APPROACHES AND TECHNIQUES
Read the following words and word combinations and use them for understanding and translation of the text:
conventional - общепринятый, традиционный
computational intelligence - вычислительный интеллект
machine learning - машинное обучение
case-based reasoning - вывод (рассуждения), основанные на прецедентах
behavior-based AI - поведенческий ИИ
referred to as - под названием, именуемый
neural networks - нейронные сети
fuzzy logic - нечеткая логика
neats versus scruffies - чистюли против нерях
ad hoc rules - ситуативные правила
inference engine - механизм логического вывода
forward chaining - прямой логический вывод
backward chaining - обратный логический вывод
directed acyclic graph - ориентированный циклический граф
arc - дуга
conditional dependence - условная зависимость
to be subject to controversy - вызывать споры
track record - послужной список, достижения
The artificial intelligence community can be roughly divided into two schools of thought: conventional AI and computational intelligence. Conventional AI is based on machine learning, which is the development of the techniques and algorithms that allow machines to “learn” or at least simulate learning. Machine learning attempts to use computer programs to generate patterns or rules from large data sets. This problem is similar to the data-mining problem (and data mining is one area where AI has found commercial success). Machine learning makes heavy use of symbolic formalism and logic, as well as statistics. Key areas in conventional AI include case-based reasoning, behavior-based AI, Bayesian networks, and expert systems. Computational intelligence, in contrast, relies more on clever algorithms (heuristics) and computation and less on formal logical systems. Computational intelligence is sometimes referred to as soft computing. It often involves iterative methods using computation to generate intelligent agents. Whereas conventional AI is considered to be a top-down approach, with the structure of solutions imposed from above, computational intelligence is more bottom-up, where solutions emerge from an unstructured initial state. Two areas of computational intelligence will be discussed further: neural networks and fuzzy logic. Hybrid intelligent systems attempt to combine the two approaches. Some proponents claim that this is appropriate, because the human mind uses multiple techniques to develop and verify results, and hybrid systems show some promise.
Another distinction within the artificial intelligence community is weak AI versus strong AI. Weak AI refers to using software to solve particular problems or reasoning tasks that do not encompass fully human intelligence. Strong AI implies creating artificial systems that are fully self-aware, the systems that can reason and independently solve problems. Current research is nowhere near creating strong AI, and a lively debate is ongoing as to whether this is even possible.
Another division in the artificial intelligence community is over the best way to design an intelligent system (Neats versus Scruffies). The Neats maintain that the solution should be elegant, obvious, and based on formal logic. The Scruffies hold that intelligence is too messy and complicated to be solved under the limitations the Neats propose. Interestingly, some good results have come from hybrid approaches, such as putting ad hoc rules (Scruffy style) into a formal (Neat) system. Not surprisingly, the Neats are often associated with conventional artificial intelligence, whereas the Scruffies are usually associated with computational intelligence.
Conventional AI has achieved success in several areas. Expert systems, or knowledge-based systems, attempt to capture the domain expertise of one or more humans and apply that knowledge. Most commonly, this is done by developing a set of rules that analyze information about a problem and recommend a course of action. Expert systems demonstrate behavior that appears to show reasoning. Expert systems work best in organizations with high levels of know-how and expertise that are difficult to transfer among staff. The simpler expert systems are all based on binary logic, but more sophisticated systems can include methods such as fuzzy logic. At the heart of an expert system is an inference engin e, a program that attempts to create answers from the knowledge base of rules provided by the expert. Knowledge engineers convert a human expert’s “rules-of-thumb” into inference rules, which are if-then statements that provide an action or a suggestion if a particular statement is true. The inference engine then uses these inference rules to reason out a solution. Forward chaining starts with the available information and tries to use the inference rules to generate more data until a solution is reached. Backward-chaining starts with a list of solutions and works backward to see if data exists that will allow it to conclude that any of the solutions are true. Expert systems are used in many fields, including finance, medicine, and automated manufacturing.
Another approach from conventional AI that has achieved some commercial success is case-based reasoning, or CBR, which attempts to solve new problems based on past solutions of similar problems. Proponents argue that case-based reasoning is a critical element in human problem solving. As formalized in computer reasoning, CBR is composed of four steps: retrieve, reuse, revise, retain. First, access the available information about the problem (Retrieve). Second, try to extend a previous solution to the current problem (Reuse). Next, test the refactored solution and revise it if necessary (Revise). Finally, store the new experience into the knowledge base (Retain).
Behavior-based artificial intelligence (BBAI) attempts to decompose intelligence into a set of distinct, semi-autonomous modules. BBAI is popular in the robotics field and is the basis for many Robocup robotic soccer teams, as well as the Sony Aibo. A BBAI system is composed of numerous simple behavior modules, which are organized into layers. Each layer represents a particular goal of the system, and the layers are organized hierarchically. A low layer might have a goal of “avoid falling,” whereas the layer above it might be “move forward.” The move forward layer might be one component of a larger “walk to the store” goal. The layers can access sensor data and send commands to the robot’s motors. The lower layers tend to function as reflexes, whereas the higher layers control more complex goal-directed behavior.
Bayesian networks are another tool in the conventional AI approach. They are heavily based upon probability theory. The problem domain is represented as a network. This network is a directed acyclic graph where the nodes represent variables, and the arcs represent conditional dependences between the variables. Graphs are easy to work with, so Bayesian networks can be used to produce models that are simple for humans to understand, as well as effective algorithms for inference and learning. Bayesian networks have been successfully applied to numerous areas, including medicine, decision support systems, and text analysis, including optical character recognition.
There is no widespread agreement yet on exactly what Computational intelligence (CI) is, but it is agreed that it includes neural networks and fuzzy computing. A neural network consists of many nodes that cooperate to produce an output. The system is trained by supplying input on the solution of known problems, which changes the weighting between the nodes. After training has tuned the parameters between the connections, neural networks can solve difficult problems in machine vision and other areas. Also known as neurocomputing, or parallel distributed processing, neural networks loosely model structures in the human brain. Neural network outputs rely on the cooperation of individual nodes. Data processing in neural networks is typically done in parallel, rather than sequentially as is the standard for nearly all modern computers. Neural nets can generalize from their training, and solve new problems, so they are self-adaptive systems. Neural networks have been criticized as “bad science” because it is difficult to explain exactly how they work. Nonetheless, neural networks have been successfully applied in areas as diverse as credit card fraud detection, machine vision, chess, and vehicle control.
Fuzzy logic, fuzzy systems, and fuzzy set theory are all ways to refer to reasoning that is based upon approximate values, rather than precise quantities. Modern computers are built upon binary, or Boolean, logic that is based on ones and zeros. The bit is zero or one, yes or no, with no middle ground. Fuzzy systems provide for a broader range of possible values. Consider the question, “Are the books in the study?” Well, yes, there are books in the study. There are also books in the office, books in the bedroom, and a pile of books in the doorway to the study. Fuzzy logic provides for an answer of 72%, meaning that 72% of the books are in the study. Fuzzy sets are based on vague definitions of sets. They are not random. Fuzzy logic is not imprecise; rather, it is a formal mathematical technique for handling imprecise data. Like neural networks, fuzzy logic is subject to controversy and criticism. But systems based on fuzzy logic have an excellent track record at certain types of problems. Antilock braking systems are based on fuzzy logic, and many appliances incorporate fuzzy logic.
Notes:
Bayesian network (Bayesian network, Bayes network, belief network, Bayes(ian) model or probabilistic directed acyclic graphical model) is a probabilistic graphical model (a type of statistical model) that represents a set of random variables and their conditional dependencies via a directed acyclic graph (DAG).
Robocup is an international robotics competition that aims to develop autonomous robots with the intention of developing research and education in the field of artificial intelligence. The best universities in the world compete in several leagues.
AIBO (Artificial Intelligence robot) is a robotic project from Sony. In Japanese, AIBO means pal or partner. AIBO was one of several types of robotic pets that were designed and manufactured by Sony. Sony Aibo is basically a robotic dog that that is able to walk and “see” its environment using the on board cameras. It is even able to recognize spoken commands in languages including Spanish and English.
Assignments
1. Translate the sentences from the texts into Russian in writing paying attention to the underlined words and phrases:
1. Machine learning makes heavy use of symbolic formalism and logic, as well as statistics.
2. Whereas conventional AI is considered to be a top-down approach, with the structure of solutions imposed from above, computational intelligence is more bottom-up, where solutions emerge from an unstructured initial state.
3. Some proponents claim that this is appropriate, because the human mind uses multiple techniques to develop and verify results, and hybrid systems show some promise.
4. Weak AI refers to using software to solve particular problems or reasoning tasks that do not encompass fully human intelligence.
5. Current research is nowhere near creating strong AI, and a lively debate is ongoing as to whether this is even possible.
6. The Neats maintain that the solution should be elegant, obvious, and based on formal logic. The Scruffies hold that intelligence is too messy and complicated to be solved under the limitations the Neats propose.
7. Expert systems, or knowledge-based systems, attempt to capture the domain expertise of one or more humans and apply that knowledge.
8. There is no widespread agreement yet on exactly what Computational intelligence (CI) is, but it is agreed that it includes neural networks and fuzzy computing.
2. Answer the following questions:
1. What are the ways to classify Artificial Intelligence?
2. How does an expert system work?
3. What are the four steps in case-based reasoning?
4. What tasks do the layers of a BBAI system perform?
5. Where are Bayesian networks applied?
6. What are the working principles of a neural network?
7. How does fuzzy logic differ from Boolean logic?
3. Translate into English:
Эпименид Кносский с острова Крит – полумифический поэт и философ, живший в VI в. до н.э., однажды заявил: «Все критяне – лжецы!». Так как он и сам был критянином, то его помнят как изобретателя так называемого критского парадокса.
В терминах аристотелевой логики, в которой утверждение не может быть одновременно истинным и ложным, и подобные самоотрицания не имеют смысла. Если они истинны, то они ложны, но если они ложны, то они истинны.
И здесь на сцену выходит нечеткая логика, где переменные могут быть частичными членами множеств. Истинность или ложность перестают быть абсолютными – утверждения могут быть частично истинными и частично ложными. Использование подобного подхода позволяет строго математически доказать, что парадокс Эпименида ровно на 50% истинен и на 50% ложен.
Таким образом, нечеткая логика в самой своей основе несовместима с аристотелевой логикой, особенно в отношении закона Tertium non datur («Третьего не дано» – лат.), который также называют законом исключения среднего. Если сформулировать его кратко, то звучит он так: если утверждение не является истинным, то оно является ложным. Эти постулаты настолько базовые, что их часто просто принимают на веру.
Более банальный пример пользы нечеткой логики можно привести в контексте концепции холода. Большинство людей способно ответить на вопрос: «Холодно ли вам сейчас?». В большинстве случаев люди понимают, что речь не идет об абсолютной температуре по шкале Кельвина. Хотя температуру в 0 K можно, без сомнения, назвать холодом, но температуру в +15 C многие холодом считать не будут.
Но машины не способны проводить такую тонкую градацию. Если стандартом определения холода будет «температура ниже +15 C», то +14,99 C будет расцениваться как холод, а +15 C – не будет!
