.1 Introduction In recent years, the excellent performance of the genetic algorithm (GA) has attracted people's concern. For the past, it is difficult to resolve the optimization problem, complex multi-objective planning problem, the piping, wiring issues in industrial and agricultural production, and machine learning, The problem of image recognition, the weight coefficient adjustment of artificial neural network, and the network structure, GA is one of the most effective ways. Although GA has successful applications in many optimization problems, they also have some shortcomings, such as local search Poor ability, unrimeted convergence and random roaming, resulting in poor convergence performance of the algorithm, it takes a long time to find the best solution, these shortcomings hinder the promotion of genetic algorithms. How to improve the search ability of genetic algorithm and improve The convergence speed of the algorithm makes it better to solve the practical problem, and it is a major topic that all scholars have been exploring. The world has launched the research and application of the genetic algorithm in the world. Problems exist and corresponding The improvement measures have long showed the powerful power of the genes. Through this mechanism, a series of intelligence, self-organized, self-trimmed organs have been produced. People must follow these biological organs in scientific research, then they must understand the genes. The concept of evolution. GA is such a way to use natural selection and evolutionary thinking in high-dimensional space, it is not necessarily to find the most advantage, but it can find more advantages, this idea is successful in human behavior The standard is very similar. For example, if you don't have to ask a military to be the best, you must overcome your opponent to be stronger than your opponent. Therefore, GA may temporarily stay in some non-optimal points until the variation happens Jump to another. An important feature of the GA search process is that it always maintains the evolution of the entire population, so that even if an individual has lost a useful feature at a certain moment, this feature will be preserved by other individuals and Continuing development. Since GA only needs to know the information of the target function, there is a wide range of adaptability. At the same time it is an intelligent search algorithm using inspirated knowledge, so it is often Searching the Search Space Highly Complex Effects than previous algorithms (such as gradient methods), DB Fogel, the evolution of DB Fogel, the concept of intelligence [1 0], although it has not been generally accepted, but evolution in the process of human survival The importance is visible, so the genetic algorithm is an embodiment of biological evolutionary thinking in engineering calculations. The future is bright. At present, GA is in engineering optimization, signal processing, pattern identification, management decision, intelligent system design and labor life. The success of the field is indicated .2. 1 Code indicates that Holland is used to use the mode theorem to analyze the encoding mechanism, it is recommended to use binary encoding, but binary coding cannot directly reflect the inherent structure of the problem, the accuracy is not high, the individual length is large, occupied Within the computer Picture. GRAY encoding is the encoding obtained by converting binary encoding through a transformation. Its purpose is to overcome the shortcomings of the Hamming cliff, and dynamic encoding GA is the accuracy of the search to increase the search when the algorithm converges to a local optimal. Make more precise search near the global optimal point, the increase in accuracy is to reduce the search area under the premise of maintaining the non-change of the string. For the problem of the problem, it can be directly encoded, so The genetic operation can be performed directly on the expressive type, thus facilitating the introduction of heuristic information related to the problem area to increase the search capabilities of the algorithm. The GA encoding [5] is to describe and solve the two-dimensional problem, the gene X Yi said;
It can also be extended to multi-dimensional problems. Multi-dimensional coding [6] GA, which makes the possibility of invalid cross-happening, and its reasonable coding length also helps algorithm to get high-precision global optimality in a short time. Solution. In combination optimization, an ordered string coding can be used, for example in the literature [7], VRP issues are subtly solved with natural digital coding. When the representation of the problem is a tree and diagram, we can also use structural code 2. 2 The adaptation function adaptation function is used to distinguish between individual good or bad standards in groups. It is the only standard for natural selection. The quality of the choice is directly affected. The introduction of adaptive value adjustment and resource sharing strategy can accelerate convergence speed and jump out. Local optimal point. Adjusting the adaptation value is the ratio relationship between the original adaptation value, the commonly used ratio transform has a linear transformation, multiplying power conversion, and index transform. What transformation is used for a problem to achieve better results , V. Kreinovich, etc., in the literature [8], which is a more detailed discussion in [9], which uses sharing technology. The formation and stability of subgroups have played a role. The main use of subsets in the text disappears. Approximate SHARING's bounds. The adaptive effect of the search progress variable is used in the literature [1 0] and is applied to the CuttingProblem to achieve better results. Adaptive selection genetic algorithm is designed in the literature [1 1]. The method of value function, the amount of calculation is much smaller than the amount of calculation of the selection operation, and effectively avoids the non-ripe convergence of the algorithm. 3 Select the selection mechanism of the strategy Yingsheng fittest makes the adaptive value larger individuals Big survival opportunities, different selection strategies have a big impact on algorithm performance. Roulet gambling is the most selected policies, but this policy may result in a large sampling error, so many improvements have been proposed. Method, such as the selection of [1 2], Boltzmann Select [1 3], etc.. But these strategies are based on the choice of adaptation value, often in the premature convergence phenomenon and stagnation phenomenon. Linear ranking selection [3], this choice can not only avoid the above problems, but also can use the original adaptation value to select the ranking, without standardizing the adaptation; however, this choice is large when the group is large, and its additional calculation Quantity (such as the overall adaptation value and sorting) is also considerable, even sometimes some synchronization restrictions are sometimes brought about by parallel implementation. Selection based on local competition mechanisms (λ
μ) Select [1 4], which allows the same survival competition opportunity to some extent to some extent to a certain extent. In [1 5], the authors are selected in a similar gradient, not only the poor chromosome A better chromosome gains more improvements, and constantly produces new individuals, which continuously expands new search space. [1 6] The author introduces Harvesting Strategies to analyze the performance of the genetic algorithm, Harvesting Strategies refers to each After one generation of intersections and mutations, twice and even multiple screens as the following group. It adopts Dismuptive Selection, which absorbs superior and inferior individuals. The experimental results show that the polarization may be easier to find the best solution. In order to improve the population Sexuality, a selection operator [1 8] based on immunological diversity, the selection operator depends on the denseness and adaptation value of the string, the larger the denseness of the string, the smaller the possibility of it, specific examples Certification improvement algorithm is effective .2. 4 Control parameter control parameters generally have group size, exchange probability, variation probability, etc., these parameters have a large effect on genetic algorithm performance. Estimated in the standard genetic algorithm, this will take Come big blindness, and affect the global optimality and convergence of the algorithm. Many scholars have realized that these parameters should adapt to changes with genetic evolution, Davis proposes adaptive operator probability method [1 9], ie The self-adaptive mechanism combines the individual adaptability of the operator probability and the operator, and the high adaptive value is dispensed with high operator probability. Whitley proposes an adaptive mutation strategy and a pair of parent string Hamming distances [2 0] The result shows the diversity of genes. Zhang Liangjie et al., By introducing I bits to improve sub-space concepts, using fuzzy reasoning techniques to determine the general principles of selecting mutation probability [2 1]. In literature [2 2] A group-variable genetic algorithm is designed. It proposes that each individual should have a concept of age and life, and eliminate individuals with age than a lifetime, making genetic algorithms to control the number of groups. This method can be found. A genetic algorithm close to the minimum price, while maintaining the size of the group in the existing level, preventing the size of the size of the group, to reduce the cost of the calculation. Ding Chengming and other proposes to use orthogonal test method to optimize the GA control parameters [2 3 ], This method utilizes the equalization of orthogonal testing so that through fewer test times can search for most of the parameter combination space, and can also determine which parameter has the most significant effect on GA results, and then accurately Search, so that the GA parameter problem has been successfully resolved. In order to ensure useful diversity of populations, the dynamic group method [2 4] is proposed, that is, when it is iterated to a certain algebra, if The value of the target function is the same, and the poor n-chromosome in the present specries are replaced by the random N-chromosome, which continuously introduces new individuals during the evolution. [2 5] Infine rules to select probability and variation probability Control, change its value online, the corresponding example indicates that there is a good performance 2. 5 genetic operator basic genetic algorithm uses a single point crossover operator and a simple mutation operator. They operate more simple, small computation However, there is a lot of limitations during use, for example, since the probability of single cross-breaking mode is small, the number of patterns is small, so that the algorithm has a lower search capability. FENG etal. Pair The hybridization diversity of GA of multidimensional continuous space is analyzed. By establishing a corresponding mathematical model, Feng explains how to explore new solution space regions in multi-dimensional continuous space and large-scale groups [2 6]. In order to make the variation to adjust the search area according to the quality adaptation of the solution, it can better improve the search capability, propose adaptive variation operator [2 7]. In order to protect the mode of high adaptive value, adaptive crossing and variation is proposed. [2 8], if the mode of adaptation value is encountered, it is protected by a bit outside the random introduction mode. In order to overcome premature, the multi-group GA [2 9], the different groups are intended to implement different control parameters, implementation Different search purposes, through immigration operators, the optimal individuals in each of the various groups through the manual selection operator. In order to prevent nearby reproduction, expand population diversity, inhibit the rapid super long individual Breeding, introducing near-relative breeding operators, whether the two individuals are judged by the Hamming distance of nearby gene fragments, the greater the distance, the smaller the nearby relatives;
In order to enhance local search capabilities, increase the drift operator, and one-half of the genes of each gene fragment of the chromosome each, respectively, the random drift of 1 probability, the larger the probability of gene drift after the rank, thereby producing A certain number of new individuals, the small-sized technologies of gene pre-selection mechanism control the drift direction [3 0]. Because the point set generated by the grid point method can be evenly distributed in the search space, and the good point is the best grid, so Can be designed with a combination of the best point in the origin of the boss [3 1], indicating that its search effect is better than pure random, and effectively avoids premature premugation. Based on the immunoassay proposed by biological immunity [3 2 ], It is possible to significantly suppress degradation phenomena during evolution, alleviate fluctuations in the late GA, thereby increasing search efficiency and convergence speed. The SRM (Self-Reproduction) operator proposed in [3 3] enhances the diversity of populations, cm (Crossove and Mutation) The operator promotes an increase in favorable variation, thereby saving the algorithm greatly saves storage space and runtime. The chaotic variation of the "scale contraction" strategy [3 4] can significantly improve the average adaptation value of the population , Improve the performance of the algorithm is an effective method to solve the optimization problem. 2. Decomposable / spliced GA coding is proposed in the integrated literature [3 5], and based on this encoding, dynamic variation is developed in the population level and the gene level, respectively. And dynamic selection operation, which greatly avoids premature maturity problem. In the enhanced GA [3 6], several new operators and new population migration strategies are introduced, and their fuzzy logic controller is designed. Get easy-to-understand fuzzy set and fuzzy rules. Multi-scale decomposition in GA in wavelet analysis, so that the length of the chromosome after decomposition is short, and genetically exchanged, variation is more thorough. Efficacy overcomes the early maturation problem caused by gene loss [3 7]. Little bodies technology not only guarantee the diversity of the group, but also has strong guiding evolutionary ability, so the introduction of small-sized technologies, improved GA treatment multi-peak The ability to optimize the problem [3 8]. Introducing analog annealing process [3 9], adding a certain "disturbance" in the process of intersection and mutation, to reach the diversity and position of the maintenance population The competition mechanism between the strings, overcomes the problem that the algorithm is easier to get into the extremely small point, so that the search is performed along the overall optimal direction. Generalized genetic algorithm [40], it is mainly multi-change operation, supplemented by gene crossover operation Really, the transfer from a local optimal state to another local optimal state is achieved, enabling the algorithm to obtain the global optimal. In order to make Ga for constraint optimization, a non-stable penalty function Ga [41], non-steady state The penalty function is a function of the genetic algebra, and when the number increases, the penalty function also increases. Big, and bring more choices to GA to promote GA to find the problem. The overall genetic algorithm is characterized by the characteristics of parallelity of neural network, the proposed genetic neural network algorithm [42], can overcome the genetic algorithm final Evolution to the optimal solution and neural network is easy to fall into the defects of local solutions. It has better global and convergence speed. It is designed with object-oriented technologies [43], this method changes in traditional GA There is only a parameter pass between the various functions, and the situation without code has improved the reusability of software from the concept, and users can more easily design and implement their own coding schemes and genetic operators. Variable genetic algorithm [44], use variation operators to perform local optimization search, and use random initialization technology to make algorithms to improve the global optimal solution while local search capabilities. Greedy genetic algorithm [45] is used in secondary distribution issues It has achieved good results, introduced new cross-operators and immigrant operators in the algorithm, ensuring the diversity of populations; and makes various groups to evolve through competition competition, and solve population diversity and Contradiction between individual preferences. 3 development trends of genetic algorithms (Ga '
s Developmen-Tal Trends) GA is full of fruitful results, making people confidence in its development prospects. The main application area is the function optimization (nonlinearity, multi-model, multi-objective, etc.), robotics (mobile robotic path planning) , Joint robot sports trajectory planning, structure optimization of cell robots, etc.), control (gas pipeline control, anti-avoidance missile control, robot control, etc.), planning (production planning, parallel mission distribution, etc.), design (VLSI layout, communication network Design, Jet Engine Design, etc.), combined optimization (TSP problem, backpack problem, diagram planning problem, etc.), image processing (pattern recognition, feature extraction, image recovery, etc.), signal processing (filter design, etc.), artificial Life (genetic evolution of life). In addition, the research of the genetic algorithm has several leading new trends: 3. 1 Based on genetic algorithm, machine learning this new research direction puts genetic algorithm from history discrete search space Optimizing the search algorithm is extended to a new machine learning algorithm with unique rule generation functions. This new learning mechanism has brought hope for the bottleneck problem of knowledge acquisition and knowledge optimization refining in artificial intelligence. Genetic algorithm as a search algorithm From the beginning, it is closely related to the machine. The classifier system CS-1 is the first genetic algorithm-based machine learning system, which is the implementation of Professor Holland, GA, etc., and the classifier system has been applied in many fields. For example, Successful application obtained in the learning multi-robot path planning system; Goldberg studies the classifier system to learn to control a gas pipeline simulation system;
Wilson studied a perception of mobile video cameras - Sports classifier system. The classifier system has a large impact in machine learning research based on genetic algorithm, but specific implementation methods and specific issues to be solved The concept of genetic algorithm is a more striking research direction in the field of machine learning in recent years. Due to the conceptual learning implicit search mechanism, the genetic algorithm has used Wuzhi in the concept of concept. Currently some embedded field knowledge Research on machine learning based on genetic algorithm, such as the unique operation of concept learning, and shows a certain advantage. In addition, the parallel implementation of the learning classification system also has considerable in machine learning research based on genetic algorithm. Component .3. 2 Genetic Algorithm and Other Intelligence Methods of Intelligence Methods and Combined Genetic Algorithms are increasing and other intelligent calculations such as neural network, fuzzy reasoning, and chaos theory are penetrated and combined with each other. This has been made in this regard, and has formed a research area of "calculation intelligence", which will have important significance for the new intelligent computing technology in the 2-1 century. GA's emergence of neural networks (including connection rights) Optimization of the coefficient, the optimization of network space structure and the optimization of network learning rules) have a new look, and the target function is neither a continuous, nor requires it, only requires that the problem can be calculated, and its search is always all over again. Solution space, it is easy to obtain global optimal solution. The combination of GA and neural network is successfully used to conduct financial budgets from time series, in these systems, training signals are blurred, data is noise, It is generally difficult to correctly give each execution quantitative evaluation. If you use GA to learn, you can overcome this difficulty, significantly improve the performance of the system. Muhlenbein analyzes the limitations of multi-layer perception machine network, and conjecture the next generation of nerves The network will be a genetic neural network. Genetic algorithm can also be used to learn fuzzy control rules and membership functions, thereby better improving the performance of fuzzy systems. The fuzzy logic, neural network and genetic algorithm are organic in literature [46]. The combination is applied to the temperature and humidity control in the greenhouse. The experimental results show that the results have been achieved. The randomness of chaotic shows is the internal randomness, known as pseudo-random, it plays an important role in biological evolution. The role is the source of system evolution and information. The combination of chaos and genetic algorithm has been attempted, such as Wu Xin Yu et al. [47] use a variety of chaotic models to construct random switches to control cross-operation to improve GA performance. More directly, use of chaotic sequence constructivore operators, and opened up new ways for genetic algorithm. 3. 3 parallel processing genetic algorithm parallel processing genetic algorithm is not only the development of genetic algorithm itself ,and It is very important for the study of a new generation of intelligent computer architecture. GA has high parallelism in operation, many researchers are exploring the high-efficiency implementation of GA on parallel. In recent years, there are many Aspects, research shows that as long as the interaction between multiple groups and the interaction between the group is properly controlled, we can improve the efficiency of the algorithm even without using parallel computers. In parallel GA research, Some parallel GA models have been performed on a specific parallel machine; parallel Ga can be divided into two categories: one is coarse particle size parallel GA, which mainly develops parallelism between groups, such as Cohoon analyzed in parallel computer The performance of the multi-group GA of division; the other is fine-grained GA, which mainly develops parallelism in a group, such as Kosak maps each individual in the group to a connecting machine's processing unit, and pointing this Method for the effectiveness of the network diagram design problem. 3. 4 Genetic algorithm and artificial penetration artificial life is a manual system with natural biological system with a natural biological system, artificial system, artificial life and genetics using computer, mechanical media simulation or construct The algorithm has a close relationship. The evolutionary model based on genetic algorithm is an important theoretical basis for studying artificial life. Although the research of artificial life is still in the enlightenment phase, the genetic algorithm has been in its evolutionary model, learning model, behavioral model, self-organizing model The preliminary application ability is displayed, and it will be more in-depth application and development. Artificial life and genetic algorithm complement each other, the genetic algorithm provides an effective tool for artificial life, and the research of artificial life will also Promote the further development of genetic algorithms. 5 Genetic algorithm and evolutionary rules and evolutionary strategies, the evolutionary rules and evolutionary strategies are three main branches of evolutionary calculations, these three typical evolutionary algorithms are creatures in nature. The evolutionary process is an object of the adaptive global optimization search process, so there is a large similarity between the three;
