Nature is evolving, and the calculation algorithm can also evolve. Evolution software allows computers to implement automatic programming, and evolution hardware can change its own structure according to changes in the environment. In this issue, Professor Kang Lishan, a doctoral tutor Kang Lishan, a Ph.D. Kanglishan, a national key laboratory of Wuhan University, will solve the mystery of evolution calculation.
Evolution and Shenghua - Research and Application of Professor Kang Lishan
-------------------------------------------------- ------------------------------
---- Q: The evolution calculation is a completely unfamiliar noun for many people, but the study in this area has a great influence in foreign countries in recent years and has a future. First, we want to explain the concept of evolution computing and its development.
---- A: The definition of evolution in the dictionary is evolving, and the process of multi-refers to the changes in nature. Nature is the source of human beings. Hundreds of years have been applied to practical problems in the actual problem for hundreds of years. They have been proven to be a successful approach, and have forming a special scientific branch of the bionics. We know that the answer provided by the nature is the result of a long adaptive process (ie the evolution process). In addition to the final result of the evolution process, we can also use this process itself to solve some of the more complex problems. That is, we don't have to clearly describe all the features of the problem, just need to generate new better solutions according to the natural law. Evolution calculation is a general problem solving method based on this idea. It uses simple encoding techniques to represent a variety of complex structures, and guides the direction of learning and determining search by simply performing simple genetic operations and natural selection of the Yinsheng Faces.
---- There are many different schools at the beginning of the evolutionary development, one is the EvolutionaryProgramming school. Americans LJFogel issued a smart research in 1996. Intelligent behavior is a state that can predict the environment. And make appropriate responses in accordance with a given object. In the study, he describes the simulation environment as a sequence consisting of symbols in a finite character set, so that the problem is transformed into how to respond to the currently observed symbol sequence to obtain the greatest income, the calculation of the income here is in the environment The next symbols will appear and the pre-defined benefits targets are determined. In 1963, Germany's I. Rechenberg and HPSchWefel et al., When the wind tunnel experiment was in the design, it was difficult to optimize the parameters of the object shape in the design, so that they use the idea of biological variation to randomly change parameters. It has achieved good results. Subsequently, they have conducted in-depth research on this method, forming another branch of evolution calculation - EvolutionaryStrategy school. The third branch of the evolutionary algorithm is a Geneticalgorithm school proposed by Johnholland, Michigan, USA, which is called a group of binary strings called populations (ie, chromosomes and individuals), each chromosome A solution corresponding to the problem. From the initial population, the selection policy based on adaptation value is used in the current population, using hybrids and variations to produce the next generation population, so that the generation is evolved until the desired termination conditions are met. In the early 1990s, these branches were integrated with each other, and the new discipline was formed.
---- Q: The essence of evolution calculations is to solve practical problems with their thoughts and principles of biological evolution. So what is the difference between the evolution algorithm compared to the traditional algorithm?
---- A: The evolutionary algorithm has many different algorithms, but the most important difference is the characteristics of the evolution and calculate the intelligence and essential parallelism.
---- The intelligence of evolution calculation includes self-organization, adaptive and self-learning. When the application evolution calculates the problem, the algorithm will search for the information obtained by the evolution process after determining the encoding scheme, the adaptive value function and the genetic operator. Since the natural selection strategy is to survive, unsatisfactory is eliminated, the individual with high adaptation has a higher probability. Usually the individual with large adaptation value has a more adaptive environment, and then genetic operations such as hybridization and genetic mutation may produce the descendants of the environment. This self-organizing, adaptive feature of the evolution algorithm also imparts its ability to automatically discover the environment and regularity of the environment according to the environment. In addition, natural selection eliminates one of the biggest obstacles in the algorithm design, that is, the full characteristics of the problem need to be described in advance, and measures the measures that should be taken for different characteristics algorithms. Thus, the method of using evolutionary calculations We can solve the complex problems that are still unintended. ---- Essential parallelity of evolution calculations in two aspects: First, the evolution calculation is inner parallel, that is, the evolution algorithm is very suitable for large-scale parallelism. The simplest parallel way is to make a few hundred or even thousands of computer each perform an independent population evolution. During the operation, no communication is performed until the end of the operation, the best individual is selected, this parallel processing method is There is no restriction and requirements in parallel system structure. It can be said that evolution computing is suitable for parallel processing on all parallel or distributed systems, and there is little impact on their parallel efficiency. Second, the evolution calculation is included, since the evolution calculates the searched manner, so that multiple regions of the solution can be searched simultaneously and communicate with each other. Although only calculations proportional to the population scale N is performed each time, approximately O (N3) times effective search is substantially carried out, which enables evolution computing to obtain a larger income at least fewer calculations.
---- Evolution calculation did not receive universal attention in the 1960s and 1970s, the main reason is that these methods itself is not mature at that time; second, because these methods need to be larger computation, and the computer is not enough at the time. Popping, and the speed also does not keep up, so that the application is limited; the third is that the artificial intelligence method based on the symbolic processing is in its peak state, which makes it difficult to recognize the effectiveness and limitations of other methods, and due to The improvement of computer speeds and the popularity of parallel computers, making evolution calculations on machine speed requirements are no longer a factor that restricts its development. In addition, due to the evolution of evolution, the evolutionary calculation has exhibited a good application prospect with the success of the evolution of the evolution and its successful development of machine learning, process control, economic forecasting, engineering optimization.
---- Q: The evolution of the evolution has been achieved in the past 30 years, but also has some phased achievements. So, in your opinion, what is the development trend of evolution calculation?
---- A: I think that evolution hardware will be an important trend in the future evolution and development, is a new direction. The so-called evolutionary hardware refers to a change in the environment based on the environment, and the hardware of the environment is changed to accommodate the hardware of its living environment. the design of. Evolutionary hardware is the combination of evolution algorithms and programmable logic devices, evolutionary calculations provide the theoretical and methodological basis for evolution hardware, and programmable integrated circuits, especially new generation-on-site programmable gate arrays (FPGAs) provide a substance foundation for evolution hardware. .
---- Evolutionary hardware is evolution to calculate a quite promising direction. It has an application in just a few years, but it has been applied in the fields of circuit design, control and robot, pattern identification, fault tolerant system and oversized integrated circuits. For example, after the engine's structure is applied, the external environment can change the strategy of action by the external environment, and the robot can change the strategy of action, and action is more flexible. Although the function of evolution hardware is now relatively simple, it has opened a new way for hardware automation design and online adaptive implementation. As the hardware complexity is increasing, the entire design is completely entered by people, while using evolution hardware, on the one hand, it can greatly reduce the burden on the designer, and on the other hand, the hardware's adaptive requirements are met. When the traditional hardware design method is designed with a large-scale circuit system, there is a comprehensive and accurate understanding of the problem in advance, first decompose the circuit system, then layered. As the complexity of the problem increases, it is quite difficult to fully grasp the entire circuit system. Even if the complex circuit system is designed, it is very difficult to analyze it. With evolution hardware, there is no need to have a deep understanding of the circuit designed, just know that its external characteristics can be. Some people even predict that once the evolution hardware is fully implemented, it will become an emerging industry with extensive application prospects. ---- Q: As an expert in domestic evolutionary calculation research, would you introduce us to the development status of evolution in China?
---- A: The evolution calculation is very fast in China. It is also very high for its application. For example, the Ministry of Agriculture, the Ministry of Machinery is applied to the practical problem that they have encountered in the field. . The research in China is mainly concentrated in three aspects. First, it is the theoretical study of evolutionary algorithms, that is, evolutionary calculations; its second is to engage in evolutionary software, mainly to do automatic programming, compiled by computer; It is a good evolution hardware, which is a very promising field. We are now looking into such a hardware, making it changed from the environment based on the environment, which is mainly using the FPGA chip, and the user can use the program to change the structure of the chip, and the evolutionary algorithm is used to control it. It becomes a wide variety of logic.
---- In general, we have entered the international leading position in certain aspects in this field. If you use differential equations to describe automatic discovery, etc., so far, no one can use often differentials. The equation is such a complex model to move the knowledge, but we can use it to describe complex structures related to time evolution. In terms of overall development, we can say that the development of evolutionary calculations can be said to be on the same horizontal line with abroad, and future development is more unlimited. ◎
(This journalist Yu Dan)
China Computer World Publishing Service Company Copyright
