Artificial intelligence research field
Artificial intelligence research is more in combination with specific fields, the main research field has expert system, machine learning, pattern identification, natural language understanding, automatic theorem, automatic programming, robotics, game, intelligent decision support system and manual Neural Networks. Artificial intelligence is an extroverted discipline, which not only requires people to learn about artificial intelligence, but also requires a more solid mathematical foundation, philosophical and biological basis, only this can make one thing I don't know. Machine simulative thinking. Because the research field of artificial intelligence is very broad, it is always for application, but also saying that someone is working, it can be used, because the most fundamental purpose of artificial intelligence is to simulate human thinking. Therefore, we can choose from a number of representative aspects from many applications to see what kind of work needs to develop artificially intelligent development. Below we specialize in the specific application expert system to see what the main research field of artificial intelligence is. Expert system is the most active and effective research area in artificial intelligence. It is a knowledge-based system that has been knowledgeable from human experts and is used to solve difficulties in solving only experts. This definition expert system: Expert system is a program system with a large number of knowledge and experience in a particular field. It applies artificial intelligence technology, simulating human experts to solve the problem of the problem of solving various problems in the field, and its level can achieve even more than Horizontal expertise. The expert system is presented at a low tide on the study of artificial intelligence. It has not only made artificial intelligence to get rid of the dilemma, but also in the development period. The classification of the expert system has interpretation, diagnostic, predictive type, design, plan, control type, monitoring, maintenance, education type, and commissioning, and from the system, it can be divided into centralized expert system, Distributed expert system, neural network expert system, symbolic system and neural network combined expert system. The name is a lot, but the basic structural map of the expert system is shown in the figure below: The human-machine interface part does not explain, it is just a user interface. Its implementation can have different forms, or it may be very complicated. People want to communicate with human experts, no longer use simple commands, but use human language to complete interactions, which requires human-machine interface to be able to understand natural language. However, the expert system can not use it, and use the key in the middle of the middle. People can think, if you want the machine to think like people, then reasoning mechanisms are essential, and it decides on a very large procedure. This expert system is efficient and usable. In reasoning, it can generally be divided into accurate reasoning and non-precise reasoning. Precise reasoning has the following features:
Accurate reasoning is the reasoning of deterministic knowledge, and the accurate reasoning is clear, it is 1 that is 1, 2 is 2, there is no blurry thing, at a point, accurate reasoning has its The strength, that is, it can be accurately reasoning, do not have to pay attention to what conclusions in the process of reasoning, each step to the next step is completely correct, there is no possibility to be wrong, it The correctness is 100% passed to the next reasoning process. Accurate reasoning and human thinking model differ, human thinking is exactly, but most of human thinking is still blurred and uncertain, human thinking results often, how to, but There is absolutely won't have anything possible in the results of accurate reasoning. Precise reasoning is a monotonic reasoning, that is, with the addition of new knowledge, the introduction of the introduction or proven proposition is only monotonous, this point and human thinking structure are also significantly different. New knowledge is likely to increase human thinking results, but it will never be monotonous. Accurate reasoning needs to know all information is probably reasonable, which is obviously different from people, people can conduct some assumptions and inference according to some situations, resulting in a result, but accurate reasoning is impossible. Because the foundation of accurate reasoning is classic logic, and classic logic can be said to be symbolized, it is concerned about the form of symbols and symbols, rather than symbols and deeper semanties after the symbols. It is also because of this, limiting the application of accurate reasoning in artificial intelligence. If this logic solves some questions, some certainty work is made, it is still, but if it makes it more complex work, it will not be heart. We can imagine examples in machine translations, some sentences in human language have no syntax at all, and it is not necessary to understand from semantics. At this time, it is not easy to reason. Let's take a look at another aspect of human thinking, non-precise. We know that the so-called reasoning is from known facts, and gradually introduces the conclusion by using relevant knowledge or proves that a hypothesis is not established. The knowledge in the expert system comes from human experts in the field, and this knowledge often has uncertainties. In this case, if you still use classic logic to do accurate reasoning, you must inevitably have an uncertainty. The objective uncertainty relationship between sexual and things is classified as a certainty, and there is no crossed boundaries between things that have a clear relationship, which will undoubtedly abandon some important properties of things, thus losing truth. Sex. Uncertainty is a reason to build in non-classic logic. It is the use and treatment of uncertainty knowledge. It is strictly that the so-called uncertainty reason is from the initial evidence of uncertainty. Deterministic knowledge, eventually launch has certain degree of uncertain, but is a reasonable (or almost reasonable) conclude of the thinking process. When you want to deal with uncertainty, it involves several basic issues that do not exist in deterministic reasoning: how to indicate this uncertainty, and how to reasonin according to how the uncertainty is reasonable, in the reasoning How to handle the uncertainty of the conclusions caused by uncertainty during the process, if the result is evaluated.
Because the computer is a device for processing a number, it is said that it represents a range of ranges, which is conducive to the calculation of uncertainty in the reason. Solving the problem is always required, and the conclusion of reasoning can not be used, is it a result, there is a need for a measure, a method of measurement, and the specific reasoning method, and the existing reasoning method is basically going. Two roads, one is based on analogism, one is based on fuzzy mathematics, the history of the former development is long, there are many ready-made results available, but since the probability is based on the product of a large sample statistics, This big sample statistics often impossible, but not fully reflects the blur, so it will not be able to handle the fuzziness; and the latter overcome the disadvantages of the former, it develops according to the fuzzy set theory, The judgment of uncertainty and the opening of the new road. Now we return to our expert system basic structural map, we know that human reasoning activities are based on certain knowledge. We always know some basic axiom (or theorem) when we explain the questions. At least a little minimum medical common sense, so that the material conditions for reasoning are based on the establishment of knowledge. Knowledge is an abstraction of facts or facts, we call it concept. Knowledge is an understanding of the property of objective things. Knowledge has its characteristics: relative correctness. Any knowledge has a certain range of applications, and it is not available to the scope of the scope; uncertainty. Due to the complexity of the real world, many facts and concepts can not be said to be absolutely correct, just like there is no absolute truth in philosophy, knowledge itself has uncertainty; representative. Similarly, if knowledge can be expressed, it is also necessary to use anything. If you want to be able to perceive, if you can't expressed, then who can understand, even expressed, it will not be applied. . Things we can't use, we know, don't know, there is no significance for applicable artificial intelligence. The facts and concepts between the concepts, the concepts, the facts and facts are related, and this connection has two, static contacts and dynamic contacts.
Static contact. For example, once we mentioned the concept of "morning", it will remember the fact that "Sun Dongsheng" or "Census", this connection is sometimes two-way, which is equivalent, and sometimes one-way, we "Sun Dongsheng" is equivalent to the "morning", and if we put "cock" and "morning", ten eight nine may be wrong. Dynamic contact. In addition to static contacts, we must also see that there will be a dynamic link between facts and concepts, which is very good in machine translation. For example, I saw a concept A, which established the connection between concept B, then we must admit that between A and B, even in life, as long as this connection is in life In the above, this connection has been recognized, we must recognize the existence of this connection. More details, when we translated an idea of science fiction, the foregoing has said that the "morning" sun is from "rising in the north", then "the sun is rising" and "the morning" concept has been established. In the translation of this article, it is necessary to pay attention, and all the sun in the morning is from "raising the north." However, it should be noted that this connection is not to be a static connection, and if this contact remembers the next article, it will be more troublesome.
As mentioned earlier is a description of an objective thing is an objective thing, and it is because objective things are interrelated, so knowledge must also be interrelated. This is the essence of the existence of knowledge links. Knowledge As part of machine intelligence, you must be able to let the machine know what is knowledge, then involve a problem with knowledge, this question is like a different method of people, such as for the deaf, you let him Contact "Morning" and "Census" together, it is not possible. For the machine, it is simply a stone that is still, and there is no feeling (more talks about feelings), it only understands the data structure specified by the figures and some people, so how to make it understand knowledge, especially The connection in knowledge is an important issue. An expert system's reasoning system is better, and there is no knowledge as a backing. At the same time, the expression of knowledge is also affecting the operation of the reasoning mechanism, the reasoning mechanism and knowledge representing the two, one of the knowledge representation can be conducive to the operation of a reasoning mechanism, and the other is not conducive to this reasoning mechanism. Operation, thus selecting knowledge representation must also choose the corresponding knowledge representation according to the specific area to be processed, and the specific knowledge represents the following: first-order word logic representation. It is a predicate representation of some knowledge as a classic logic, because it is in the form of a predicate, and it is of course convenient for reasoning, but there are many knowledge can't be expressed as predicates, and the key reason is because the predicate is Executive knowledge is expressed, and the uncertain things cannot be effectively expressed; and this representation does not reflect the inherent linkage of knowledge, finding the task within the knowledge to be handed over to the reasoning system or additional system, this It is more troublesome. Production expression. Its basic form is similar to the form of our IF statement because some ready-made statements are similar to some of the computer, so it is more convenient to handle it. It noted the application range of contact and knowledge, but it is in a congenital understanding of structural knowledge. Frame representation. Its basic practice is to put many things together, constitute a collection, then expressed in this collection and facts, this representation is much more scientific than the first two science. In the machine translation, if an old lady said VC, do we have to link it with Microsoft, and you must be with Vitamin C. This indication limits the occasion of the concept, which may be its shortcomings, but relative to the previous two representations, it is a representation of human thinking, which embodies the knowledge applicable. range. More importantly, it can be inherited, at this point, it is more close to people's thinking. Semantic Network Method. We can imagine the knowledge system they own, there is a structure, but in another point of view, it is a network, universally contacted the network, and the semantic network representation is in the network of human knowledge systems. One side, and it enables losing reasoning to get a good performance on it, and laid a solid foundation for complex reasoning. It is very close to human thinking, but it does not correctly represent the classic relationship, which reflects network, but it ignores the class properties of things. Frame representation and semantic network representation are complementary in this. Script representation. This representation begins to be applied in terms of natural language understanding, because of the particularity requirements of natural language understanding, there is a representation. It correctly expressed the contextual relationship, indicating the static relationship between things, dynamic relationships, taking into account the scene (context), but the world's scenes can be too much, and it is almost impossible to save these scenes. This limits its application range.
In view of the knowledge representation, the expression of knowledge is close to human beings, and there is a method of representing the difference with humanity, and we can see a feature: approaching human thinking, Let the computer represent a certain amount of trouble, but the representation of the machine is close to the machine, but it cannot fully represent the human knowledge structure. There is a difference between machines and people, which may begin to develop a computer that starts to study the new structure, allowing the difference between machines and human thinking to minimize. But because the current human thinking structure, the structure of the human brain can not be very clear, so it is unknown to whisper to the thoughts of people and machines on such a machine. Moreover, it is currently, replacing so many computers are also unrealistic, so there is a need to use additional methods to make machines and thinking and human thinking more close. Some people say that artificial intelligence is a database plus search. From a certain extent, this sentence can indeed explain the status quo of artificial intelligence. Whether it is in the knowledge base, it is still in the reasoning machine, it should be related to the process. In general, search is divided into two kinds, one is a non-heuristic search, and the other is heuristic search. Non-heuristic search does not change the search strategy during the search process, which disabate the intermediate information obtained by the search, which blindly is very large, the efficiency is poor, used for small issues, it is impossible to use for large problems; During the search process, the inspiration information related to the problem is added to guide the search to the process of being performed in a relatively small range and accelerate the resulting result. We all know that there is an NP complete problem in the computer, which is precisely because this makes non-heuristic search unavailable in many occasions, but the heuristic search is used to reduce the search quantity, it looks better than Non-heuristic search, what is the solution it is not the best solution that is often a headache. In general, non-heuristic searches take advantage of the search, the space you need to search quickly; heuristic searches With the search, the space you need to search, but the increase is far less than non-inspiration Search. Some places in the problem space are not searching for the acquisition of intermediate information. With the continuous improvement of computer hardware performance and the actual system's needs, it is now possible to use non-heuristic searches, so non-heuristic search is still very wide in practical applications.
With a search method, then we can now look at what you want to search. Data Structure Decision Algorithm. For problems we know, we can use state space or representation of the representation of or trees to represent a problem space to be searched. Due to engineering practices, the results of the search can sometimes be the best solution (sometimes the optimal solution has not decided), but the secondary solution, we can think of the translation of the machine translation may have many kinds of translations. Where is the optimal thing about it? Therefore, there are many similar depth priorities we are familiar with, and there are many evolutionary search algorithms, such as genetic algorithms, analog 煺 fire algorithms, etc., and other search methods are independent of the problem. And it is possible to find the optimal solution (or secondary solution) in a short period of time, which is particularly adapted to use the problem space comparison. Take the genetic algorithm, I think more importantly, we don't have to care about how it does, but just care about it, this is the greatest difference from traditional search algorithms. The artificial intelligence is pursuing to make the machine have a similar human intelligence. If you can tell a computer to do what, it can do it yourself, and you don't need to tell it how to do it, the artificial intelligence has been implemented. We will now return to the basic structural map of the above expert system. We have a throttle, with a knowledge base, can achieve the user's functionality, but we should also notice that another important part, knowledge acquisition, one human expert, so it can be an expert, it is it In the practice process, we have continuously enriched our knowledge, let the conclusions of yourself feed back themselves after combining practice, let yourself modify the mistake, people are a negative feedback system, and the expert system that we have not mentioned above There is no feedback at all, this machine expert is now this level, will be this level in the future, what it knows, does not change because of his own practice. So it cannot apply practical needs. The job of knowledge is to establish such a feedback mechanism, feedback the result feedback to the knowledge base, modify the known knowledge, allowing it to get more accurate, more availability. If it makes it possible to learn, it is better, the programmer can write to the expert system without writing a rule of a bar, but only hand it over to the computer with a computer that is tagged with a computer. It will generate a knowledge base so that it is more like a person. Therefore, if an expert system has its own function, it will be greatly improved in terms of system maintenance and system availability. Machine learning is produced under such demand. The method of machine learning has the following:
Mechanical learning. Its another name deadbook can directly reflect its characteristics, this is the simplest, most primitive learning method, and the strength of the machine, people's weaknesses. Guidance learning. This way of learning is to provide general instructions or suggestions to the system to the system. The system is specifically converted into detail knowledge and sent into the knowledge base. In the learning process, it is necessary to evaluate the knowledge in the process of repeated knowledge, so that it continues to improve . Summary study. We see that the machine is not inductive, but interpretation, it is applicable to the special to general, not adaptable from general to special, from special to general summary is a human, is a sign of wisdom. There are many specific induction methods, but their essence is to make computer learning regularly. Category ratio. The class ratio is also a learning that is compared to similar things. Its basis is a ratio, that is, comparing new things and old things in memory. If some attributes are found, some of them can be subsequently inferred in their other properties. Explanation. This is a new way of learning in recent years. It is not to learn by summarizing or subjectivity, but by using the relevant domain knowledge and a training instance, it is a general description of this goal concept, and this general description is a formal representation. General knowledge. Through the above learning methods are to get knowledge, they have known knowledge by a convenient way. As mentioned earlier, because the mechanical thinking and human thinking methods are different, so that the machine is to learn from themselves to understand and use knowledge, and it is not one of the goals of machine learning. In the field of artificial intelligence, the system can be said to be the following, the system type is in the controlled feedback system in the control, and the results are re-acting on the knowledge base, so the knowledge base is constantly Correct to adapt to the needs of the system. But we noticed what results obtained if the result is applied to the reasoning opportunity. The expert system we discussed earlier, the reasoning machine is also good, machine learning is good, and the gap between people is what we also need to tell them, not just tell them what, they will do it. Two research methods of artificial intelligence, one is to find the mathematical explanation of human intelligence, as long as you find mathematical explanation, then artificial intelligence can be achieved; the other is to simulate the human brain with a structure of software or hardware. The structure is simulated human thinking by a similar method of imitation. The neural network is based on the latter idea. In a sense, for the neural network, the result is that it is not a knowledge base, but the structure of the reasoning machine, it is also an important way to study artificial intelligence. The neural network is also the function of simulating the neurons in the human brain, and it is desirable to simulate the function of the human brain by simulating the most basic unit neuron function of the human brain. It is like a neural network composed of a certain example training, just like teaching a child, after training, this neural network can complete a specific function. It is the study of the examples of the example, modify the structure of the knowledge base and the thrilling machine to achieve the purpose of realizing artificial intelligence. Finally, there is an application area, which is the model identification. I think it should be applied in the knowledge excavation, because the data obtained in the project is getting more and more, I want to be able to determine a certain law from these data. Easy, not to mention the new laws in these data, therefore it is necessary to perform data mining, and its application will have a huge significance for decision support systems. People can think, artificial intelligence also needs to think, this is reasonable; people can learn, artificial intelligence needs to learn; people can have knowledge, then artificial intelligence requires knowledge.
