Uneven allocation of elements in linear multi-state sliding window system Gregory Levitin The Israel Electric Corporation Ltd., Reliability Department, P. O. Box 10, Haifa 31000, Israel Tel. +972-48183726, Fax. +972-48183790, E-mail: levitin@iec.co.il Abstract The linear multi-state sliding window system consists of n linearly ordered positions and m statistically independent elements with different characteristics that are to be allocated at these positions. Each element can have different states: from complete failure up to perfect functioning. A performance rate is associated with each state. The system fails if the sum of the performance rates of elements located at any r consecutive positions is lower than a demand w. An algorithm based on the universal generating function method is suggested for determination of the linear multi-state sliding window system reliability. This algorithm can handle cases in which any number of multistate elements are allocated in the same position while some positions remain empty. It is shown that such uneven allocation provides greater system reliability than the even one. A genetic algorithm is used as optimization tool in order to solve the optimal element allocation problem. Keywords: sliding window system, consecutive k-out-of-r-from-n:F system, multi-state element, universal moment generating function, genetic algorithm. 1 Nomenclature Acronyms SWS linear multi-state sliding window system ME multi-state element u-function universal moment generating function Nomenclature n number of positions in SWS m number of MEs in SWS r number of consecutive positions in SWS sliding window w minimal allowable cumulative performance of a set of MEs allocated at r consecutive positions ej j-th ME of SWS Ci i-th position of SWS E set of all available MEs: E={e1,…,em} Ei set of MEs located at position i h(j) number of position the ME ej is allocated at: ejEh(j) H vector representing allocation of MEs in SWS: H={h(j),1jm} F SWS failure probability R SWS reliability Kj number of different states of ME j Vj random performance rate of ME j vj,k performance rate of ME j in state k pj,k probability of state k of ME j Si number of different states of group i 2 Gi random performance rate of group i (sum of performance rates of MEs allocated at position Ci) gi,k performance rate of group i in state k i ,k probability of state k of group i Ni number of different states of i-th set of r groups Gi random vector representing performance rates of i-th set of r consecutive groups gi,k vector representing performance rates of i-th set of r consecutive groups in state k Qi,k probability of state k of i-th set of r consecutive groups j(z) u-function representing distribution of random value Vj ui(z) u-function representing distribution of random value Gi Ui(z) vector-u-function representing distribution of random vector Gi ,, composition operators over u-functions 1(x) 1, x is True 0, x is False Pr(x) probability of event x sum(y) sum of elements of vector y 1. Introduction This paper considers a linear multi-state sliding window system (SWS) consisting of n linearly ordered positions at which m multi-state elements (MEs) are allocated. Each ME j can have Kj different states: from complete failure up to perfect functioning. A performance rate is associated with each state. The SWS fails if the sum of the performance rates of MEs allocated at any r consecutive positions is lower than the demand w. 3 The SWS model was suggested in [1] as generalization of the consecutive k-out-of-r-fromn:F system to the multi-state case. The linear consecutive k-out-of-r-from-n:F system has n ordered binary-state elements and fails if at least k out of r consecutive elements fail. This system was formally introduced by Griffith [2], but had been mentioned previously by Tong [3], Saperstein [4,5], Naus [6] and Nelson [7] in connection with tests for non-random clustering, quality control and inspection procedures, service systems, and radar problems. The introduction of the SWS model was motivated by examples from manufacturing, quality control and service systems [1]. Consider for example an industrial heating system that should provide certain temperature along a line with moving parts (Fig. 1). The temperature at each point of the line is determined by a cumulative effect of r closest heaters. Each heater consists of several electrical heating elements. The heating effect of each heater depends on the availability of its heating elements and therefore can vary discretely (if the heaters are different, the number of different levels of heat radiation and the intensity of the radiation at each level are specific to each heater). In order to provide the temperature, which is not less than some specified value at each point of the line, any r adjacent heaters should be in states where the sum of their radiation intensity is greater than an allowed minimum w. A variety of other systems also fit the model: radar systems that should provide certain radiation intensity in different distances, combat systems that should provide certain fire density along a defense line, etc. It was shown in [8] that when the MEs are not identical, their allocation strongly affects the entire SWS reliability. The optimal ME arrangement problem was formulated and an algorithm for its solution was suggested for the case when m=n and only one ME is located in each position (actually the problem formulated in [8] is not allocation but sequencing of the MEs). In 4 many cases, even for m=n, greater reliability can be achieved if some of MEs are gathered in the same position than if all the MEs are evenly distributed between all the positions. Consider, for example, a simple case in which four MEs should be allocated within a SWS with four positions (m=n=4). Each ME j has two states: failure state with performance vj0=0 and normal state with performance vj1=1. The probability of normal state is pj, the probability of failure is qj=1-pj. For r=3 and w=2 the system succeeds if each three consecutive positions contain at least two elements in normal state. Consider two possible allocations of the MEs within the SWS (Fig. 2): A. MEs are evenly distributed among the positions. B. Two MEs are allocated at second position and two MEs are allocated at third position. In case A, the SWS succeeds either if no more than one ME fails or if MEs in first and fourth positions fail and MEs in second and third positions are in normal state. Therefore, the system reliability is RA=p1p2p3p4+q1p2p3p4+p1q2p3p4+p1p2q3p4+p1p2p3q4+q1p2p3q4. For identical MEs with pj=p RA=p4+4qp3+q2p2. In case B, the SWS succeeds if at least two MEs are in normal state. The system reliability in this case is RB=p1p2p3p4+q1p2p3p4+p1q2p3p4+p1p2q3p4+p1p2p3q4+ q1q2p3p4+q1p2q3p4+q1p2p3q4+p1q2q3p4+p1q2p3q4+p1p2q3q4. For identical MEs RB=p4+4qp3+6q2p2. One can see that the uneven allocation B is more reliable: RB-RA=5q2p2=5(1-p)2p2. 5 Consider now the same system when w=3. In case A the system succeeds only if it does not contain any failed ME: RA=p1p2p3p4. In case B it succeeds if it contains no more than one failed element: RB=p1p2p3p4+q1p2p3p4+p1q2p3p4+p1p2q3p4+p1p2p3q4. For identical MEs: RA=p4, RB=p4+4qp3 and RB-RA=p4+4qp3- p4=4(1-p)p3. Observe that even for w=4 when in case A the system is unable to meet the demand (R A=0) because w>r, in case B it still succeeds with probability RB=p1p2p3p4. This paper considers a general optimal allocation problem in which the number of MEs is not necessarily equal to number of positions (mn) and an arbitrary number of elements can be allocated at each position (some positions may be empty). In order to evaluate the reliability of SWS with any given allocation of MEs the procedure based on a universal generating function technique is developed. A genetic algorithm based on principles of evolution is used as an optimization engine. The integer string solution encoding technique is adopted to represent element allocation in the GA. Section 2 of the paper presents a formulation of the optimal allocation problem. Section 3 describes the technique used for evaluating the SWS reliability for the given allocation of different MEs with the specified performance distribution. Section 4 describes the optimization approach used and its adaptation to the problem formulated. In the fifth section, illustrative examples are presented in which the best-obtained allocation solutions are presented for two different SWSs. 6 2. Problem formulation The SWS consists of n consecutively ordered positions Ci (1in). At each position C1,…,Cn, elements from a set E={e1,…,em} can be allocated. The functioning of each element ej is characterised by its performance Vj. In each state k (1kKj) the performance rate of the element ej is vjk. Thus the performance Vj is a discrete random value which has the following distribution: Kj Pr{Vj v j, k } p j, k , p j, k 1. (1) k 1 The MEs allocation problem can be considered as a problem of partitioning a set E of m elements into a collection of n mutually disjoint subsets Ei (1in), i.e. such that n E i E, (2) i 1 E i E j ø, ij. (3) Each subset Ei, corresponding to SWS position Ci, can contain from 0 to m elements. Further we refer to subset Ei as i-th group. The partition of the set E can be represented by the vector H={h(j),1jm}, where h(j) is the number of the group to which element ej belongs. One can easily obtain the sum of performance rates of the MEs belonging to i-th group (allocated at Ci) as m G i Vj1(h ( j) i). (4) j1 The linear SWS with n positions contains n-r+1 sets of r consecutive positions. The sum of performance rates of elements located at positions belonging to each one of such sets should not be less than w. Therefore, the entire SWS reliability can be defined as follows: r r 1 i 1 i2 R= Pr( G i w G i w ... n n r 1 x r 1 G i w)=Pr( i n r 1 7 x 1 G i w). ix (5) For the given n, r, w and set of m MEs with specified performance distributions, the only factor influencing the SWS reliability is the allocation of its elements H. Therefore, the optimal allocation problem can be formulated as follows: Find vector H* maximizing the SWS reliability R: H*(r,m,n,w)=arg{R(H, r,m,n,w)max}. (6) 3. SWS reliability estimation based on a universal generating function The procedure used in this paper for SWS reliability evaluation is based on the universal ztransform (also called u-function or universal generating function) technique, which was introduced in [9] and which proved to be very effective for reliability evaluation of different types of multi-state systems [10-14]. The u-function technique can straightforwardly handle cases in which any number of multistate elements are allocated in the same position while some positions remain empty. 3.1. u-functions for individual MEs and group of MEs located at the same position The u-function of a discrete random variable X is defined as a polynomial K u ( z) q k z x k , (7) k 1 where the variable X has K possible values and qk is the probability that X is equal to xk. The ufunctions can also be used for representing distributions of more complex mathematical objects [10]. For example the u-function of a discrete random vector X can take the form (7) where x k is k-th realization of X. In this case the exponents are not scalar variables, but vectors and ufunctions (named vector-u-functions) are not longer polynomials. 8 In our case, the u-function can define a ME performance rate distribution, i.e. it represents all of the possible states of the ME j by relating the probabilities of each state pjk to the performance rate vjk of the ME in the form: Kj j (z) p j,k z v j,k (8) k 1 Consider i-th group of MEs (group of MEs located at position Ci). The random performance of this group Gi is equal to cumulative performance of all of its MEs. In order to obtain ufunction ui(z) representing distribution of random value Gi one can use a composition operator that determines ui(z) using simple algebraic operations on the individual u-functions of MEs belonging to i-th group. The composition operator for a pair of MEs ej and ef takes the form: Kj ( j (z), f (z)) ( p j,k z v j,k k 1 Kf , p f ,s z vf ,s s1 K j Kf ) p j,k p f ,s z v j,k vf ,s . (9) k 1s1 The resulting polynomial relates probabilities of each of the possible combinations of states of the two independent MEs (obtained by multiplying the probabilities of corresponding states of each ME) with cumulative performance of the pair of MEs in these states. One can see that the operator satisfies the following condition: {u1 (z),..., u t (z), u t 1 (z),..., u m (z)} {{u1 (z),..., u t (z)}, {u t 1 (z),...., u m (z)}} (10) for arbitrary t. Therefore, it can be applied recursively to obtain the u-function for an arbitrary group of MEs. The i-th group of MEs can be now considered as a single ME with random performance Gi. This performance takes values from the set {gi,1,…,gi,Si}. The total number of possible values of group performance Si is usually much less than the product of numbers of states of the elements belonging to the group since different combination of ME states can produce the same cumulative performance. The distribution of Gi is represented by u-function ui(z): Si u i (z) ( j (z)) = i,k z jEi k 1 9 gi ,k . (11) Note that the absence of any ME at position Ci implies that performance of i-th group is equal to 0 with probability 1. In this case, the corresponding u-function takes the form ui(z)=z0. (12) Note that for any j(z) (z0, j(z))=j(z). (13) 3.2. u-function for set of MEs located at r adjacent positions In order to represent the performance distribution of i-th set consisting of r groups (the set of groups numbered from i to i+r-1) one has to modify the u-function by replacing the random value X with the random vector Gi={Gi,…,Gi+r-1} consisting of random performance values corresponding to all the groups belonging to the set (this replacement produces a vector-ufunction). Each combination of states of groups belonging to the set constitutes a state of the set. The total number of different states of the i-th set is equal to the number of possible combinations of the states of the individual groups belonging to the set. Since all the groups are statistically r independent and each group j has Sj states, the total number of states of the set is Ni = Si j1 . j1 Therefore the vector-u-function Ui(z) corresponding to i-th set of r groups consists of Ni different terms. Consider a state k of i-th set that corresponds to state sj of each individual group j: iji+r-1. The performance values of the groups of the i-th set in state k are represented by the realization gi,k of the random vector Gi: gi,k={ g i,si ,..., g i r 1,si r 1 }. The probability of any state of the set is equal to the product of the probabilities of the corresponding states of the individual groups. For example, the vector-u-function for a set of two groups i and i+1 (for r=2) takes the form: 10 Si Si1 {gi ,s ,gi1,s } i i1 U i (z) q si ,si1 z , (14) si 1 si11 where q si ,si 1 is a probability of an event in which group i is in state si and group i+1 is in state si+1. It can be easily seen that for the statistically independent groups q si ,si1 i,si i1,si1 . Therefore, the vector-u-function of the set can be obtained by applying the following operator over u-functions of individual groups: Si U i (z) (u i (z), u i1 (z)) ( i,si z Si Si 1 gi ,s i si 1 i,si i1,si1 Si 1 , i1,si1 z gi 1,s i 1 ) si 11 {g g } z i,si , i1,si1 . (15) si 1 si 11 Applying the operator over u-functions of r consecutive groups one obtains the vector-ufunction corresponding to the set containing these groups: Si Si 1 U i (z) Si r 1 (u i (z), u i1 (z),..., u ir-1 (z)) ... i,si i1,si1 ...ir 1,sir 1 z si 1si 11 {gi ,si ,gi 1,si 1 ,...,gi r 1,si r 1 } (16) si r 11 Simplifying this representation one obtains: Ni U i ( z ) Q i ,k z gi , k , (17) k 1 where Qi,k is the probability that the i-th set is in state k and vector gi,k consists of values of performance rates of groups at state k. The obtained vector-u-function defines all of the possible states of the i-th set of r groups. Having the vectors gi,k representing performance rates of groups belonging to the i-th set in any state k one can obtain the sum of performance rates of all the MEs belonging to this set: r 1 sum(gi,k)= g i j,si j . j0 11 (18) By summing the probabilities of all of the states in which this sum is less than the demand w, one can obtain the probability of failure of the i-th set of r consecutive groups. To do so one can use the following operator : ( U i (z)) Ni Q k 1(sum( g i, k ) w) . (19) k 1 3.3. u-functions for all the sets of r consecutive groups Note that the SWS considered contains exactly n-r+1 sets of r consecutive groups and each group can belong to no more than r such sets. To obtain the u-function corresponding to all the sets of r consecutive groups the following procedure is introduced: 1. Define u-function U1-r(z) as follows U1-r(z)= z g0 , (20) where the vector g0 consists of r zeros. 2. Define the following operator over vector-u-function Ui(z) and u-function of individual group ui+r(z): Ni Ψ( U i (z), u ir (z)) Ψ( Qi,k z k 1 gi ,k Si r , ir ,s z s1 g j,s Ni Si r ) Q i,k ir ,s z ( g ,g ) i ,k i r ,s , (21) k 1 s1 where operator over arbitrary vector y and value x shifts all the vector elements one position left: y(s-1)=y(s) for 1<sr and assigns the value x to the last element of y: y(r)=x (the first element of vector y disappears after applying the operator). The operator removes the performance value of the first group of the set and adds the performance value of the next (not considered yet) group to the set preserving the order of groups belonging to the set. Therefore, applying the operator over vector-u-function representing performance distribution of the i-th set of r groups (represented by the random vector Gi) one obtains the vector-u-function representing the performance distribution of the i+1-th set (represented by the random vector Gi+1). 12 3. Using the operator in sequence as follows: Ui1r (z) Ψ(Uir (z), u i (z)) (22) for i=1,…,n one obtains vector-u-functions for all of the possible sets of r consecutive groups: U1(z), …, Un-r+1(z). Note that the vector-u-function for the first set U1(z) is obtained after applying the operator r times. Applying the operator (19) to the vector-u-functions U1(z), …, Un-r+1(z) one can obtain the failure probability for each set of r consecutive groups. Note that if for some combination of MEs states i-th set of r groups fails, the entire SWS fails independently of the states of the MEs that do not belong to this set of groups. Therefore the terms corresponding to the failure of i-th set can be removed from the vector-u-function Ui(z) since they should not participate in determining further state combinations that cause system faults. This consideration lies at the base of the following algorithm. 3.4. Algorithm for SWS reliability evaluation for a given allocation of MEs H 1. Assign ui(z)=z0 for each i=1,…,n. 2. According to the given vector H for each 1jm, determine j(z) using Eq. (8) and modify uh(j)(z): uh(j)(z)=(uh(j)(z), j(z)). 3. Assign F=0 and U1-r(z)= z g0 . 4. Repeat the following for i=1,…,n: 4.1. Obtain Ui1r (z) (Uir (z), u i (z)) . 4.2. If ir add value (Ui1r (z)) to F and remove all the terms with sum(gi+1-r,k)<w from Ui1r (z) . 13 5. Obtain the SWS reliability as R=1-F. Alternatively, the system reliability can be obtained as the sum of the coefficients of the last vector-u-function U n1r (z) . 3.5. Example In order to illustrate the procedure, we will obtain the reliability of the SWS presented in the introduction (Fig. 2). In this SWS n=m=4, r=3, w=2, Kj=2, Pr{Vj=1}=pj, Pr{Vj=0}=1-pj=qj for any ME ej 1j4. The u-functions of the individual MEs are: 1(z)=q1z0+p1z1, 2(z)=q2z0+p2z1, 3(z)=q3z0+p3z1, 4(z)=q4z0+p4z1. First, consider the case A. The ME allocation in this case is represented by vector H={1,2,3,4}. The u-functions representing distribution of random values G1, G2 G3 and G4 for the groups of MEs allocated at the same positions are: u1(z)=(z0,q1z0+p1z1)=q1z0+p1z1, u2(z)=(z0,q2z0+p2z1)=q2z0+p2z1, u3(z)=(z0,q3z0+p3z1)=q3z0+p3z1, u4(z)=(z0,q4z0+p4z1)=q4z0+p4z1, Following step 3 of the algorithm we assign F=0, U-2(z)=z0,0,0. Following step 4 of the algorithm we obtain: U-1(z)=[U-2(z),u1(z)]= [z0,0,0, q1z0+p1z1]=q1z0,0,0+p1z0,0,1, U0(z)=[U-1(z),u2(z)]=[ q1z0,0,0+p1z0,0,1, q2z0+p2z1]= q1q2z0,0,0+p1q2 z0,1,0+q1p2z0,0,1+p1p2z0,1,1. U1(z)=[U0(z),u3(z)]=[q1q2z0,0,0+p1q2z0,1,0+q1p2z0,0,1+p1p2z0,1,1, q3z0+p3z1]= q1q2q3z0,0,0+p1q2q3z1,0,0+q1p2q3z0,1,0+p1p2q3z1,1,0+ q1q2p3z0,0,1+p1q2p3z1,0,1+q1p2p3z0,1,1+p1p2p3z1,1,1 The terms of U1(z) with sum(g1,k)<2 are marked in bold. After removing the marked terms, U1(z) takes the form: 14 U1(z)=p1p2q3z1,1,0+p1q2p3z1,0,1+q1p2p3z0,1,1+p1p2p3z1,1,1. U2(z)=[U1(z),u4(z)]=[p1p2q3z1,1,0+p1q2p3z1,0,1+q1p2p3z0,1,1+p1p2p3z1,1,1, q4z0+p4z1]= p1p2q3q4z1,0,0+p1q2p3q4z0,1,0+q1p2p3q4z1,1,0+p1p2p3q4z1,1,0 p1p2q3p4z1,1,1+p1q2p3p4z0,1,1+q1p2p3p4z1,1,1+p1p2p3p4z1,1,1. The terms with sum(g2,k)<2 are marked in bold. After removing these terms the vector-ufunction takes the form: U2(z)=q1p2p3q4z1,1,0+p1p2p3q4z1,1,0+p1p2q3p4z1,1,1+p1q2p3p4z0,1,1+q1p2p3p4z1,1,1+p1p2p3p4z1,1,1. The SWS reliability is equal to the sum of the coefficients of vector-u-function U2(z): RA=q1p2p3q4+p1p2p3q4+p1p2q3p4+p1q2p3p4+q1p2p3p4+p1p2p3p4. The ME allocation in case B is represented by vector H={2,2,3,3}. The u-functions representing distribution of random vales G1, G2 G3 and G4 for the groups of MEs allocated at the same positions are after simplification: u1(z)=z0, u2(z)=(1(z), 2(z))=(q1z0+p1z1, q2z0+p2z1)=q1q2z0+(p1q2+q1p2)z1+p1p2z2, u3(z)=(3(z), 4(z))=(q3z0+p3z1, q4z0+p4z1)=q3q4z0+(p3q4+q3p4)z1+p3p4z2, u4(z)=z0, Following step 3 of the algorithm we assign F=0, U-2(z)=z0,0,0. Following step 4 of the algorithm we obtain: U-1(z)=[U-2(z),u1(z)]= [z0,0,0, z0]=z0,0,0, U0(z)=[U-1(z),u2(z)]=[z0,0,0, q1q2z0+(p1q2+q1p2)z1+p1p2z2]= q1q2z0,0,0+(p1q2+q1p2)z0,0,1+p1p2z0,0,2. U1(z)=[U0(z),u3(z)]=[q1q2z0,0,0+(p1q2+q1p2)z0,0,1+p1p2z0,0,2, q3q4z0+(p3p4+q3p4)z1+p3p4z2]= q1q2q3q4z0,0,0+(p1q2+q1p2)q3q4z0,1,0+p1p2q3q4z0,2,0+ q1q2(p3q4+q3p4)z0,0,1+(p1q2+q1p2)(p3q4+q3p4)z0,1,1+p1p2(p3q4+q3p4)z0,2,1 q1q2p3p4z0,0,2+(p1q2+q1p2)p3p4z0,1,2+p1p2p3p4z0,2,2. 15 The terms of U1(z) with sum(g1,k)<2 are marked in bold. After removing the marked terms, U1(z) takes the form: U1(z)=p1p2q3q4z0,2,0+(p1q2+q1p2)(p3q4+q3p4)z0,1,1+p1p2(p3q4+q3p4)z0,2,1+ q1q2p3p4z0,0,2+(p1q2+q1p2)p3p4z0,1,2+p1p2p3p4z0,2,2. U2(z)=[U1(z),u4(z)]=[ p1p2q3q4z0,2,0+(p1q2+q1p2)(p3q4+q3p4)z0,1,1+p1p2(p3q4+q3p4)z0,2,1+ q1q2p3p4z0,0,2+(p1q2+q1p2)p3p4z0,1,2+p1p2p3p4z0,2,2, z0]= p1p2q3q4z2,0,0+(p1q2+q1p2)(p3q4+q3p4)z1,1,0+p1p2(p3q4+q3p4)z2,1,0+ q1q2p3p4z0,2,0+(p1q2+q1p2)p3p4z1,2,0+p1p2p3p4z2,2,0. U1(z) does not contain terms with sum(g2,k)<2, therefore no terms are removed in this vector-ufunction. The SWS reliability is equal to the sum of the coefficients of the vector-u-function U2(z). RB= p1p2q3q4+(p1q2+q1p2)(p3q4+q3p4)+p1p2(p3q4+q3p4)+ q1q2p3p4+(p1q2+q1p2)p3p4+p1p2p3p4= p1p2q3q4+p1q2p3q4+p1q2q3p4+q1p2p3q4+q1p2q3p4+p1p2p3q4+p1p2q3p4+ q1q2p3p4+p1q2p3p4+q1p2p3p4+p1p2p3p4. 4. Optimization technique Finding the optimal ME allocation in SWS is a complicated combinatorial optimization problem having nm possible solutions. An exhaustive examination of all these solutions is not realistic even for a moderate number of positions and elements, considering reasonable time limitations. As in most combinatorial optimization problems, the quality of a given solution is the only information available during the search for the optimal solution. Therefore, a heuristic search algorithm is needed which uses only estimates of solution quality and which does not require derivative information to determine the next direction of the search. 16 The recently developed family of genetic algorithms is based on the simple principle of evolutionary search in solution space. GAs have been proven to be effective optimization tools for a large number of applications. Successful applications of GAs in reliability engineering are reported in [10,15-21]. It is recognized that GAs have the theoretical property of global convergence [22]. Despite the fact that their convergence reliability and convergence velocity are contradictory, for most practical, moderately sized combinatorial problems, the proper choice of GA parameters allows solutions close enough to the optimal one to be obtained in a short time. 4.1. Genetic Algorithm Basic notions of GAs are originally inspired by biological genetics. GAs operate with "chromosomal" representation of solutions, where crossover, mutation and selection procedures are applied. "Chromosomal" representation requires the solution to be coded as a finite length string. Unlike various constructive optimization algorithms that use sophisticated methods to obtain a good singular solution, the GA deals with a set of solutions (population) and tends to manipulate each solution in the simplest manner. A brief introduction to genetic algorithms is presented in [23]. More detailed information on GAs can be found in Goldberg’s comprehensive book [24], and recent developments in GA theory and practice can be found in books [21, 22]. The steady state version of the GA used in this paper was developed by Whitley [25]. As reported in [26] this version, named GENITOR, outperforms the basic “generational” GA. The structure of steady state GA is as follows: 17 1. Generate an initial population of Ns randomly constructed solutions (strings) and evaluate their fitness. (Unlike the “generational” GA, the steady state GA performs the evolution search within the same population improving its average fitness by replacing worst solutions with better ones). 2. Select two solutions randomly and produce a new solution (offspring) using a crossover procedure that provides inheritance of some basic properties of the parent strings in the offspring. The probability of selecting the solution as a parent is proportional to the rank of this solution. (All the solutions in the population are ranked by increasing order of their fitness). 3. Allow the offspring to mutate with given probability Pm. Mutation results in slight changes in the offspring structure and maintains diversity of solutions. This procedure avoids premature convergence to a local optimum and facilitates jumps in the solution space. The positive changes in the solution code created by the mutation can be later propagated throughout the population via crossovers. 4. Decode the offspring to obtain the objective function (fitness) values. These values are a measure of quality, which is used in comparing different solutions. 5. Apply a selection procedure that compares the new offspring with the worst solution in the population and selects the one that is better. The better solution joins the population and the worse one is discarded. If the population contains equivalent solutions following the selection process, redundancies are eliminated and, as a result, the population size decreases. Note that each time the new solution has sufficient fitness to enter the population, it alters the pool of prospective parent solutions and increases the average fitness of the current population. The average fitness increases monotonically (or, in the worst case, does not vary) during each genetic cycle (steps 2-5). 6. Generate new randomly constructed solutions to replenish the population after repeating steps 2-5 Nrep times (or until the population contains a single solution or solutions with equal 18 quality). Run the new genetic cycle (return to step 2). In the beginning of a new genetic cycle, the average fitness can decrease drastically due to inclusion of poor random solutions into the population. These new solutions are necessary to bring into the population new "genetic material" which widens the search space and, like a mutation operator, prevents premature convergence to the local optimum. 7. Terminate the GA after Nc genetic cycles. The final population contains the best solution achieved. It also contains different nearoptimal solutions, which may be of interest in the decision-making process. 4.2. Solution representation and basic GA procedures To apply the genetic algorithm to a specific problem, one must define a solution representation and decoding procedure, as well as specific crossover and mutation procedures. As it was shown in section 2, any arbitrary m-length vector H with elements h(i) belonging to the range [1,n] represents a feasible allocation of MEs. Such vectors can represent each one of the possible nm different solutions. The fitness of each solution is equal to the reliability of SWS with allocation, represented by the corresponding vector H. To estimate the SWS reliability for the arbitrary vector H, one should apply the procedure presented in section 3. The random solution generation procedure provides solution feasibility by generating vectors of random integer numbers within the range [1,n]. It can be seen that the following crossover and mutation procedures also preserve solution feasibility. The crossover operator for given parent vectors P1, P2 and the offspring vector O is defined as follows: first P1 is copied to O, then all numbers of elements belonging to the fragment between a and b positions of the vector P2 (where a and b are random values, 1a<bm) are copied to the corresponding positions of O. The following example illustrates the crossover procedure for m=6, n=4: 19 P1=2 4 1 4 2 3 P2=1 1 2 3 4 2 O=2 4 2 3 4 3 The mutation operator moves a randomly chosen ME to the adjacent position (if such a position exists) by modifying a randomly chosen element h(i) of H using rule h(i)=max{h(i)-1,1} or rule h(i)=min{h(i)+1,n} with equal probability. The vector O in our example can take the following form after applying the mutation operator : O=2 3 2 3 4 3. 5. Illustrative examples 5.1. SWS with identical MEs Consider a SWS with n=10 positions in which m=10 two-state identical MEs are to be allocated. The performance distribution of each ME is Pr(Vj=1)=0.9, Pr(Vj=0)=0.1 for 1j10. Table 1 presents allocation solutions obtained for different r and w (number of identical elements in each position). The reliability of the SWS corresponding to the obtained allocations is compared with its reliability corresponding to the case when the MEs are evenly distributed among the positions. One can see that the reliability improvement achieved by the free allocation increases with the increase of r and w. On the contrary, the number of occupied positions in the best obtained solutions decreases when r and w grow. Fig. 3 presents the SWS reliability as a function of demand w for r=2, r=3 and r=4 for even ME allocation and for unconstrained allocation obtained by the GA. 5.2. SWS with different MEs 20 Consider the ME allocation problem presented in [8], in which n=m=10, r=3 and w=1. The performance distributions of MEs are presented in Table 2. The best ME allocation solutions obtained by the GA are presented in Table 3 (list of elements located at each position). The best even allocation solution obtained in [8] considerably improves when the even allocation constraint is removed. One can see that the best unconstrained allocation solution obtained by the GA in which only 4 out of 10 positions are occupied by the MEs provides 42% reliability increase over even allocation. The system reliability as a function of demand for the obtained even and unconstrained allocations is presented in Fig. 4. Table 3 presents also the best allocations of the first m MEs from Table 2 (for m=9, m=8 and m=7). Observe that free allocation of 9 MEs in the SWS still provides greater reliability than does even allocation of 10 MEs. 5.3. Computational Effort and Algorithm Consistency The C language realization of the algorithm was tested on a Pentium II PC. The chosen parameters of GA were NS=100, Nrep=2000, Nc=10 and Pm=1. The time taken to obtain the bestin-population solution (time of the last modification of the best solution obtained) did not exceed 20 seconds for the optimization problems presented in the previous sections when r=3. The parameter r has the greatest influence on the computational time since with the growth of r the number of possible states for each set of r groups increases dramatically. For instance, solving the second optimization problem for r=4 takes 120 seconds. It should be noted also that increase of the demand w leads to reduction of the computational time because of more intensive truncation of the u-functions in step 4.2 of the SWS reliability evaluation algorithm. To demonstrate the consistency of the suggested algorithm, we repeated the GA 25 times with different starting solutions (initial population) for the second problem. The coefficient of variation (CV) was calculated for fitness values of best-in-population solutions obtained during 21 the genetic search by different GA search processes. The variation of this index during the GA procedure is presented in Fig. 5. One can see that the standard deviation of the final solution fitness does not exceed 0.6 % of its average value. References [1] G. Levitin, Linear multi-state sliding window systems, IEEE Trans. Reliability, 52, 2003, pp.263-269. [2] W. Griffith, On consecutive k-out-of-n failure systems and their generalizations, A. P. Basu (ed), Reliability and Quality Control, 1986, pp. 157-165; Elsevier (North-Holland). [3]. Y. Tong, A rearrangement inequality for the longest run with an application in network reliability, J. Applied Probability, vol. 22, 1985, pp. 386-393. [4] B. Saperstain, The generalized birthday problem, J. Amer. Statistical Assoc, vol. 67, 1972, pp. 425-428. [5] B. Saperstain, On the occurrence of n successes within N Bernoulli trails, Technometrics, vol. 15, 1973, pp. 809818. [6] J. Naus, Probabilities for a generalized birthday problem, J. Amer. Statistical Assoc, vol. 69, 1974, pp. 810-815. [7] J. Nelson, Minimal-order models for false-alarm calculations on sliding windows, IEEE Trans. Aerospace Electronic Systems, vol. AES-14, 1978, pp. 351-363. [8] G. Levitin, "Optimal allocation of elements in linear multi-state sliding window system", Reliability Engineering and System Safety, 76, 2002, pp. 245-254. [9] I. A. Ushakov, Universal generating function, Sov. J. Computing System Science, vol. 24, No 5, 1986, pp. 118129. [10] A. Lisnianski, G. Levitin, Multi-state system reliability. Assessment, Optimization and Applications, World Scientific, 2003. [11] G. Levitin, "Reliability evaluation for linear consecutively-connected systems with multistate elements and retransmission delays", Quality and Reliability Engineering International, vol. 17, 2001. [12] G. Levitin, "Evaluating correct classification probability for weighted voting classifiers with plurality voting", European journal of operational research, vol. 141, pp. 596-607, 2002. [13] G. Levitin, "Reliability evaluation for acyclic consecutively-connected networks with multistate elements", Reliability Engineering & System Safety, vol. 73, pp. 137-143, 2001. [14] G. Levitin, "Incorporating Common-Cause Failures into Non-Repairable Multi-State Series-Parallel System Analysis", IEEE Transactions on Reliability, vol. 50, pp. 380-388, 2001. 22 [15] L. Painton and J. Campbell, "Genetic algorithm in optimization of system reliability", IEEE Trans. Reliability, 44, 1995, pp. 172-178. [16] D. Coit and A. Smith, "Reliability optimization of series-parallel systems using genetic algorithm", IEEE Trans. Reliability, 45, 1996, pp. 254-266. [17] D. Coit and A. Smith, "Redundancy allocation to maximize a lower percentile of the system time-to-failure distribution", IEEE Trans. Reliability, 47, 1998, pp. 79-87. [18] Y. Hsieh, T. Chen, D. Bricker, "Genetic algorithms for reliability design problems", Microelectronics and Reliability, 38, 1998, pp. 1599-1605. [19] J. Yang, M. Hwang, T. Sung, Y. Jin, "Application of genetic algorithm for reliability allocation in nuclear power plant", Reliability Engineering & System Safety, 65, 1999, pp. 229-238. [20] M. Gen and J. Kim, "GA-based reliability design: state-of-the-art survey", Computers & Ind. Engng, 37, 1999, pp. 151-155. [21] M. Gen and R. Cheng, Genetic Algorithms and engineering design, John Wiley & Sons, New York, 1997. [22] T. Back, Evolutionary Algorithms in Theory and Practice. Evolution Strategies. Evolutionary Programming. Genetic Algorithms, Oxford University Press, 1996. [23] S. Austin, "An introduction to genetic algorithms", AI Expert, 5, 1990, pp. 49-53. [24] D. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison Wesley, Reading, MA, 1989. [25] D. Whitley, The GENITOR Algorithm and Selective Pressure: Why Rank-Based Allocation of Reproductive Trials is Best. Proc. 3th International Conf. on Genetic Algorithms. D. Schaffer, ed., pp. 116-121. Morgan Kaufmann, 1989. [26]. G. Syswerda, “A study of reproduction in generational and steady-state genetic algorithms, in G.J.E. Rawlings (ed.), Foundations of Genetic Algorithms, Morgan Kaufmann, San Mateo, CA, 1991. 23 Table 1. Solutions of the first allocation problem. Position C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 r=2, w=1 r=3, w=2 2 1 3 Free allocation Even allocation Improvement 0.951 0.920 3.4% 2 r=4, w=3 5 3 2 5 2 3 2 Reliability 0.941 0.866 8.7% 24 0.983 0.828 18.7% Table 2. MEs' performance distributions for the second allocation problem. No of ME State 1 2 p V p V 0.03 0.0 0.10 0.0 1 0.22 0.2 0.10 0.1 2 0.75 0.5 0.40 0.2 3 - 0.40 0.4 4 5 No of ME 6 7 State p V p V 0.01 0.0 0.20 0.0 1 0.22 0.4 0.10 0.3 2 0.77 0.5 0.10 0.4 3 - 0.60 0.5 4 3 4 5 p V p V p 0.17 0.0 0.05 0.0 0.08 0.83 0.6 0.25 0.3 0.20 - 0.40 0.5 0.15 - 0.30 0.6 0.45 - 0.12 8 9 10 p V p V p 0.05 0.0 0.20 0.0 0.05 0.25 0.4 0.10 0.3 0.25 0.70 0.6 0.15 0.4 0.70 - 0.55 0.5 - 25 V 0 1 2 4 5 V 0 2 6 - Table 3. Solutions of the second allocation problem (n=10, r=3, w=1). Position C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 Reliability m=10 Even allocation Free allocation 5 9 3 6, 7, 10 1 4 2, 5 7 1, 4 10 8 3, 8, 9 6 2 0.536 0.765 26 m=9 m=8 m=7 2, 5, 8, 9 3, 6 1 7 7 1, 4 5, 7, 8 3 6 3 6 4 1, 2 4, 5 2 0.653 0.509 0.213 Fig. 1 Figure 1: Example of SWS with r=3. 27 Fig. 2 e1 e2 e3 e4 C1 C2 C3 C4 A C1 e1,e2 e3,e4 C2 C3 B Figure 2: Two possible allocations of MEs in SWS with n=m=4. 28 C4 Fig. 3 1 0.8 0.6 R 0.4 0.2 0 0 1 2 3 4 5 W even r=2 even r=4 unconstrained r=3 even r=3 unconstrained r=2 unconstrained r=4 Figure 3: Reliability of SWS with identical MEs for different r and ME allocations. 29 Fig. 4 1 0.8 0.6 R 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 W even unconstrained Figure 4: Reliability of SWS with different MEs for optimal even and uneven ME allocations. 30 Fig. 5 15 Coefficient of Variation (%) 12 9 6 3 0 0 2 4 6 8 10 No of Crossovers (Thousands) Figure 5: CV of best-in-population solution fitness obtained by 25 different search processes as function of number of crossovers. 31