Evolution in Time and Space – The Parallel Genetic Algorithm

In this paper, a Parallel Genetic Algorithm has been developed and mapped onto a coarse grain MIMD multicomputer whose processors have been configured in a fully connected chordal ring topology. In this way, parallel diffusion processes of good local information among processors have been carried out. The Parallel Genetic Algorithm has been applied, specifically, to the Travelling Salesman Problem. Many experiments have been performed with different combinations of genetic operators; the test results suggest that PMX crossover can be avoided by using only the inversion genetic operator and that a diffusion process leads to improved search in Parallel Genetic Algorithms.

[EN] In this research, a new model of computing is presented, within the framework of natural computing and computer models inspired by biology: Networks of Genetic Processors (NGP). This new model is based on the one hand, by the family of networks of bio-inspired processors, specifically Networks of Evolutionary Processors (NEP) and Networks of Splicing Processors (NSP), and on the other hand by Genetic Algorithms. We can define the new model as a network of biologically-inspired processors where operations used by processors are crossover and mutation. One of the major interests studying the NGP is the computational power of the operations of crossover and mutation acting together. The NEP is a complete model that uses operations of symbol mutation: insertion, substitution and deletion, the NSP is a complete model that uses splicing operations, but the NEP is no longer a complete model using only substitution operations, as it happens to the NSP if we restrict the context of its splicing rules to the empty context. The study of the new model presented here responds to what happens when we put together in a single model the substitution rules from the NEP (called mutation rules) and the splicing rules with empty context from the NSP (called crossover rules).\n\nWhen we work with networks of biologically-inspired processors there are two basic types of networks, accepting networks and generating networks. These types of networks are mainly used to work at a theoretical level or to solve decision problems. To work on a more practical level such as solving optimization problems, we propose a new type of network, Networks of Genetic processors as Parallel Genetic Algorithms, inspired by the Parallel Genetic Algorithms.\n\nIn this work we prove the computational completeness of our new model by showing that it is equivalent to the Turing machine. We prove the computational completeness of Parallel Genetic Algorithms by using this result and the similarity between the NGP and Parallel Genetic Algorithms. Moreover, we propose a characterization of the Chomsky hierarchy by using the NGP. Here, we simulate every grammar in the language classes of the Chomsky's hierarchy by using a NGP with an small number of processors required for each simulation. Hence, it gives an appreciable idea of the descriptional complexity of the different families of languages.\n\nFinally, in this work there is an experimental study of the behavior of the model for the resolution of some practical problems. First, we design and implement a simulator that allows the execution of networks of Genetic Processors in any of its three defined types: accepting networks, generating networks or as Parallel Genetic Algorithms. This allows us to test the model with different optimization problems. Then, we make a study to see if the new model could solve NP problems in polynomial time. We use the decision problem of Hamiltonian cycle in a graph. Finally, we test the simulator with two optimization problems showing a good computational behavior. The problems are the Multidimensional Knapsack problem and the Traveling Salesman problem.

With the introduction of Splash, Splash 2, PAM, and other reconfigurable computers, a wide variety of algorithms can now be feasibly constructed in hardware. In this paper, we describe the Splash 2 Parallel Genetic Algorithm (SPGA), which is a parallel genetic algorithm for optimizing symmetric traveling salesman problems (TSPs) using Splash 2. Each processor in SPGA consists of four Field Programmable Gate Arrays (FPGAs) and associated memories and was found to perform 6.8 to 10.6 times the speed of equivalent software on a state-of-the-art workstation. Multiple processor SPGA systems, which use up to eight processors, find good TSP solutions much more quickly than single processor and software-based implementations of the genetic algorithm. The four-processor island-parallel SPGA implementation out performed all other SPGA configurations tested. We conclude noting that the described parallel genetic algorithm appears to be a good match for reconfigurable computing machines and that Splash 2's various interconnect resources and support for linear systolic and MIMD computing models was important for the implementation of SPGA.

Parallel generic algorithms (PGAs) have been developed to reduce the large execution times that are associated with serial generic algorithms (SGAs). They have also been used to solve larger problems and to find better solutions. A comparative analysis of five different coarse-grained PGAs is conducted using the traveling salesman problem as the basis of this case study. To make fair comparisons, all of these PGAs are based on the same baseline SGA, implemented on the same parallel machine (IBM SP2), tested on the same set of traveling salesman problem instances, and started from the same set of initial populations. As a result of the experiments conducted in this study, a particular PGA that combines a new subtour technique with a known migration approach is identified to be the best for the traveling salesman problem among the five PGAs being compared.

The class of MIMD parallel computers includes multiprocessor architectures, such as UMA, COMA, and NUMA, as well as multicomputer architectures, such as MPP and COW. Despite huge technological differences, machines with different architectures may be used to run certain similar applications. An interesting category of applications that may be successfully executed on a variety of parallel architectures is that of genetic algorithms (GAs) that provide more and more attractive solution of many complex engineering problems for which classical optimization methods cannot be used. GAs prove to be parallelizable in a natural way. Parallel genetic algorithms (PGAs) highlight good balance between calculations and communications that allows obtaining good results on different classes of parallel architectures. The paper describes the results obtained in a comparative study of PGAs, implemented on a cluster of SunBlade 100 workstations (COW) and on a Sun Enterprise E10000 (UMA) computer. The study demonstrates suitability of these different architectures for the development and execution of this class of applications. Also, some relevant facts concerning the performance of PGAs are presented.

Deceptive problems have always been considered difficult for Genetic Algorithms. To cope with this characteristic, the literature has proposed the use of Parallel Genetic Algorithms (PGAs), particularly multi-population island-based models. Although the existence of multiple populations encourages population diversity, these problems are still difficult to solve. This paper introduces a new initialization mechanism for each of the populations of the islands based on Voronoi cells. In order to analyze the results, a series of different experiments using several real-value deceptive problems and a set of representative parameters (migration ratio, migration frequency and connectivity) have been chosen. The results obtained suggest that the Voronoi initialization method improves considerably the performance obtained with a traditionally uniform random initialization.

The genetic algorithm is a robust search and optimization technique based on the principles of natural genetics and survival of the fittest. Genetic algorithms (GA) are a promising new approach to global optimization problems, and are applicable to a wide variety of problems. HYPERGEN was developed as a research tool for investigating parallel genetic algorithms applied to combinatorial optimization problems. It provides the user with a wide variety of options to test the particular problem at hand. In addition, HYPERGEN is modular enough for a user to insert routines of his own for special needs, or for doing further research studies on parallel GAs. HYPERGEN was used successfully to find new 'best' tours on three 'standard' TSP problems, and out-performed a parallel simulated annealing algorithm on various package placement problems. The authors found it fairly easy to fine tune the parameters that drive a parallel GA for near optimal performance (population size, migration rate, and migration interval). >

This chapter presents an evolutionary approach for solving the traveling salesman problem (TSP) and the TSP with backhauls (TSPB). We propose two evolutionary algorithms for solving the difficult TSPs. Our focus is on developing evolutionary operators based on conventional heuristics. We rely on a set of detailed computational experiments and statistical tests for developing an effective algorithm.The chapter starts with a careful survey of the algorithms for the TSP and the TSPB, with a special emphasis on crossover and mutation operators and applications on benchmark test instances. The second part addresses our first evolutionary algorithm. We explore the use of two tour construction heuristics, nearest neighbor and greedy, in developing new crossover operators.We focus on preserving the edges in the union graph constructed by edges of the parent tours.We let the heuristics exploit the building blocks found in this graph. This way, new solutions can inherit good blocks from both parents. We also combine the two crossover operators together in generating offspring to explore the potential gain due to synergy. In addition, we make use of 2-edge exchange moves as the mutation operator to incorporate more problem specific information in the evolution process. Our reproduction strategy is based on the generational approach. Experimental results indicate that our operators are promising in terms of both solution quality and computation time.In the third part of the chapter, we present the second evolutionary algorithm developed. This part can be thought of as an enhancement of the first algorithm. A common practice with such algorithms is to generate one child or two children from two parents. In the second implementation, we investigate the preservation of good edges available in more than two parents and generate multiple children.We use the steady-state evolution as a reproduction strategy this time and test the replacement of the worst parent or the worst population member to find the better replacement strategy. Our two mutation operators try to eliminate the longest and randomly selected edges and a third operator makes use of the cheapest insertion heuristic. The algorithm is finalized after conducting a set of experiments for best parameter settings and testing on larger TSPLIB instances. The second evolutionary algorithm is also implemented for solving randomly generated instances of the TSPB. Our experiments reveal that the algorithm is significantly better than the competitors in the literature. The last part concludes the chapter.KeywordsNetworks-GraphsTraveling Salesman ProblemsEvolutionary AlgorithmsCrossover OperatorMutation OperatorHeuristics

Wireless sensor network (WSN) technology is widely used in environment monitoring, health care, surveillance systems and unmanned space or planet exploration. This paper focuses on routing problems with data gathering by a mobile robot in a WSN, also referred to as a Traveling Salesman Problem with Neighborhoods (TSPN) or NP-hard problem. In this paper, we propose a clustering-based parallel genetic algorithm with migration (CBPGA), so that the mobile robot can gather all data from all sensors and the travel costs of the mobile robot clearly decrease. First, a clustering algorithm is used to effectively reduce the number of visited nodes, especially in situations with dense sensor distributions or large sensing radii. Next, the set of visited nodes is encoded as chromosomes by a chromosome generation algorithm (CGA), and the master-slave parallel genetic algorithm with migration is performed to more efficiently generate the near-optimal route. Lastly, a travel cost-reduction scheme is used to remove redundant travel costs. Simulation results confirm that the clustering-based parallel genetic algorithm with migration more efficiently generates a near-optimal route that reduces the travel costs of a mobile robot in robot routing problems with WSN.

Genetic algorithms are adaptive search algorithms that have been shown to be robust optimization algorithms for multimodal real-valued functions and a variety of combinatorial optimization problems. In contrast to more standard search algorithms, genetic algorithms base their progress on the performance of a population of candidate solutions, rather than on a single candidate solution. The authors will concentrate on the application of genetic algorithms to the traveling salesman problem. For this problem, there exist several such algorithms, ranging from pure genetic algorithms to genetic algorithms that incorporate heuristic information. These algorithms will be reviewed and their performance contrasted. A serious drawback of genetic algorithms is their inefficiency when implemented on a sequential machine. However, due to their inherent parallel properties, they can be successfully implemented on parallel machines, resulting in considerable speedup. Parallel genetic algorithms will be reviewed and their uses in the traveling salesman problem will be indicated.

This paper discusses a parallel genetic algorithm (GA) which focuses on the local operator for Traveling salesman problem (TSP). The local operator is a simple GA named as Local Genetic Algorithm (LGA). The LGA is combined to another GA named as Global Genetic Algorithm (GGA). It increases the computational time running a GA as a local operator in another one. To solve this problem, we build a parallel system based on our previous works for running the LGA to speed up the process. The results show that LGA improve the search quality significantly and it is more efficient running LGA with parallel system than single CPU.

Optimization problem like Travelling Salesman Problem (TSP) can be solved by applying Genetic Algorithm (GA) to obtain perfect approximation in time. In addition, TSP is considered as a NP-hard problem as well as an optimal minimization problem. Selection, crossover and mutation are the three main operators of GA. The algorithm is usually employed to find the optimal minimum total distance to visit all the nodes in a TSP. Therefore, the research presents a new crossover operator for TSP, allowing the further minimization of the total distance. The proposed crossover operator consists of two crossover point selection and new offspring creation by performing cost comparison. The computational results as well as the comparison with available well-developed crossover operators are also presented. It has been found that the new crossover operator produces better results than that of other cross-over operators.

This paper presents a parallel artificial immune model termed as tower master-slave model (TMSM) for solving optimising problems. Based on TMSM, the parallel immune memory clonal selection algorithm (PIMCSA) is also proposed. TMSM is a two level coarse-grained parallel artificial immune model with distributed immune response and distributed immune memory. In PIMCSA, vaccines are extracted and migrated between populations rather than individual migration as has been done in parallel genetic algorithms. It is a good balance between population diversity and the convergent speed. Experimental results on the numerical optimization and TSP problems show that PIMCSA achieves good performance in terms of both solution quality and computation time.

Genetic Algorithm (GA) is a population-based stochastic optimizer that mimics natural selection to solve complex, non-differentiable, and multimodal problems. This paper first reviews the biological foundations and canonical operators- selection, crossover and mutation--before surveying major variants such as Hybrid GA, Parallel GA and the Predator--Prey Dynamic GA that inject local search, island parallelism or ecological balance to mitigate premature convergence. This paper then benchmarks these algorithms on continuous function optimisation and combinatorial tasks, using the sphere function and the travelling-salesman problem as illustrative cases. Results show that Hybrid GA with simulated-annealing local search achieves the best accuracy--speed trade-off, while PDGA offers the most stable evolution curve in high-dimensional multimodal landscapes. Finally, this paper identifies remaining challenges--expensive fitness evaluation, parameter sensitivity and interpretability--and outline future directions including surrogate-assisted evolution, GPU/quantum acceleration and lifelong on-device adaptation. The study confirms that GA, when properly hybridised and parallelised, remains a competitive backbone for next-generation AI systems.

In the past few years, CUDA and OpenCL are developed in full use of the GPU, which is a significant topic in high performance computing. In this paper, we have proposed an implementation of the genetic algorithm for the traveling salesman problem on the parallel OpenCL architecture. Population initialization, fitness evaluation, selection, crossover and mutation operators are implemented on the GPU by using individual to thread mapping. Moreover we have evaluated our algorithm using a set of benchmark instances from the TSPLIB library. The comparison results shows that GPU computations provide better performance than traditional CPU implementation.