ABSTRACT
Computational Intelligence (CI) is a successor of artificial intelligence. CI relies on heuristic algorithms such as in fuzzy systems, neural networks, and evolutionary computation. In addition, computational intelligence also embraces techniques that use Swarm intelligence, Fractals and Chaos Theory, Artificial immune systems, Wavelets, etc. Computational intelligence is a combination of learning, adaptation, and evolution used to intelligent and innovative applications. Computational intelligence research does not reject statistical methods, but often gives a complementary view of the implementation of these methods. Computational intelligence is closely associated with soft computing a combination of artificial neural networks, fuzzy logic and genetic algorithms, connectionist systems such as artificial intelligence, and cybernetics. CI experts mainly consider the biological inspirations from nature for
implementations, but even if biology is extended to include all psychological and evolutionary inspirations then CI includes only the neural, fuzzy, and evolutionary algorithms. The Bayesian foundations of learning, probabilistic and possibilistic reasoning, Markovian chains, belief networks, and graphical theory have no biological connections. Therefore genetic algorithms is the only solution to solve optimization problems. CI studies problems for which there are no effective algorithms, either because it is not possible to formulate them or because they are complex and thus not effective in real life applications. Thus the broad definition is given by: computational intelligence is a branch of computer science studying problems for which there are no effective computational algorithms. Biological organisms solve such problems every day: extracting meaning from perception, understanding language, solving ill-defined computational vision problems thanks to evolutionary adaptation of the brain to the environment, surviving in a hostile environment. However, such problems may be solved in different ways. Defining computational intelligence by the problems that the field studies there is no need to
The exploration of CI is concerned with subordinate cognitive func-
tions: perceptual experience, object identification, signal analysis, breakthrough of structures in data, simple associations, and control. Solution for this type of problems can be obtained using supervised and unsupervised ascertaining, not only neural, fuzzy, and evolutionary overtures but also probabilistic and statistical overtures, such as Bayesian electronic networks or kernel methods. These methods are used to solve the same type of problems in various fields such as pattern recognition, signal processing, classification and regression, data mining. Higher-level cognitive functions are required to solve non-algorithmizable problems involving organized thinking, logical thinking, complex delegacy of knowledge, episodic memory, projecting, realizing of symbolic knowledge. These jobs are at present puzzled out by AI community using tech-
niques based on search, symbolic cognition representation, logical thinking with frame-based expert systems, machine learning in symbolic domains, logics, and lingual methods. Although the jobs belong to the class of non-algorithmic problems, there is a slight overlap between jobs solved using low and high-level cognitive functions. From this aspect AI is a part of CI concentrating on problems that require higher cognition and at present are more comfortable to solve using symbolic knowledge representation. It is possible that other CI methods will also find applications to these problems in the future. The main overlap areas between low and high-level cognitive functions are in sequence learning, reinforcement learning, machine learning, and distributed multi-agent systems. All tasks that require rule based reasoning based on perceptions, such as robotics and automation, machine-driven automobile, etc., require methods for solving both low and high-level cognitive problems and thus are a raw meeting ground for AI experts with the rest of CI community. “Symbol manipulation is the source for all intelligence” - this idea
was proposed by Newell and Simon and they declared that the theory of intelligence was about physical symbols rather than symbolic variables. Symbolic representation of physical are based on multi-dimensional approach patterns comprising states of the brain. Representative models of brain processes do not offer precise approximation of any problem that is described by continuous variables rather than symbolic variables. Estimations to brain processes should be done at a proper level to obtain similar functions. Symbolic dynamics may provide useful information on dynamical systems, and may be useful in modeling transition between low to high level processes. The division between low, and high level cognitive functions is only a rough approximation to the processes in the brain. Incarnated cognition has been intensively studied in the last decade, and developmental estimates depicting how higher processes
In philology it is admitted that real meaning of linguistic terms comes
from body-based metaphors and is equivalently true in mathematics also. New CI methods that go beyond pattern identification and help to solve AI problems may eventually be developed, starting from distributed knowledge representation, graphical methods, and activations networks. The dynamics of such models will probably allow for reasonable symbolic approximations. It is instructive to think about the spectrum of CI problems and various approximations needed to solve them. Neural network models are inspired by brain processes and structures at almost the lowest level, while symbolic AI models by processes at the highest level. The brain has a very specific modular and hierarchical structure, it is not a huge neural network. Perceptron model of a neuron has only one internal parameter, the firing threshold, and a few synaptic weights that determine neuron-neuron interactions. Individual neurons credibly act upon brain information processing in an undistinguished manner. The basic processors used for neural network modelling include bigger
neural structures, such as microcircuits or neural cell gatherings. These structures have more complex domestic states and more complex interactions between elements. An electronic network from networks, hiding the complexness of its processors in a hierarchical way, with different aborning properties at each level, will get increasingly more inner knowledge and additional complex interactions with other such systems. At the highest-level models of whole brains with a countless number of potential internal states and very complex interactions may be obtained. Computational intelligence is certainly more than just the study of the design of intelligent agents; it includes also study of all non-algoritmizable processes that humans (and sometimes animals) can solve with various degrees of competence. CI should not be dealt as a bag of tricks without deeper basis. Challenges from good numerical approaches in various applications should be solicited, and knowledge and search-based methods should complement the core CI techniques in problems involving intelligence. Goldberg and Harik view computational intelligence more as a way of thinking about problems, calling for a “broader view of the scope of the discipline.” They have examined the restrictions to build up computational design,
finding the exemplars of human behaviors to be most useful. Although this is surely worthy delineating the problems that CI wants to solve and welcoming all methods that can be used in such solutions, independent of their inspirations, is even more important. Arguably, CI comprises of those paradigms in AI that relate to some
kind of biological or naturally occurring system. General consensus sug-
FIGURE 1.1: Computational Intelligence Paradigms
gests that these paradigms are neural networks, evolutionary computing, swarm intelligence, and fuzzy systems. Neural networks are based on their biological counterparts in the human nervous system. Similarly, evolutionary computing draws heavily on the principles of Darwinian evolution observed in nature. Swarm intelligence, in turn, is modeled on the social behavior of insects and the choreography of birds flocking. Finally, human reasoning using imprecise, or fuzzy, linguistic terms is approximated by fuzzy systems. This chapter describes these paradigms of CI briefly. Following the discussion, other paradigms of CI such as Granular Computing, Chaos Theory, and Artificial Immune Systems are also dealt. Hybrid approaches of CI and the challenges that CI faces are elaborated in this chapter. Figure 1.1 shows these four primary branches of CI and illustrates that
hybrids between the various paradigms are possible. More precisely, CI is described as the study of adaptive mechanisms to enable or facilitate intelligent behavior in complex and changing environments. There are also other AI approaches, that satisfy both this definition as well as the requirement of modeling some naturally occurring phenomenon, that do not fall neatly into one of the paradigms mentioned thus far. A more pragmatic approach might be to specify the classes of problems that are of interest without being too concerned about whether or not the solutions to these problems satisfy any constraints implied by a particular definition for CI. The following section identifies and describes four primary problem
classes for CI techniques. A compendious overview of the main concepts behind each of the widely recognized CI paradigms is presented in this
Optimization, defined in the following section is undoubtedly the most important class of problem in CI research, since virtually any other class of problem can be re-framed as an optimization problem. This transformation, particularly in a software context, may lead to a loss of information inherent to the intrinsic form of the problem. The major classes of problems in CI are grouped into five categories as Control Problems, Optimization Problems, Classification Problems, Regression Problems, and NP Complete Problems. In the following sections Optimization and NP Complete Problems are discussed.
