ROLUL METODELOR BAZATE PE INTELIGENA COMPUTAIONAL IN PROCESUL DECIZIONAL

Computational intelligence based approaches in decision-makingCernianu Nicolae, Conf.univ.dr.

Matei Cernianu Alice Dalina, Lect.univ.dr.

Titu Maiorescu University

Facultatea de tiine Economice Tg-Jiu

Abstract

Based on statistical information, the decision appears to be a dynamic process management, rational choice consciously, a line of action, from any number of possibilities to achieve a particular goal or purpose, and whose application influence the activity of at least one person other than the makers. For this reason the management decision is linked, first, the forecasting ability of the manager, who in turn would be empty without statistical data and information content of a given volume and a certain structure. Presentation of information also plays an important role. There is the same whether the information received is unclear or confusing, if the transition from information to knowledge is easy or difficult. It must be possible that information from the initial calculations, to obtain information derived. One of the principles that underlie modern information system is to obtain maximum information derived starting from a minimum of primary information. Perception of the full message of statistical information is subject to a number of factors such as level of education of the beneficiary, its degree of information that is reflected in the information.The problem of modeling nonlinear processes with complex structure and a large number of variables remains a challenge, even against the most advanced econometric estimation techniques.Traditional methods, based mainly on linear regression, usually does not provide an adequate solution to this problem, especially when the functional form of the model is unknown and should be intuitively selected from a wide range of potentially very possibilities. According to traditional collectively time, estimating a model can not be achieved without its functional form to be specified ex ante. Usually, however, required knowledge of such a specification is not available and therefore only recourse is to experience or intuition.Econometricians have considered several relevant criteria for selection of functional forms: theoretical consistency, the scope, flexibility, and compliance with facts calculator facilities. In particular, the definition of "flexibility" of a focused research effort. Classical way to address this challenge has been to refine methods to estimate parameters and to provide a priori specification of functional forms elastic enough. In these terms, flexibility is the ability of the functional form (defined parameters) - is this a production function, a function or a function of cost profit - to approximate the behavior different but theoretically consistent, by appropriate choice of parameters.One of the most popular functional forms have been proposed for modeling the behavior of producers and whose flexibility has been proven theoretically and practically justified, is the function TRANSLOG (Logarithmic transcendental function).

More recently, some researchers have studied the nonparametric estimation methods designed to achieve adjustment of the regression functions subject to structural restrictions monotony and concavity. Unlike the mentioned approaches and directions, using computational intelligence techniques based on different ways to meet the requirement of flexibility: neural modeling (neural networks), fuzzy logic-based modeling (fuzzy inference systems) and neuro-fuzzy model that resulting from hybridization techniques with neural fuzzy. According DEX, intelligence is the ability to easily understand and correct to refer to what is essential to resolve the situation and new problems based on previous experience.More generally, intelligence is a property of all systems driven by purpose and decision of such a system is the ability to adapt their behavior to achieve their objectives in a context that involves interaction with the environmentA system is adaptive if it can change the parameters in order to better fulfill their tasks. Adaptability is characteristic of any process in which a structure is progressively modified to provide better performance in interaction with two essential attributes medium.Those intelligent systems - the ability to learn and adapt - underlying the emergence and development of computational intelligence .It is a methodology involving calculations of a system that provides the ability to learn and / or to address new situations - perception being that this system has certain attributes of reasoning:-Generalization-Discovery-Association-Abstract

Computational Intelligence integrates also a set of practical concepts that enable or facilitate the achievement of appropriate actions (intelligent behavior) in complex and changeable (variable time) paradigms, algorithms, mechanisms of self-organization and implementation. Self-organization can be seen as a test system continues to organize itself into complex structures, even in the presence of permanent forces which tend to de-structured, described by the second law of thermodynamics (entropy). The main areas of computational intelligence are:- Fuzzy logic (fuzzy inference systems) - Neural networks - Calculation of evolutionary (genetic algorithms, etc.) - Hybrid Systems

These areas are in a relationship of complementarily (not competition), in which each contributes its own advantages and techniques to solve problems. Fuzzy logic enables approximation, genetic algorithms search conducted "systematic", neural networks have the ability to learn and adapt.Fuzzy logic is the logic of reasoning about, being an extension of multivalent logic and a generalization of conventional logic. At its heart lies the notion of generalization of the classical variety of fuzzy sets.Fuzzy sets refer to classes of objects neprecise boundaries in which membership of a class is gradual between 0 (total non-possession) and 1 (full membership). According to Zadeh, the father of fuzzy logic, as complexity increases, precise formulations lose meaning and meaningful formulations lose precision. Artificial neural networks are biologically inspired systems for parallel processing of information that shapes the structure of the brain mass. They simulate a parallel computing structure, highly interconnected, which contains a large number of relatively simple processing elements (called neurons). Processing elements are connected by weighted connections.Connection weights store information (knowledge), the network is thus able to adapt by modifying these weights. Evolutionary calculation is also a set of methods inspired by biology. It allows finding optimal solutions to problems by stochastic search techniques based on natural biological mechanisms of evolution: selection, recombination, mutation, etc.. During evolution, individuals in the population becomes more adapted to the environment than the individuals from which they were created - like a natural fit. Genetic algorithms, for example, operates with populations of potential solutions and apply the best survival (evolutionary theory, Darwin) to provide approximations of becoming the best solution

Evolutionary calculation includes a number of important areas such as:-Genetic algorithms;Evolutionary programming;-Evolutionary strategies;-Genetic programming.Hybrid systems combine two or more basic computational intelligence techniques to increase performance and allow their mutual potentiation. Examples of hybrid systems can be mentioned:-Neuro-fuzzy systems;-Neuro-genetic systems;-Fuzzy-genetic systems;-Neuro-fuzzy-genetic systems.Methods based on computational intelligence are currently considered the best technique for estimating the nonlinear models, both in terms of predictive power (superior) and in terms of proven flexibility nonlinear process modeling. They offer such a convenient and efficient way to overcome the difficult problem of specifying a priori high form econometric model, replacing it by the means of architecture and fine tuning type neural networks, namely the structure fuzzy partitions, membership functions of the form that define the fuzzy sets and fuzzy inference types suitable for nonlinear process modeling subject. that the structure

Building the model is accompanied by a phase of training, learning. Details depend on the technique used computational intelligence. But they all go through two distinct moments: learning and testing.

Learning is based on a complete set of examples, which are identified based on relations linking the values of input variables, output respectively. The learning process ends when the results provided by the model is close enough for data solutions by which to learn.

But there is certainty that the model will perform equally well in other situations. For this reason, it is tested with data different from those used for learning, but of the same population. if necessary, a phase of readjustment necessary to make to provide good results compared with test data. Only after this process model can be considered operational.

Therefore, it is necessary to develop three sets of data: learning, testing and evaluation. Both neural networks and fuzzy inference systems undergoing a phase estimation model. Unlike traditional functional models, which refers to the model parameters estimates if synaptic weights of neural networks to estimate the network, and fuzzy inference systems are estimated parameters of local models that comprise the global model by a fuzzy aggregation process. In all cases, estimating aims to assign a value / output variables based on values taken by the input variables.

Results from estimation are real values. It has been shown that both neural networks and fuzzy inference systems are universal approximates, that can estimate any nonlinear function defined on a compact domain with an accuracy as good, if it has sufficient data and model structure is well chosen.

The evaluation model is designed to establish the model's ability to determine accurate values for new cases. In general, model performance is evaluated using a set of predictions, which compares the actual output values provided by model

To obtain a relevant evaluation of the model predictive power, it applies a data set distinct from that used in the learning process. These are called data evaluation and are usually the last part of the available data retained for this purpose.

Predictive power calculated for the data evaluation (e.g., mean square prediction error) may be accepted as valid for the new data value and express the overall quality of the model. The integration model in decision-making more generally is the stage that completes the process of modeling.

Conclusion

Computational Intelligence is at the confluence of several subject areas. In principle, it can be seen as a subfield of artificial intelligence (AI), but differs from it in a few key issues. Thus, the typical approach for Artificial Intelligence classical type are descending (top-down) and are based mainly on formal logic, ie intelligent system designer shall forward its whole set of knowledge required to solve a problem, since the construction phase (development).Conversely, if computational intelligence approach is more bottom-up (bottom-up), meaning that algorithms are designed to teach themselves (through accumulation of experience) what they need to know to solve a problem.

In addition, computational intelligence includes all classes of biologically inspired algorithms: neural networks are inspired by biological and cognitive processes in the brain, the calculation of evolutionary (genetic algorithms, etc..) Are based on Darwinian Theory of evolution inspired and natural selection, systems multi-agent is based on self-organizing mechanisms at community, species or ecosystems, biological.Also an effect of the bill biological inspiration is that the emergence and development in computational intelligence are some key attributes of natural intelligent systems: the ability to learn, to adapt, to organize themselves and make generalizations (to respond adequately to environmental stimuli in new situations not explicitly derived from experience).

Remarkable properties listed above justify why the methods based on computational intelligence are currently considered the best technique for estimating the nonlinear models, both in terms of predictive power (superior) and in terms of proven flexibility in modeling nonlinear processes.BIBLIOGRAPHY

1.Georgescu V. - "Flexible estimation of cost functions based on fuzzy logic modeling

approach", Fuzzy Economic Review, No. 2, Vol.III, 1998, p. 49-68.

2. Georgescu V. - "Flexible estimation procedures for fuzzy logic modeling, with

economic applications", Proceedings of the 5th Congress of SIGEF, October, 1998,

Lausanne3. Andrei, A, Oprescu, Gh., Marin, D., Mitru, D., Roman, M. (2001). Modele dinamice de conducere optimal a activitii firmei, Editura ASE, Bucureti4. Scarlat, E., Mrcine, Virginia (2002), Dinamica economic. Dinamica sitemelor haotice, Editura MatrixRom, Bucureti.5. Scarlat, E., Chiri, Nora (2001). Macroeconomie dinamic, Editura Economic, Bucureti