A Ca-Svm Based Monte Carlo Approach for Evaluating Complex Network Reliability Yuan-peng Ruan1, Zhen He1 1 School of Management, Tianjin University, Tianjin, China (ruanyuanpeng@sina.com) Abstract – Many real-world complex systems can be modeled as networks. Evaluation of network reliability plays an important role in engineering applications. When evaluating the S-T complex network reliability, the traditional approaches may bring about the problems of increasing computational complexity or decreasing the calculation accuracy. This paper proposes a CA-SVM based Monte Carlo approach based on the drawbacks of traditional approaches. Support Vector Machine (SVM) is a fast and efficient algorithm to ascertain the network connectivity in simulation process. Cellular automata (CA) is used for creating training data points, which speeds up the computing process. Particle swam optimization (PSO) is used for parameters selection of SVM, which increases the accuracy of the result. An example is shown to illustrate the proposed approach. Keywords – Complex network reliability, support vector machine, cellular automata, particle swarm optimization I. INTRODUCTION Many real-world complex systems such as communication systems [1], power transmission and distribution systems [2] and transportation systems [3] can be modeled as networks. The traditional approaches [4] to evaluate the reliability of network are based on the minimal cut sets or minimal path sets. Nevertheless these approaches will lead to NP-hard problems due to the increasing complexity of the network [4]. As a result, some works put forth Monte Carlo Simulation (MCS) based approaches, which are now recognized as playing an important role in the evaluation of network reliability [4, 5]. Ascertaining the connectivity of S-T network is an important problem under the above mentioned MCS based approaches. The approaches frequently used before are based on depth-first procedure or breadth-first procedure [5]. With the increasing complexity of the network, the above mentioned approaches will be time consuming. Cellular automata (CA) based approach, which can overcome the above difficulty, is proposed to the application in network reliability evaluation [6, 7]. Recently, some works extend the use of CA to problems of computing the availability of the renewable network [8], evaluating the K-terminal reliability of the network [9] and This paper is supported by National Natural Science Foundation of China (70931004). computing the maximum unsplittable flow of the network [9] . Since MCS based approaches require a large number of connectivity evaluations of S-T network, it may be convenient to replace this evaluation by an approximated but fast algorithm. Because the S-T network has two states (operating or failed), it’s a two-category classification problem to ascertain the connectivity of the network. Support Vector Machine (SVM), as a artificial intelligence technique, was developed by Vapnik [10]. Because of the superiority of structural risk minimization principle (SRM) over empirical risk minimization principle (ERM), SVM has been widely used for network reliability evaluation [11, 12]. The quality of SVM classification models depends on a proper setting of the parameters (SVM hyper-parameters and SVM kernel parameters), so the main issue to apply SVM is how to set these parameters. The works published before chose grid search method to set parameters [11, 12]. However, when using this method, one should increase the search range or decrease the step size to make the optimal solution accurate, which may result in a highly time-consuming search process [13]. To overcome this problem, we substitute the grid search method with particle swarm optimization (PSO). In this work, we proposed a CA-SVM based Monte Carlo approach to evaluate the complex network reliability. Firstly, establish the training data with CA; secondly, train the SVM with PSO and cross-validation; lastly, ascertain the connectivity of each simulation with the SVM trained and compute the S-T network reliability. Section II briefly describes CA algorithm. Section III briefly describes SVM and parameters selection using PSO, then proposes a CA-SVM based Monte Carlo approach for evaluating complex network reliability. An example is presented in section IV to illustrate the proposed approach. II. ASCERTAINMENT OF THE CONNECTIVITY OF S-T NETWORK USING CA A. Description of CA Cellular automata(CA), a kind of approach to simulate the behavior of dynamic discrete systems, was originally conceived by Ulam and Von Neumann in the 1950s. CA consists of some cells, usually assumed to be homogeneous and with limited discrete states. Each cell’s action at a given time t relies on its state at the time t-1, those of its neighborhood at the time t-1 and a transition rule. As shown in Fig. 1, CA can be mainly classified into one-dimensional CA and two-dimensional CA based on its cells’ dimensions. Two-dimensional CA also can be classified into two categories: Von Neumann neighborhood and Moore neighborhood [14]. i,j-1 i-1,j i,j i+1,j i-1 i i+1 (a)One-dimensional type Suppose a set of N training data points {(X1, y1), (X2, y2), … , (XN, yN)}, where yi = {1, -1}. For linear SVM, as shown in Fig.2, consider the separating hyperplane i-1,j-1 i,j-1 i+1,j-1 i-1,j i,j H2 i+1,j H H1 i,j+1 i-1,j+1 i,j+1 i+1,j+1 (b)Von Neumann type (c) Moore type Fig.1. Types of CA B. CA Algorithm of Ascertaining the Connectivity Let G=(N,A) be a network graph, where N is the set of n nodes, A is the set of directed arcs. The S-T network connectivity evaluation refers to finding if there is a path from a source node S to a terminal node T. It’s assumed that Ei is the neighborhood of the node i, defined as Ei = {j∈ N s.t. (j, i) ∈ A} and w (i, t) is the state of node i at the time t. The state w (i, t) of each node is binary, assuming the value of 1 when node i is active and of 0 when passive. Each node i follows an OR Boolean transition function w(i,t+1)=OR(w(j, t),…,w(k, t),w(i, t)), j,…,k∈Ei (1) As some works have shown, S-T connectivity can be computed in O(n) time using CA, which is an advantage over the traditional approaches [6,9]. The basic algorithm proceeds as follows: 1. t = 0 2. Set all the cells state values to 0 3. Set w(S, 0) = 1 4. t = t + 1 5. Update all cells states by function (1) 6. If w (T, t) = 1, then stop: c=1 and there is a path between S and T 7. If t < n – 1 go to step 4. Else 8. c = 0 and there is no path between S and T Fig. 2. Linear SVM H: y = w X b 0 (1) where w is normal to the hyperplane H. The two hyperplanes H1: y = w X b 1 and H2: y = w X b 1 (2) is parallel to H and the data points closest to the two parallel are called support vectors. To partition the two groups completely, the optimal separator can be obtained by a constrained optimization formulation [15]: 1 2 Min w (3) 2 w,b s.t. yi (w X i b) 1, i 1,2,..., N . However, sometimes it is impossible to separate the training data points linearly. To solve this problem, imperfect separation should be considered and the formulation (3) will be transformed as follows: N 1 2 Min w C i w,b 2 i 1 s.t. (4) yi (w X i b) 1 i , i 0, i 1,2,..., N . The Lagrangian formulation for the dual problem of formulation (4) is as follows: N 1 N N i j yi y j X i X j k 2 i 1i 1 k 1 Max (5) s.t. 0 i C, i 1,2,..., N III. RELIABILITY EVALUATION OF COMPLEX NETWORK USING CA-SVM BASED MONTE CARLO APPROACH A. Description of SVM Classifier Support vector machine (SVM), which is desired to find a separator to partition data-set as far as possible, provides a novel approach to the two-category classification problem. m y i i 0 i 0 where i represents the Lagrangian multiplier of Xi and C is the penalty parameter. After the solution has been obtained, the decision function for new Xi is as follows: N f ( X i ) sgn( j 1 j y j Xi X j b) (6) For non-linear SVM, the decision function for new Xi can be obtained through replacing X i X j with kernel Start algorithm function. There are many kernel functions which can be used [15]. Among them, the commonest are Gaussian radial basis function and polynomial function, which is as follows: Xi X j / 2 No 2 , k ( X i , X j ) ( X i X j 1) d . (7) Set fitness function(cross validation accuracy) Create a network state randomly based on arc reliability Yes Initialize position and velocity of each particle Ascertain the connectivity of the created network using SVM (8) Because the reliability evaluation is a non-linear problem and the parameter of Gaussian radial basis function is continuous, which is easy to be tuned [13], this paper selects Gaussian radial basis function to train SVM. Perform SVM on each particle and compute the fitness function Max. Max. simulation simulation iteration iteration M? M? Max. Max. iteration iteration F? F? Yes B. Parameter Optimization Using PSO Compute the network reliability IV. EXAMPLE DISCUSSION The complex network shown in Fig.4 was proposed by Yoo and Deo [17]. It consists of 21 arcs which have the following reliabilities: r7=0.81, r4=r12=r13=r19=0.981, and other ri=0.9. 3 5 16 7 9 15 10 10 5 9 14 2 1 3 11 T 20 S 8 8 1 pijt and gijt respectively represent the best previous 11 7 12 13 6 Fig. 4. A complex network TABLE I COMPARISON OF ALGORITHMS C. Proposed Approach The proposed approach for reliability evaluation of complex network is illustrated in Fig.3. CA and PSO are fast and efficient alternatives for DFS and grid search approach. MCS, as a kind of simulation technology, has a big advantage over the tradition approaches, which can result in NP-hard problems, for complex network reliability evaluation. 17 18 (10) position of each particle and the best particle in the whole swarm within the iteration t. c1 r1 and c2 r2 determine the weights of two parts. c1 and c2 are learning rates which are nonnegative constants. r1 and r2 are generated random numbers in the interval [0, 1]. w is the inertia coefficient which is a constant in the interval [0, 1]. 4 2 4 The Fig. 3. Proposed algorithm flow diagram 6 and each particle will position with the equation xijt xijt 1 vijt 1. End algorithm 21 Since the SVM generalization performance heavily depends on the setting of C and σ, these parameters should be set properly. Particle swarm optimization (PSO), a population based optimization algorithm, was first introduced by Kennedy and Eberhart [16]. In PSO, each particle represents a potential solution to the optimization problem. The performance of each particle depends on the pre-defined fitness function. Each particle flies according to its own experience and the experience of its neighboring particles with a certain velocity. The flying velocity of each particle can be updated during each iteration with the equation (9) vijt wvijt 1 c1r1 ( pijt 1 xijt 1 ) c2 r2 ( g ijt 1 xijt 1 ) No Update position and velocity of each particle 19 k(Xi , X j ) e 2 Create N training data points using CA Train SVM with optimal C and σ Reliability Relative error rate Proposed approach 0.9968 CA-MCS based approach[6, 7] 0.9973 -0.039% 0.01% The exact result 0.997186 / After 100000 simulation iterations, the results shown in Table I illustrate that proposed approach gives a result which is not better than CA-MCS based approach but accurate enough for engineering applications. Because proposed approach substitutes SVM for CA when ascertaining network connectivity during each simulation iteration, it will be a fast and efficient substitute for CAMCS based approach when a large number of simulation iterations are needed. V. CONCLUSION Complex network reliability has become a hot research topic recently. S-T network reliability evaluation, as the basis of this research topic, should be focused on. Some traditional approaches can not solve complex network reliability evaluation well due to inconvenience or inaccuracy. This paper proposes a CA-SVM based Monte Carlo approach for reliability evaluation of complex network, which is a fast and efficient substitute for some CA-MCS and DFS-SVM based approaches. This paper combines CA algorithm, which can create training data points instead of DFS, and PSO, which can be substituted for grid search algorithm to select the parameters of SVM, into the proposed approach. The proposed approach also can be extended to other areas such as evaluations of network availability and K-terminal network reliability. ACKNOWLEDGMENT The authors thank the editor and the anonymous referees for their comments and suggestions. REFERENCES [1] K. K. Aggarwall, “A simple method for reliability evaluation of a communication system,” IEEE Trans. Commun., COM-23, pp. 563–566, May, 1975. [2] W.C. Yeh, “A revised layered-network algorithm to search for all d-minpaths of a limited-flow acyclic netwok,” IEEE Trans. Reliab., vol. 47, no. 4, pp. 436–442, Dec., 1998. [3] T. Aven, “Availability evaluation of oil/gas production and transportation systems,” Reliab. Engng., vol. 18, no. 2, pp. 35–44, Sep., 1987. [4] R. Billinton, R.N. Allan, Reliability evaluation of engineering systems, concepts and techniques, 2nd ed. New York: Plenum Press, 1992. [5] G. Fishman, “A comparison of four Monte Carlo methods for estimating the probability of s-t connectedness,” IEEE Trans. Reliab., vol. 35, no. 2, pp. 145–155, June, 1986. [6] C.M.S. Rocco, J.A. Moreno, “Network reliability assessment using a cellular automata approach,” Reliability Engineering and System Safety, vol. 78, no. 3, pp. 289–295, June, 2002. [7] W.C. Yeh, Y.C. Lin, Y.Y. Chung, “Performance analysis of cellular automata Monte Carlo Simulation for estimating network reliability,” Expert Systems with Applications, vol. 37, no. 5, pp. 3537–3544, May, 2010. [8] Z. Enrico, P. Luca, Z. Valérie, “A combination of Monte Carlo simulation and cellular automa for computing the availability of complex network systems,” Reliability Engineering and System Safety, vol. 91, no. 2, pp. 181–190, Feb., 2006. [9] C.M.S. Rocco, Z. Enrico, “Solving advanced network reliability problems by means of cellular automata and Monte Carlo sampling,” Reliability Engineering and System Safety, vol. 89, no. 2, pp. 219–226, August, 2005. [10] V.N. Vapnik, Statistical Learning Theory. New York: John Wiley Press, 1998. [11] C.M. Rocco, J.A. Moreno, “Fast Monte Carlo reliability evaluation using support vector machine,” Reliability Engineering and System Safety, vol. 76, no. 3, pp. 237–243, June, 2002. [12] F.Q Yuan, U. Kumar, K.B. Misra, “Complex system reliability evaluation using support vector machine for incomplete data-set,” International Journal of Performability Engineering, vol. 7, no. 1, pp. 32–42, January, 2010. [13] Y. Ren, G.C. Bai, “Determination of optimal SVM parameters by using GA/PSO,” Journal of Computers, vol. 5, no. 8, pp. 32–42, August, 2010. [14] M. Tomassini, M. Perrenoud, “Cryptography with cellular automata,” Appl. Soft Comput., vol. 1, no. 2, pp. 151–160, August, 2001. [15] B. Scholkopf, A.J. Smola, Learning with Kernels. London: The MIT Press, 2000. [16] J. Kennedy, R.C. Eberhart, “Particle Swarm Optimization,” in Proceedings of IEEE International Conference on Neural Networks, NJ, pp. 1942–1948. [17] Y.B. Yoo, N. Deo, “A comparison of algorithms for terminal-pair reliability,” IEEE Trans. Reliab., vol. 37, no. 2, pp. 210–215, June, 1988.