| The term Artificial Intelligence (AI) was
| |
| | as the 'No free lunch theorem'. Various
|
| first used by John McCarthy who used it
| |
| | empirical tests have been performed to
|
| to mean "the science and engineering of
| |
| | compare classifier performance and to
|
| making intelligent machines". It can also
| |
| | find the characteristics of data that
|
| refer to intelligence as exhibited by an
| |
| | determine classifier performance.
|
| artificial (man-made, non-natural,
| |
| | Determining a suitable classifier for a
|
| manufactured) entity. The terms strong
| |
| | given problem is however still more an
|
| and weak AI can be used to narrow the
| |
| | art than science.
|
| definition for classifying such systems.
| |
| | The most widely used classifiers are the
|
| AI is studied in overlapping fields of
| |
| | neural network (multi-layer perceptron),
|
| computer science, psychology, philosophy,
| |
| | support vector machines, k-nearest
|
| neuroscience and engineering, dealing
| |
| | neighbors, Gaussian mixture model,
|
| with intelligent behavior, learning and
| |
| | Gaussian, naive Bayes, decision trees and
|
| adaptation and usually developed using
| |
| | radial basis functions. Van der Walt and
|
| customized machines or computers.
| |
| | Barnard[2] investigated very specific
|
| Research in AI is concerned with
| |
| | artificial data sets to determine
|
| producing machines to automate tasks
| |
| | conditions under which certain
|
| requiring intelligent behavior. Examples
| |
| | classifiers perform better and worse than
|
| include control, planning and scheduling,
| |
| | others.
|
| the ability to answer diagnostic and
| |
| | Conventional AI
|
| consumer questions, handwriting, natural
| |
| | Conventional AI mostly involves methods
|
| language, speech and facial recognition.
| |
| | now classified as machine learning,
|
| As such, the study of AI has also become
| |
| | characterized by formalism and
|
| an engineering discipline, focused on
| |
| | statistical analysis. This is also known
|
| providing solutions to real life
| |
| | as symbolic AI, logical AI, neat AI and
|
| problems, knowledge mining, software
| |
| | Good Old Fashioned Artificial
|
| applications, strategy games like
| |
| | Intelligence (GOFAI). (Also see
|
| computer chess and other video games. One
| |
| | semantics.) Methods include:
|
| of the biggest difficulties with AI is
| |
| | Expert systems: apply reasoning
|
| that of comprehension. Many devices have
| |
| | capabilities to reach a conclusion. An
|
| been created that can do amazing things,
| |
| | expert system can process large amounts
|
| but critics of AI claim that no actual
| |
| | of known information and provide
|
| comprehension by the AI machine has taken
| |
| | conclusions based on them.
|
| place.
| |
| | Case based reasoning: stores a set of
|
| enerally speaking AI systems are built
| |
| | problems and answers in an organized data
|
| around automated inference engines. Based
| |
| | structure called cases. A Case Based
|
| on certain conditions ("if") the system
| |
| | Reasoning system upon being presented
|
| infers certain consequences ("then"). AI
| |
| | with a problem finds a case in its
|
| applications are generally divided into
| |
| | knowledge base that is most closely
|
| two types, in terms of consequences:
| |
| | related to the new problem and presents
|
| classifiers ("if shiny then diamond") and
| |
| | its solutions as an output with suitable
|
| controllers ("if shiny then pick up").
| |
| | modifications.
|
| Controllers do however also classify
| |
| | Bayesian networks
|
| conditions before inferring actions and
| |
| | Behavior based AI: a modular method
|
| therefore classification form a central
| |
| | building AI systems by hand.
|
| part of most AI systems.
| |
| | Computational intelligence
|
| Classifiers make use of pattern
| |
| | Computational intelligence involves
|
| recognition for condition matching. In
| |
| | iterative development or learning (e.g.
|
| many cases this does not imply absolute,
| |
| | parameter tuning e.g. in connectionist
|
| but rather the closest match. Techniques
| |
| | systems). Learning is based on empirical
|
| to achieve this divides roughly into two
| |
| | data and is associated with non-symbolic
|
| schools of thought: Conventional AI and
| |
| | AI, scruffy AI and soft computing.
|
| Computational intelligence (CI)[citation
| |
| | Methods mainly include:
|
| needed]
| |
| | Neural networks: systems with very strong
|
| Classifiers
| |
| | pattern recognition capabilities.
|
| Classifiers are functions that can be
| |
| | Fuzzy systems: techniques for reasoning
|
| tuned according to examples, making them
| |
| | under uncertainty, have been widely used
|
| very attractive for use in AI. These
| |
| | in modern industrial and consumer product
|
| examples are known as observations or
| |
| | control systems.
|
| patterns. In supervised learning, each
| |
| | Evolutionary computation: applies
|
| pattern belongs to a certain predefined
| |
| | biologically inspired concepts such as
|
| class. A class can be seen as a decision
| |
| | populations, mutation and survival of the
|
| that has to be made. All the observations
| |
| | fittest to generate increasingly better
|
| combined with their class labels are
| |
| | solutions to the problem. These methods
|
| known as a data set.
| |
| | most notably divide into evolutionary
|
| When a new observation is received, the
| |
| | algorithms (e.g. genetic algorithms) and
|
| observation is classified based on
| |
| | swarm intelligence (e.g. ant algorithms).
|
| previous experience. A classifier can be
| |
| | With hybrid intelligent systems attempts
|
| trained in various ways, there are mainly
| |
| | are made to combine these two groups.
|
| statistical and machine learning
| |
| | Expert inference rules can be generated
|
| approaches.
| |
| | through neural network or production
|
| A wide range of classifiers are
| |
| | rules from statistical learning such as
|
| available, each with its strengths and
| |
| | in ACT-R. It is thought that the human
|
| weaknesses. Classifier performance
| |
| | brain uses multiple techniques to both
|
| depends greatly on the characteristics of
| |
| | formulate and cross-check results. Thus,
|
| the data to be classified. There is no
| |
| | systems integration is seen as promising
|
| single classifier that works best on all
| |
| | and perhaps necessary for true AI.
|
| given problems, this is also referred to
| |
| |
|
