2003 Program Mon 14Tue 15Wed 169:00 -10:30 R.G. StronginR. FletcherYa.D. Sergeyev.. COFFE BREAK .11:00 - 12:30R. FletcherR.G. Strongin11:00 - 11:40 M. Gaviano 11:40 - 12:10 G. Liuzzi 12:10- 12:40 G. Fasano ________________________ LUNCH ____________________________ 14:30 - 15:10F. SchoenS. LeyfferYu.G. Evtushenko15:10 -15:50P. ContucciL. ZanniD. Lera15:50 -16:20B. AddisB. ColsonD. Kvasov. COFFE BREAK .16:40 -17:10S. ButenkoT. SerafiniB. Morini17:10 -17:40G. SpalettaS. BonettiniG.Landi17:40-16:10 F. TintiF. Zama Thu 17Fri 189:00 - 10:30V. Ruggiero 9:00 - 10:00 E. Galligani 10:00 - 11:00 V. Ruggiero COFFE BREAK ..11:00 - 12:30Ya.D. Sergeyev11:30 - 12:30 E. Galligani_____________________ LUNCH ___________________14:30 - 15:10 15:10 - 15:50 15:50 - 16:20L. Luksan G. Toraldo S. Pieraccini.. COFFE BREAK .16:40 - 17:10 17:10 - 17:40L. Bergamaschi V. De Simone 2 July 14-20, 2003 OPT2003 main lectures Main lectures Roger Fletcher University of Dundee (Schotland, UK). The lectures will focus around the description of SQP, SLP and Augmented Lagrangian Methods, and issues relating to the globalization of these methods. This will start by referring to some early techniques, and extending to include filter techniques and possibly some comments on line search techniques such as are used in SNOPT. Recent ideas currently under development for non-monotonic filter techniques will also be included. References [1] R. Fletcher, Practical Methods of Optimization, 2nd Edn, Wiley, Chichester, 1987. [2] R. Fletcher, N.I.M. Gould, S. Leyffer, Ph.L. Toint, A. Wachter, Global Convergence of aTrust-Region SQP-Filter Algorithms for Nonlinear Programming, SIAM J. Optimization 13 (2002), 635-659. [3] R. Fletcher, S. Leyffer, Ph.L. Toint, On the Global Convergence of a Filter-SQP Algorithm, Tech. Rep. Dundee Numerical Analysis Report NA/197, (October 2000) and in SIAM J. Optimization 13 (2002), 44-59. [4] C.M. Chin, R. Fletcher, On the Global Convergence of an SLP-Filter Algorithm that takes EQP steps, Tech. Rep. Dundee Numerical Analysis Report NA/199, (2001), accepted for publication in Mathematical Programming. Valeria Ruggiero University of Ferrara (Italy). We consider the solution of nonlinear systems with restriction on the sign of some variables, as those arising from the Karush-Kuhn-Tucker optimality conditions for Nonlinear Optimization problems. These systems can be solved by the Newton Interior Point iterative method. At each iteration the method computes a direction by solving a perturbed Newton equation and determines a damping parameter to state a convenient steplength. Crucial features of the method are the analysis of the perturbation parameter, the solution of the inner linear system and the analysis of the damping parameter. We consider the inexact solution of the linear system by different kinds of methods. A set of numerical results on optimal control problems are reported. References [1] C. Durazzi, V. Ruggiero, Indefinitely Preconditioned Conjugate Gradient Method for Large Sparse Equality and Inequality Constrained Quadratic Problems, Numerical Linear Algebra with Applications (2003), to appear. [2] C. Durazzi, V. Ruggiero, Numerical Solution of Special Linear and Quadratic Programs via a Parallel Interior-Point Method, Parallel Computing 29(4) (2003), 485-503. [3] C. Durazzi, V. Ruggiero, G. Zanghirati, Parallel Interior-Point Method for Linear and Quadratic Programs with Special Structure, JOTA 110(2) (2001), 289-313. [4] C. Durazzi, V. Ruggiero, A Newton Inexact Interior-Point method for large scale nonlinear optimization problems, Annali dell'Universita di Ferrara (2003), to appear. Yaroslav D. Sergeyev University of Calabria (Italy) and Nizhni Novgorod State University (Russia). The lectures deal with univariate and multivariate Lipschitz global optimization problems. The new concept of non-redundant parallelism is introduced to construct global search algorithms for Cortona, Italy 3 main lectures OPT2003 multiprocessor systems. Efficient sequential methods that can be parallelized in the framework of this concept are developed. Theoretical estimates of efficiency of parallel methods constructed on their basis are given. A special attention is paid to elaboration of efficient adaptive schemes for estimating unknown Lipschitz constants. References [1] R.G. Strongin, Ya.D. Sergeyev, Global Optimization with non-convex constraints: sequential and parallel algorithms, Kluwer Academic Publ., 2000. [2] V.A. Grishagin, Ya.D. Sergeyev, R.G. Strongin, Parallel characteristical global optimization algorithms, Journal of Global Optimization 10 (1997), 185-206. [3] Ya.D. Sergeyev, V.A. Grishagin, Parallel asynchronous global search and the nested optimization scheme, J. Comput. Anal. Appl. 3(2) (2001), 123-145. [4] R.G. Strongin, Ya.D. Sergeyev, Global optimization: fractal approach and non-redundant parallelism, Journal of Global Optimization 27(1) (2003), 25-50. [5] Ya.D. Sergeyev, P. Pugliese, D. Famularo, Index information algorithm with local tuning for solving multidimensional global optimization problems with multiextremal constraints, Mathematical Programming 96(3) (2003), 489-512. Roman G. Strongin University of Nizhni Novgorod (Russia). Multiextremal global optimization problems are under consideration. All functions employed in the problem formulations are supposed to be Lipschitzian. It is assumed that the basic operation available for collecting information needed to assess the sought global optimizer is a trial, which is in running some (often) cumbersome black box to evaluate the left-hand sides of the constraints and the function to be optimized at a given trial point. Each trial is to be immediately terminated if the violation of some constraint is encountered. This economizes the search effort and allows to tackle the problems with partially defined characteristics. Multidimensional problems are first reduced to the one-dimensional equivalents by employing Peano-type space filling curves. This is supported by some software realizing (theoretically justified) algorithms. Multiple Peano scannings are suggested for better translation of metric properties from many dimensions to one-dimensional scales. Constrained one-dimensional problems are then reduced to the unconstrained ones (with separate account of each constraint). Solving of these reduced problems is performed by applying specially derived efficient one-dimensional search procedures (index algorithms). Convergence theory is presented. The possibility to accelerate the rate of convergence by adaptively changing the order of the constraint's examination in the Search process is exhibited. Regularity conditions (in the form of the e-reserved solutions) are introduced and implemented to further accelerate the search. Numerical examples are provided. It is demonstrated how all the needed unknown parameters (including the Lipschitz constants) could be substituted by their running estimates. References [1] R.G. Strongin, Ya.D. Sergeyev, Global Optimization with Non-Convex Constraints, Sequential and Parallel Algorithms, Kluwer Academic Publishers, Dordrecht, 2000. [2] K.A. Barkalov, R.G. Strongin, A Global Optimization Technique with an Adaptive Order of Checking for Constraints, Computational Mathematics and Mathematical Physics 42(9) (2002), 1289-1300. 4 July 14-20, 2003 OPT2003__________________________________________________________________seminars Seminars Local and Global Combinatorial Optimization: a Statistical Mechanics Approach Pierluigi Contucci, University of Bologna (Italy). mon 15, 15:10 We will introduce a class of combinatorial optimization problems that emerge in a variety of domains among which: condensed matter physics, theory of financial risks and error correcting codes in information transmissions. We will show the performances of two algorithms: the "greedy" (quick decrease along the gradient) and the "reluctant" (slow decrease close to the level curves) as well as those of a "stochastic convex interpolation" of the two. Concepts like the average relaxation time and the wideness of the attraction basin will be analised and their system size dependence illustrated. Application of Theorems of alternative to numerical methods Yury G. Evtushenko, Dorodnicyn Computing Centre of Russian Academy of Sciences (Russia). A.I. Golikov, Dorodnicyn Computing Centre of Russian Academy of Sciences (Russia). wed 16, 14:30 Theorems of the alternative (TA) lie at the heart of mathematical programming. TA were used to derive necessary optimality conditions for LP and NLP problems and for various other pure theoretical investigations. We show that TA give us an opportunity to construct new numerical methods for solving linear systems with equalities and inequalities, to simplify computations arising in the steepest descent method, to propose new methods for solving LP problem, to construct the separating plane, etc. With original linear system we associate an alternative system such that one and only one of these systems is consistent. Moreover an alternative system is such that the dimension of its variable equals to the total amount of equalities and inequalities (except constraints on the signs of variables) in the original system. If the original system is solvable then numerical method for solving this system consists of minimization of the residual of the alternative inconsistent system. Prom the results of this minimization we determine a normal solution of the original system. Since the dimensions of the variables in original and alternative systems are different, the passage from the original consistent system to the minimization problem for the residual of the alternative inconsistent system may be very reasonable. This reduction may lead to the minimization problem with respect to variable of lower dimension and makes it possible to determine easy a normal solution of the original system. Proposed technique does not need an a priori assumption regarding the consistency of the original system. The essence of the method is based on the duality theory. This Research has been supported by the Russian Foundation for Basic Research (Grants No 01-01-00804 and No 00-15-96080). References [I] A.I. Golikov, Yu.G. Evtushenko, A New Method for Solving Systems of Linear Equalities and Inequalities,,- Doklady Mathematics 64(3) (2001), 370-373. [2] A.I. Golikov, Yu.G. Evtushenko, Theorems of the Alternative and Their Applications in Numerical Methods, Computational Mathematics and Mathematical Physics 43(3) (2003), 338-358. Cortona, Italy 5 seminar _________________________________________________________________________OPT2003 Additive Operator Splitting (AOS) methods for solving systems of nonlinear equations in finite differences Emanuele Galligani,, University of Modena and Reggio Emilia (Italy). fri 18, 9:00 Operator splitting is an important concept used in different fields of Mathemat ics in order to design effective numerical methods. There exists al lot of lit erature on the development, analysis and implementation of multiplicative and additive operator splitting methods for solving large and sparse systems of differ ence equations arising in the discretization of partial differential equations. The drawback of the high computational complexity for the implicit Euler and Crank- Nicolson methods for solving the higher dimensional diffusion equations has been known since the early numerical -studies on parabolic equations. The work of Peaceman and Rachford on J. SIAM (1955) is a famous example to overcome this difficulty by operator splitting methods. The Alternating Direction Implicit (ADI) method belongs to the class of Multiplicative Operator Splitting methods; to this class also belong the Alternating Direction Explicit (ADE) method and the Fractional Step or Local One Dimensional (LOD) method. The Arithmetic Mean method proposed in [1] belongs to the class of Additive Operator Split ting (AOS) methods. These methods are ideally suited for an implementation on parallel computers; besides, the additivity is essential to make the splitting symmetric. At present, there exists a lot of interest in AOS methods for solving nonlinear systems of difference equations [2,3,4]. This work is concerned with a convergence analysis of the Arithmetic Mean method for solving a large class of nonlinear difference equations arising from the discretization of steady-state and time-dependent diffusion processes. References [1] V.K. Sau'lyev, Integration of Equations of Parabolic Type by the Method of Nets, Pergamon Press, Oxford, 1964. [2] Z.Z. Bai, A class of two-stage iterative methods for systems of weakly nonlinear equations, Numerical Algorithms 14 (1997), 295-319. [3] D. Wang, Z.Z. Bai, D.J. Evans, On the monotone convergence of multisplit-ting method for a class of systems of weakly nonlinear equations, Intern. J. Computer Math. 60 (1996), 229-242. [4] E. Galligani, The Newton-arithmetic mean method for the solution of systems of nonlinear equations, Appl. Math. Comput. 134 (2003), 9-34. Complexity analysis in minimization problems Marco Gaviano, University of Cagliari (Italy). wed 16, 11:00 The investigation of the numerical complexity of algorithms that minimize functions from Rn into R is a very difficult issue. In recent years new results have been found giving useful guidelines for a better understanding of the algorithm behaviors and even for improving their performances. Many results established for general or specific minimization problems are reviewed. The case of algorithms for global minimization is also considered. Stochastic global optimization methods Daniela Lera, University of Cagliari (Italy). 6 July 14-20, 2003 OPT2003 seminars wed 16, 15:10 In this seminar we will present some stochastic techniques for solving global optimization problems. Stochastic methods are techniques that contain some stochastic elements. This means that either the outcome of the method is itself a random variable (i.e. algorithms in which the placement of observations is based on the generation of random points in the domain of the objective function), or we consider the objective function to be a realization of a stochastic process. Excellent surveys on the subject are in [1-4]. Here we will discuss, in particular, the so-called two-phase methods, i.e. methods in which both a global phase and a local phase can be distinguished. In the global phase, the function is evaluated in a number of randomly sampled points. During the local phase the sample points are "manipulated" in order to yield a candidate global minimum. We will give a brief presentation of clustering techniques and finally we will show random search methods and Simulated Annealing algorithms. References [1] B. Betro, Bayesian methods in global optimization, Journal of Global Optimization 1(1) (1991), 1-14. [2] C.G.E. Boender, H.E. Romeijn, Stochastic methods, Handbook of global optimization, in Nonconvex Optim. Appl., 2 (1995), 829-869. [3] F. Schoen, Stochastic tecniques for global optimization: a survey of recent advances, Journal of Global Optimization 1 (1991), 207-228. [4] A.A. Torn, A. Zilinskas, Global Optimization, Springer-Verlag, 1989. Solving Complementarity Constraints via Nonlinear Optimization Sven Leyffer, Argonne National Laboratory (USA). tue 15, 14:30 An exciting new application of nonlinear programming techniques is mathematical programs with complementarity constraints (MPCC). Problems of this type arise in many engineering and economic applications. Recently, it has been shown that MPCCs can be solved efficiently and reliably as nonlinear programs. This talk reviews these developments and examines alternative nonlinear formulations of the complementarity constraints. Unlike standard smoothing techniques, the new reformulations do not require the control of a smoothing parameter. Thus they have the advantage that the smoothing is exact, in the sense that Karush-Kuhn-Tucker points of the reformulation correspond to strongly stationary points of the MPCC. Interior-point methods for large-scale nonlinear nonconvex optimization Ladislav Luksan, Institute of Computer Science, Praha, and Technical University of Liberec, Liberec (Czech Republic). thu 17, 14:30 The contribution is devoted to interior-point methods for largescale nonlinear nonconvex optimization. We describe two types of such methods based on line-search and trust-region approaches. In these methods, inequality constraints are split into active and inactive subsets to overcome problems with instability. Inactive constraints are eliminated directly while active constraints are used to define symmetric indefinite linear system. Inexact solution of this system is obtained iteratively using indefinitely preconditioned conjugate gradient methods. We show that a suitable scaling of the trust-region subproblem leads to preconditioners used in line search methods. We also study a merit function, Cortona, Italy 7 seminars OPT2003 which includes effect of possible regularization. This regularization can be used to overcome problems with near linear dependence of active constraints. References [1] L. Luksan, J. Vlacek, Indefinitely Preconditioned Inexact Newton Method for Large Sparse Equality Constrained Nonlinear Programming Problems, Numerical Linear Algebra with Applications 5 (1998), 219-247. [2] L. Luksan, J. Vlacek, Numerical experience with iterative methods for equality constrained nonlinear programming problems, Optimization Methods and Software 16 (2001), 257-287. [3] L. Luksan, M. Ctirad, J. Vlacek, Interior-point method for nonlinear noncon-vex optimization, to appear in Numerical Linear Algebra with Applications. [4] L. Luksan, M. Ctirad, J. Vlacek, Nonsmooth equation method for nonlinear nonconvex optimization, Proceedings of the conference on Finite Element Methods, Jyvaskyla 2002. Stochastic methods for global optimization and molecular conformation problems Fabio Schoen, University of Florence. mon 14, 14:30 Some stochastic techniques for global optimization will be introduced, and their applicability to nonconvex minimization in the field of minimum energy molecular conformation problems will be discussed. The systems which will be considered range from atomic clusters interacting through pair potentials (with particular reference to Lennard-Jones and Morse potential functions) and pairs of large bio-molecules (and, in particular, proteins) which interact in such a way as to form a complex. It is widely recognized that the best available methods for this kind of highly structured problems are stochastic ones, composed of a suitable mix of stochastic exploration techinques coupled with deterministic local searches. This talk will outline the main problems within this framework, the best algorithmic approaches, the results obtained so far, in particular those obtained by the optimization group in Florence, which recently led to the discovery of 5 new putative global optima for Morse clusters. References [1] M. Locatelli, F. Schoen, Efficient Algorithms for Large Scale Global Optimization: Lennard-Jones Clusters, Computational Optimization and Applications (2003), to appear. [2] M. Locatelli, F. Schoen, Fast Global Optimization of Difficult Lennard-Jones Clusters, Computational Optimization and Applications 21 (2002), 55-70. [3] M. Locatelli, F. Schoen, Minimal interatomic distance in Morse clusters, Journal of Global Optimization 22 (2002), 175-190. [4] B. Addis, F. Schoen, .A randomized global optimization method for rigid protein-protein docking, (2003), submitted. Parallel computing in interior-point methods Gerardo Toraldo, University of Naples "Federico II" (Italy). Marco D'Apuzzo, Second University of Naples (Italy). thu 17, 15:10 The effect of parallel computing in the field of the numerical computation, and more specifically of the linear and nonlinear optimization has been extremely 8 July 14-20, 2003 OPT2003 seminars intense in the last decade, mainly drawn from the availability of low cost high performance computers machines. In this talk a very short overview of such impact will be outlined. Then the talk will focus on the parallel computational issues arising in implementing Interior Point Algorithms. The usually few but computationally expensive iterations of the interior point algorithms turn out to be an advantage, since they can be efficiently parallelized by using standard (parallel) linear algebra software tools. In the context of parallel processing, iterative methods are very attractive because of their features, therefore their use in Interior Point Methods framework definitely deserves to be investigated, including the use of ad hoc strategies needed to deal with the ill conditioning naturally arising in such methods. References [1] M. D'Apuzzo, M. Marino, P.M. Pardalos, G. Toraldo, A Parallel Implementation of a Potential Reduction Algorithm for Box-Constrained Quadratic Programming, in Lecture Notes in Computer Science, 1900 - EuroPar 2000, Springer-Verlag, 2000, 839-848. [2] R.B. Schnabel, A view of the limitations, opportunities and challenges in parallel nonlinear optimization, Parallel Computing 21(6) (1995), 875-905. [3] M.H. Wright, III-conditioning and computational error in interior methods for nonlinear programming, SIAM J. Optim. 9(1) (1998), 84-111. Gradient projection methods for quadratic programs Luca Zanni, University of Modena and Reggio Emilia (Italy). tue 15, 15:10 Gradient projection methods for quadratic programming problems with simple constraints are considered. Starting from the analysis of the classical versions of these schemes, some recent gradient projection algorithms are introduced and the importance of using appropriate linesearch strategies and steplength selection rules is stressed. Linesearch techniques based on both limited minimization rules and nonmonotone strategies are considered. For the steplength selection, the Barzilai-Borwein spectral rules are discussed and some new suggestions are presented. Numerical results showing the behaviours of the proposed approaches are reported. References [1] J. Barzilai, J.M. Borwein, Two Point Step Size Gradient Methods, IMA J. Numer. Anal. 8 (1988), 141-148. [2] E.G. Birgin, J.M. Martinez, M. Raydan, Nonmonotone Spectral Projected Gradient Methods on Convex Sets, SIAM J. Optim.10(4) (2000), 1196- 1211. [3] T. Serafini, G. Zanghirati, L. Zanni, Gradient Projection Methods for Quadratic Programs and Applications in Training Support Vector Machines, Tech. Rep. 48, Dept. of Math., University of Modena and Reggio Emilia, (2003). Cortona, Italy 9 contributed abstracts OPT2003 Contributed abstracts A procedure of global/local smoothing for global optimization Bernardetta Addis, University of Florence (Italy). 15:50 mon A large number of applications can be modelled as global optimization problems in which the objective function has a large amount of local, but not global minima. Frequently these functions appear to have a peculiar behaviour, a funnel structure, i.e. the objective can be viewed as the sum of two functions, a simpler one with one or a few number of local minima and the other, a bounded one with a large number of oscillations. In other words these functions can be viewed as signal perturbed by some kind of "noise" .In this work we develop a procedure to cut off the noise. The method works on the constant piecewise function obtained by substituting the value at each point with the value of the minimum reached by a local search starting from the point itself. To amend the loss of continuity, and so the possibility to have information of the local behaviour of the function from a point, we applied an approximated local smooth on this new function. The smoothing procedure produce two effects: the reduction of possible remaining oscillations, and, most important, the possibility of finding descent directions and so of moving from a point to another reducing the value of the function. References [1] M. Locatelli, On the multilevel structure of global optimization problems, (2003), submitted. [2] J.J. More, Z. Wu, Issues in Large Scale Global Molecular Optimization, in Large Scale Optimization with Applications. Part III: Molecular Structure and Optimization, Springer, New York, 1997, 99-121. Preconditioners for Iterative Solvers Used in Interior Point Methods Luca Bergamaschi, University of Padua (Italy). Giovanni Zilli, University of Padua (Italy). Jacek Gondzio, University of Edinburgh (UK). 17, 16:40 thu Every iteration of the interior point method for linear, quadratic or nonlinear programming requires the solution of a possibly very large and almost always very sparse linear system. In this communication we concentrate on the advantages of iterative solution techniques applied to this system. A number of methods has been implemented including BiCGSTAB, GMRES and QMR and the classical Conjugate Gradient method. Since the linear systems tend to be very ill-conditioned as the interior point iteration approaches the solution, iterative methods fail to provide sufficient accuracy unless appropriately preconditioned. In this communication we concentrate on a number of indefinite preconditioners. Our concern [2] is to find a significantly sparser factorization than that of original system and still capture most of the numerical properties of this system. We illustrate our approach by applying it to a number of public domain large quadratic problems with sizes reaching 100,000 variables. From the numerical results we conclude that the solution times for^such problems on a modern PC are measured in minutes when direct methods are used and drop to seconds when iterative methods with appropriate preconditioners are used. References [1] A. Altman, J. Gondzio, Regularized symmetric indefinite systems in interior point methods for linear and quadratic optimization, Optimization Methods and Software 11-12 (1999), 275-302. [2] L. Bergamaschi, J. Gondzio, G. Zilli, Preconditioners for Iterative  10July 14-20, 2003 OPT2003 contributed abstracts Solvers Used in Interior Point Methods, (2003), in preparation. A Nonmonotone Inexact Newton Method Silvia Bonettini, University of Modena e Reggio Emilia (Italy). tue In this work we describe a variant of the inexact Newton method for solving nonlinear systems of equations. We define a nonmonotone inexact Newton step and a nonmonotone backtracking strategy. For this nonmonotone scheme we present the convergence theorems. Finally, we show how we can apply these strategies to Newton inexact interior-point method and we present some numerical examples. References [1] S.C. Eisenstat, H.F. Walker, Globally Convergent Inexact Newton Method, SIAM Journal on Optimization 4 (1994), 393-422. [2] L. Grippo, F. Lampariello, S. Lucidi, A nonmonotone line search technique for Newton's method, SIAM Journal of Numerical Analysis 23 (1986), 707-716. 15, 15:50 Graph Optimization Approach to Studying the Market Structure Sergiy Butenko, University of Florida (USA). 14, 17:10 mon In the modern world of information, one often encounters serious challenges of dealing with massive datasets. In many practically important cases, a massive dataset can be represented as a very large graph with certain attributes associated with its vertices and edges. Studying the structure of this graph is essential for understanding the structural properties of the application it represents. Well-known examples of applying this approach are the Internet graph, the Web graph, and the Call graph, all of which obey the power-law model. Here we consider another important application: representing the stock market as a graph. Stock markets generate huge amounts of data, which can be used for constructing the market graph that reflects the market behaviour. We show that this graph also follows the power-law model. Moreover, we detect cliques and independent sets in this graph. These special formations have a clear practical interpretation, and their analysis gives us a valuable tool of classifying financial instruments into different groups, which provides a deeper insight into the internal structure of the stock market. References [1] V. Boginski, S. Butenko, P.M. Parda-los, On Structural Properties of the Market Graph, in Innovation in Financial and Economic Networks, A. Nagurney, ed., Edward Elgar Publishers, 2003, to appear. On some Derivative-Free Optimization algorithms using structure Benoit Colson, University of Namur (Belgium). tue We study some derivative-free optimiza-tion algorithms for solving mathematical programs involving expensive functions. The main ingredients of these algorithms are trust regions, suitable interpolation models (Newton fundamental polynomials) and procedures for improving the geometry of the interpolation set. A basic algorithm is extended to solve problems where the Hessian is known to have a particular structure. The case of banded and sparse Hessian is studied, as well as partially separable functions. Finally, we present a further extension of this algorithm in view of solving constrained optimization problems using an SQP-filter approach. The presentation includes implementation details and practical issues as well as complete numerical results obtained with dedicated softwares. References  Cortona, Italy  11 contributed abstracts OPT2003 [1] B. Colson, Ph.L. Toint, A derivative-free algorithm for sparse unconstrained optimization problems, in Trends in Industrial and Applied Mathematics, A.H. Siddiqi and M. Kocvara, eds., Proceedings of the-1st International Conference on Industrial and Applied Mathematics of the Indian Subcontinent, 131-147, Kluwer Academic Publishers, Dordrecht, 2002. [2] B. Colson, Ph.L. Toint, Exploiting band structure in unconstrained optimization without derivatives, Optimization and Engineering 2(4) (2001), 399-412. [3] A.R. Conn, K. Scheinberg, Ph.L. Toint, A derivative free optimization algorithm in practice, Tech. Rep. TR98/11, Department of Mathematics The University of Namur, Namur, Belgium, 1998. [4] R. Fletcher, N.I.M. Gould, S. Leyffer, Ph.L. Toint, A. Wachter, Global convergence of a trust-region SQP-filter algorithm for general nonlinear programming, Tech. Rep. 99/03, Department of Mathematics FUNDP, Namur, Belgium, 1999. Revised version, July 2001. An Interior Point Solver for Large-Scale Quadratic Programs Valentino, De Simone, Second University of Naples (Italy). S. Cafieri, A. Mucherino, Second University of Naples (Italy). 17, 17:10 thu We present a primal-dual potential reduc-tion solver for convex quadratic programming problems. We focus on the large and sparse case, and on the use of iterative techniques to solve the linear system arising at each iteration of the method. In order to deal with the increasing ill-conditioning of such systems when the iterates get close to the boundary of the feasible region, we use a preconditioning technique based on incomplete factorizations with predictable memory requirements. Moreover, we propose suitable termination rules allowing to adapt the accuracy requirement in solving the linear systems on the quality of the current iterate, in order to avoid unnecessary inner iterations when the iterates are far from the solution. We show some results of computational experiments carried out an a IBM SP system and on a cluster of PC's, variyng the preconditioning and the adaptive termination strategies. We also compare the proposed solver, based on the best combination of the considered strategies, with the package MOSEK. References [1] C.J. Lin, J.J. More, Incomplete Cholesky factorizations with limited memory, SIAM J. Sci. Comput. 21 (1999), 24-45. [2] M. D'Apuzzo, M. Marino, P.M. Pardalos, G. Toraldo, A Parallel Implementation of a Potential Reduction Algorithm for Box-Constrained Quadratic Programming, in Lecture Notes in Computer Science, 1900, A. Bode et al, eds., Springer-Verlag, 2000, 839-848. [3] S. Cafieri, M. D'Apuzzo, M. Marino, A. Mucherino, G. Toraldo, An Interior Point Solver for Sparse Quadratic Programs with Bound Constraints, Tech. Rep. CPS TR-06-02 Center for Research on Paralr lei Computing and Supercomputers, CPS-CNR, (2002), submitted. , [4] M. D'Apuzzo, M. Marino, Parallel Computational Issues of an Interior Point Method for Solving Large Bound-Constrained Quadratic Programming Problems, Parallel Computing 29(4) (2003), 467-483. GKLS-Generator for Testing Global Optimization Algorithms Dmitri Kvasov, University of Rome "La Sapienza" (Italy) and University of Nizhni Novgorod (Russia). Marco Gaviano, Daniela Lera, University of Cagliari (Italy).  12 July 14-20, 2003 OPT2003 contributed abstracts Yaroslav D. Sergeyev, University of Calabria (Italy) and University of Nizhni Novgorod (Russia). wed Development of efficient global optimiza-tion algorithms is impossible without examination of their behaviour on sets of sophisticated test problems. The lack of complete information (such as number of local minima, their locations, attraction regions, local and global values, ecc.) describing global optimization tests taken from real-life applications does not allow to use them for studying and verifying validity of the algorithms. That is why a significant effort is made to construct test functions [1-4]. In this communication, a procedure for generating three types (non-differentiable, continuously differentiable, and twice continuously differentiable) of classes of multidimensional and multiextremal test functions with known local and global minima is presented. The procedure consists of defining a convex quadratic function sj^stematically distorted by polynomials. Each test class provided by the GKLS-generator consists of 100 functions and is defined by the following parameters: (i) problem dimension, (ii) number of local minima, (iii) global minimum value, (iv) radius of the attraction region of the global minimizer, (v) distance from the global minimizer to the quadratic function vertex. The other necessary parameters are chosen randomly by the generator for each test function of the class. A special notebook with a complete description of all functions is supplied to the user. References [1] C.A. Floudas, P.M. Pardalos, C. Ad-jiman, W.R. Esposito, Z.H. Gms, S.T.Harding, J.L. Klepeis, C.A. Meyer, C.A. Schweiger, Handbook of Test Problems in Local and Global Optimization, Kluwer Academic Publishers, Dordrecht, 1999. [2] M. Gaviano, D. Lera, Test functions with variable attraction regions for global optimization problems, J. Global Optim. 13(2) (1998), 207-223. [3] J. Pinter, Global optimization: Software, test problems, and applications, in Handbook of Global Optimization, 2 P.M. Pardalos, H.E. Romeijn, eds., Kluwer Academic Publishers, Dordrecht, 2002, 515-569. [4] F. Schoen, A wide class of test functions for global optimization, J. Global Optim. 3 (1993), 133-137. Fruitful use of Negative Curvatures for the convergence to Second Order Critical Points in Large Scale Unconstrained Optimization Giovanni Fasano, University of Roma "La Sapienza" (Italy). Stefano Lucidi, University of Roma "La Sapienza" (Italy). wed In this talk we are concerned with ef-16, ficiently solving the unconstrained opti- 12:10 mization problem min((x), x ( Rn, n large, by means of iterative algorithms which converge towards second order stationary points. The convergence to second order points requires an accurate knowledge of the properties of the Hessian matrix of ((x}. More and Sorensen [1] proved that the algorithms converging to second order stationary points compute search directions which imply proper "factorizations" for the Hessian matrix. However, direct techniques usually need the storage of matrices. On the contrary, in this paper, we assume that n is large and we are aimed at proposing efficient algorithms, which both exploit the second order information associated to ((x) and avoid the large memory storage. On this purpose, first, we prove that the use of suitable CG-type algorithms can supply factorizations which verify the properties pointed out by More and Sorensen, without requiring the storage of any matrix. Then, we describe a new algorithm converging to second order stationary points. The proposed method adopts a Truncated Newton scheme based on a CG-type  Cortona, Italy 13 contributed abstracts OPT2003 algorithm [2,3]. Some significant numerical results, which confirm the theoretical conclusions, are reported. References [1] J.J. More, D.C. Sorensen, On the use of directions of negative curvature in a modified Newton method, Mathematical Programming 16 (1979), 1-20. [2] G. Fasano, A Planar-Conjugate Gradient algorithm for Large Scale Unconstrained Optimization, Part I: Theory, Tech. Rep. 04-02, Diparti-mento di Informatica e Sistemistica "A. Ruberti" Universita di Roma "La Sapienza", (2002). [3] G. Fasano, A Planar-Conjugate Gradient algorithm for Large Scale Unconstrained Optimization, Part II: Application, Tech. Rep. 13-02, Dipar-timento di Informatica e Sistemistica "A. Ruberti" Universita di Roma "La Sapienza", (2002). The Total Variation Regularization in dynamic MR Imaging Germana Landi, University of Bologna (Italy). Elena Loli Piccolomini, University of Bologna (Italy). 16, 17:10 wed Total Variation (TV) regularization method, proposed by Rudin, Osher and Fatemi in [1], has recently become a very popular technique in image restoration problems. The motivation for its success is that the TV regularization performs well for denoising and deblurring while preserving sharp discontinuities. The good performance of the TV model makes it particularly attractive for biomedical imaging. We study the use of TV regularization technique in the reconstruction of dynamic Magnetic Resonance images [2]. Several iterative schemes have been proposed in the literature for solving the TV minimization problem. We compare a Newton and a fixed point method [3] performances on both test problems and real MR data. References [1] L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms, Physica D. 60 (1992), 259-268. [2] Z.P. Liang, P.C. Lauterbur, An efficient method for dynamic Magnetic Resonance Imaging, IEEE Trans. Med. Imag. 13(4) (1994),. [3] C.R. Vogel, M.E. Oman, Nonlinear total variation based noise removal algorithms, Physica D. 60 (1996), 227-238. A truncated Newton algorithm for large- scale constrained optimization problems converging to second order stationary points Giampaolo Liuzzi, University of Roma "La Sapienza" (Italy). Gianni Di Pillo, Stefano Lucidi, Laura Palagi, University of Roma "La Sapienza" (Italy). 16, 11:40 wed We consider the problem (P) min((x), subject to g(x) ( 0, where ( : Rn ( R and g : Rn ( Rm are three times continuously differentiate functions. For the solution of Problem (P), we define a primal-dual algorithm which generates a sequence converging to points satisfying the second order necessary conditions for optimality. This property can be enforced by using an efficient local algorithm globalized by means of a particular merit function. More in particular, the local algorithm, which is based on a truncated Newton's method, produces two directions. A first order direction, which ensures local superlinear convergence toward a KKT pair, and a second order direction which is used to enforce convergence toward second order stationary points of the overall algorithm. These two directions are employed to compute the new primal-dual iterate by means of a linesearch procedure based on a continuously differentiable exact augmented Lagrangian function. In this way we get a twofold result. On the one hand, we can  14 July 14-20, 2003 OPT2003 contributed abstracts assess the goodness of the search directions produced by the local algorithm and measure their progress toward a KKT pair. On the other hand, by exploiting the non-convexity of the merit function, iterates can escape from first-order stationary points and converge to second order ones. Preliminary numerical results are reported which shows the viability of the proposed method. Enhanced Newton methods for bound-constrained nonlinear systems Benedetto, Morini, University of Florence (Italy). Stefania Bellavia, Maria Macconi, University of Florence (Italy). wed The topic of this work is the reliable so-lution of bound-constrained systems of nonlinear equations. These problems are very frequent in the numerical solution of mathematical models of real life problems, see e.g. cross-sectional properties of structural elements, dimensions of mechanical linkages, concentrations of chemical species [2,3]. In order to solve bound-constrained systems, in [1] we generalized the trust-region strategy for unconstrained systems. The proposed method handles the bounds in a general and reliable way by using an affine scaling approach, exhibits global convergence and the ultimate rate of convergence to a solution within the box is quadratic. Here, we review the features of the method and show its computational performance on a wide range of real problems. Further, we point out the pitfalls of our affine scaling approach when the method approaches solutions on the border. Detailed numerical experiments will be given to support our theoretical remarks. References [1] S. Bellavia, M. Macconi, B. Morini, An affine scaling trust region approach to bound-constrained nonlinear systems, Applied Numerical Mathematics 44 (2003), 257-280. [2] L.G. Bullard, L.T. Biegler, Iterative linear programming strategies for constrained simulation, Computers and Chemical Engineering 15 (1991), 481-491. [3] W. Morton, Robustness, improvements for Newton's method: bounds, step restriction, restart strategies and solving singular linearisations, Tech. Rep. (2002). An Inexact Infeasible Interior Point method for Semidefinite Programming Sandra Pieraccini, Politecnico of Torin (Italy). Stefania Bellavia, University of Florence (Italy). thu SDP arises in a wide variety of areas as optimal control, structural optimization, eigenvalue optimization [1]. In this talk, we present a globally convergent infeasi-ble Interior Point (IP) method for SDP, obtained generalizing to SDP the infeasi-ble IP method for LP given in [2]. This generalization allows the use of inexact search directions. In fact, at each iteration of an IP method for SDP a linear system must be solved. Solving this system is an expensive task when the size of the problem is large. Moreover, when the current iterate is far from the solution it may be unnecessary to compute the search directions with a high accuracy. Then, it can be convenient to use iterative methods for solving the linear systems with an accuracy that increases as the solution is approached. Further, the coefficient matrix is generally dense even if the data matrices of the original problem are sparse. In this context, Krylov methods may be useful. In fact, Krylov methods only require the action of the matrix onto a vector and not the matrix itself and this allows to exploit the sparsity of the original problem. References [1] L. Vandenberghe, S. Boyd, Semidefinite Programming, Siam Review 38 (1996), 263-280.  Cortona, Italy 15 contributed abstracts OPT2003 [2] M. Kojima, N. Megiddo, S. Mizuno, A primal-dual infeasible interior-point algorithm for linear programming, Mathematical Programming 61 (1993), 263-280. Parallel Training of Support Vector Machines Thomas Serafini, University of Modena and Reggio Emilia (Italy). Gaetano Zanghirati, University of Ferrara (Italy). Luca Zanni, University of Modena and Reggio Emilia (Italy). tub We present a parallel approach for the solution of quadratic programming (QP) problems with box constraints and a single linear constraint, arising in the training of Support Vector Machines. In this kind of application the problem is dense and generally large-sized (((104). An iterative decomposition technique has been presented in [1], which solves the problem by splitting it into a sequence of very small QP (inner) subproblems (generally with size less than 102). The approach proposed in [3] follows this decomposition idea, but it is designed to split the whole problem into QP subproblems of sufficiently large size (> 103), so that they can be efficiently solved by special gradient projection methods [2]. On scalar architectures this new technique allows for comparable performance with those of the algorithm in [1], but it is much more suited for parallel implementations. In fact, parallel versions of the gradient projection methods can be applied to solve the large QP inner subproblems and the other expensive tasks of each decomposition iteration can be easily distributed among the available processors. We present several improvements of this approach and we evaluate their effectiveness on large-scale benchmark problems both in scalar and parallel environments. References [1] T. Joachims, Making Large-Scale SVM Learning Practical, in Ad- vances in Kernel Methods - Support Vector Learning, MIT Press, Cambridge, Mass., 1998. [2] T. Serafini, G. Zanghirati, L. Zanni, Gradient Projection Methods for Quadratic Programs and Applications in Training Support Vector Machines, Tech. Rep. 48, Dept. of Math., University of Modena and Reggio Emilia, Italy, (2003). [3] G. Zanghirati, L. Zanni, A Parallel Solver for Large Quadratic Programs in Training Support Vector Machines, Parallel Computing 29(4) (2003), 535-551. Symmetric composition of symmetric numerical integration methods Giulia Spaletta, University of Bologna (Italy). Mark Sofroniou, Wolfram Research, Inc., Champaign, Illinois (USA). mon This work focuses on the derivation of composition methods for the numerical integration of ordinary differential equations. In contrast to the Aitken-Neville algorithm used in extrapolation, composition can conserve desirable geometric properties of a base integration method, such as symplecticity [2]. We survey existing composition methods and describe results of a numerical search for new methods [3,4]. The optimization procedure that has been adopted [1] can be very intensive, especially when deriving high order composition schemes. To overcome this, we make use of parallel computation. Numerical examples indicate that the new methods perform better than previously known methods. References [1] P.E. Gill, W. Murray, M.H. Wright, Practical Optimization, Academic Press, London, 1981. [2] E. Hairer, Ch. Lubich, G. Wanner, Geometric Numerical Integration. Structure-Preserving Algorithms for Ordinary Differential Equations, in 16 July 14-20, 2003 OPT2003 contributed abstracts Springer Ser. Comput. Math., 31, Springer-Verlag, New York, 2002. [3] R.I. McLachlan, G.R.W. Quispel, Splitting methods, Acta Numerica 11 (2002), 341-434. [4] H. Yoshida, Construction of high order symplectic integrators, Phys. Lett. A. 150 (1990), 262-268. Numerical Solution for Pseudomonotone Variational Inequality Problems by Ex-tragradient Methods Federica Tinti, University of Padua (Italy). Valeria Ruggiero, University of Ferrara (Italy). tue In this work we have analyzed from the numerical point of view the class of projection methods for solving variational inequality problems. We focus on some specific extragradient methods making use of different choices of the steplengths a. Subsequently we have analyzed the hy-perplane projection methods in which we construct an appropriate hyperplane which strictly separates the current iterate from the solutions of the problem. Finally we have included a collection of test problems implemented in Matlab to illustrate the effectiveness of the proposed methods. References [1] E.N. Khobotov, Modification of the extra-gradient method for solving variational inequalities of certain optimization problems, U.S.S.R. Computational Mathematics and Matem-atical Physics 27 (1987), 120-127. [2] P. Marcotte, Application of Khobotov's algorithm to variational inequalities and network equilibrium problems, Inform. Systems Oper. Res. 29 (1991), 258-270. [3] M.V. Solodov, B.F. Svaiter, A new projection method for variational inequality problems, SI AM Journal Control Optimization 37(3) (1999), 756-776. Truncated conjugate gradient iterations for solving ill-posed problems Fabiana Zama, University of Bologna (Italy). Elena Loli Piccolomini, University of Bologna (Italy). wed A large variety of applications give raise naturally to ill-posed problems. Whenever the underlying physical or technical problem is modelled by an integral equation of the first kind with a smooth kernel. The data usually stem from measurements with a limited precision, i.e., only perturbed data are available. The inverse problem is ill-posed and requires regularization methods. In this work we describe an iterative algorithm for finding the solution and the regularization parameter. Truncated Conjugate Gradients Iterations are implemented for computing the solution, while the value of the regularization is determined in order to have decreasing values of the objective functional. We develop a stopping criterion for the CG-iterates which is linked to the noise level and the current value of the regularization parameter. References [1] A. Bjorck, Numerical methods for least squares problems, SIAM, 1996. [2] M.E. Kilmer, D.P. OLeary, Choosing regularization parameters in iterative methods for ill-posed problems, SIAM J. Matrix Anal. Appl. 22 (2001), 1204-1221. [3] A. Frommer, P. Maass, Fast CG-based methods for Tikhonov-Phillips regularization, SIAM J. Sci. Comput. 20 (1999), 1831-1850. [4] W. Gander, Least squares with a quadratic constraint, Numer. Math. 36 (1981), 291-307.  Gortona, Italy 17 LIST OF PARTICIPANTS OPT2003 List of participants Bernardetta Addis (University of Florence, Italy) Luca Bergamaschi (University of Padova, Italy) Silvia Bonettini (University of Modena and Reggio Emilia, Italy) Margherita Bresco (University of Salerno, Italy) Sergiy Butenko (University of Florida, USA) Sonia Cafieri (Second University of Naples, Italy) Benoit Colson (University of Namur, Belgium) Pierluigi Contucci (University of Bologna, Italy) Marco D'Apuzzo (Second University of Naples, Italy) Valentina De Simone (Second University of Naples, Italy) Carla Durazzi (University of Ferrara, Italy) Yury G. Evtushenko (Russian Academy of Science, Russia) . Giovanni Fasano (University of Roma "La Sapienza", Italy) Roger Fletcher (University of Dundee, UK) Elena Franchini (University of Padova, Italy) Emanuele Galligani (University of Modena and Reggio Emilia, Italy) Marco Gaviano (University of Cagliari, Italy) Nicola Guglielmi (University of L 'Aquila, Italy) Dmitri Kvasov (University Roma "La Sapienza", Italy) Germana Landi (University of Bologna, Italy) Daniela Lera (University of Cagliari, Italy) Sven Leyffer (Argonne National Lab., USA) Giampaolo Liuzzi (University of Roma "La Sapienza", Italy) Ladislav Luksan (National Academy of Sciences, Czech Republic) Maria Macconi (University of Florence, Italy) Benedetta Morini (University of Florence, Italy) Antonio Mucherino (Second University of Naples, Italy) Sandra Pieraccini (Politecnico of Turin, Italy) Stefania Ragni (University of Bari, Italy) Valeria Ruggiero (University of Ferrara, Italy) Fabio Schoen (University of Florence, Italy) Thomas Serafmi (University of Modena and Reggio Emilia, Italy) Yaroslav D. Sergeyev (Univ. of Calabria, Italy, and Nizhni Novgorod State Univ., Russia) Mark Sofroniou (Wolfram Research, USA) Giulia Spaletta (University of Bologna, Italy) Roman G. Strongin (University of Nizhni Novgorod, Russia) Federica Tinti (University of Padua, Italy) > Gerardo Toraldo (University of Naples, Italy) Fabiana Zama (University of Bologna, Italy) Gaetano Zanghirati (University of Ferrara, Italy) Luca Zanni (University of Modena and Reggio Emilia, Italy) Giovanni Zilli (University of Padova, Italy) 18 July 14-20, 2003 14, 15, 17:10 16, 15:50 16, 16:40 17, 15:50 15, 16:40 14, 17:40 15, 17:40 16, 17:40 >?Ust>萁ogXUUC#6@B*CJ]aJmH phsH CJ(jCJUmHnHsHu6]mH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 556@B*CJ.OJQJ\]^JaJ.mH phsH 556@ B*CJ.OJQJ\]^JaJ.mH phsH 6B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH ?Ut >w  d-DM ]^ ` d-DM ]^r -DM ^r$d-DM ]^a$$d-DM a$  -DM -DM ^ >  !n\L:L:L#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH -56@B*CJRHq\]aJmH phsH 1j56@B*U\]aJmHnHphu#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphumH sH #6@B*CJ]aJmH phsH !-K|eL  d-DM ^   d-DM ^   d-DM ^   d-DM ^   d-DM ^  m-DM ^ $ V-DM a$ J d-DM ]^ `J !),-FHJK_acd7hirsuv  ݴآݴ؎|jVIIݢB*CJaJmH phsH &6:@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH Kdis 0oo_ -DM ^ M d-DM ]M^  d;-DM ^ % d-DM ^% e d-DM ]e^ `f d-DM ]f^ ` #-DM   d-DM ^ /01ACJ  | 謚lZH6H#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 156B*CJOJQJ\]^JaJmH phsH )6:>*@B*CJ]aJmH phsH "6>*B*CJ]aJmH phsH &6>*@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphumH sH #6@B*CJ]aJmH phsH 0BCg >,sT & F (d-DM ^`( & F (d-DM ^`( & F (d-DM ^`($d-DM ]a$$"-DM ^"a$ ! -DM ^ $&f-DM ]&a$| 5 Y  f g =>+,:ɷۥzjXFjzjjjX#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH T-wbKds-DM ^`dF-DM ]^ dFH-DM ^  & F (d-DM ^`( & F (d-DM ^`( & F .-DM ^. & F (d-DM ^`(KSTM,-.˹xskVC%6:>*B*CJ]aJmH phsH )6:>*@B*CJ]aJmH phsH 6]mH sH mH sH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphu#6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH -$If$If$If ##-DM ^-DM ^ -DM %&12@ABCOZfghuvǽǽ߽we`[ǽe@CJ@CJ#6@B*CJ]aJmH phsH 6@B*]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6CJ]mH sH #6@B*CJ]aJmH phsH >*>*@6]mH sH mH sH &6>*@B*CJ]aJmH phsH #&2AndWWW $<<$Ifa$ <<$If$$Ifl\o "' Ij  064 laABCgnh[ $xx$Ifa$$If$$Ifl\o "' Ij  064 laghv~~ <<$If <<$Ifl$$Ifl40o ' 064 la"n$h[ $xx$Ifa$$If$$Ifl\o "' Ij  064 la!"#1:;EFVWXdepqyzʸܖʖrܖ`rܖP6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6]mH sH #6@B*CJ]aJmH phsH 6CJ]mH sH "#1;FW{{{ $<<$Ifa$ <<$Ifl$$Ifl40o ' 064 laWXeqzndWWWndWWW $<<$Ifa$ <<$If$$Ifl\o "' Ij  064 la nh[ $xx$Ifa$$If$$Ifl\o "' Ij  064 la23?AJRUW^efrsųͮśtb͛RD6B*]aJmH phsH 6@B*]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @6B*CJ]aJmH phsH CJ@CJ#6@B*CJ]aJmH phsH 6]mH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6@B*]aJmH phsH {{{ $<<$Ifa$ <<$Ifl$$Ifl40o ' 064 la*23@AndWWWndW $<<$Ifa$ <<$If$$Ifl\o "' Ij  064 laAJRSTUVa__G  cdv-DM ^c$$Ifl\o "' Ij  064 la $<<$Ifa$VW^efsN<H$If~$$IflFj  N 06    4 la  dv$If$  dv$Ifa$  dv$Ifq^$  dv$Ifa$  dv$If~$$IflFj  N 06    4 la r  dv$If  dv$Ifl$$Ifl40j  a064 la  >?@ijtʸ練zp`̈́[VVmH sH @CJ6@B*]aJmH phsH 6CJ]mH sH @CJ@CJ#6@B*CJ]aJmH phsH 6@B*]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH CJ6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH ! ?q^$  dv$Ifa$  dv$If~$$IflFj  N 06    4 la?@N\jt<xriY  dv$If $$Ifa$$If   $If  $Ifl$$Ifl40j  a064 laq^$  dv$Ifa$  dv$If~$$IflFj  N 06    4 laxrb  dv$If$If   $If  $Ifl$$Ifl40j  a064 lalUCC " -DM   8Pdv-DM ^P  Pdv-DM ^P~$$IflFj  N 06    4 laĮxdxdME6@B*CJaJmH phsH 5\mH sH -56@B*CJRHx\]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphumH sH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH 'j6B*U]aJmHnHphuySrZZ$7Wd-DM ^7`Wa$$7Wd -DM ]^7`Wa$d-DM ^%-DM ^$ ds-DM ] ^a$+-DM ^+-DM ^  "! -DM ^ 1K(}ɷɰo]N]?@B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 5\mH sH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH Y*RScd?Q ̽ۤtbP>P#6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH 5@CJ\mH sH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH Sdn!y!T""#K$`$$t%%k & K-DM ^ $dj-DM ]^a$ =-DM $7Wd-DM ^7`Wa$ *-DM ^ $dn-DM ]a$+-DM ^+ K-DM Q !m!n!x!y!!!"."T"q""""*#X#­ucۓTB3@B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 5\mH sH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH X####$J$K$_$`$$$ %s%t%u%%%%Ͻv`N>9mH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphu#6@B*CJ]aJmH phsH B*CJaJmH phsH 5@CJ\mH sH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH %%%H'S''(#)uW$ d-DM ]^ `a$$d -DM ^`a$$d-DM ]^`a$ *-DM ^ $  d-DM ] ^ a$ 6#-DM ^ & K-DM ^ %%%%%%V&&G'H'R'S''''(J(ñvoZvoH9@B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH "6:B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH +j6@B*U]aJmHnHphuJ(V((((((")#)F)~)))))*u*******Ŷ׃tgRM>@B*CJaJmH phsH mH sH )56@B*CJ\]aJmH phsH B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #))***11e2a$0d-DM ]0^`a$=-DM ^$ds-DM ]^a$5-DM ^53-DM ^$d -DM ]^`a$$d-DM ^`a$*=+++Z,,;..S//0m00111112d2e22223ɷ핃|id|UF@B*CJaJmH phsH @B*CJaJmH phsH mH sH %56B*CJ\]aJmH phsH  CJmH sH #6@ B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 333&3'3(3*3:3<3C33333ƿp^K:!6:>*B*]aJmH phsH %6:>*B*CJ]aJmH phsH "6>*B*CJ]aJmH phsH &6>*@B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH 'j6B*U]aJmHnHphu CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH  e2'3;3<3334667i7{hWW d-DM ^ d-DM ^$ECd-DM ]E^C`a$ d-DM ^ $ -DM ] a$ u#-DM ^  -DM $d-DM ]^`a$ 33333344444*4z4ʵygSC1#6@B*CJ]aJmH phsH 6@B*]aJmH phsH &6:@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #5@B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH 556@B*CJOJQJ\]^JaJmH phsH  z4456b6~6666677h7i7j7m7įxf_K9"6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH m7z778h889T999G:::A;;5<<<!=m== >W>w>x>>>ͻݻݗ݅݅sݗaZEZ)56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6@B*]aJmH phsH i7x>>???;@<@@{e{S S-DM ^ &r-DM ^r &-DM ^$E#d+-DM ]E^#`a$$Ed-DM ]E^`a$EV-DM ]E^V$EMd-DM ]E^M`a$>>>>>>???A?g???????ϽtbP:P+j6@B*U]aJmHnHphu#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH ?:@<@=@>@D@@@@@@AAAEAϻldOd=.@B*CJaJmH phsH #5@B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH 5\mH sH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH 5j56@B*CJU\]aJmHnHphumH sH 6B*CJ]aJmH phsH @@AFAGG.HH}_$d0-DM ]^`a$$d-DM ]^`a$H-DM ]^H"$ j d-DM ]^`a$"-DM ^"-DM ^-DM ^EAFAGAJARAVAAAEBBB=CCC#DDE]EEѿweweSA/ww/#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ ]aJmH phsH  CJmH sH EEGFF9GGGGG H-H.HPt,b#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH mH sH )6:>*@B*CJ]aJmH phsH &6>*@B*CJ]aJmH phsH PPEQQQQQQQ R$R?R@RRRRRSSPSQSSSɷxfWHCmH sH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH -56@B*CJRHn\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH MQQ@RRQSS Ts_H`dh-DM ]`^<&-DM ]<^$<6d&-DM ]<^6`a$$<#d+-DM ]<^#`a$$<d-DM ]<^`a$<V-DM ]<^V$dVd-DM ]d^V`a$SSS T T T TTTT UWUUUBV®vdTdBd0#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #5@B*CJ\aJmH phsH )56B*CJRHn\]aJmH phsH BVVVVV5W6WGWWWWWWWW XVXԽtbN<*#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH mH sH -56@B*CJRHn\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  TV6WWZZ[H\S\up`C-DM ^C$a$ X#-DM ^ $Z-DM ^"$ 'Rd-DM ]^R`a$< d-DM ]<^ dc-DM $Rd-DM ]^R`a$VXXXEYY2Z~ZZZZZZ[[[H\ɷɥzu`K74CJ&6>*@B*CJ]aJmH phsH )6:>*@B*CJ]aJmH phsH )6:>*@B*CJ]aJmH phsH mH sH 6B*CJ]aJmH phsH 'j6B*U]aJmHnHphu CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH H\R\S\m\\\\#]$]>]]]]]]]*^<^t^u^^^^ƷӀn_Pn_A@B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH %56B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH  CJmH sH )56B*CJRHn\]aJmH phsH S\$]]u^#__pccxZB$ >>d-DM ^>a$$:d-DM ]^:`a$dK-DM ^$ d"-DM ]^ `a$$d-DM ]^`a$$d-DM ]^`a$$d-DM ^`a$^ _"_#_r___________H``ֿ}m}]M=+#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH 6B*CJ]aJmH phsH 6B*CJ]aJmH phsH 6:B*]aJmH phsH "6:B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH -56@B*CJRHn\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ``7aa|bbcocpcqcrcccccccd4dNdOdɷ۞|eSD5@B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH -56@B*CJRHn\]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ccOddheeffw`K d-DM ] ^ dF-DM ] ^$d-DM ]^`a$$:d-DM ]:^`a$$#d-DM ]^#`a$$+d-DM ]+^`a$>-DM ^>Odlddddddde5eTeWegeheeeer]rK<-@B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH eeeffKfffffffffϽvbN<*#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH -56@B*CJRHn\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH  ff,g-g.g@gBgggggGhhh0ii²֞tbP@b.#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )6:>*@B*CJ]aJmH phsH )6:>*@B*CJ ]aJmH phsH &6>*@B*CJ]aJmH phsH 6B*CJ]aJmH phsH 'j6B*U]aJmHnHphumH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH f-gAgBgg5k@k;llkI"$ EROVd-DM ]R^O`Va$$REd-DM ]R^E`a$R-DM ]R^$ }Rd-DM ]R^a$ #R-DM ^R  ! U-DM ^ $d-DM ]^`a$ii!jrjjjjk4k5k?k@kwkkkl˹j`N?0@B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 5CJ\mH sH )56@ B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH l:l;lNllllll#m>mambmmmmmmmŶ椶n\H4&6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH #5@B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH lbmmfpqppp(q|_B 5&OVd-DM ]^O`V 5&Ud-DM ]^`U 5&-DM ]^ 5&-DM ]^$xd-DM ]^x`a$ [d=-DM ] ^[$R2d-DM ]R^2`a$m*n{nnokoo pXpepfpppqppppppppyo]N];%56B*CJ\]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 5CJ\mH sH )56@ B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH p q'q(qIqWqYqjqkqrqqqqrrrrȶ׏}n_M8)56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH (qkqrqqrTrgrhrrmWG-DM ] F&V-DM ^V F%V-DM ^V 5&@-DM ]^@ 5&J-DM ]^J 5&d-DM ]^` 5&cd-DM ]^`c 5&b-DM ]^brSrTrUrcrfrhrirjrurvr~rrrrо}i}iTB#6@B*CJ]aJmH phsH )6:>*@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ ]aJmH phsH .j6:@B*U]aJmHnHphumH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphu CJmH sH #6@B*CJ]aJmH phsH rrrrrsswwscF-DM ^$d-DM ^`a$ &+D-D.:/&M d;&`#$+D-D.:/&M ^d;&`#$+D-D.:/&M  {-DM ^ -DM rrrrrrr sssssssssʵ~wiQ=)=&6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH .j6:@B*U]aJmHnHphu6@B*]mH phsH  CJmH sH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH 556@B*CJOJQJ\]^JaJmH phsH sssst?tttu0u[uuuv)vvKwtwwwwwwɷqlWP>#6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH mH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH w-x?x@xVxfxxxxx yy=y>y\y]yŶsa\G@+\)56@B*CJ\]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH B*CJaJmH phsH @ B*CJaJmH phsH #6@ B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH w@xy>y]yyyyyyppU$4& +D-D.:/&M a$ d-DM ^ d-DM ^ d-DM ^ d-DM dP-DM ^$dV-DM ]^`a$$dh-DM ]^`a$ ]ytyyyyyyyyyyyyyy zGzpzzz׶}i}WE3#6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH 6@B*]mH phsH #6@B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH  CJmH sH @B*CJaJmH phsH #5@B*CJ\aJmH phsH yy~~}}oc "-DM -DM ]=$&d-DM ]&^`a$$d}-DM ]^`a$;-DM ^;$6 d-DM ^6` a$4& +D-D.:/&M zz {1{u{{{{|C|m||||}:}}}}~/~T~|~~~~~ɷۓہ큷oۓ_ZmH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ~~~~(Ldq³u`uN<*#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH   "$HMcdeĽ듁ohYL<6B*CJ]aJmH phsH B*CJaJmH phsH @B*CJaJmH phsH  CJmH sH "6:B*CJ]aJmH phsH #6@B*CJ]aJmH phsH /j6@B*CJU]aJmHnHphu#6@B*CJ]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH 6CJ]aJmH sH jUmHnHumH sH @B*CJaJmH phsH  !#$eĀۀio$s7d-DM ]s^7`a$ d-DM ^ d-DM ^ d-DM dK-DM $ 3d-DM ] ^3a$  -DM -DM eÀĀڀۀ܀݀߀ 徯~gP<,6B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH 4j56:@B*U\]aJmHnHphu@B*CJaJmH phsH  CJmH sH @ B*CJaJmH phsH #5@ B*CJ\aJmH phsH )56@ B*CJ\]aJmH phsH mH sH )56@B*CJ\]aJmH phsH  2[ρHhist߂>S핎yl]Kݎl<@ B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH itԃ ~obF$3*&#$+D-D.9/(M a$ d-DM dF-DM [8&#$+D-D/(M $[8&#$+D-D/(M a$$sd"-DM ]s^`a$$s d-DM ]s^ `a$s]-DM ]s^]S]wzԃރ ۴weSD5. CJmH sH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH #5@B*CJ\aJmH phsH )56@ B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH 6@B*]mH phsH mH sH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH gى|i\ d-DM d-DM ]$d-DM ]^`a$b-DM ]^b$]d-DM ]^]a$$nSd-DM ]n^S`a$3*&#$+D-D.9/(M C 6b܅)NuҾvdRv@.v#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH uĆ: 8`ĈɈϽweP>#6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH  CJmH sH #6@ B*CJ]aJmH phsH Ɉ-Mfg҉؉ىډ܉̺yj[T=&,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH  CJmH sH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH  ܉މ.͊Gl;c݌ ʸʦڸڦڦڔ}ha}U<1j56@B*U\]aJmHnHphu6CJ]aJmH sH  CJmH sH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ى  #$;DEGH -DM -DM ^-DM ^ -DM  -DM BX-DM ^X$Sd.-DM ]^S`a$ "$%:<CFHdqՍӫӅyi\M\i;#6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH 6CJ]aJmH sH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH .j6:@B*U]aJmHnHphu6B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH Ս_ˎ֎<LMqxɷ폂saO?-O#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH @B*CJaJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH HM @S~veT+d-DM ^+ d-DM ^ dB-DM ^$R d-DM ]R^ `a$$e d-DM ]e^ `a$$V d-DM ]V^ `a$$d-DM ]^`a$xڏ 8Q\ɐ ?@RSiϽ϶Ϧxc^I^7#5@B*CJ\aJmH phsH )56@ B*CJ\]aJmH phsH mH sH )56@ B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @ B*CJaJmH phsH i}~֑̑͑בؑڑܑ.S|鶱r`P>,#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH "6:B*CJ]aJmH phsH (56:B*CJ\]aJmH phsH (56:B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH  CJmH sH @B*CJaJmH phsH ~͑ёבqS$ hd-DM ] ^h`a$7-DM ^7$7d -DM ^7a$$}@d-DM ]}^@`a$$eB&+D]-D.:/&M a$d-DM d-DM  d-DM ^ |В"Jt“ߔ,U~ܕɹɃ~ۧɕlZHA CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH (3X`x̖(3µp`QB0#6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH 3NT\cz˗Η ,PƷƥzk\M\;)#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @ B*CJaJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH XBCFGJnnjj` -DM 3>d -DM ^>9d-DM ^99d-DM ^9$+Fd5-DM ]+^F`a$$Pd5-DM ]^P`a$$ Ud0-DM ] ^U`a$ Pr֘-:WX}³toZE3#5@B*CJ\aJmH phsH )56@ B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @ B*CJaJmH phsH #6@ B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 0ABCDEGIJK[]^eg˹ˢ{tdN<#6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphu6B*CJ]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH 6CJ]aJmH sH jUmHnHumH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH J\]f}~uhf$d-DM ]^`a$#-DM ]^#$d-DM ]^`a$#$d9&`#$+DH-D.:/&M ^a$ -DM -DM &-DM g|~֚ߚȸ}bK4,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH 4j56:@B*U\]aJmHnHphu CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH 6B*CJ]aJmH phsH %56B*CJ\]aJmH phsH 6CJ]aJmH sH mH sH &6:@B*CJ]aJmH phsH  8`ۛ/W}ϜHp9c؞ȶȶpȔȔ^ȶLp#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ؞)U}̟!KtuǠ%Egh۷o]K۠9#6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 5CJ\mH sH )56@B*CJ\]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ԡ֡ء)PtϢТߢ~o[G[G3'6@B*CJRHl]aJmH phsH '6@B*CJRHl]aJmH phsH '6@B*CJRHl]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @ B*CJaJmH phsH hסТ;3sYC$|d-DM ]|a$$| d -DM ]|^ a$$"d-DM ]"^a$$|d&-DM ]|^`a$$|*d"-DM ]|^*`a$$|3d-DM ]|^3a$$d-DM ]^`a$ߢ:;cͣɵiRH4#!@B*CJRHlaJmH phsH '5@B*CJRHl\aJmH phsH 5CJ\mH sH -56@ B*CJRHl\]aJmH phsH -56@ B*CJRHl\]aJmH phsH -56@ B*CJRHl\]aJmH phsH -56@ B*CJRHl\]aJmH phsH  CJmH sH '6@B*CJRHl]aJmH phsH '6@B*CJRHl]aJmH phsH !@B*CJRHlaJmH phsH !@B*CJRHlaJmH phsH  23467]_o[I4I) j6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6@B*CJH*]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH &6:@ B*CJ]aJmH phsH ,56:@ B*CJ\]aJmH phsH ,56:@ B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH  CJmH sH 3ʨ˨ΨϨިuu $dN  -DM -DM -DM P$&;d-DM ]&^;`a$$|;d-DM ]|^;`a$$|}d-DM ]|^}`a$ /U|ͥ޴~lZE3#6@B*CJ]aJmH phsH ) j6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6@B*CJH*]aJmH phsH )56@B*CJ\]aJmH phsH ) j6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH ͥFn459=e0V˹q___M#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ) j6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ɨ˨̨ͨϨݨߨ 2]wxضؔ؂pض^LE CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH "6:B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6CJ]aJmH sH +j6@B*U]aJmHnHphumH sH 6B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH  x(s`d8-DM ]$d-DM ]^`a$$d-DM ]^`a$$ d-DM ]^ `a$J-DM ]^J$Ed-DM ]^Ea$xө '>aƷ~lZK<-@B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH ɪ&Ily۫˹ufT?)56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH '(<T\]Ȭ³㓎w`L:(#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH mH sH 6@B*]mH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH (]ukM$!d5-DM ]!^`a$$!d-DM ]!^`a$!N-DM ]!^N$@d-DM ]^@`a$B&+Dl-D.:/&M ] d-DM ]^ d-DM ]^>gܭ-ˮ@gtv˹崙݃q_ݹXCX4@B*CJaJmH phsH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH ίI\h˰װϿvdRC6B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH Gmst߱ ѼsdR@0+mH sH 6@B*]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH %56B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH t }fH$;d-DM ]^;`a$7&#$+D-D/9M $!7&#$+D-D/9M ]!a$$!d-DM ]!a$$! d -DM ]!^ a$$!d-DM ]!a$$!d0-DM ]!^`a$67<GLMNWX[\]^ҾvavL8&6@B*CJH*]aJmH phsH )56@B*CJ\]aJmH phsH ) j6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH ) j6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH ,56:@B*CJ\]aJmH phsH ,56:@B*CJ\]aJmH phsH ^_`abfhijklmnopqԲرydyR@#6@ B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH ) j6@B*CJ]aJmH phsH &6@B*CJH*]aJmH phsH )56@B*CJ\]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH ) j6@B*CJ]aJmH phsH Բ$Nvų:cٴ,S{ʵ˹ۧqq_˧XH<6CJ]aJmH sH j6U]aJmHnHu CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH "45>VWYZٷP}d-DM ^d-DM ^$]d-DM ^]a$O-DM ^O -DM -DM h$;d-DM ]^;`a$ !#356=?UXZж Irͻp^L:p#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6CJ]aJmH sH &6:@B*CJ]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH jUmHnHu#6@B*CJ]aJmH phsH mH sH 6B*CJ]aJmH phsH rķطٷ0GOPįyjX?*(56:B*CJ\]aJmH phsH 0j56:B*U\]aJmHnHphu#6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH )56@ B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH P:׽zξj[N d-DM d=-DM $*d-DM ^*`a$$*d-DM ^*`a$$8d-DM ]^8`a$p-DM ]^p$7d-DM ]^7`a$d-DM ^9b߹ &'Ѻ"LƻiɷɁo`oN<6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH 1j56@B*U\]aJmHnHphu#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH x>[\l 3]$Mǵǥo]K99#6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH "6B*CJH*]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6@B*CJH*]aJmH phsH #6@B*CJ]aJmH phsH K,oo^d-DM ^dc-DM ^$d+-DM ]^`a$$d -DM ]^a$d-DM ^`6-DM ^6$6 d-DM ]^6` a$%BJL 19ôyi\M>/@ B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@ B*CJ\]aJmH phsH  CJmH sH 9Qq+,RoIJvdR=()56@ B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@ B*CJ]aJmH phsH o#IsȶȎzhzVD2#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH "6:@B*]aJmH phsH &6:@B*CJ]aJmH phsH .j6:@B*U]aJmHnHphu6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH @B*CJaJmH phsH #5@B*CJ\aJmH phsH vjj "-DM -DM B$&d"-DM ]&^`a$$>d-DM ]>^`a$@-DM ^@$7d-DM ]^7`a$d-DM ^ s<d'Py Ef˹˃qjUjjF@B*CJaJmH phsH )56@B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH fŹ~lVB&6:@B*CJ]aJmH phsH +j6@B*U]aJmHnHphu#6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH 'j6B*U]aJmHnHphu6CJ]aJmH sH mH sH  CJmH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 3Yxf-DM ^f$&Sd-DM ]&^S`a$ d-DM dK-DM Vd-DM ^`V3-DM ^3 -DM -DM [n|ǵq^L:+@B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH %56B*CJ\]aJmH phsH 6B*CJ]aJmH phsH B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH 6CJ]aJmH sH mH sH &6:@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH DYiŰteJ3,56:@B*CJ\]aJmH phsH 4j56:@B*U\]aJmHnHphu@ B*CJaJmH phsH #5@ B*CJ\aJmH phsH )56@B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH )56@ B*CJ\]aJmH phsH  CJmH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH  7_"Lvųn^L:(#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH (56:@B*\]aJmH phsH -v EOöÕn_PA@B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH B*CJaJmH phsH 6B*CJ]aJmH phsH )56@ B*CJ\]aJmH phsH  CJmH sH 6B*CJ]aJmH phsH ~^wYIJ-DM ^J$qEd(-DM ]q^E`a$ dR-DM ^ R-DM ^d -DM $3d"-DM ]^3`a$$3d&-DM ^3`a$$3d-DM ]^3`a$Oejm}~ ?]^ۿn\WB-)56@ B*CJ\]aJmH phsH )56@B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH ^ljųטlXF4#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH &6:@B*CJ]aJmH phsH (56:@B*\]aJmH phsH ,56:@B*CJ\]aJmH phsH 4j56:@B*U\]aJmHnHphu#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #5@B*CJ\aJmH phsH  j2Z'Qy?bc˹ݧݹ遼~ibbPA@B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@ B*CJ\]aJmH phsH mH sH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH ciee[O -DM -DM  $ds-DM ]^`a$$ds-DM ]^`a$$d}-DM ]^`a$$d}-DM ]^`a$$ d-DM ]^ `a$ $HXmpϽϡqbP;)#6@B*CJ]aJmH phsH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH )56@B*CJ\]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH  |n`R@#j6B*U]mHnHphu6@B*]mH phsH 6@B*]mH phsH 6B*CJ]mH phsH 6B*]mH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH 6CJ]aJmH sH jUmHnHumH sH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH *\)xc 3 0d-DM ^0 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3 d.-DM ^$`-DM ^`a$ -#-DM ^ -DM )*;<=[\mnͽtbPA@B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH mH sH 556@B*CJOJQJ\]^JaJmH phsH  ()67[\ij'(56˼ݨwhhhV#6@B*CJ]aJmH phsH @B*CJaJmH phsH @B*CJaJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH B*CJaJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH )\(UlW 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3  d-DM ^  3 d-DM ^ 3 *d-DM ^*67TUi"#56efstuǸהǸׂǦp^Opp^O@B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH 6B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #f(UlW 3 %d-DM ^% 3 d-DM ^ 3 jd-DM ^j 3 d-DM ^ 3 "d-DM ^" 3 md-DM ^m 3 d-DM ^ 3 Vd-DM ^V '(456TUab~˼ݔ݂sa݂sOݔݔ#6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH U'XlW 3 [d-DM ^[ 3 2d-DM ^2 3 d-DM ^ 3 6d-DM ^6 3 Cd-DM ^C 3 d-DM ^ 3 ;d-DM ^; 3 Ud-DM ^U&'789WXj&'(FGǸǸٸٗمvdǸdǸ#6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH !G5olW 3 d-DM ^ 3 ^d-DM ^^ 3 5d-DM ^5 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3 d-DM ^ 3 d-DM ^GUVW45FGHno}˼ݔ݂˼˼pcp݂^˼mH sH B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH #6@B*CJ]aJmH phsH  CJmH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH $)d$a$  -DM  3 d-DM ^ 3 d-DM ^ 3 cd-DM ^c 3 d-DM ^ 3 d-DM ^ ()345cdrstג~hVF>>6]mH sH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH +j6@B*U]aJmHnHphumH sH @B*CJaJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH #6@B*CJ]aJmH phsH @B*CJaJmH phsH 6B*CJ]aJmH phsH #6@B*CJ]aJmH phsH  CJmH sH 6]mH sH   -DM $a$* 00P. A!"#$h% <'0P. A!"#$h% <'0P. A!"#$h% <'0P. A!"#$h% <'0P. A!n"T#$h% <'0P. A! "#$h% <'0P. A!`"#$h% <'0P. A!"#$h% <'0P. A!A"#$h% <'0P. A!"#$h% <'0P. A!"#$h% <90P. A!"#$h% P0 * 00P. A!"#$h% << 00P. A!"#$h% P0* 00P. A!"8#$h% << 00P. A!"8#$h% P090P. A!"x#$h% P0 * 00P. A! "x#$h% << 00P. A! "x#$h% P0B* 00P. A! "#$h% << 00P. A! "#$h% P0:!90P. A!"#$h% P0 >* 00P. A!"3#$h% << 00P. A!"3#$h% P0v< * 00P. A!"g#$h% << 00P. A!"g#$h% P090P. A!n"T#$h% P0 * 00P. A!n"#$h% << 00P. A!n"#$h% P0P* 00P. A!n"#$h% << 00P. A!n"#$h% P0"!90P. A!"#$h% P0c * 00P. A!"#$h% <* 00P. A!"#$h% << 00P. A!"#$h% P0* 00P. A!"#$h% << 00P. A!"#$h% P090P. A!n"#$h% P0FF* 00P. A!"#$h% << 00P. A!n"#$h% P0* 00P. A!" #$h% << 00P. A!n" #$h% P0OX90P. A!""#$h% P0 r* 00P. A!"#$h% << 00P. A!"#$h% P0** 00P. A!"D#$h% << 00P. A!"D#$h% P090P. A!o"$#$h% P0 * 00P. A!j"#$h% << 00P. A!j"#$h% P03 * 00P. A!t"3#$h% << 00P. A!t"3#$h% P06'0P. A!"#$h% < iD@D 1KG=K91$7$8$H$6]_HmHsHtH@ 03>;>2>: 1;$md-D@&M ]^`m&6:@B*CJ]aJmH phsH @ 03>;>2>: 2+$$-D@&M ^a$556@B*CJOJQJ\]^JaJmH phsH @ 03>;>2>: 3)$$d-D@&M a$556@B*CJ.OJQJ\]^JaJ.mH phsH \@\ 03>;>2>: 4$$@&a$&6>*@B*CJ]aJmH phsH Z@Z 03>;>2>: 5$$@&a$#6@B*CJ]aJmH phsH Z@Z 03>;>2>: 6$$@&a$#6@B*CJ]aJmH phsH T@T 03>;>2>: 7$@&#6@B*CJ]aJmH phsH j@j 03>;>2>: 8"$$  dv@&a$6B*CJ]aJmH phsH  @ 03>;>2>: 9+ $$-D@&M ]a$156B*CJOJQJ\]^JaJmH phsH :A@: A=>2=>9 H@8DB 0170F0C@ A=>2=>9 B5:AB A >BABC?><%>d-DM ^>#6@B*CJ]aJmH phsH &1<GR]$%(),-&1<GR]`C!!-Kdis 0BCg > , T-&2ABCghv"#1;FWXeqz*23@AJRSTUVW^efs ?@N\jtySdnyTK ` t!!!!H#S##$#%%&&&--e.'/;/u]uuuuuuuzz{{{{{{{ |!|#|$|e|||||i~t~~ gم  #$;DEGHM @S~͍э׍XBCFGJ\]f}~uhםО;3ʤˤΤϤޤ x(]ut "45>VWYZٳP:׹zκ:C,01HQRTU!bK,o3Y~^c*\)\(U#f(U'XG5o)d00(00?0?0?0?00000`000000000000000000`0000 0 0 0 0 0 0 00000`000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000`00 000000000000000000000`0000000000000000`00000000000000`000000000000000`0000000000000@00000000000000000000`0000000000000000000`00`00`00000@0@0@0@0@0@0@0@0@00@00000`00`00`00`00`00000000000@0000@000000000`00`00000`000`00`00000000000000000000000`00`000`000`0`00000000000000`00`000`000`0@0@`0@0@`00@0@0@0@000@00@00@0@0@0@00@0@00@00`0@00@0@`000`00@`0@000@00@00@0@0@0@0@0@0@0@00@0@0`00@`00@0@`000`00`0000000000000000000`00`000`000`00`00000000000000000000`00`000`00000000000000000000000000000000000000000000000000000000000000@000@000@000>! | Q X#%J(*33z4m7>?EAEIJ+MPSBVVXH\^`Odefilmprrsw]yz~e SuɈ܉Սxi|3Pg؞ߢͥx^ԲriaξǿUKx9osfO^j6Gvy{}     !#$&')+-/12K0-Ag"WAV ?S%#)e2i7@HM TS\cfl(qrwyiىH~Jh3(Pξ,)Uwz|~ "%(*,.03x8/1@+0p(  NB  S D1NB  S D1NB @ S D1HB @ C DHB  C DHB   C DNB   S D1HB   C DNB   S D1HB  C D HB  C D HB  C D HB  C D HB  C DHB  C DHB  C DT  C  HB  C DHB  C D f  s * HB  C DHB  C DHB  C DHB  C Df  s * HB   C DHB ! C DHB " C D HB # C Df $ s * f % s * NB & S D1  HB ' C D f ( s *! f ) s *" HB * C D#HB + C D$f , s *% f - s * &  HB . C D'HB / C D)HB 0 C D(B S  ?0-t!!'/;<<HV-cTnhno{{| $C]ˤ5C1,7%7t,%t&tJ&Jt(%(t 3&t O&t 8O&Ot &tn&ntUM%t&%)&)t?w&?t%t0/&/ty%t[&[tq tB&Bt%tt &t%t1>)&>tO&tsCt N&Nt!c='ct"#%#t$gNt%t#B$Bt' }&t(t):t* &t+%t,t-Wh't.)&t0&t/'&t-ChWfMN[\~K ` t!!!"2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2+D:\Data\NNGU\NewSite\StrongSource\INdAM.doc"C@;0?>2@C:\WINDOWS\Application Data\Microsoft\Word\2B>:>?8O opt2003.asd"C@;0?>2@C:\WINDOWS\Application Data\Microsoft\Word\2B>:>?8O opt2003.asd"C@;0?>2-D:\Data\NNGU\NewSite\StrongSource\opt2003.doc~*쩠俟@ OJQJo(" &2ABCghv"#1;FWXeqz*23@AJRSVW^efs ?@j@UU UUp@UnknownG:Times New Roman5Symbol3& :Arial"hww  _!xxd2QINdAM"C@;0?>2"C@;0?>2 Oh+'0  8 D P \hpxINdAMfNdA  Normal.dot t13Microsoft Word 9.0@G@@|!R@^~5T  ՜.+,0 hp|  _ INdAM    !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijkmnopqrsuvwxyz{Root Entry F@/5T1Table:cWordDocumentrSummaryInformation(lDocumentSummaryInformation8tCompObjjObjectPool@/5T@/5T  F Microsoft Word MSWordDocWord.Document.89q