AI research field

Artificial intelligence research is carried out in combination with specific fields. The main research fields include expert systems, machine learning, pattern recognition, natural language understanding, automatic theorem proof, and automatic programming, robotics, game, and Intelligence decide to support systems and artificial neural networks.

Artificial intelligence is an outward-oriented discipline. It not only requires people who study it to understand the knowledge of artificial intelligence, but also requires a solid mathematical foundation, philosophical and biological foundation, only in this way can a computer without any knowledge simulate human thinking.

Because artificial intelligence has a broad field of research, it is generally oriented to applications, that is, where someone is working, and where it can be used, because the most fundamental purpose of artificial intelligence is to simulate human thinking. Therefore, we can select several representative aspects from many application fields to see what needs to be done in the Development of AI.

Next we will look at the specific application expert system to see what is the main research area of artificial intelligence.

Expert systems are currently the most active and effective research field in artificial intelligence. They are a knowledge-based system that obtains knowledge from human experts, and used to solve the difficult problems that only experts can solve. This defines an Expert System: An Expert System is a program system with a large amount of knowledge and experience in a specific field, it uses artificial intelligence technology to simulate the thinking process of human experts to solve various problems in the field. Its level can reach or even exceed the level of human experts. The expert system was proposed when the research on artificial intelligence was at a low tide. Its appearance and its great potential not only brought artificial intelligence out of the predicament, it has entered a period of development.

The classification of expert systems includes interpretation, diagnosis, prediction, design, planning, control, monitoring, maintenance, education, and debugging, it can be divided into a centralized expert system, a distributed expert system, a neural network expert system, and an expert system that combines a symbolic system with a neural network. There are many names, but the basic structure of the expert system is shown in:

The man-machine interface is just a user interface. Its implementation can be in different forms or complicated. People want to communicate with human experts in the same way, instead of using simple commands. Instead, they want to use human languages to complete interaction. This requires that human-computer interfaces have the function of natural language understanding. But whether the expert system can be used or not is the key layer in the middle. People can think about it. If they want machines to think like humans, then the reasoning mechanism is essential, in addition, it determines the efficiency and availability of the expert system in a large program.

In terms of reasoning, it can generally be dividedPrecise reasoningAndNon-precise reasoning. Precision reasoning has the following features:

• Precision reasoning is based on deterministic knowledge. The knowledge based on precision reasoning is clear, that is, 1 is 1, that is, 2 is 2. There is no fuzzy thing, at one point, precision reasoning has its advantages. That is to say, it can accurately reason. In the process of reasoning, it does not have to worry about whether there will be any questions about the accuracy of conclusions, every step to the next step is completely correct, and there is no possibility of right or wrong. Its correctness is 100% passed to the next reasoning process.

• Precise reasoning differs greatly from human thinking models. human thinking has a precise aspect. However, most human thinking is vague and uncertain, the results of human thinking are often what is possible and probably what is possible, but there is absolutely no such thing in the results of precise reasoning.

• Precision reasoning is a monotonic reasoning, that is, with the addition of new knowledge, the conclusion or proof proposition will only increase monotonically, this is also significantly different from the human thinking structure. New knowledge may increase the outcome of human thinking, but it will never increase monotonically.

• Accurate reasoning requires knowing all the information before it can be used for reasoning. This is obviously different from people. People can make assumptions and inferences based on some situations to generate a result, precise reasoning is impossible.

Because the basis of precise reasoning is classic logic, classic logic can be said to be a symbolicFormal ReasoningIt is concerned with the formal relationship between symbols, rather than the deep semantic relationship between symbols and symbols. This restricts the application of precise reasoning in artificial intelligence. If you want this logic to solve some questions and perform some deterministic work, it is still possible, but it cannot be used to perform more complex work. We can imagine the examples in Machine Translation. Some sentences in human language have no syntax at all, so we have to understand them in terms of semantics. At this time, precise reasoning is not easy to use.

Let's take a look at the non-precision aspect of human thinking. We know that reasoning is a process of thinking that starts from known facts and uses relevant knowledge to draw conclusions or prove that a hypothesis is not true. The knowledge in the expert system comes from human experts in the field, and this kind of knowledge is often subject to uncertainty. In this case, if you still use classical logic for exact reasoning, it is necessary to classify the original uncertainty of the objective thing and the relationship between the objective existence of the matter into certainty, and artificially draw a line between the things that originally do not have a clear similarity relationship, this will undoubtedly discard some important attributes of things, thus losing authenticity. Uncertainty Reasoning is a kind of reasoning based on non-classical logic. It is the application and processing of Uncertainty Knowledge. Strictly speaking, the so-called uncertainty reasoning is based on the initial evidence of uncertainty, through the use of Uncertainty Knowledge, the final launch of a certain degree of uncertainty, but it is a reasonable (or almost reasonable) thinking process.

When uncertainty is to be dealt with, there are several basic problems that do not exist in deterministic reasoning: how to express this uncertainty and how to obtain one (or more conclusions) based on the uncertainty ), in the process of reasoning, how to deal with the uncertainty of the conclusion brought about by uncertainty should we evaluate the result.

• Because the computer is a device for processing numbers, the representation of uncertainty is eventually expressed as a range value, which is conducive to the calculation of uncertainty in conclusion in reasoning.

• There is always a need for reasoning to solve the problem. If the reasoning result can be used or is not a result, a measurement method is required. The measurement method is different from the specific reasoning method, the existing Reasoning Methods basically follow two steps: one is based on the Introduction theory and the other is based on fuzzy mathematics. The former has a long history and has many ready-made results available, however, because probability is the result of a large sample statistics, such large sample statistics are often impossible and fuzzy, therefore, we cannot effectively deal with the ambiguity. The latter overcomes the shortcomings of the former and develops based on the fuzzy set theory, opening up a new way for the determination and acquisition of uncertainty.

Now we are back to the basic structure of our expert system. We know that human reasoning activities are based on certain knowledge, when solving geometric questions, we always need to know some basic principles (or theorems). When a doctor sees a doctor, we need at least a bit of medical knowledge, so that we can have material conditions for reasoning, reasoning is based on knowledge.

Knowledge is the abstraction of facts or facts, which we call a concept. Knowledge is an understanding of certain attributes of objective things. Knowledge has its own characteristics:

• Relative correctness. Any knowledge has a certain scope of application and cannot be used out of scope;

• Uncertainty. Due to the complexity of the real world, many facts and concepts cannot be said to be absolutely correct. Just as there is no absolute truth in philosophy, there is uncertainty in the knowledge itself;

• Expressive. Similarly, knowledge can be expressed, spoken, recorded, and perceived. If it cannot be expressed, who can understand it, even the representation is not displayed, so there is no application. Something that cannot be used. We know or don't know it. It doesn't make sense for application-oriented artificial intelligence.

There are links between facts and concepts, between concepts, and between facts and facts,Static ContactAndDynamic Contact.

• Static Contact. For example, once we mention the concept of "Morning", we will think of the fact that "the sun is rising" or "the cock is called". Such connections are sometimes bidirectional, Which is equivalent, sometimes it is one-way. We can compare "the sun rises East" with "the morning". If we say "cock" to "the morning, it may be wrong.

• Dynamic Contact. In addition to static connections, we must also see that there is also a dynamic relationship between facts and concepts, which is well reflected in machine translation. For example, if we see concept A in the above text and it establishes the relationship between concept B, we must acknowledge that there is a connection between concept A and concept B in the following text, even if this kind of connection does not exist in our lives, as long as we acknowledge it in the above, we must acknowledge its existence. To be more specific, when we were translating a sci-fi novel, the previous article had already said that the "Morning" sun rose from the "North, "The sun rises in the North" and establishes a connection with the concept of "the morning". In this article, we should note that everything about the morning sun rises in the north. However, you must note that this kind of contact cannot be taken to the next article as a static one. If you remember this contact to be taken to the next article, it will be a big headache.

We have already said that knowledge is a description of a certain attribute of objective things. Because objective things are interrelated, knowledge must also be interrelated. This is the essential reason for the existence of Knowledge connections.

Knowledge, as part of machine intelligence, must be able to let machines know what knowledge is. That involves the problem of knowledge representation. This problem is like a different method for people to record a specific thing, for example, it is impossible for a deaf man to associate "Morning" with "Rooster. For a machine, it is simply a deaf, blind, and insensitive stone. It only understands numbers and some artificially defined data structures, so how to make it understand the knowledge, especially the connection in the knowledge, is an important issue. The reasoning system of an expert system is doing well, and it cannot do anything without the knowledge.

At the same time, the expression of Knowledge affects the operation of the reasoning mechanism. The reasoning mechanism and the expression of knowledge are related. A kind of knowledge expression can be conducive to the operation of a reasoning mechanism, the other is not conducive to the operation of this reasoning mechanism. Therefore, when selecting knowledge representation, You must select the corresponding Knowledge Representation Method Based on the specific fields to be processed. The specific knowledge representation includes the following types:

• Logical notation of level 1 predicates. It expresses some knowledge as the predicate expression in the classic logic. Because it is expressed as a predicate, it is convenient to make inferences, however, there is a lot of knowledge that cannot be expressed as a predicate at all. The key reason is that a predicate can only represent accurate knowledge, but cannot effectively represent uncertain things; at the same time, this representation does not reflect the internal connection of knowledge. The task of finding the internal connection of knowledge must be handed over to the reasoning system or another system, which is troublesome.

• Generative notation. Its basic form is similar to the form of our if statement, because it is similar to some existing statements in the computer, so it is much easier to process it. It notices the application scope of connection and knowledge, but it is inherently inadequate in expressing structural knowledge.

• Framework notation. The basic practice is to put a lot of things together to form a set, and then represent the links and facts in the set. This representation method is much more scientific than the first two. In machine translation, if an old lady says

VC, we should not associate it with Microsoft, but be equivalent to vitamin C. This kind of representation limits the appearance of concepts, which may be its deficiency. However, compared with the previous two kinds of representation, it is a kind of representation that is very accepted by human thinking, it virtually reflects the scope of knowledge. More importantly, it can be inherited. At this point, it is closer to human thinking.

• Semantic Network Method. We can imagine that our knowledge system is structured. However, from another perspective, it is a network with universal connection, semantic Network representation represents the networking of the human knowledge system, and it can make full use of contextual reasoning and lay a solid foundation for complicated reasoning. It is very close to human thinking, but it does not correctly represent the class relationship, it reflects the network, but it ignores the class attributes of things. Framework representation and semantic network representation are complementary in this regard.

• Script notation. This kind of representation is applied in natural language understanding because the special nature of natural language understanding requires such a representation. It correctly represents the context, the static and dynamic relationships between things, and fully considers the scenario (context). However, there are too many scenarios in the world, it is almost impossible to save these scenarios. This limits its application scope.

In terms of several methods of knowledge representation, the representation of knowledge is similar to that of humans, and there is a representation that is far different from that of humans. In general, we can see a feature: the representation that is close to human thinking makes it difficult for computers to express themselves, but the representation that is close to machines cannot fully represent the human knowledge structure. There is a difference between machines and people, which may start to encourage people to study new structures of computers and minimize the difference between machines and human thinking. However, because of the current human thinking structure, the structure of the human brain cannot be clearly understood, it is unknown to the extent to which such a machine can narrow the thinking gap between humans and machines. In addition, it seems that it is unrealistic to replace so many computers. Therefore, it is necessary to adopt another method to bring machines closer to human thinking and thinking.

Some people say that artificial intelligence is a database and a search. To some extent, this statement can indeed illustrate the current situation of artificial intelligence. Both the knowledge base and the inference engine involve the search process.

In general, there are two types of Search: Non-heuristic search and heuristic search. Non-heuristic search does not change the search strategy during the search process, and does not use the intermediate information obtained by the search. It is blind and inefficient, and can be used for small problems. It is impossible to use it for large-scale problems; heuristic Search adds question-related information to the search process to guide the search process to a relatively small scope and accelerate the process of obtaining results. We all know that there is an NP-completeness problem in computers, which makes non-heuristic search unavailable in many scenarios. However, although heuristic search uses the intermediate search results, it reduces the search volume, it seems better than non-heuristic search. It is often a headache if the solution obtained is the optimal solution. In general, non-heuristic search requires a rapid increase in search space as the search progresses; heuristic search requires an increase in search space as the search progresses, however, the increase is far smaller than that of non-heuristic search. Some of the problems in the space do not need to be searched because of the acquisition of intermediate information. With the continuous improvement of computer hardware performance and the demand of the actual system, it seems that non-heuristic search cannot be used now. Therefore, non-heuristic search is still widely used in practical applications.

With the search method, we can now see what the search looks like. The data structure determines the algorithm implementation. For the problems we know, we can use the state space or the representation of the or tree to represent a problem space to be searched.

Because of the need of engineering practice, the search results can sometimes be not the optimal solution (sometimes the optimal solution is not the criterion), but the sub-optimal solution, we can think about the many ways that a sentence can be translated in Machine Translation. Where can we talk about what is the best. Therefore, with some similar deep-priority and breadth-priority algorithms we are familiar with, many of the algorithms we are currently studying also have an evolutionary search algorithm, such as genetic algorithms, some of their search methods are independent of the problem, and they can find the optimal solution (or sub-optimal solution) in a short period of time ), it is especially suitable for use when the problem space is relatively high.

Taking genetic algorithms as an example, I think the more important thing is that we don't have to worry about how it is done, but only about what it is. This is the biggest difference from traditional search algorithms. What AI is pursuing is to make machines have intelligence similar to humans. If you can tell a computer what to do, it can do it on its own, instead of telling it how to do it, then artificial intelligence has been implemented.

Now let's go back to the basic structure of the above expert system. With the inference engine and knowledge base, we can implement the user's functions as per the statement, but we should also note that the knowledge acquisition part is another important component, A human expert can only become a human expert, that is, it can continuously enrich its own knowledge in the process of practice, so that its conclusions can be fed back to itself after combined with practice, let yourself modify errors. People are a negative feedback system, and the expert system we mentioned above has no feedback at all. This machine expert is now at this level, in the future, it will be at this level. What it knows will not be changed because of its own practices. Therefore, it cannot meet the actual needs of the project. The knowledge acquisition part is to establish such a feedback mechanism to feedback the obtained results to the knowledge base, modify the known knowledge, and make the results more accurate and available. It would be even better if it could be learned on an instance. Programmers don't have to write individual rules to this expert system, instead, you only need to hand over the instances marked with computers to the computer, and it will generate the knowledge base itself, so that it is more like a person. Therefore, if an expert system has a self-learning function, the system maintenance and system availability will be greatly improved.

Machine Learning has come into being under such a demand. Machine learning methods include:

• Mechanical learning. Another name of rote learning can directly reflect its characteristics. This is the simplest and most primitive learning method. It is also the strength of machines and the weakness of people.

• Guided Learning. This learning method is to provide general instructions or suggestions to the system from the external environment. The system converts them into details and sends them to the knowledge base, in the course of learning, we should repeat the knowledge to improve it.

• Inductive learning. We can see that what machines are good for is not induction, but deduction. It applies from special to general, but not from general to special. from special to general induction is unique to humans, is a sign of wisdom. There are many specific inductive learning methods, but their essence is to allow computers to learn from general patterns.

• Analogy learning. Analogy means learning by comparing similar things. It is based on analogy, that is, comparing new things with old things in memory. If we find that some attributes are the same between them, we can (hypothetical) we can infer that their other attributes are the same.

• Explanation-based learning. This is a new learning method that has emerged in recent years. Instead of learning through induction or analogy, it uses relevant domain knowledge and a training instance to learn a certain target concept and generate a general description of this concept, this general description is a general knowledge that can be formally expressed.

The above learning method is to get knowledge, and to get knowledge through a convenient method. As we have already said before, because machine thinking methods differ greatly from human thinking methods, it allows machines to learn and generate their own knowledge that is easy to understand and use, it is also one of the goals of machine learning.

In the field of AI research, the system mentioned above can be shown as follows:

This system is a type of Negative Feedback System in control, and the result is re-applied to the knowledge base. Therefore, the knowledge base is constantly corrected to meet the needs of the system. However, we have noticed what the result will be if the result is applied to the reasoning opportunity.

The difference between the expert system, inference engine, and machine learning discussed earlier is that we still need to tell them how to do it, instead of simply telling them what to do, they will do it. Two Methods of artificial intelligence research: one is to find the mathematical interpretation of human intelligence. As long as the mathematical interpretation is found, artificial intelligence can be achieved; another is to use a software or hardware structure to simulate the structure of the human brain, and simulate human thinking through a method similar to bionic. Neural Networks are based on the latter idea. In a sense, for a neural network, the result is not the knowledge base, but the structure of the inference engine. It is also an important way to study artificial intelligence.

Neural networks are used to simulate the function of neurons in the human brain. They hope to simulate the function of brain by simulating the basic unit neuron function. It uses a neural network trained by a certain number of examples. Just like teaching a child, after training, the neural network can complete specific functions. It modifies the structure of the knowledge base and inference engine through the learning of examples to achieve the goal of artificial intelligence.

Finally, another application area is model recognition. I think it should be applied in knowledge mining because there are more and more data in the project, it is not easy to manually identify a rule from the data, let alone discover a new rule in the Data. Therefore, it is necessary to perform data mining, its application will be of great significance to the decision support system.

People can think, AI also needs to think, this is reasoning; people can learn, AI also needs to learn; people can possess knowledge, then AI also needs to possess knowledge.

Artificial intelligence is designed to simulate the activity of the human brain. Humans can already use many new technologies and materials to replace many functions of the human body. As long as they simulate the human brain, people can complete the research of human life and create themselves. This is not only scientific, but also epoch-making in philosophy.

Finally, let's summarize the various fields of AI research. With reference to the functions of people in various activities, we can get the field of artificial intelligence, that is, replacing people's activities. In which field is the intellectual activity carried out by someone, and in which field is the field of AI research. Artificial intelligence is to use the strengths of machines to help humans carry out intellectual activities. The purpose of artificial intelligence research is to simulate the functions of human neural systems.