TRAVELLING SALESMAN PROBLEM Epiteo ey FEDERICO Greco I-TechPublished by In-Teh In-Teh is Croatian branch of I-Tech Education and Publishing KG, Vienna, Austra, Abstracting and non-profit use of the material is permitted with credit to the source. Statements and opinions expressed in the chapters are these of the individual contributors and not necessarily those of the editors or publisher. No responsibility is accepted for the accuracy of information contained in the Published articles. Publisher assumes no responsibility liability for any damage or injury to persons or property arising out of the use of any materials, instructions, methods or ideas contained inside. After this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in any publication of which they are an author or editor, and the make other personal use of the work, 1008 In-teh, www.in-tehorg Additional copies can be obtained from: [email protected] First published September 2008 Printed in Croatia A catalogue record for this book is available from the University Library Rijeka under no. 111225072 Travelling Salesman Problem, Edited by Federico Greco pe cm. ISBN 978-953-7619-10-7 1, Travelling Salesman Problem, Federico GrecoPreface In the middle 1930s computer science was yet a not well defined academic discipline. Actually, fundamental concepts, such as ‘algorithm’, or ‘computational problem’, has been formalized just some year before. In these years the Austrian mathematician Karl Menger invited the research community to consider from a mathematical point of view the following problem taken from the every day life. A traveling salesman has to visit exactly once each one of a list of m cities and then return to the home city. He knows the cost of traveling from any city i to any other city j. Thus, which is the tour of least possible cost the salesman can take? ‘The Traveling Salesman Problem (for short, TSP) was born. More formally, a TSP instance is given by a complete graph G on a node set V= {1,2,...mf, for some integer m, and by a cost function assigning a cost cj to the arc (i) , for any i, jin V. TSP is a representative of a large class of problems known as combinatorial optimization problems. Among them, TSP is one of the most important, since it is very easy to describe, but very difficult to solve. Actually, TSP belongs to the NP-hard class. Hence, an efficient algorithm for TSP (that is, an algorithm computing, for any TSP instance with m nodes, the tour of least possible cost in polynomial time with respect to m) probably does not exist. More precisely, such an algorithm exists if and only if the two computational classes P and NP coincide, a very improbable hypothesis, according to the last years research developments. From a practical point of view, it means that it is quite impossible finding an exact, algorithm for any TSP instance with m nodes, for large m, that has a behaviour considerably better than the algorithm which computes any of the (n-1)! possible distinct tours, and then returns the least costly one. If we are looking for applications, a different approach can be used. Given a TSP instance with m nodes, any tour passing once through any city is a feasible solution, and its cost leads to an upper bound to the least possible cost. Algorithms that construct in polynomial time with respect to m feasible solutions, and thus upper bounds for the optimum value, are called heuristics. In general, these algorithms produce solutions but without any quality guarantee as to how far is their cost from the least possible one. If it can be shown that the cost of the returned solution is always less than k times the least possible cost, for some real number k>1, the heuristic is called a k-approximation algorithm.vw Unfortunately, k-approximation algorithm for TSP are not known, for any k>1. Moreover, in a paper appeared in 2000, Papadimitriou, and Vempala have shown that a k- approximation algorithm for TSP for any 97/96>k>1 exists if and only if P=NP. Hence, also fincling a good heuristic for TSP seems very hard. Better results are known for NP-Hard subproblem of TSP. For example, a 3/2- approximation algorithm is known for Metric TSP (in a metric TSP instance the cost function verifies the triangular inequality). Anyway, the extreme intractability of TSP has invited many researchers to test new heuristic technique on this problem. The harder is the problem you test on, the more significant are the result you obtain. A large part of this book is devoted to some bio-inspired heuristic techniques that have been developed in the last years. Such techniques take inspiration from the nature. Actually, the animals that usually form great groups behave by instinct trying to satisfy the group necessity in the best possible way. Similarly, the natural systems develop in order to (locally) minimize their potential by finding a stationary point. In chapter 1 [Population-Based Optimization Algorithms for Solving the Travelling Salesman Problem] the following bio-inspired algorithmic techniques are considered: Genetic Algorithms, Ant Colon Optimization, Particle Swarm Optimization, Intelligent Water Drops, Artificial Immune Systems, Bee Colony Optimization, and Electromagnetism- like Mechanisms. Every section briefly introduces one of these techniques and an algorithm applying it for solving TSP. In the last section the obtained experimental results are compared. Chapter 2 [Bio-inspired Algorithms for TSP and Generalized TSP] is divided into two parts. In the first part, a new algorithm using the Ant Colon Optimization technique is considered. The obtained experimental results are then compared with other two algorithms using the same technique. In the second part, the combinatorial optimization problem called Generalized TSP (GTSP) is introduced, and a Genetic Algorithm for solving is proposed. We recall that a GSTP instance provides a complete graph G = (V,E), and a cost function (as in a TSP instance), together with a partition of the node set V into p subsets. A feasible solution for GTSP is a tour passing at least once from each one of the p subsets of V. Clearly, GTSP is a generalization of TSP. In Chapter 3 [Approaches to the Travelling Salesman Problem Using Evolutionary Computing Algorithms] an algorithm for TSP using the Genetic Local Search is considered. It is a hybrid technique, as it combines a genetic algorithm approach by a local search technique: As in a genetic algorithm the fitness of a population is the target, but a local search optimization phase is applied whenever a new individual is created during the evolutionary process. At the end of the chapter some experimental results are discussed. Chapter 4 [Particle Swarm Optimization Algorithm for the Traveling Salesman Problem] and Chapter 5 [A Modified Discrete Particle Swarm Optimization Algorithm for the Generalized Traveling Salesman Problem] deals with the Particle Swarm Optimization (PSO) technique. In a PSO algorithm the current solution is seen as a particle whose movement in the solution space is controlled by a certain velocity operator. As the solution space of a TSP instance is discrete, it is more correct referring to discrete PSO approach for TSP.vl In Chapter 4 the authors propose some velocity operators for a discrete PSO algorithm for TSP, and compare by computational experiments the results of the proposed approach with other known PSO heuristics for TSP. In Chapter 5 a discrete PSO approach is considered for Generalized TSP. Afterwards, the proposed algorithm is hybridized with a local search improvement heuristic. In the last section some the computational results compare the proposed algorithm, and its improvement with other known discrete PSO algorithm for GTSP. In Chapter 6 [Solving TSP via Neural Networks] and in Chapter 7 [A Recurrent Neural Network to Traveling Salesman Problem] Neural Network techniques for solving TSP are considered. In particular, Chapter 6 is devoted to the recent progress in the transiently chaotic neural network (TCNN), a discrete-time neural network model, are presented. An algorithm for TSP using such technique is then introduced, and the obtained results are compared with other neural networks algorithms. In Chapter 7 a technique based on the Wang's Recurrent Neural Networks with the “Winner Takes All” principle is used to solve the Assignment Problem (AP). By lightly modifying such technique, an algorithm for TSP is derived. Finally, some TSP instances taken from the TSP library are chosen for comparing the proposed algorithm with some other algorithms using different techniques. Chapter 8 [Solving the Probabilistic Travelling Salesman Problem Based on Genetic Algorithm with Queen Selection Scheme] treats an extension of TSP, the Probabilistic TSP (PTSP). A PTSP instance provides a complete graph G=(V,E), and a cost function (as in a TSP instance), together with a real number 0 < P;< 1 for each node i in V. P; represents the probability of the node i to be visited by a tour. Clearly, the goal of PTSP is to find a tour of minimal expected cost. In this chapter an optimization procedure based on a Genetic Algorithm framework is presented. In Chapter 9 [Niche Pseudo-Parallel Genetic Algorithms for Path Optimization of Autonomous Mobile Robot - A Specific Application of TSP] an application of TSP to the Path Optimization of Autonomous Mobile Robot is considered. An autonomous mobile robot has to find a non-collision path from initial position to objective position in an obstacle space trying to minimize the path cost. This problem can be modelled as a TSP instance. The authors consider a genetic algorithm, called Niche Pseudo-Parallel Genetic Algorithm, for solving TSP. The last Chapter [The Symmetric Circulant Traveling Salesman Problem] gives an example of a theoretical research on TSP. Actually, it is interesting to investigate if TSP becomes easier or remains hard (from a computational complexity point of view) when it is restricted to a particular class of graphs. In this chapter the case in which the graph in the instance is symmetric, and circulant is deeply analyzed, and an overview on the most recent resulls is given. By summing up, in this book the problem of finding algorithmic technique leading to good/ optimal solutions for TSP (or for some other strictly related problems) is considered. An important thing has to be outlined here. As already said, TSP is a very attractive problem for the research community. Anyway, it arises as a natural subproblem in many applications concerning the every day life. Indeed, each application, in which an optimal ordering of avill number of items has to be chosen in a way that the total cost of a solution is determined by adding up the costs arising from two successively items, can be modelled as a TSP instance. Thus, studying TSP can be never considered as an abstract research with no real importance. Itis time to start with the book. Enjoy the reading! September 2008 EditorContents Preface v . Population-Based Optimization Algorithms for Solving the Travelling 004 Salesman Problem Mohammad Reza Bonyadi, Mostafa Rahimi Azghadi and Hamed Shah-Hosseini Bio-inspired Algorithms for TSP and Generalized TSP 035 Zhifeng Hao, Han Huang and Ruichu Cai ‘Approaches to the Travelling Salesman Problem Using Evolutionary 063 Computing Algorithms Jyh-Da Wei . Particle Swarm Optimization Algorithm for the Traveling Salesman 075 Problem Elizabeth F. G. Goldbarg, Marco C. Goldbarg and Givanaldo R. de Souza . A Modified Discrete Particle Swarm Optimization Algorithm for the 097 Generalized Traveling Salesman Problem ‘Mehmet Fatih Tasgetiren, Yun-Chia Liang, Quan-Ke Pan and P. N. Suganthan Solving TSP by Transiently Chaotic Neural Networks 417 Shyan-Shiou Chen and Chih-Wen Shih . A Recurrent Neural Network to Traveling Salesman Problem 135 Paulo Henrique Siqueira, Sérgio Scheer, and Maria Teresinha Ams Steiner Solving the Probabilistic Travelling Salesman Problem Based on 157 Genetic Algorithm with Queen Selection Scheme YurHsin Liu9. Niche Pseudo-Paralle! Genetic Algorithms for Path Optimization of 173 Autonomous Mobile Robot - A Specific Application of TSP Zhihua Shen and Yingkai Zhao 10. The Symmetric Circulant Traveling Salesman Problem 181 Federico Greco and Ivan GeracePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem Mohammad Reza Bonyadi, Mostafa Rahimi Azghadi and Hamed Shah-Hosseini Department of Electrical and Computer Engineering, Shahid Beheshti University, Tehran, Iran 1. Introduction The Travelling Salesman Problem or the TSP is a representative of a large class of problems known as combinatorial optimization problems. In the ordinary form of the TSP, a map of cities is given to the salesman and he has to visit all the cities only once to complete a tour such that the length of the tour is the shortest among all possible tours for this map. The data consist of weights assigned to the edges of a finite complete graph, and the objective is to find a Hamiltonian cycle, a cycle passing through all the vertices, of the graph while having the minimum total weight. In the TSP context, Hamiltonian cycles are commonly called tours. For example, given the map shown in figure I, the lowest cost route would be the one written (A, B,C, E, D, A), with the cost 31. © Fig. 1. The tour with A=>B =>C =>E =>D => A is the optimal tour. In general, the TSP includes two different kinds, the Symmetric TSP and the Asymmetric TSP. In the symmetric form known as STSP there is only one way between two adjacent cities, ie., the distance between cities A and B is equal to the distance between cities B and A (Fig. 1). But in the ATSP (Asymmetric TSP) there is not such symmetry and it is possible to have two different costs or distances between two cities. Hence, the number of tours in the ATSP and STSP on 1 vertices (cities) is (7-1)! and (n-1)/2, respectively. Please note that the graphs which represent these TSPs are complete graphs. In this chapter we mostly consider the STSP. It is known that the TSP is an NP-hard problem (Garey é& Johnson, 1979) and is often used for testing the optimization algorithms. Finding Hamiltonian cycles or traveling2 Travelling Salesman Problem salesman tours is possible using a simple dynamic program using time and space O(2" n0), that finds Hamiltonian paths with specified endpoints for each induced subgraph of the input graph (Eppstein, 2007). The TSP has many applications in different engineering and optimization problems. The TSP is a useful problem in routing problems eg. in a transportation system, There are different approaches for solving the TSP. Solving the TSP was an interesting problem during recent decades. Almost every new approach for solving engineering and optimization problems has been tested on the TSP as a general test bench. First steps in solving the TSP were classical methods. These methods cor of heuristic and exact methods. Heuristic methods like cutting planes and branch and bound (Padherg & Rinaldi, 11987), can only optimally solve small problems whereas the heuristic methods, such as 2-opt (Lin & Kernighan, 1973), 3-opt, Markov chain (Martin et al., 1991), simulated annealing (Kirkpatrick et al, 1983) and tabu search are good for large problems. Besides, some algorithms based on greedy principles such as nearest neighbour, and spanning tree can be introduced as efficient solving methods. Nevertheless, classical methods for solving the TSP usually result in exponential computational complexities. Hence, new methods are required to overcome this shortcoming. These methods include different kinds of optimization techniques, nature based optimization algorithms, population based optimization algorithms and etc. In this chapter we discuss some of these techniques which are algorithms based on population. Population based optimization algorithms are the techniques which are in the set of the nature based optimization algorithms. The creatures and natural systems which are working and developing in nature are one of the interesting and valuable sources of inspiration for designing and inventing new systems and algorithms in different fields of science and technology. Evolutionary Computation (Eiben & Smith, 2003), Neural Networks (Haykin, 99), Time Adaptive Self-Organizing Maps Ghah-Hosseini, 2006), Ant Systems (Dorigo & Stutzle, 2004), Particle Swarm Optimization (Eberhart & Kennedy, 1995), Simulated Annealing, (Kirkpatrik, 1984), Bee Colony Optimization (Teodorovic et al., 2006) and DNA Computing (Adleman, 1994) are among the problem solving techniques inspired from observing nature. In this chapter population based optimization algorithms have been introduced. Some of these algorithms were mentioned above. Other algorithms are Intelligent Water Drops (IWD) algorithm (Shah-Hosseini, 2007), Artificial Immune Systems (AIS) (Dasgupta, 1999) and Electromagnetism-like Mechanisms (EM) (Birbil & Fang, 2003). In this chapter, every section briefly introduces one of these population based optimization algorithms and applies them for solving the TSP. Also, we try to note the important points of each algorithm and every point we contribute to these algorithms has been stated. Section nine shows. experimental results based on the algorithms introduced in previous sections which are implemented to solve different problems of the TSP using well-known datasets, 2. Evolutionary algorithms 2.4 Introduction Evolutionary Algorithms (EAs) imitates the process of biological evolution in nature. These are search methods which take their inspiration from natural selection and survival of the fittest as exist in the biological world. EA conducts a search using a population of solutions. Each iteration of an EA involves a competitive selection among all solutions in thePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 3 population which results in survival of the fittest and deletion of the poor solutions from the population. By swapping parts of a solution with another one, recombination is performed and forms the new solution that it may be better than the previous ones. Also, a solution can be mutated by manipulating a part of it. Recombination and mutation are used to evolve the population towards regions of the space which good solutions may reside. Four major evolutionary algorithm paradigms have been introduced during the last 50 years: genetic algorithm is a computational method, mainly proposed by Holland (Holland, 1975). Evolutionary strategies developed by Rechenberg (Rechenberg, 1965) and Schwefel (Schwefel, 1981). Evolutionary programming introduced by Fogel (Fogel et al., 1966), and finally we can mention genetic programming which proposed by Koza (Koza, 1992). Here we introduce the GA (Genetic Algorithm) for solving the TSP. At the first, we prepare a brief background on the GA. 2.2 Genetic algorithms Genetic Algorithms focus on optimizing general combinatorial problems. GAs have long been studied as problem solving tools for many search and optimization problems, specifically those that are inherent in NP-Complete problems. Various candidate solutions are considered during the search procedure in the system, and the population evolves until a candidate solution satisfies the predefined criteria. In most GAs, a candidate solution, called an individual, is represented by a binary string (Goldberg, 1989) ie. a string of 0 or 1 elements. Each solution (individual) is represented as a sequence (chromosome) of elements (genes) and is assigned a fitness value based on the value given by an evaluation function. The fitness value measures how close the individual is to the optimum solution. A set of individuals constitutes a population that evolves from one generation to the next through the creation of new individuals and deletion of some old ones. The process starts with an initial population created in some way, e.g. through a random process. Evolution can take ‘two forms: Crossover: Two selected chromosomes can be combined by a crossover operator, the result of which will replace the lowest fitness chromosome in the population. Selection of each chromosome is performed by an algorithm to ensure that the selection probability is proportional to the fitness of the chromosome. A new chromosome has the chance to be better than the replaced one. The process is oriented towards the sub-regions of the search space, where an optimal solution is supposed to exist (Goldberg, 1989). Mutation: In mutation process, a gene from a selected chromosome is randomly changed. This provides additional chances of entering unexplored sub-regions. Finally, the evolution is stopped when either the goal is reached or a maximum CPU time has been spent (Goldberg, 1989). In the following the GA operation pseudo code has been written: 1. Start 2. Population initialization 3. Repeat until (satisfying termination criteria) * Selection * Cross over + Mutation4 Travelling Salesman Problem * Making new population with the fittest solutions + Evaluation * Checking the termination criterion 4, Take the best solution as output 5. End 2.3 Solving the TSP using GA As mentioned earlier, the TSP is known as a classical NP-complete problem, which has extremely large search spaces and is very difficult to solve (Louis & Gong, 2000). Hence, classical methods for solving TSP usually result in exponential computational complexities. These methods consist of heuristic and exact methods. Heuristic methods like cutting planes and branch and bound (Padherg & Rinaldi, 1987), can only optimally solve small problems while the heuristic methods, such as 2-opt (Lin & Kernighan, 1973), 3-opt, Markov chain (Martin et al., 1991), simulated annealing (Kirkpatrick et al., 1983) and tabu search are good for large problems. Besides, some algorithms based on greedy principles such as nearest neighbour, and spanning tree can be used as efficient solving methods. Nevertheless, because of the tremendous number of possible solutions and large search spaces, GAs seem to be wise approaches for solving the TSP especially when they are accompanied with carefully designed genetic operators (Jiao & Wang, 2000). GAs search the large space of solutions toward best answer and the operators can help the search process become faster and also they prepare the ability to avoid being trapped in local optima. In recent years, solving the TSP using evolutionary algorithms and specially GAs has attracted a lot of attention. Many studies have been performed and researchers try to contribute to different parts of solving process. Some of researchers pose different forms of GA operators (Yan et al., 2005) in comparison to the former ones and others attempt to combine GA with other possible approaches like ACO (Lee, 2004), PSO and ete. In addition, some authors implement a new evolutionary idea or combine some previous algorithms and idea to create a new method (Bonyadi et al., 2007). Here we investigate some of these works and compare their results. Due to the spread of related works we can not mention all of them here. But The reader is referred to the prepared references for further information. In all of the performed works, two instances are mentionable. First: all of the proposed algorithms work toward finding the nearest answer to the best solution. Second: solving the TSP in a more little time is a key point in this problem because of its special application which require, finding the best feasible answer fast. In (Bonyadi et al., 2007), the authors made some changes to two previous local search algorithms ie. the Shuffled Frog Leaping (SFL) and the Civilization and Society (CS) and combined these two algorithms with the GA idea. In this study, as it is common in a conventional GA, at first the elements of the population perform mutation or crossover in random order. Then for every element of this population, a local search algorithm, which is a mix of both SFL and CS, is performed. The results demonstrate significant improvements in terms of time complexity and reaching better solutions in comparison to the GAs which apply only SFL or CS in their usual forms. Hence, the main contribution in this work is combining two previous search methods and using them with the GA, simultaneously. The evaluation results of the proposed algorithm have been prepared in section nine. In another work (Yan et al., 2005) a new algorithm based on Inver-over operator, for combinatorial optimization problems has been proposed. Inver-over is based on simplePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 5 inversion; however, knowledge taken from other individuals in the population influences its action. In this algorithm some new strategies including selection operator, replace operator and some new control strategy have been applied. The results prove that these changes are very efficient to accelerate the convergence. A consequence, it is inferred that, one of the points for contribution is operators. Suitable changes in the conventional form of operators might lead to major differences in the search and optimization procedure. Through the experiments, GAs are global search algorithms appropriate for problems with huge search spaces. In addition, heuristic methods can be applied for search in local areas. Hence, combination of these two search algorithms can result in producing high quality solutions. Cooperation between Speediness of local search methods in regional search and robustness of evolutionary methods in global search can be very useful to obtain the global optimum. Recently, (Nguyen et al, 2007) proposed a hybrid GA to find high-quality solutions for the TSP. The main contribution of this study is to show the suitable combination of a GA as a global search with a heuristic local search which are very promising for the TSP. In addition, the considerable improvements in the achieved results prove that the effectiveness and efficiency of the local search in the performance of hybrid GAs. Through these results, one of other points where it can be kept in mind is the design of the GA in a case that it balances between local and global search. Moreover, many other studies have been performed that all of them combine the local and global search mechanisms for solving the TSP. As mentioned earlier, one of the points that solving the TSP can contribute is recombination operators i.e. mutation and crossover. Based on (Takahashi, 2005) there are two kinds of crossover operators for solving the TSP. Conventional encoding of the TSP which is an array representation of chromosomes where every element of this array is a gene that in the TSP shows a city. The first kind of crossover operator corresponds to this chromosome structure. In this operator two parents are selected and with exchanging of some parts in parents the children are reproduced. The second type performs crossover operation with mentioning epistasis. In this method it is tried to retain useful information about links of parent's edges which leads to convergence. Also, in (Tsai et al., 2004) another work on genetic operators has been performed which resulted in good achievements. 3. Ant colony optimization (ACO) 3.4 Introduction The ACO (Ant Colony Optimization) heuristic is inspired by the real ant behaviour (figure 2) in finding the shortest path between the nest and the food (Beckers et al., 1992). This is achieved by a substance called pheromone that shows the trace of an ant. In its searching the ant uses heuristic information which is its own knowledge of where the smell of the food comes from and the other ants’ decision of the path toward the food by pheromone information (Holldobler & Wilson, 1990). Fig. 2. Real ant behaviour in finding the shortest path between the nest and the food6 Travelling Salesman Problem In fact the algorithm uses a set of artificial ants (individuals) which cooperate to the solution of a problem by exchanging information via pheromone deposited on graph edges. The ACO algorithm is employed to imitate the behaviour of real ants and is as follows: Initialize Loop Each ant is positioned on a starting node Loop Each ant applies a state transition rule to incrementally build a solution and a local pheromone updating rule Until all ants have built a complete solution A global pheromone updating rule is applied Until end condition 3.2 State transition Consider n is the city amount; m is the quantity of the ants in an ACO problem; dij is the length of the path between adjacent cities i and j; rj (t) is the intensity of trail on edge (i ) at time t . At the beginning of the algorithm, an initialization algorithm determines the ants positions on different cities and initial value tj (0), a small positive constant c for trail intensity are set on edges. The first element of each ant's tabu list is set to its starting cit The state transition is given by equation 1, which ant k in city i chooses to move to city j: (yO yo” it / allowed he org’ w kealiowedy 0, otherwise where allowed = {N-tabuj}, which is the set of cities that remain to be visited by ant k positioned on city i (to make the solution feasible) « and are parameters that determine the relative importance of trail versus visibility, and 1)= 1/d is the visibility of edge (i) 3.3 Trial updating In order to improve future solutions, the pheromone trails of the ants must be updated to reflect the ant’s performance and the quality of the solutions found. The global updating rule is implemented as follows. Once all ants have built their tours, pheromone is updated onall edges according to the following formula (equations 2 to 4) mt THD = pry + Bary 2 where Q 1 _|2, itedgecs, pissed by thekthamatcuneneyste ach = 4 by ® 0, otherwisePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem Zz wok ary = Sah ® p (0). I ay which computes the time taken to travel such that time svel! max(e,,, vel from city i to city j with the velocity vel!” . Here, the function e(.) represents the two14 Travelling Salesman Problem dimensional positional vector for the city. The function max(.,.) returns the maximum value among its arguments, which is used here to threshold the negative velocities to a very small positive number e,, = 0.0001. 8. Foreach IWD, update the soil of the path traversed by that IWD using equation 10. soil(i, J) = (\~ p).soil(i, f) ~ p.Asoil(i, j) ,) Vy (10) soit” = soit! + asoil(i, j) WD ; , . where soil!” represents the soil that the IWD carries. The IWD goes from city i to city j. ‘The parameter pis a small positive number less than one. Here we use p = 0.9. 9. For each WD, complete its tour by using steps 4 to 8 repeatedly. Then, calculate the length of the tour traversed by the IWD, and find the tour with the minimum length among all IWD tours in this iteration. We denote this minimum tour by Ty - 10. Update the soils of paths included in the current minimum tour of the IWD, denoted by Ty which is computed based on equation 11 2.soit” a j Wi ET y a) “r= TI. 11. If the minimum tour 7\, is shorter than the best tour found so far denoted by Ty soil(i, j) = (I~ p).soil(i, j)~ p. then we update the best tour by applying equation 12. p= Ty and LenTp) = LenTyy) (2) 12. Go to step 2 unless the maximum number of iterations is reached or the defined termination condition is satisfied. 13. The algorithm stops here such that the best tour is kept in 7p and its length is Len(’g) It is reminded that it is also possible to use only Ty and remove step 11 of the IWD algorithm. However, it is safer to keep the best tour 7 of all iterations than to count on only the minimum tour T\, of the last iteration. The IWD algorithm is experimented by artificial and some benchmark TSP environments. The proposed algorithm converges fast to optimum solutions and finds good and promising results. This research (Shah-Hosseini, 2007) is the beginning of using water drops ideas to solve engineering problems. So, there is much space to improve and develop the IWD algorithm. 6. Arti 6.1 Introduction Recently, there was an increasing interest in the area of Artificial Immune System (AIS) and its application for solving various problems specifically for the TSP (Zeng & Gu, 2007), (Lu jal immune systemsPopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 15 et al., 2007). AIS is inspired by natural immune mechanism and uses immunology idea in order to develop systems capable of performing different tasks in various areas of research such as pattern recognition, fault detection, diagnosis and a number of other fields including, optimization. Here we want to know the AIS completely. To start with, it might be useful to become more familiar with natural immune system, Natural immune systems consist of the structures and processes in the living body that provide a defence system against invaders and also altered internal cells which lead to disease. In a glance, immune system's main tasks can be divided into three parts; recognition, categorization and defence. As recognition part, the immune system firstly has to recognize the invader and foreign antigens e.g. bacteria, viruses and etc. After recognition, classification must be performed by immune systems, this is the second part. And appropriate form of defence must to be applied for every category of foreign aggressive phenomenon as the third part. The most significant aspect of the immune systems in mammals is learning capability. Namely, the immune systems can grow during the life time and is capable of using learning, memory and associative retrieval in order to solve mentioned recognition and classification tasks. In addition, the studies show that the natural immune systems are useful phenomena in information processing and can be helpful in inspiration for problem solving and various optimization problems (Keko et al., 2003). 6.2 Artificial immune system Like the natural immune systems the AIS is a set of techniques, which try to algorithmically mimic natural immune systems’ behaviour (Dasgupta, 99). As mentioned earlier, the immune system is susceptible to all of the invaders, also the outer influences, like vaccines which are artificial ways of raising individual's immunity. Vaccines are other factors that can stimulate the immune system’s susceptibility. This feature is the key point of the AIS structure. The vaccines in the AIS are abstracted forms of the preceding information. Vaccination modifies genes based on the useful knowledge of the problem to achieve higher fitness in comparison to the fitness that obtained from a random process when for example a classical GA is applied. Once again it is necessary to point out that, vaccines contain some important information about the problem and in consequence the vaccination process employed in a right manner can be very useful in the performance of the algorithm. Like classical GA and based on its structure the AIS can work. The GA operators (crossover and mutation) search the problem space randomly and hence they don’t have enough capability of meeting the actual problem at the local level. GAs are known as incapable of search fine local tuning because they are global search algorithms. Immune method through vaccination tries to overcome such blindness of crossover and mutation (Keko et al., 2003). After vaccination, the immune method might leads to deterioration. This case happens when vaccination leads to smaller fitness values than previous ones. Hence, another important part of immune algorithm is prevention of deterioration when inserting vaccine. In short, immune operators perform in four steps: firstly, an individual is selected, randomly. Now as the second step, the vaccine is inserted at the individual's randomly chosen place. Vaccine insertion might leads to deterioration, the third step is checking for deterioration. And finally the forth step is discarding every individual that shows degeneration right after vaccine. This way of checking could be dangerous for diversity and could result in algorithm's inability to avoid local optima, especially when combined with small populations. The studies show that the use of immune systems resulted in faster16 Travelling Salesman Problem convergence when population is large enough and diversity is secured. The combination of immune algorithm and GA, form the immune genetic algorithm (IGA). Many of previous works that are performed on the TSP used IGA. Now, we first investigate the IGA and its structure in detail and after that we have a look at some previous works around the TSP. In summery, the IGA consists of these steps: 1. Creation of initial population in some way, e.g. through a random process Abstract vaccines according to the former information Checking the termination criterion (if it is satisfied go to step 10 and else go to next step) Crossover on the randomly selected individuals Mutation on the produced children Vaccination on the former step outcome Deterioration checking Discarding every individual that shows degeneration right after vaccine 9. Gotostep3 10. End As it is obvious from the ten steps which have been mentioned above, the IGA is very similar to the conventional GA, but they are different in operators. The IGA has vaccine operator to overcome the universality problem of the conventional GA. For more information on IGA you can refer to many cited papers which are prepared at the end of this chapter. SN ATELY 6.3 Solving the TSP using AIS and IGA The first work in investigating potential application of the immune system in solving numerical optimization problems was the study by Bersini and Varela (Bersini & Varela, 90), who proposed immune employment mechanism. After that, many studies have been performed that focus on the AIS and IGA. Also, the IGA and AIS have been applied for solving the TSP in many cases. In (Jiao & Wang, 2000) the IGA and its parts have been introduced in detail and the IGA has been shown as an algorithm that accomplished in two steps: 1) a vaccination and 2) an immune selection. These phases are completely similar to that we mentioned about IGA and AIS in this section. In the mentioned paper, it is proved that the IGA theoretically converges with probability one. Besides, strategies and methods of selecting vaccines and constructing an immune operator are also given. Also, the IGA has been applied to the TSP and the results which are presented in this study illustrate that IGA is able to restrain the degenerate phenomenon effectively during the evolutionary process and can improve the searching ability, adaptability and greatly increase the converging speed. Recently, some works have been performed on the TSP which employ IGA. In (Zeng & Gu, 2007), a novel genetic algorithm based on immunity and growth for the TSP is presented. In this paper at first, a reversal exchange crossover and mutation operator is proposed which lead to preservation of the good sub-tours and making individuals various. At the next part, a new immune operator is proposed to restrain individuals’ degeneracy. In addition, a novel growth operator is proposed to obtain the optimal solution with more chances. Results and investigations that performed in this study show that the algorithm is feasible and effective as it is claimed. In addition, in another study (Lu et al., 2007), a modified immune genetic algorithm is applied to solve the Travelling Salesman Problem This method called an improved IGA by its authors. In this paper, at first, a new selection strategy is incorporated into the conventional genetic algorithm to improve the performancePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 17 of genetic algorithm. Besides the authors changed the selection strategy and in a new form it includes three computational procedures: evaluating the diversity of genes, calculating the percentage of genes, and computing the selection probability of genes. Based on the prepared results it is inferred that, by incorporating inoculating genes into conventional procedures of genetic algorithm, the number of evolutional iterations to reach an optimal solution can be significantly reduced and in consequence it results in faster answer in comparison to conventional IGA. In addition to the mentioned works, the biological immune idea can be combined with other population based optimization algorithms which all of them are prepared in this chapter. As an instance, the paper (Qin et al., 2006) proposes a new diversity guaranteed ant colony algorithm by adopting the method of immune strategy to ant colony algorithm and simulating the behaviour of biological immune system. This method has been applied to the TSP benchmarks and results show that the presented algorithm has strong capability of optimization; it has diversified solutions, high convergence speed and succeeds in avoiding the stagnation and premature phenomena. Based on the performed studies some points can be inferred as mentioned in the following (Keko et al., 2003): The simulation results show that the variation in population size has the same effect on the GA and IGA. In both of the mentioned techniques, large population sizes require more generation to achieve higher fitness, resulting in relatively slow rate of convergence. Hence new ideas are required for faster convergence. Some of these new ideas had been presented in some works as you see in some investigated papers. Also, based on the simulation results, the running time of the IGA and the regular GA do not have large differences, since in the IGA all the vaccines are determined before the algorithm starts and when they are required they can be loaded from a look up table. Combining immune operator with another local improving operator can be an additional idea for getting better answers from the IGA, One of the advantages of the IGA over the plain GA is that itis less susceptible to changing control parameters such as crossover or mutation probability. The simulation results demonstrate that changing these parameters has slight influence to the overall performance. It is worth mentioning that more studies and attentions in the AIS and IGA are employing other AIS features like adaptive vaccine selection. 7. Bee colony optimization 7.1 Introduction Similar to other natural inspired and collective intelligence based algorithms such as PSO which is taken from the bird’s life and ACO based on the ant colony social life, another kind of artificial intelligence systems that can be useful in solving many engineering, management, control and computational problems, is an algorithm inspired from Bee colonies in nature. The Bee Colony Optimization (BCO) algorithms are interesting metaheuristic algorithms that represent another direction in the field of swarm intelligence. Here, firstly we introduce the bee system and bee colony optimization briefly and then some recent works on the TSP which have used bee systems are investigated. 7.2 Bee colony optimization The bee colony’s function according to nature is as follows. At first, each bee belonging to a colony looks for the feed individually. When a bee finds the feed, it informs other bees by18 Travelling Salesman Problem dancing, Other bees collect and carry the feed to the hive. After relinquishing the feed to the hive, the bee can take three different actions, 1. Abandon the previous food source and become again uncommitted follower. 2. Continue to forage at the food source without recruiting the nestmates. 3. Dance and thus recruit the nestmates before the return to the food source. With a certain probability that is dependent on the obtained feed quality, its distance from the hive and the number of the bees which are now engaged with this feed resource, a bee selects one of the stated actions and follows its work in a similar repetitive form (Todorovic & Dell'Orco, 2005). This behaviour can be applied to many complicated engineering problems including computational, control, optimization, transportation, etc. In the following we study such a method that focuses on the TSP solving. 7.3 BCO applic ‘The BCO algorithm can be a significant method in local search applications. One of the most primary works on the bees and their life is (Gato & Hagiwara, 97). In this study, the authors applied bee system along with GA and introduced a modified and improved form of the conventional GA. Based on this fact that the regular GA lacks the global search ability; the improvement is regarding to overcome this shortcoming. Hence, a new GA inspired by the bee colony's function has been presented, the authors called it, bee system. The main Purpose of this modified GA (bee system) is to improve the local search ability of GAs without degrading the global search ability. In the proposed bee system, firstly global search is performed using the simple GA structure. Through this global search step, some chromosomes with reasonable high fitness produced which are called superior chromosome. These superior chromosomes are kept for the local search procedure and each of them corresponds to a local population. At the beginning of the local search all of the chromosomes in each local population make couple (cross over) with its population superior chromosome. This crossover is named concentrated crossover which tries to search concentratedly around the related superior chromosome. Another difference between the bee system and ordinary GA is migration among the population. In this method, the bee system selects one individual per predetermined generation, and transfers it to the neighbouring population which is called migration. Using this migration technique, each population tries to search independently and cooperatively. Moreover, for a more effective search, a simplified Simplex Method named Pseudo-Simplex Method is introduced and employed in the proposed bee system. All of the mentioned operators are in the local search part. After passing the predetermined generations, the local search stops. If the best solution found so far does not suffice the ending condition, the global search starts again and the algorithm is repeated (Sato & Hagiwara, 97). It was a kind of application based on the bee colony’s function which is used to solve the TSP. Simulation results depict that the introduced method has a good potential to solve the TSP and other complicated problems. 7.4 Solving the TSP using BCO Another study around bee colony and its applications is a work performed for transportation modelling with focus on artificial life (ALife) approach (Lucic & Teodorovic, 2002). This paper shows that the ALife models that have been developed for solving complex transportation problems are inspired by social insect’s behavior. Interaction between individual insects in a colony of social insects has been well documented. ThePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 19 examples of such interactive behavior are bee dancing during the food procurement, ants’ pheromone secretion, and performance of specific acts which signal the other insects to start performing the same actions. Based on these studies we can construct the artificial systems such as bee systems. In the mentioned study, the artificial bee system has been applied to solve the TSP. Assume that, the graph in which the traveling salesman route should be discovered is shown by G = (N, A) that N= nodes (cities) and A= links connecting these nodes. This graph can correspond to the network that the artificial bees are collecting nectar. The hive can also be placed randomly in one of the network's nodes. For solving the TSP using the bee system it is necessary that two parameters correspond to each others, tour length and nectar quantity. Here, it is assumed that the nectar quantity that is possible to collect flying along a certain link is inversely proportional to the link length. In other words, the shorter the link, the higher the nectar quantity along that link. The artificial bees collect the nectar during the predetermined time interval. After that, the hive position is changed randomly and artificial bees start to collect the nectar from the new hive location. Each iteration is composed of a certain number of stages. The stage is an elementary time unit in the bees’ environment. During one stage the artificial bee will visit nodes, create partial traveling salesman tour, and after that return to the hive (the number of nodes to be visited within one stage is prescribed by the analyst at the beginning of the search process). In the hive the bee will participate in a decision making process. The artificial bee will decide whether to abandon the food source and become again an uncommitted follower, continue to forage at the food source without recruiting nestmates, or dance and thus recruit nestmates before returning to the food source (Lucic & Teodorovic, 2002). During any stage, bees are choosing nodes to be visited in a random manner. The randomness in not useful here and the mentioned paper’s authors assumed that the probability of choosing node j by the k-th bee, located in node i (during stage u +1 and iteration z) equals to equation 13: -ad; y nr) @ remax(z-bl) ston “ay i= gy (ez) FEN, (2), Vhot 2 o 5 (wh snl) ty) . r=max(z-b.)) TEN ya. 0, otherwise Where: j- Node indexes (i, j= 1,2, ..., N), di - Length of link (i,j), k - Bee index (k = 1, 2, ..., B), B - The total number of bees in the. z- Iteration index (z = 1, 2, ..., M), M - Maximum number of iteration, u-Stage index (u = 1,2,...|(s0/s]), - Number of nodes visited by every artificial bee during one stage,20 Travelling Salesman Problem 1ny(r) - total number of bees that visited link (i, 1) in r-th iteration, b - Memory length, gx(u, 2) - Last node that bee k visits at the end of stage u in iteration z, Ni(u, 2) ~ Set of unvisited nodes for bee k at stage u in iteration z (in one stage bee will visits s nodes; we have |Nk(u, z) | = |N] - us), a- Input parameter given by analyst. This equation is based on some simple rules in solving the TSP using the bee system. These rules have been prepared as follows: The greater distance between nodes i and j leads to the lower probability that the k-th bee located in the node i will choose node j during stage u and iteration z. The greater number of iterations (z) makes the higher influence of the distance. In other words, at the beginning of the search process, artificial bees have “more freedom of flight”. It means that, the bees have more chance to search the entire solution space. But when more iterations have been performed the bees have less freedom to explore the solution space such as the search at first, because, near the end of the search process, with a high probability the solution is in our neighbourhood. Probability of selecting a new link by a bee is related to the total number of the last bees which had been visited this link, before this. The greater total number of bees results in a higher probability of choosing that link in the future. All of the above mentioned points have been employed in the equation 13. Another important point in this problem is the bee decision about the following of the search process. After relinquishing the food, the bee is making a decision about abandoning the food source or continuing the foraging at the food source. It is assumed that every bee can obtain the information about nectar quantity collected by every other bee. The probability that, at the beginning of stage u + 1, bee k will use the same partial tour that is defined in stage u in iteration z is equal to the following (equation 14): Lg (u.z)-min( Lp, . rewlu. a4 pyuttz)=e Where Ly (u, 2) is the length of partial route that is discovered by bee k in stage u in iteration z. Based on equation 14 if a bee has discovered the shortest partial travelling salesman tour in stage u in iteration z, the bee will fly along the same partial tour with the probability equal to one. Besides, the longer tour has the smaller chance to choose based on this equation. For having a global search it is better that the individual bees have interaction with each others. ‘To follow this purpose the probability of that the artificial bee continues foraging at the food source without recruiting nestmates is tuned to a very low value and hence the probability of that the bee flies to the dance floor and dance with other bees becomes low. In other words, when at the beginning of a new stage, the bee does not follow the previous partial travelling salesman tour, it will follow other bees and interacts to their dancing, But the bee must select one of the advertised dancing arenas (partial travelling salesman tour) in the dancing area, and hence another selection must be performed. This selection can be carried out in terms of two conditions: 1) the length of that partial tour and 2) the number of bees which are engaged in that partial tour. It is clear that the selection can be done based on thePopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 24 smaller tour length and also the greater number of bees. Based on these conditions the authors prepare a relation as it is shown in equation 15, where: P.@ ~ Parameters given by the analyst, a (u,z)~ The normalized value of the partial route length Bgla Y(u, 2) - The set of partial tours that were visited by at least one bee. The normalized value of the number of bees advertising the partial tour, PP eit Haga ) Pelt. (15) Ze re¥(u.z) As it is shown in the mentioned work (Lucic & Teodorovic, 2002), this bee system has been tested ona large number of well known test benches such as Eil51.tsp, Berlind2.tsp, St70.tsp, Pr76.tsp, Kroal00.tsp, Bil101.tsp, Tsp225.tsp and A280.tsp. Also, for improving the results in each step, the 2-opt or 3-opt algorithms have been applied. The results reveal that the mentioned method is very efficient. In all instances with less than 100 nodes, the bee system achieves the optimal solution and in the large cases it has a significant improvement in comparison to the other prevalent methods. The simulation results have been organized in section nine. One of the recent work for solving the TSP using bee's behaviour and BCO algorithm is (Teodorovic et al., 2006). In this paper the authors propose the Bee Colony Optimization Metaheuristic (BCO). Moreover, this study, describes two BCO algorithms that the authors call them, the Bee System (BS) and the Fuzzy Bee System (FBS). In the case of EBS the agents (artificial bees) use approximate reasoning and rules of fuzzy logic in their communication and acting. In this way, the FBS is capable to solve deterministic combinatorial problems, as well as combinatorial problems characterized by uncertainty. In this paper, The BCO as a new computational paradigm is described in detail at first. After that the TSP as a case study has been solved using the proposed bee system. The proposed bee system is similar to that had been seen in the previous investigated study but in this paper the BCO algorithm has been described completely. For further information about the BCO algorithm please refer to the related resources prepared at the end of the chapter. 8. Electromagnetism 8.1 Introduction ‘The Electromagnetism-like mechanism is a heuristic that was introduced by (Birbil & Fang, 2003). The method utilizes an attraction-repulsion mechanism to move the sample points towards the optimality. In other words, EM simulates the attraction-repulsion mechanism of electromagnetism theory which is based on Coulomb's law. The main concentration of the first introduction of this heuristic was on the problems with bounded variables on the form equal to equation 16. Min(£(x)) st. x €[1u] (16)22 Travelling Salesman Problem where / and w are defined as the following form (equation 17): [u]=xe &, |), 0.5 then Ye~ yi Jo(Length) else ue y~Jo(Length) end if if f(y) 0 then % ag eayl + FY (tuk - 34) & else % oa cay + il Ol = hd) 10: end if m end for 12 endif 13: end for26 Travelling Salesman Problem 8.6 Termination criteria There are 2 termination criteria introduced by Birbil et al. for electromagnetism as follow: 1. Maximum number of iterations. They claim that in general, 25 iterations per dimension (ie., MAXITER=25n) is satisfactory for converging to the optimum point for moderate difficulty functions. 2. Successive number of iterations spent without changing the current best point. In other word, if the current best point is not improved for certain number of iterations, the algorithm may be stopped. 8.7 Solving the TSP using EM-like mechanism One of the most attractive approaches for solving TSP using EM is cited in (Wu et al., 2006). In this study, a hybrid algorithm based on EM and K-OPT is introduced. They used a revised EM-like algorithm which proposed by (Birbil & Fang, 2005). In this version of EM, a parameter v belong to (0, 1) is introduced. The perturbed point x is selected as the farthest point from the current best point, Xt, in the current population. The calculation of the total force vector remains the same for all points except xj. For xp, the component forces are perturbed by a random number 4, where \ is uniformly distributed between 0 and 1. The directions of the component forces are perturbed; that is, if the random variable 2 is less than the parameter v then the direction of the component force is reversed. Besides, they use a formulation for calculating the forces which proposed in (Maenhout & Vanhoucke, 2005) for solving the Nurse Scheduling problem. As we know, TSP is an integer value problem but the EM algorithm works in real valued problems (continues space). This problem makes the transformation very significant. In the proposed approach in (Wu et al., 2006), one of the well-known algorithms (Random Key (RK)) for transforming the continuous domain into the discrete domain has been used. The concept of RK technique is simple and can be applied easily. When we obtain a k-dimensional solution, we sort the value corresponding to each dimension. Any sorting algorithm can be used in the method. The indices of the sorted list will be the solution in discrete space. By applying the RK algorithm, any continuous algorithm like EM will be able to work in a discrete space. 9. Experimental results In this section some results of discussed population based methods for solving the TSP have been prepared. At each subsection the mentioned algorithms and studies based on the some cited paper have been compared. 9.1 Evolutionary algorithms The first study that has been cited in section 2 was (Bonyadi et al., 2007). In this work some changes to two previous local search algorithms i.e. Shuffled Frog Leaping (GFL) and Civilization and Society (CS) have been made and these algorithms are combined with the GA idea. The shown results illustrate that the mentioned hybrid algorithm has better results in comparison to the GA using the SFL method. The results have been shown in Table 1. In another work (Yan et al., 2005) a new algorithm based on Inver-over operator, for combinatorial optimization problems has been proposed. The shown results prove that these changes are very efficient to accelerate the convergence speed. As a consequence, it isPopulation-Based Optimization Algorithms for Solving the Travelling Salesman Problem 27 inferred that, one of the points for contribution is operators. Suitable changes in the conventional form of operators might lead to major differences in search and optimization procedure. The mentioned results have been prepared in Table 2. Algorithm Average path value for 80 point input (million) GA 26 GA using SFL method 19 GA using Proposed approach ty Exact solution 10 ‘Table 1. (Bonyadi et al, 2007) simulation results Instance | Result in TSBLIB | Optimum in TSBLIB | Results by (Yan et al., 2005) Eil76 538 345.387, 544.369) Pr136 96772 9672 96770.924 Table 2. (Yan et al., 2005) simulation results As mentioned earlier, one of the points that solving the TSP can contribute is recombination operators i.e. mutation and crossover. Based on (Takahashi, 2005) there are two kinds of crossover operators for solving the TSP. Takahashi tries to retain useful information about links of parent’s edges which leads to convergence. The Takahashi’s experimental results suggest that changing crossover operators at arbitrary time according to city data structure is available to improve the performance of GAs. 9.2 ACO algorithms The algorithm presented in (Dorigo & Gambardella, 1997) is listed in Table 3. As it is ‘mentioned, the paper uses an algorithm based on ACS for solving the TSP. Problem name | ACS results (Dorigo & Gambardella 1997) | Optimum E150 427.96 425 Eil75 5a237, 535) Table 3. (Dorigo & Gambardella, 1997) simulation results 9.3 PSO algorithms Table 4 illustrates the results of the paper presented in (Yuan et al., 2007) which works based on ACO in combination with PSO. Problem name | Best Worst | Average Oliver30__ | 425.6542 | 457.2354 | 432.2231 Attas 33534_| 39679 | 34556 Table 4. (Yuan et al., 2007) simulation results28 Travelling Salesman Problem 9.4 IWD algorithms Based on the observation on the behavior of water drops, (Shah-Hosseini, 2007) develops an artificial water drop which possesses some of the remarkable properties of the natural water drop. The IWD algorithm is experimented by artificial and some benchmark TSP environments. The results show that the proposed algorithm converges fast to optimum solutions and finds good and promising results. Figures 8 and 9 depict the results of running this algorithm on some TSP benchmarks. Fig. 8. The best tour found by the proposed algorithm after 300 iterations for the 76-city problem eil76. The algorithm gets a good local optimum with the tour length 559 which is quite close to the global optimum 538. Fig. 9. The best tour found by the proposed algorithm after 1500 iterations for the 100-city problem kroA100. The algorithm gets a good local optimum with the tour length 23156 which is quite close to the global optimum 21282. Figure 10 shows the average length of the best tours of the IWD algorithm in 10 independent runs for the TSP problems in which the cities are on a circle. The number of cities is increased from 10 to 100 by the value of five, and in each case the best average tour length over 10 runs is depicted. Based on the simulation results, it is inferable that the IWD algorithm converges fast to optimum solutions and finds good and promising results.Population-Based Optimization Algorithms for Solving the Travelling Salesman Problem 29 Best Tour Length a Problom size (numberof cites 5-1) Fig. 10, The dotted lines show the global optimum tour length whereas the solid lines are the best tour lengths obtained by the IWD algorithm. 9.5 AIS algorithms The IGA and AIS have been applied for solving the TSP in many cases. In (Jiao & Wang, 2000) it is proved that the IGA is theoretically convergent with probability one. Besides, strategies and methods of selecting vaccines and constructing an immune operator are also given. The simulation results illustrate that IGA is able to restrain the degenerate phenomenon effectively during the evolutionary process and can improve the searching ability and adaptability, while greatly increase the converging speed. In another work, (Zeng & Gu, 2007), a novel genetic algorithm based on immunity and growth for the TSP is presented. The value obtained by the mentioned algorithm is prepared in Table 5. Results and investigations that performed in this study show that the algorithm is feasible and effective as itis claimed. Problem | Result in TSBLIB | Results by (Zeng & Gu, 2007) EAS 429.983, 428.872 Pr136 9672 9670.9 Table 5. (Zeng & Gu, 2007) simulation results 9.6 BCO algorithms In section 7, one of the main work that had been studied was the study around bee colony and its applications for transportation modelling with focus on artificial life (ALife) approach (Lucie & Teodorovic, 2002). This paper shows that the ALife models that have been developed for solving complex transportation problems are inspired by social insect’s behavior. The proposed algorithm in this work has been tested on a large number of well known TSP test benches such as Eil51.tsp, Berlin52.tsp, St70.tsp, Pr76.tsp, Kroal00.tsp, Eil101.tsp, Tsp225.tsp and A280.tsp. Also, for improving the results in each step, the 2-opt or 3-opt algorithms have been applied. Table 6 demonstrates the algorithm results. The results reveal that the mentioned method is very efficient. In all instances with less than 100 nodes, the bee system achieves the optimal solution and in the large cases it has a significant improvement in comparison to the other prevalent methods.30 Travelling Salesman Problem n Best value in (Lucic & Average value in (Lucic Problem | Optimalvalue | “Teodorovic, 2002) & Teodorovic, 2002) Fist BBS 228.87 428.87 Pr76 108159 108159 108159 Table 6. (Lucic & Teodorovic, 2002) simulation results obtained by the Bee System enriched with 3-opt heuristic. Another work on the field of BCO and for solving the TSP is (Teodorovic et al., 2006). In this paper the authors propose the Bee Colony Optimization Metaheuristic (BCO). Moreover, this study, describes two BCO algorithms that the authors call them, the Bee System (BS) and the Fuzzy Bee System (FBS). In the case of FBS the agents (artificial bees) use approximate reasoning and rules of fuzzy logic in their communication and acting. The simulation results of the BS can be seen in Table 7. Problem name | Optimal value | Best value by (Teodorovic et al, 2006) Eis. 429.983 431.121 Pr76 108159 108790 Table 7. (Teodorovic et al., 2006) simulation results 9.7 Electromagnetism-like mechanisms. Table 8 illustrates the results for EM which introduced in (Wu et al., 2006). 9.8 Comparison of various algorithms Figure 11 shows a comparison among, various methods for two standard TSP problems named st70 and kroa00. Problem name | Best | Optimal | Average 14 cities 30.879 | 30.879__| 31.80731 16 cities 32, 32 | 330889 Table 8. (Wu et al., 2006) simulation results Saat | |= ual FP OP te gg o @) © Fig, 11. @) The results of various algorithm applied on Kroai100, (b) The results of various algorithm applied on st70 (The vertical axes show the best tour length obtained by each algorithm).Population-Based Optimization Algorithms for Solving the Travelling Salesman Problem 31 10. Conclu: Maybe this chapter is the first versatile study on the population based optimization algorithms focused on solving the TSP. In this study, the state of the art of population based optimization algorithms such as Evolutionary Algorithms (EA), Ant Colony Optimization Algorithms (ACO), Particle Swarm Optimization Algorithms (PSO), Intelligent Water-Drops Algorithm (IWD), Artificial Immune Systems (AIS), Bee Colony Optimization Algorithms (BCO) and finally Electromagnetism-like Mechanisms (EM) has been introduced and investigated. The chapter includes nine parts before this; first one is introduction on the TSP and optimization algorithm, seven sections are about mentioned population based optimization algorithms and some related works which use these methods for solving the ‘TSP, and finally the last part encompasses experimental results on the perused studies. All the sections try to introduce the related population based algorithm truly. Then the authors attempt to explore some useful studies that have been done by other researchers for solving the TSP. In addition some important points, where contribution or innovation in different parts of the related algorithm or in solving the TSP can be applied, have been pointed. The experimental results demonstrate a brief comparison among the various population based optimization methods. In this section you can find some tables, graphs and figures which compare the presented methods with their counterparts in terms of efficiency using some well known benchmarks on the TSP. The performed study shows that all of the stated methods have some weakness and some strength points which are noticed at the related section. As a consequence, the further research can focus on these points for amplification of strengths and eliminating or improving the weaknesses. In addition, an innovative population based method inspired by natural water drops behaviour is reviewed in this, chapter. 11. References Adleman, L. M. (1994). Molecular computation of solutions to combinatorial problem. Science, 1994, pp. 1021-1023. Beckers, R.; Deneubourg, J-L.; Goss, S. (1992). Trails and U-turns in the selection of the shortest path by the ant Lasius Niger. Journal of Theoretical Biology, Vol. 159, pp. 397-415, 1992. Bersini, H. & Varela, F. J., (1990). Workshop on Parallel Problem Solving from Nature. LNCS, Springer-Verlag 496 343, Proc. Ist. Birbil, S. & Fang, Sh. (2003). An Electromagnetism-like Mechanism for Global Optimization. Journal of Global Optimization, , Kluwer Academic Publishers, Vol. 25, (2003), pp. 263-282, ISSN 263-282, 2003. Birbil, S. & FANG, Sh. (2005). Convergence of a Population Based Global Optimization Algorithms. Journal of Global Optimization Bonyadi, R. M.; Rahimi Azghadi S, M. & Shah-Hosseini, H. (2007). Solving Traveling Salesman Problem Using Combinational Evolutionary Algorithm. In: IFIP International Federation for Information Processing, Volume 247, Artificial Intelligence and Innovations 2007: From Theory to Applications, eds. Boukis, C, Pnevmatikakis, L., Polymenakos, L., pp. 37-44, (Boston: Springer). Cowan, E. W. (1968), Basic Electromagnetism, Academic Press, New York.