6th International Science, Social Sciences, Engineering and Energy Conference 17-19 December, 2014, Prajaktra Design Hotel, Udon Thani, Thailand I-SEEC 2014 http//iseec2014.udru.ac.th A teaching-learning-based optimization algorithm for sizing optimization of trusses Anan Nimtawat Faculty of Technology, Udon Thani Rajabhat University, Thailand anannimt@gmail.com Abstract In this study, a modified teaching learning based optimization algorithm is applied to sizing design of truss structures. Non-linear constrained engineering problems, such as truss design problems, are popularly solved by using meta-heuristic search algorithms such as Genetic algorithms (GA), Particle swarm optimization (PSO), and teaching learning based optimization (TLBO). It is because the metaheuristic search algorithms only need to iteratively evaluate problem functions without requiring gradient information of the problem functions. The TLBO has been more widely used in that it can give good solution with smaller numbers of predetermined parameters than other algorithms. Originally the TLBO consists of two searching steps: teacher phase and learner phase. Firstly, in the teacher phase, all searching points in the defined population, learners in the defined class, are encouraged to move toward the promoted teacher that is the best searching point in the iteration. Secondly, all searching points are motivated to move by randomly interacting with other points in the population. A modified TLBO algorithm is proposed and validated by applying to truss design problems. The results show that the modified TLBO can find good solutions of the design examples. Keywords:Optimization ; teaching learning based optimization ; truss design 1. Introduction The optimization of trusses is one of the most active researches in engineering [1-5]. The truss optimization problems can be classified into three groups including sizing optimization, shaping optimization and topology optimization. The sizing optimization is to find the optimal size of all truss elements without changing its original shape and configuration. The shaping optimization is to optimize the truss design by allowing to change the element size and the truss shape but no configuration changes otherwise it is called the topology optimization. The sizing optimization of trusses is a kind of non-linear 2 optimization problem that may be hard to find the gradient of the objective function. Although there exists a number of recognized gradient-based method to solve non-linear function such as the steepest descent method and the conjugate gradient method, it is not easy to explore possible solutions in a quite large search space. A number of the population-based evolutionary algorithms such as Genetic algorithms (GA) and Particle swarm optimization (PSO) are recently employed to solve large scaled constrained non-linear problems, for example, the optimization of trusses [4, 5]. However the performance of the mentioned algorithms greatly depend on required arbitrary parameters used in the algorithms. For example, the parameters required in the GA include the probabilistic of crossover and the probabilistic of mutation [5]. The parameters required in the PSO include inertia weight, and social and cognitive parameters [4]. Recently a novel algorithm called the teaching learning based optimization (TLBO) proposed by Rao et al. [1] is employed to solve non-linear problems, and truss optimization problems [1-3]. The TLBO algorithm mimics the teaching behaviors between a teacher and learners, and the learning behaviors among learners in the classroom. The philosophy of the TLBO is that the teacher tries to share her knowledge to her learners in the class. The teacher will try to improve the mean of grade of the class. In additions, each learners in the class will also share their knowledge among other learners in the class. The required input parameters of TLBO are only the population size and the number of generations that are the common parameter of the population-based evolutionary algorithms such as GA and PSO. Although there is a teaching factor that is arbitrary required in the TLBO algorithm, it can be randomly generated in the algorithm. Therefore it may be said that the TLBO is a parameter-less algorithm [1]. In this study, the original TLBO is modified for finding the global optimum solution of sizing optimization of trusses. The proposed modified TLBO will be described in the next section. 2. Teaching-learning based optimization The basic concept of the TLBO algorithm is that it attempts to move the searching point toward the best current point or called the teacher, and attempts to move the searching point by getting some information from the other point in the same population [1]. There is only one parameter required in the TLBO algorithm that is the teaching factor that will be explained later. Like a Genetic algorithm (GA), the TLBO finds the best point or solution by evolving initial solutions to better solutions. The TLBO also uses a population of learners or a population of individuals to search the best solution by a number of iterations or generations. Each learner consists of a vector of design variables representing marks of subjects. Unlike a GA, the TLBO has no crossover or mutation operators. The TLBO has two operators called teacher phase and learner phase. The TLBO processes can be summarized here. First, the number of starting points (learners) is uniformly randomly initialized including setting up the termination criterion such as the number of iterations. A searching point is a vector of design variables ( x ). Then, the mean of each component (design variables) of x is calculated. Next, the best point, the point with minimum value of the objective function at the current iteration, is identified and promoted as the teacher of the class (population). The teacher will try to move the current mean to its own point, so the new mean will be the current teacher. Then the difference value between the current mean of each direction and the corresponding current best point are calculated. This different value is called the difference mean. After that all points in the population will be updated in two step called the teacher phase and the learner phase. Let N is the number of x in the population. At the current iteration t , the mean of x t is the calculated by Eq. (1) designated as x mean ,t . As mentioned above, the new mean of x t is the current best point or the teacher designated as xteacher ,t . Then, the difference mean Δt can be calculated using Eq. (2). 3 1 N x k ,t N k 1 Δt rt (xteacher ,t TF x mean ,t ) xmean,t (1) (2) Here, Δt is the difference mean, rt is the random number in the range [0, 1], and TF is the teaching factor that can be either 1 or 2. In the teacher phase, all searching points (learners) are encouraged to move toward the best point (teacher) of the current population. All points x old ,t are moved to the new positions x new,t using Eq. (3). However, the point will move to the new position only when the new position is better than its current position. x new,t xold ,t Δt (3) In the learner phase, the point is moved by randomly interacting with others. For the minimization problem, the point will be moved using Eq. (4). However, the point will be moved to the new position only when the new position is better than its current position. x A, new,t x A,old ,t rt (x A,old ,t x B ,t ) if f (x A,t ) f (x B ,t ) x A, new,t x A,old ,t rt (x B ,t x A,old ,t ) if f (x A,t ) f (x B ,t ) (4) Here f ( x A,t ) and f ( x B ,t ) is the objective function value of the x A , t and x B , t respectively. All points are iteratively moved until the termination criterion is reached, and the best point is identified as the global optimum solution. 3. The formulation of the truss optimization problem The truss optimization problem is to find the minimum weight of truss structures while the configuration of the truss structure is fixed. The optimization problem can be written as: find x A1 , A2 , A3 , minimize W (x) i Li Ai , An m (5) i 1 subjected to L i U L j U AL Ai AU where W is the total weight of the design truss, m is the total number of element, i is the material unit weight of element i , Li is the length of element i , and Ai is the cross sectional area of element i . The design truss must satisfy all three constraints: the element stress constraint i , the nodal deflection constraint j , and the cross sectional element area constraint Ai . The cross sectional area constraints can be satisfied just by limiting the search space of cross-sectional areas of elements. 4 The function for calculating the total weight of the truss design as shown in Eq. (5) is defined as the objective function of the truss optimization problem. However, the truss design with very small weight may violate any required constraint that is defined as an infeasible truss design. To handle the infeasible design, a penalty function is multiplied to the objective function W to insure that the feasible design will have the smaller weight than the infeasible design. Here, the penalized objective function W p is defined as Eq. (6). Wp WE (6) Here, E is the total degree of constraint violation that is composed of the total degree of stress constraint violation E , and the total degree of displacement constraint violation E of the truss design. The total degree of constraint violation E is defined as Eq. (7). E 1 E E (7) The total degree of stress constraint violation E of the truss design is the summation of the degree of stress constraint violation, and can be written as Eq. (8). m E e ,i (8) i 1 Here e ,i is the degree of element stress constraint violation of element i , and m is the number of elements of the truss design. The degree of stress constraint violation e ,i of each element is defined as Eq. (9). if L i U otherwise e ,i 0 e ,i i a (9) a Here i is the element stress of element i , and a is the lower bound L or upper bound U of the allowable element stress of element i . The total degree of displacement constraint violation E of the truss design is the summation of the degree of displacement constraint violation, and can be written as Eq. (10). n E e , j j 1 (10) Here, e , j is the degree of nodal displacement constraint violation in each direction of node j , and n is the number of nodes of the truss design. If the planar truss design is considered, two directions x and y are considered. The degree of deflection constraint violation e , j of each node is defined as Eq. (11). 5 if L i U otherwise e , j 0 e , j (11) j a a Here i is the nodal displacement in each direction of node j , and a is the lower bound L or upper bound U of the allowable nodal displacement of node j . 4. A modified TLBO algorithm for sizing optimization of trusses A modified TLBO algorithm is proposed in this study for sizing optimization of planar trusses. A searching point or a design solution is a vector of cross-sectional area of a truss design. The best solution is the vector of cross-sectional area that configures the optimal (minimum weight) truss design. The configuration of the truss is not varied, but only the cross-sectional area of each element of the truss is set as a design variable. 4.1. Initialization Let a solution x is a vector of cross-sectional areas. The population of x is randomly generated according to the number of population size N . All truss structures generated using cross-sectional areas defined by x are analyzed by using the finite element method. The details of the finite element method for truss structures can be found elsewhere [6]. Then the result of the penalized objective function W p are calculated using Eq. (6). 4.2. Teacher phase At the iteration t , the current best design xteacher ,t is identified, and the current mean x mean ,t is calculated using Eq. (1), and the difference mean Δt is calculated using Eq. (2). The adaptive teaching factor TF as shown in Eq. (12) is employed. It is noted TF is normally large in the early step of iterations and will decreasingly approach to 1 in the later step of iterations. TF ,t f (x worstfeas ,t ) f (xteacher ,t ) (12) Here f (x worstfeas ,t ) and f (xteacher ,t ) is the result of penalized objective function W p according to x worstfeas ,t and xteacher ,t respectively when x worstfeas ,t is the worst feasible design of the current population, and xteacher ,t is the current best design in the tth iteration. If the f (xteacher ,t ) is zero, TF is set to 1. The sign of the difference mean Δt is modified before calculating x new,t using Eq. (3). The concept is the x new,t will always move toward the xteacher ,t . The x new,t will be accepted only when the x new,t is better than the x old ,t . The rules for identifying that which x k ,t is better is proposed as followed: if x A , t is feasible and x B , t is infeasible, x A , t is better than x B , t 6 if both are feasible, x A , t is better than x B , t when x A , t has smaller objective value than x B , t if both are infeasible, x A , t is better than x B , t when x A , t has smaller penalty value than x B , t . It is noted that an infeasible x k ,t representing a truss design that has some constraint violations. 4.3. Learner phase After the teacher phase, all solutions x k ,t in the population will be updated again using Eq. (4). At first, a solution x A , t will be randomly mated to x B , t . Then x A, new,t is calculated using Eq. (4). The Eq. (4) is also modified in such a way that the x A, mew,t will move toward the x B , t only when x B , t is better than x A, old ,t otherwise x A, mew,t will move away x B , t . In addition, similar to the teacher phase, the new solution x A, mew,t will be accepted only when it is better than the old one x A, old ,t by employing the proposed comparison rules mentioned above. 4.4. Termination After the process of the learner phase is applied, the algorithm will be getting in the teacher phase again, and the learner phase will be applied again. The algorithm will be operated iteratively until the termination criterion is accomplished. In this study, the termination criterion is simply set as the required number of iterations that is given as the required input of the algorithm. 5. Design problem The proposed modified TLBO algorithm is implemented and is verified by testing a problem of sizing optimization of the planar ten-bar cantilever truss. The test results are compared with results in literatures that also used the TLBO as an optimizer. The truss configured as shown in Fig. (1) is one of the benchmark problem in sizing optimization of trusses [2, 3]. The range of the cross-sectional area used is 0.1 in 2 to 35.0 in 2 . For the two types of constraint, the stress for each truss element is limited to 25 ksi , and the displacement in horizontal and vertical directions of all nodes of the truss design is limited to 2 in. The unit weight of the material is set as 0.1 lb/in3 , and the modulus of elasticity is set as 107 psi. (5) [1] [8] 360 in (3) [7] [10] [2] (1) [9] [5] [6] 360 in 360 in y x [3] (6) [4] (4) 100 kips (2) 100 kips Fig. 1. The configuration of the planar ten-bar truss. 7 The number of populations used is 20, and the required number of iterations is 2,000. The number of runs is 30. The results are shown in Table 1. The minimum weight obtained is 5061.1295 lb without any constraint violation. However, it is noted that the results from Degertekin SO and Hayalioglu MS [2] has some degree of displacement constraint violation 0.000051, and the result from Camp CV [3] has some degree of constraint violation 0.000035. Table 1. Results after 30 runs of the example Elements No. This study Degertekin and Hayalioglu [2] Camp [3] 1 30.53329 30.4286 30.6684 2 0.10000 0.1000 0.1000 3 23.15025 23.2436 23.1584 4 15.27638 15.3677 15.2226 5 0.10002 0.1000 0.1000 6 0.55563 0.5751 0.5421 7 21.04019 20.9665 21.0255 8 7.46202 7.4404 7.4654 9 0.10001 0.1000 0.1000 10 21.51140 21.5330 21.4660 Minimum weight (lb) 5061.1295 5060.96 5060.973 Average weight (lb) 5061.3946 5062.08 5064.808 Standard Deviation (lb) 0.2236 0.79 6.3707 Average number of structure analyses 37,600 16,872 13,767 6. Conclusions and future researches The teaching-learning-based optimization (TLBO) mimics the behavior of learners in a classroom. The learners or the searching points move toward the best learner of the population, and move toward the other better learner. This study found that the adaptive teaching factor is recommended. The modified TLBO moves searching points quite fast in the early iterations and then gradually move searching points finely to explore the better point in a quite small step. A topology optimization of trusses using TLBO is one of the interesting future researches. References [1] Rao RV, Savasani VJ, Vakharia DP. Teaching-learning-based optimization: a novel for comstrained mechanical design optimization problems. Computer-Aided Design 2011;43:303-315. [2] Degertekin SO, Hayalioglu MS. Sizing truss structures using teaching-learning-based optimization. Computers and Structures 2013;119:177-88. [3] Camp CV, Farshchin M. Design of space trusses using modified teaching-learning based optimization. Engineering Structures 2014;62-63:87-97. [4] Schutte JF, Groenwold AA. Sizing design of truss structures using particle swarms. Struct Multidisc Optim 2003;25:261-9. [5] Camp CV, Pezeshk S, Cao G. Optimized design of two-dimentional structures using a genetic algorithm. J Struct Eng 1998;124(5):551-9. [6] Chandrupatla TR, Belegundu AD. Introduction to Finite Elements in Engineering. 3rd ed. New Jersey: Prentice Hall; 2002.