|
What is Computational Intelligence (CI) and what are its relations with Artificial Intelligence (AI)? A brief survey of the scope of CI journals and books with “computational intelligence” in their title shows that at present it is an umbrella for three core technologies (neural, fuzzy and evolutionary), their applications, and selected fashionable methods. At present CI has no comprehensive foundations and is more a bag of tricks than a solid branch of science. The change of focus from methods to challenging problems is advocated, with CI defined as a part of computer science devoted to solution of non-algoritmizable problems. In this view AI is a part of CI focused on problems related to higher cognitive functions, while the rest of the CI community works on problems related to perception and control, or lower cognitive functions. Grand challenges on both sides of this spectrum are addressed. Introduction What exactly is Computational intelligence (CI)? How is it related to other branches of computer science, such as artificial intelligence (AI), classification, cognitive informatics, connectionism, data mining, graphical methods, intelligent agents and intelligent systems, knowledge discovery in data (KDD), machine intelligence, machine learning, natural computing, parallel distributed processing, pattern recognition, probabilistic methods, soft computing, multivariate statistics, optimization and operation research? This is a very confusing issue, hotly debated, but with no consensus in sight. Computational intelligence became a new buzzword that means different things to different people. Branches of science are not defined, but slowly develop in the process of sharing and clustering of common interests. In CI these interest generally focus on problems that only humans and animals can solve, problems requiring intelligence. Specific interests also focus on methods and tools that are applicable to this type of problems. Starting with seminal s, special sessions, growing into separate conferences and specialized journals, different branches of CI evolve in many directions, frequently quite far from original roots and inspirations. New communities are formed and need to establish their identity by defining borders distinguishing them from other scientific communities. Artificial Intelligence (AI) was the first large scientific community, established already in the mid 1950s, working on problems that require intelligence to be solved. Its evolution has been summarized in the 25th anniversary issue of the AI Magazine by Mackworth [1]: “In AI’s youth, we worked hard to establish our paradigm by vigorously attacking and excluding apparent pretenders to the throne of intelligence, pretenders such as pattern recognition, behaviorism, neural networks, and even probability theory. Now that we are established, such ideological purity is no longer a concern. We are more catholic, focusing on problems, not on hammers. Given that we do have a comprehensive toolbox, issues of architecture and integration emerge as central.” IEEE Computational Intelligence Society defines its subjects of interest as neural networks, fuzzy systems and evolutionary computation, including swarm intelligence. The approach taken by the journals and by the book authors is to treat computational intelligence as an umbrella under which more and more methods will be added. A good definition of the field is therefore impossible, because different people include or exclude different methods under the same CI heading. Chess programs based on heuristic search are already in the superhuman computational intelligence category, but they do not belong to CI defined in such a way. In the early days of CI some experts tried to explicitly exclude such problems. Take for example this definition: “A system is computationally intelligent when it: deals only with numerical (low level) data, has a pattern recognition component, and does not use knowledge in the AI sense” [2]. As in the case of AI the need to create strong identity by emphasizing specific methods defining Computational Intelligence as a field should be replaced by focus on problems to be solved, rather than hammers. Below some remarks on the current state of CI are made, based on analysis of journals and books with “computational intelligence” in their title. Then a new definition of CI is proposed and some remarks on what should computational intelligence really be are made. Finally grand challenges to computational intelligence are discussed. Grand Challenges to Computational Intelligence A number of grand challenges for AI has been formulated, starting with the famous Turing Test for machine intelligence, This requires a very-large knowledge base and efficient retrieval of structures. While the CyC project [18] has created such knowledge base manually coding it over a period of more than 30 years the retrieval mechanisms that it offers are too inefficient to use it in large-scale dialog systems. A grand challenge for CI community is to propose more efficient knowledge representation and retrieval structures, perhaps modeled on the associative memory of the brain, perhaps using different knowledge representations for different purposes [19]. Vector and similaritybased models cannot yet replace complex frames in reasoning processes. Semantic networks, although in principle could provide efficient association and inference mechanisms, have never been used on a large scale. Feigenbaum [20] proposed a reasoning test which should be simpler for computers than the Turing Test. Instead of a general dialog that has to be based on extensive knowledge of the world, this test is based on the expert knowledge in a narrow domain. Reasoning in some field of mathematics or science by human expert and artificial system is evaluated by another expert in the same field who is posing problems, questions, and asking for explanations. This could be achieved with super-expert systems in various domains, giving some measures of progress towards intelligent reasoning systems. The World Championship for 1st Order Automated Theorem Proving organized at the Conference on Automated Deduction (CADE) could be organized not only between computers, but could also involve humans, although much longer time to complete the proofs may be required. Other grand AI challenges [20] are concerned with large-scale knowledge bases, bootstraping on the knowledge resources from the Internet and creating semantic Internet. The 20-questions game could also be a good challenge for AI, much easier than the Turing test, requiring extensive knowledge about objects and their properties, but not about complex relations between objects. In fact some simple vectorspace techniques may be used to play it [19], making it a good challenge not only for AI, but also for the broader CI community. What would be a good grand challenge for non-AI part of computational intelligence? This has been the subject of a discussion panel on the challenges to the CI in the XXI century, organized at the World Congress on Computational Intelligence in Anchorage, Alaska, in 1998. The conclusion was that a grand challenge for CI is to build an artificial rat, an artificial animal that may survive in a hostile environment. The intermediate steps require solution to many problems in perception, such as object recognition, auditory and visual scene analysis, spatial orientation, memory, motor learning, behavioral control, but also some reasoning and planning. The ultimate challenge may be to build not only an animal, but a human-like system that in addition to survival will be able to pass the Turing test. Imagine the future in which superintelligence based on some form of computations has been realized. What we would like it to do and to be? In the long run everything seems to be possible. Computational intelligence should be human-centered, helping humans not only to solve their problems, but also to formulate meaningful goals, leading to a true personal fulfillment. It should protect us starting from birth, not only monitoring the health hazards, but also observing and guiding personal development, gently challenging children at every step to reach their full physical as well as mental potential. It should be a technology with access to extensive knowledge, but it also should help humans to make wise decisions presenting choices and their possible consequences. Although it may seem like a dangerous utopia perhaps deeper understanding of developmental processes, cognitive and emotional brain functions, real human needs, coupled with a technology that can recognize behavioral patterns, make sense of observations, understand natural language, plan and reason with extensive background knowledge, will lead to a better world in which no human life is wasted. Intelligence with wisdom is perhaps an ultimate goal for human-oriented science. Such utopia is worth dreaming of, although we are still very far from this level (see some speculations on this topic in [21–23]). A long-term goal for computational intelligence is to create cognitive systems that could compete with humans in large number of areas. So far this is possible only in restricted domains, such as recognition of specific patterns, processing of large amount of numerical information, memorization of numerous details, high precision control with small number of degrees of freedom, and reasoning in restricted domains, for example in board games. Brains are highly specialized in analysis of natural patterns, segmentation of auditory and visual scenes, and control of body movements, mapping perceptions to actions. Despite great progress in computational intelligence artificial systems designed to solve lower level cognitive functions are still far behind the natural ones. Situation is even worse when higher-level cognitive functions, involving complex knowledge structures necessary for understanding of language, reasoning, problem solving or planning, are considered. Human semantic and episodic memory is vastly superior to the most sophisticated artificial systems, storing complex memory patterns and rapidly accessing them in an associative way. So far CI understood as a collection of different methods had no clear challenges of the AI magnitude. Improving clusterization, classification and approximation capabilities of CI systems is incremental and there are already so many methods that it is always possible to find alternative solutions. At the technical level fusion of different CI techniques is considered to be a challenge, but attempts to combine evolutionary and neural methods, to take just one example, have a long history and it is hard to find results that are significantly better than those achieved by other techniques. The challenge is at the meta-level, to find all interesting solutions automatically, especially in difficult cases. Brains are flexible, and may solve the same problem in many different ways. Different applications – recognition of images, handwritten characters, faces, analysis of signals, mutimedia streams, texts, or various biomedical data – usually require highly specialized methods to achieve top performance. This is a powerful force that leads to compartmentalization of different CI branches and creation of meta-learning systems competitive with the best methods in various applications will be a great challenge. If we acknowledge that CI should be defined as the science of solving non-algorithmizable problems the whole field will be firmly anchored in computer science and many technical challenges may be formulated. Focusing on problems instead of tools will allow for greater integration of AI community, and enable competition with other methods for various applications, facilitating real progress towards more difficult problems. Broad foundations for CI that go beyond pattern recognition need to be constructed, including solving problems related to the higher cognitive functions (see [24], this volume). Inspirations drawn from cognitive and brain sciences, or biology in general, will continue to be very important, but at the end of the road CI will become a solid branch of computer science.() 网站原创范文除特殊说明外一切图文作品权归所有；未经官方授权谢绝任何用途转载或刊发于媒体。如发生侵犯作品权现象，英语毕业论文，英语论文题目，保留一切法学追诉权。() 更多范文欢迎访问我们主页 当然有需求可以和我们 联系交流。-X() |
